1 //
2 //                     The LLVM Compiler Infrastructure
3 //
4 // This file is distributed under the University of Illinois Open Source
5 // License. See LICENSE.TXT for details.
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the interfaces that NVPTX uses to lower LLVM code into a
10 // selection DAG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "NVPTXISelLowering.h"
15 #include "NVPTX.h"
16 #include "NVPTXTargetMachine.h"
17 #include "NVPTXTargetObjectFile.h"
18 #include "NVPTXUtilities.h"
19 #include "llvm/CodeGen/Analysis.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GlobalValue.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/MC/MCSectionELF.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include <sstream>
39 
40 #undef DEBUG_TYPE
41 #define DEBUG_TYPE "nvptx-lower"
42 
43 using namespace llvm;
44 
45 static unsigned int uniqueCallSite = 0;
46 
47 static cl::opt<bool> sched4reg(
48     "nvptx-sched4reg",
49     cl::desc("NVPTX Specific: schedule for register pressue"), cl::init(false));
50 
51 static cl::opt<unsigned>
52 FMAContractLevelOpt("nvptx-fma-level", cl::ZeroOrMore, cl::Hidden,
53                     cl::desc("NVPTX Specific: FMA contraction (0: don't do it"
54                              " 1: do it  2: do it aggressively"),
55                     cl::init(2));
56 
IsPTXVectorType(MVT VT)57 static bool IsPTXVectorType(MVT VT) {
58   switch (VT.SimpleTy) {
59   default:
60     return false;
61   case MVT::v2i1:
62   case MVT::v4i1:
63   case MVT::v2i8:
64   case MVT::v4i8:
65   case MVT::v2i16:
66   case MVT::v4i16:
67   case MVT::v2i32:
68   case MVT::v4i32:
69   case MVT::v2i64:
70   case MVT::v2f32:
71   case MVT::v4f32:
72   case MVT::v2f64:
73     return true;
74   }
75 }
76 
77 /// ComputePTXValueVTs - For the given Type \p Ty, returns the set of primitive
78 /// EVTs that compose it.  Unlike ComputeValueVTs, this will break apart vectors
79 /// into their primitive components.
80 /// NOTE: This is a band-aid for code that expects ComputeValueVTs to return the
81 /// same number of types as the Ins/Outs arrays in LowerFormalArguments,
82 /// LowerCall, and LowerReturn.
ComputePTXValueVTs(const TargetLowering & TLI,Type * Ty,SmallVectorImpl<EVT> & ValueVTs,SmallVectorImpl<uint64_t> * Offsets=nullptr,uint64_t StartingOffset=0)83 static void ComputePTXValueVTs(const TargetLowering &TLI, Type *Ty,
84                                SmallVectorImpl<EVT> &ValueVTs,
85                                SmallVectorImpl<uint64_t> *Offsets = nullptr,
86                                uint64_t StartingOffset = 0) {
87   SmallVector<EVT, 16> TempVTs;
88   SmallVector<uint64_t, 16> TempOffsets;
89 
90   ComputeValueVTs(TLI, Ty, TempVTs, &TempOffsets, StartingOffset);
91   for (unsigned i = 0, e = TempVTs.size(); i != e; ++i) {
92     EVT VT = TempVTs[i];
93     uint64_t Off = TempOffsets[i];
94     if (VT.isVector())
95       for (unsigned j = 0, je = VT.getVectorNumElements(); j != je; ++j) {
96         ValueVTs.push_back(VT.getVectorElementType());
97         if (Offsets)
98           Offsets->push_back(Off+j*VT.getVectorElementType().getStoreSize());
99       }
100     else {
101       ValueVTs.push_back(VT);
102       if (Offsets)
103         Offsets->push_back(Off);
104     }
105   }
106 }
107 
108 // NVPTXTargetLowering Constructor.
NVPTXTargetLowering(const NVPTXTargetMachine & TM,const NVPTXSubtarget & STI)109 NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
110                                          const NVPTXSubtarget &STI)
111     : TargetLowering(TM), nvTM(&TM), STI(STI) {
112 
113   // always lower memset, memcpy, and memmove intrinsics to load/store
114   // instructions, rather
115   // then generating calls to memset, mempcy or memmove.
116   MaxStoresPerMemset = (unsigned) 0xFFFFFFFF;
117   MaxStoresPerMemcpy = (unsigned) 0xFFFFFFFF;
118   MaxStoresPerMemmove = (unsigned) 0xFFFFFFFF;
119 
120   setBooleanContents(ZeroOrNegativeOneBooleanContent);
121   setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
122 
123   // Jump is Expensive. Don't create extra control flow for 'and', 'or'
124   // condition branches.
125   setJumpIsExpensive(true);
126 
127   // By default, use the Source scheduling
128   if (sched4reg)
129     setSchedulingPreference(Sched::RegPressure);
130   else
131     setSchedulingPreference(Sched::Source);
132 
133   addRegisterClass(MVT::i1, &NVPTX::Int1RegsRegClass);
134   addRegisterClass(MVT::i16, &NVPTX::Int16RegsRegClass);
135   addRegisterClass(MVT::i32, &NVPTX::Int32RegsRegClass);
136   addRegisterClass(MVT::i64, &NVPTX::Int64RegsRegClass);
137   addRegisterClass(MVT::f32, &NVPTX::Float32RegsRegClass);
138   addRegisterClass(MVT::f64, &NVPTX::Float64RegsRegClass);
139 
140   // Operations not directly supported by NVPTX.
141   setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
142   setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
143   setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
144   setOperationAction(ISD::SELECT_CC, MVT::i8, Expand);
145   setOperationAction(ISD::SELECT_CC, MVT::i16, Expand);
146   setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
147   setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
148   setOperationAction(ISD::BR_CC, MVT::f32, Expand);
149   setOperationAction(ISD::BR_CC, MVT::f64, Expand);
150   setOperationAction(ISD::BR_CC, MVT::i1, Expand);
151   setOperationAction(ISD::BR_CC, MVT::i8, Expand);
152   setOperationAction(ISD::BR_CC, MVT::i16, Expand);
153   setOperationAction(ISD::BR_CC, MVT::i32, Expand);
154   setOperationAction(ISD::BR_CC, MVT::i64, Expand);
155   // Some SIGN_EXTEND_INREG can be done using cvt instruction.
156   // For others we will expand to a SHL/SRA pair.
157   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i64, Legal);
158   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
159   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
160   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
161   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
162 
163   setOperationAction(ISD::SHL_PARTS, MVT::i32  , Custom);
164   setOperationAction(ISD::SRA_PARTS, MVT::i32  , Custom);
165   setOperationAction(ISD::SRL_PARTS, MVT::i32  , Custom);
166   setOperationAction(ISD::SHL_PARTS, MVT::i64  , Custom);
167   setOperationAction(ISD::SRA_PARTS, MVT::i64  , Custom);
168   setOperationAction(ISD::SRL_PARTS, MVT::i64  , Custom);
169 
170   if (STI.hasROT64()) {
171     setOperationAction(ISD::ROTL, MVT::i64, Legal);
172     setOperationAction(ISD::ROTR, MVT::i64, Legal);
173   } else {
174     setOperationAction(ISD::ROTL, MVT::i64, Expand);
175     setOperationAction(ISD::ROTR, MVT::i64, Expand);
176   }
177   if (STI.hasROT32()) {
178     setOperationAction(ISD::ROTL, MVT::i32, Legal);
179     setOperationAction(ISD::ROTR, MVT::i32, Legal);
180   } else {
181     setOperationAction(ISD::ROTL, MVT::i32, Expand);
182     setOperationAction(ISD::ROTR, MVT::i32, Expand);
183   }
184 
185   setOperationAction(ISD::ROTL, MVT::i16, Expand);
186   setOperationAction(ISD::ROTR, MVT::i16, Expand);
187   setOperationAction(ISD::ROTL, MVT::i8, Expand);
188   setOperationAction(ISD::ROTR, MVT::i8, Expand);
189   setOperationAction(ISD::BSWAP, MVT::i16, Expand);
190   setOperationAction(ISD::BSWAP, MVT::i32, Expand);
191   setOperationAction(ISD::BSWAP, MVT::i64, Expand);
192 
193   // Indirect branch is not supported.
194   // This also disables Jump Table creation.
195   setOperationAction(ISD::BR_JT, MVT::Other, Expand);
196   setOperationAction(ISD::BRIND, MVT::Other, Expand);
197 
198   setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
199   setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
200 
201   // We want to legalize constant related memmove and memcopy
202   // intrinsics.
203   setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
204 
205   // Turn FP extload into load/fextend
206   setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
207   setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
208   setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
209   // Turn FP truncstore into trunc + store.
210   setTruncStoreAction(MVT::f32, MVT::f16, Expand);
211   setTruncStoreAction(MVT::f64, MVT::f16, Expand);
212   setTruncStoreAction(MVT::f64, MVT::f32, Expand);
213 
214   // PTX does not support load / store predicate registers
215   setOperationAction(ISD::LOAD, MVT::i1, Custom);
216   setOperationAction(ISD::STORE, MVT::i1, Custom);
217 
218   for (MVT VT : MVT::integer_valuetypes()) {
219     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
220     setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
221     setTruncStoreAction(VT, MVT::i1, Expand);
222   }
223 
224   // This is legal in NVPTX
225   setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
226   setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
227 
228   // TRAP can be lowered to PTX trap
229   setOperationAction(ISD::TRAP, MVT::Other, Legal);
230 
231   setOperationAction(ISD::ADDC, MVT::i64, Expand);
232   setOperationAction(ISD::ADDE, MVT::i64, Expand);
233 
234   // Register custom handling for vector loads/stores
235   for (MVT VT : MVT::vector_valuetypes()) {
236     if (IsPTXVectorType(VT)) {
237       setOperationAction(ISD::LOAD, VT, Custom);
238       setOperationAction(ISD::STORE, VT, Custom);
239       setOperationAction(ISD::INTRINSIC_W_CHAIN, VT, Custom);
240     }
241   }
242 
243   // Custom handling for i8 intrinsics
244   setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom);
245 
246   setOperationAction(ISD::CTLZ, MVT::i16, Legal);
247   setOperationAction(ISD::CTLZ, MVT::i32, Legal);
248   setOperationAction(ISD::CTLZ, MVT::i64, Legal);
249   setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16, Legal);
250   setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Legal);
251   setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Legal);
252   setOperationAction(ISD::CTTZ, MVT::i16, Expand);
253   setOperationAction(ISD::CTTZ, MVT::i32, Expand);
254   setOperationAction(ISD::CTTZ, MVT::i64, Expand);
255   setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16, Expand);
256   setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
257   setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
258   setOperationAction(ISD::CTPOP, MVT::i16, Legal);
259   setOperationAction(ISD::CTPOP, MVT::i32, Legal);
260   setOperationAction(ISD::CTPOP, MVT::i64, Legal);
261 
262   // PTX does not directly support SELP of i1, so promote to i32 first
263   setOperationAction(ISD::SELECT, MVT::i1, Custom);
264 
265   // We have some custom DAG combine patterns for these nodes
266   setTargetDAGCombine(ISD::ADD);
267   setTargetDAGCombine(ISD::AND);
268   setTargetDAGCombine(ISD::FADD);
269   setTargetDAGCombine(ISD::MUL);
270   setTargetDAGCombine(ISD::SHL);
271 
272   // Now deduce the information based on the above mentioned
273   // actions
274   computeRegisterProperties(STI.getRegisterInfo());
275 }
276 
getTargetNodeName(unsigned Opcode) const277 const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const {
278   switch (Opcode) {
279   default:
280     return nullptr;
281   case NVPTXISD::CALL:
282     return "NVPTXISD::CALL";
283   case NVPTXISD::RET_FLAG:
284     return "NVPTXISD::RET_FLAG";
285   case NVPTXISD::Wrapper:
286     return "NVPTXISD::Wrapper";
287   case NVPTXISD::DeclareParam:
288     return "NVPTXISD::DeclareParam";
289   case NVPTXISD::DeclareScalarParam:
290     return "NVPTXISD::DeclareScalarParam";
291   case NVPTXISD::DeclareRet:
292     return "NVPTXISD::DeclareRet";
293   case NVPTXISD::DeclareRetParam:
294     return "NVPTXISD::DeclareRetParam";
295   case NVPTXISD::PrintCall:
296     return "NVPTXISD::PrintCall";
297   case NVPTXISD::LoadParam:
298     return "NVPTXISD::LoadParam";
299   case NVPTXISD::LoadParamV2:
300     return "NVPTXISD::LoadParamV2";
301   case NVPTXISD::LoadParamV4:
302     return "NVPTXISD::LoadParamV4";
303   case NVPTXISD::StoreParam:
304     return "NVPTXISD::StoreParam";
305   case NVPTXISD::StoreParamV2:
306     return "NVPTXISD::StoreParamV2";
307   case NVPTXISD::StoreParamV4:
308     return "NVPTXISD::StoreParamV4";
309   case NVPTXISD::StoreParamS32:
310     return "NVPTXISD::StoreParamS32";
311   case NVPTXISD::StoreParamU32:
312     return "NVPTXISD::StoreParamU32";
313   case NVPTXISD::CallArgBegin:
314     return "NVPTXISD::CallArgBegin";
315   case NVPTXISD::CallArg:
316     return "NVPTXISD::CallArg";
317   case NVPTXISD::LastCallArg:
318     return "NVPTXISD::LastCallArg";
319   case NVPTXISD::CallArgEnd:
320     return "NVPTXISD::CallArgEnd";
321   case NVPTXISD::CallVoid:
322     return "NVPTXISD::CallVoid";
323   case NVPTXISD::CallVal:
324     return "NVPTXISD::CallVal";
325   case NVPTXISD::CallSymbol:
326     return "NVPTXISD::CallSymbol";
327   case NVPTXISD::Prototype:
328     return "NVPTXISD::Prototype";
329   case NVPTXISD::MoveParam:
330     return "NVPTXISD::MoveParam";
331   case NVPTXISD::StoreRetval:
332     return "NVPTXISD::StoreRetval";
333   case NVPTXISD::StoreRetvalV2:
334     return "NVPTXISD::StoreRetvalV2";
335   case NVPTXISD::StoreRetvalV4:
336     return "NVPTXISD::StoreRetvalV4";
337   case NVPTXISD::PseudoUseParam:
338     return "NVPTXISD::PseudoUseParam";
339   case NVPTXISD::RETURN:
340     return "NVPTXISD::RETURN";
341   case NVPTXISD::CallSeqBegin:
342     return "NVPTXISD::CallSeqBegin";
343   case NVPTXISD::CallSeqEnd:
344     return "NVPTXISD::CallSeqEnd";
345   case NVPTXISD::CallPrototype:
346     return "NVPTXISD::CallPrototype";
347   case NVPTXISD::LoadV2:
348     return "NVPTXISD::LoadV2";
349   case NVPTXISD::LoadV4:
350     return "NVPTXISD::LoadV4";
351   case NVPTXISD::LDGV2:
352     return "NVPTXISD::LDGV2";
353   case NVPTXISD::LDGV4:
354     return "NVPTXISD::LDGV4";
355   case NVPTXISD::LDUV2:
356     return "NVPTXISD::LDUV2";
357   case NVPTXISD::LDUV4:
358     return "NVPTXISD::LDUV4";
359   case NVPTXISD::StoreV2:
360     return "NVPTXISD::StoreV2";
361   case NVPTXISD::StoreV4:
362     return "NVPTXISD::StoreV4";
363   case NVPTXISD::FUN_SHFL_CLAMP:
364     return "NVPTXISD::FUN_SHFL_CLAMP";
365   case NVPTXISD::FUN_SHFR_CLAMP:
366     return "NVPTXISD::FUN_SHFR_CLAMP";
367   case NVPTXISD::IMAD:
368     return "NVPTXISD::IMAD";
369   case NVPTXISD::MUL_WIDE_SIGNED:
370     return "NVPTXISD::MUL_WIDE_SIGNED";
371   case NVPTXISD::MUL_WIDE_UNSIGNED:
372     return "NVPTXISD::MUL_WIDE_UNSIGNED";
373   case NVPTXISD::Tex1DFloatS32:        return "NVPTXISD::Tex1DFloatS32";
374   case NVPTXISD::Tex1DFloatFloat:      return "NVPTXISD::Tex1DFloatFloat";
375   case NVPTXISD::Tex1DFloatFloatLevel:
376     return "NVPTXISD::Tex1DFloatFloatLevel";
377   case NVPTXISD::Tex1DFloatFloatGrad:
378     return "NVPTXISD::Tex1DFloatFloatGrad";
379   case NVPTXISD::Tex1DS32S32:          return "NVPTXISD::Tex1DS32S32";
380   case NVPTXISD::Tex1DS32Float:        return "NVPTXISD::Tex1DS32Float";
381   case NVPTXISD::Tex1DS32FloatLevel:
382     return "NVPTXISD::Tex1DS32FloatLevel";
383   case NVPTXISD::Tex1DS32FloatGrad:
384     return "NVPTXISD::Tex1DS32FloatGrad";
385   case NVPTXISD::Tex1DU32S32:          return "NVPTXISD::Tex1DU32S32";
386   case NVPTXISD::Tex1DU32Float:        return "NVPTXISD::Tex1DU32Float";
387   case NVPTXISD::Tex1DU32FloatLevel:
388     return "NVPTXISD::Tex1DU32FloatLevel";
389   case NVPTXISD::Tex1DU32FloatGrad:
390     return "NVPTXISD::Tex1DU32FloatGrad";
391   case NVPTXISD::Tex1DArrayFloatS32:   return "NVPTXISD::Tex1DArrayFloatS32";
392   case NVPTXISD::Tex1DArrayFloatFloat: return "NVPTXISD::Tex1DArrayFloatFloat";
393   case NVPTXISD::Tex1DArrayFloatFloatLevel:
394     return "NVPTXISD::Tex1DArrayFloatFloatLevel";
395   case NVPTXISD::Tex1DArrayFloatFloatGrad:
396     return "NVPTXISD::Tex1DArrayFloatFloatGrad";
397   case NVPTXISD::Tex1DArrayS32S32:     return "NVPTXISD::Tex1DArrayS32S32";
398   case NVPTXISD::Tex1DArrayS32Float:   return "NVPTXISD::Tex1DArrayS32Float";
399   case NVPTXISD::Tex1DArrayS32FloatLevel:
400     return "NVPTXISD::Tex1DArrayS32FloatLevel";
401   case NVPTXISD::Tex1DArrayS32FloatGrad:
402     return "NVPTXISD::Tex1DArrayS32FloatGrad";
403   case NVPTXISD::Tex1DArrayU32S32:     return "NVPTXISD::Tex1DArrayU32S32";
404   case NVPTXISD::Tex1DArrayU32Float:   return "NVPTXISD::Tex1DArrayU32Float";
405   case NVPTXISD::Tex1DArrayU32FloatLevel:
406     return "NVPTXISD::Tex1DArrayU32FloatLevel";
407   case NVPTXISD::Tex1DArrayU32FloatGrad:
408     return "NVPTXISD::Tex1DArrayU32FloatGrad";
409   case NVPTXISD::Tex2DFloatS32:        return "NVPTXISD::Tex2DFloatS32";
410   case NVPTXISD::Tex2DFloatFloat:      return "NVPTXISD::Tex2DFloatFloat";
411   case NVPTXISD::Tex2DFloatFloatLevel:
412     return "NVPTXISD::Tex2DFloatFloatLevel";
413   case NVPTXISD::Tex2DFloatFloatGrad:
414     return "NVPTXISD::Tex2DFloatFloatGrad";
415   case NVPTXISD::Tex2DS32S32:          return "NVPTXISD::Tex2DS32S32";
416   case NVPTXISD::Tex2DS32Float:        return "NVPTXISD::Tex2DS32Float";
417   case NVPTXISD::Tex2DS32FloatLevel:
418     return "NVPTXISD::Tex2DS32FloatLevel";
419   case NVPTXISD::Tex2DS32FloatGrad:
420     return "NVPTXISD::Tex2DS32FloatGrad";
421   case NVPTXISD::Tex2DU32S32:          return "NVPTXISD::Tex2DU32S32";
422   case NVPTXISD::Tex2DU32Float:        return "NVPTXISD::Tex2DU32Float";
423   case NVPTXISD::Tex2DU32FloatLevel:
424     return "NVPTXISD::Tex2DU32FloatLevel";
425   case NVPTXISD::Tex2DU32FloatGrad:
426     return "NVPTXISD::Tex2DU32FloatGrad";
427   case NVPTXISD::Tex2DArrayFloatS32:   return "NVPTXISD::Tex2DArrayFloatS32";
428   case NVPTXISD::Tex2DArrayFloatFloat: return "NVPTXISD::Tex2DArrayFloatFloat";
429   case NVPTXISD::Tex2DArrayFloatFloatLevel:
430     return "NVPTXISD::Tex2DArrayFloatFloatLevel";
431   case NVPTXISD::Tex2DArrayFloatFloatGrad:
432     return "NVPTXISD::Tex2DArrayFloatFloatGrad";
433   case NVPTXISD::Tex2DArrayS32S32:     return "NVPTXISD::Tex2DArrayS32S32";
434   case NVPTXISD::Tex2DArrayS32Float:   return "NVPTXISD::Tex2DArrayS32Float";
435   case NVPTXISD::Tex2DArrayS32FloatLevel:
436     return "NVPTXISD::Tex2DArrayS32FloatLevel";
437   case NVPTXISD::Tex2DArrayS32FloatGrad:
438     return "NVPTXISD::Tex2DArrayS32FloatGrad";
439   case NVPTXISD::Tex2DArrayU32S32:     return "NVPTXISD::Tex2DArrayU32S32";
440   case NVPTXISD::Tex2DArrayU32Float:   return "NVPTXISD::Tex2DArrayU32Float";
441   case NVPTXISD::Tex2DArrayU32FloatLevel:
442     return "NVPTXISD::Tex2DArrayU32FloatLevel";
443   case NVPTXISD::Tex2DArrayU32FloatGrad:
444     return "NVPTXISD::Tex2DArrayU32FloatGrad";
445   case NVPTXISD::Tex3DFloatS32:        return "NVPTXISD::Tex3DFloatS32";
446   case NVPTXISD::Tex3DFloatFloat:      return "NVPTXISD::Tex3DFloatFloat";
447   case NVPTXISD::Tex3DFloatFloatLevel:
448     return "NVPTXISD::Tex3DFloatFloatLevel";
449   case NVPTXISD::Tex3DFloatFloatGrad:
450     return "NVPTXISD::Tex3DFloatFloatGrad";
451   case NVPTXISD::Tex3DS32S32:          return "NVPTXISD::Tex3DS32S32";
452   case NVPTXISD::Tex3DS32Float:        return "NVPTXISD::Tex3DS32Float";
453   case NVPTXISD::Tex3DS32FloatLevel:
454     return "NVPTXISD::Tex3DS32FloatLevel";
455   case NVPTXISD::Tex3DS32FloatGrad:
456     return "NVPTXISD::Tex3DS32FloatGrad";
457   case NVPTXISD::Tex3DU32S32:          return "NVPTXISD::Tex3DU32S32";
458   case NVPTXISD::Tex3DU32Float:        return "NVPTXISD::Tex3DU32Float";
459   case NVPTXISD::Tex3DU32FloatLevel:
460     return "NVPTXISD::Tex3DU32FloatLevel";
461   case NVPTXISD::Tex3DU32FloatGrad:
462     return "NVPTXISD::Tex3DU32FloatGrad";
463   case NVPTXISD::TexCubeFloatFloat:      return "NVPTXISD::TexCubeFloatFloat";
464   case NVPTXISD::TexCubeFloatFloatLevel:
465     return "NVPTXISD::TexCubeFloatFloatLevel";
466   case NVPTXISD::TexCubeS32Float:        return "NVPTXISD::TexCubeS32Float";
467   case NVPTXISD::TexCubeS32FloatLevel:
468     return "NVPTXISD::TexCubeS32FloatLevel";
469   case NVPTXISD::TexCubeU32Float:        return "NVPTXISD::TexCubeU32Float";
470   case NVPTXISD::TexCubeU32FloatLevel:
471     return "NVPTXISD::TexCubeU32FloatLevel";
472   case NVPTXISD::TexCubeArrayFloatFloat:
473     return "NVPTXISD::TexCubeArrayFloatFloat";
474   case NVPTXISD::TexCubeArrayFloatFloatLevel:
475     return "NVPTXISD::TexCubeArrayFloatFloatLevel";
476   case NVPTXISD::TexCubeArrayS32Float:
477     return "NVPTXISD::TexCubeArrayS32Float";
478   case NVPTXISD::TexCubeArrayS32FloatLevel:
479     return "NVPTXISD::TexCubeArrayS32FloatLevel";
480   case NVPTXISD::TexCubeArrayU32Float:
481     return "NVPTXISD::TexCubeArrayU32Float";
482   case NVPTXISD::TexCubeArrayU32FloatLevel:
483     return "NVPTXISD::TexCubeArrayU32FloatLevel";
484   case NVPTXISD::Tld4R2DFloatFloat:
485     return "NVPTXISD::Tld4R2DFloatFloat";
486   case NVPTXISD::Tld4G2DFloatFloat:
487     return "NVPTXISD::Tld4G2DFloatFloat";
488   case NVPTXISD::Tld4B2DFloatFloat:
489     return "NVPTXISD::Tld4B2DFloatFloat";
490   case NVPTXISD::Tld4A2DFloatFloat:
491     return "NVPTXISD::Tld4A2DFloatFloat";
492   case NVPTXISD::Tld4R2DS64Float:
493     return "NVPTXISD::Tld4R2DS64Float";
494   case NVPTXISD::Tld4G2DS64Float:
495     return "NVPTXISD::Tld4G2DS64Float";
496   case NVPTXISD::Tld4B2DS64Float:
497     return "NVPTXISD::Tld4B2DS64Float";
498   case NVPTXISD::Tld4A2DS64Float:
499     return "NVPTXISD::Tld4A2DS64Float";
500   case NVPTXISD::Tld4R2DU64Float:
501     return "NVPTXISD::Tld4R2DU64Float";
502   case NVPTXISD::Tld4G2DU64Float:
503     return "NVPTXISD::Tld4G2DU64Float";
504   case NVPTXISD::Tld4B2DU64Float:
505     return "NVPTXISD::Tld4B2DU64Float";
506   case NVPTXISD::Tld4A2DU64Float:
507     return "NVPTXISD::Tld4A2DU64Float";
508 
509   case NVPTXISD::TexUnified1DFloatS32:
510     return "NVPTXISD::TexUnified1DFloatS32";
511   case NVPTXISD::TexUnified1DFloatFloat:
512     return "NVPTXISD::TexUnified1DFloatFloat";
513   case NVPTXISD::TexUnified1DFloatFloatLevel:
514     return "NVPTXISD::TexUnified1DFloatFloatLevel";
515   case NVPTXISD::TexUnified1DFloatFloatGrad:
516     return "NVPTXISD::TexUnified1DFloatFloatGrad";
517   case NVPTXISD::TexUnified1DS32S32:
518     return "NVPTXISD::TexUnified1DS32S32";
519   case NVPTXISD::TexUnified1DS32Float:
520     return "NVPTXISD::TexUnified1DS32Float";
521   case NVPTXISD::TexUnified1DS32FloatLevel:
522     return "NVPTXISD::TexUnified1DS32FloatLevel";
523   case NVPTXISD::TexUnified1DS32FloatGrad:
524     return "NVPTXISD::TexUnified1DS32FloatGrad";
525   case NVPTXISD::TexUnified1DU32S32:
526     return "NVPTXISD::TexUnified1DU32S32";
527   case NVPTXISD::TexUnified1DU32Float:
528     return "NVPTXISD::TexUnified1DU32Float";
529   case NVPTXISD::TexUnified1DU32FloatLevel:
530     return "NVPTXISD::TexUnified1DU32FloatLevel";
531   case NVPTXISD::TexUnified1DU32FloatGrad:
532     return "NVPTXISD::TexUnified1DU32FloatGrad";
533   case NVPTXISD::TexUnified1DArrayFloatS32:
534     return "NVPTXISD::TexUnified1DArrayFloatS32";
535   case NVPTXISD::TexUnified1DArrayFloatFloat:
536     return "NVPTXISD::TexUnified1DArrayFloatFloat";
537   case NVPTXISD::TexUnified1DArrayFloatFloatLevel:
538     return "NVPTXISD::TexUnified1DArrayFloatFloatLevel";
539   case NVPTXISD::TexUnified1DArrayFloatFloatGrad:
540     return "NVPTXISD::TexUnified1DArrayFloatFloatGrad";
541   case NVPTXISD::TexUnified1DArrayS32S32:
542     return "NVPTXISD::TexUnified1DArrayS32S32";
543   case NVPTXISD::TexUnified1DArrayS32Float:
544     return "NVPTXISD::TexUnified1DArrayS32Float";
545   case NVPTXISD::TexUnified1DArrayS32FloatLevel:
546     return "NVPTXISD::TexUnified1DArrayS32FloatLevel";
547   case NVPTXISD::TexUnified1DArrayS32FloatGrad:
548     return "NVPTXISD::TexUnified1DArrayS32FloatGrad";
549   case NVPTXISD::TexUnified1DArrayU32S32:
550     return "NVPTXISD::TexUnified1DArrayU32S32";
551   case NVPTXISD::TexUnified1DArrayU32Float:
552     return "NVPTXISD::TexUnified1DArrayU32Float";
553   case NVPTXISD::TexUnified1DArrayU32FloatLevel:
554     return "NVPTXISD::TexUnified1DArrayU32FloatLevel";
555   case NVPTXISD::TexUnified1DArrayU32FloatGrad:
556     return "NVPTXISD::TexUnified1DArrayU32FloatGrad";
557   case NVPTXISD::TexUnified2DFloatS32:
558     return "NVPTXISD::TexUnified2DFloatS32";
559   case NVPTXISD::TexUnified2DFloatFloat:
560     return "NVPTXISD::TexUnified2DFloatFloat";
561   case NVPTXISD::TexUnified2DFloatFloatLevel:
562     return "NVPTXISD::TexUnified2DFloatFloatLevel";
563   case NVPTXISD::TexUnified2DFloatFloatGrad:
564     return "NVPTXISD::TexUnified2DFloatFloatGrad";
565   case NVPTXISD::TexUnified2DS32S32:
566     return "NVPTXISD::TexUnified2DS32S32";
567   case NVPTXISD::TexUnified2DS32Float:
568     return "NVPTXISD::TexUnified2DS32Float";
569   case NVPTXISD::TexUnified2DS32FloatLevel:
570     return "NVPTXISD::TexUnified2DS32FloatLevel";
571   case NVPTXISD::TexUnified2DS32FloatGrad:
572     return "NVPTXISD::TexUnified2DS32FloatGrad";
573   case NVPTXISD::TexUnified2DU32S32:
574     return "NVPTXISD::TexUnified2DU32S32";
575   case NVPTXISD::TexUnified2DU32Float:
576     return "NVPTXISD::TexUnified2DU32Float";
577   case NVPTXISD::TexUnified2DU32FloatLevel:
578     return "NVPTXISD::TexUnified2DU32FloatLevel";
579   case NVPTXISD::TexUnified2DU32FloatGrad:
580     return "NVPTXISD::TexUnified2DU32FloatGrad";
581   case NVPTXISD::TexUnified2DArrayFloatS32:
582     return "NVPTXISD::TexUnified2DArrayFloatS32";
583   case NVPTXISD::TexUnified2DArrayFloatFloat:
584     return "NVPTXISD::TexUnified2DArrayFloatFloat";
585   case NVPTXISD::TexUnified2DArrayFloatFloatLevel:
586     return "NVPTXISD::TexUnified2DArrayFloatFloatLevel";
587   case NVPTXISD::TexUnified2DArrayFloatFloatGrad:
588     return "NVPTXISD::TexUnified2DArrayFloatFloatGrad";
589   case NVPTXISD::TexUnified2DArrayS32S32:
590     return "NVPTXISD::TexUnified2DArrayS32S32";
591   case NVPTXISD::TexUnified2DArrayS32Float:
592     return "NVPTXISD::TexUnified2DArrayS32Float";
593   case NVPTXISD::TexUnified2DArrayS32FloatLevel:
594     return "NVPTXISD::TexUnified2DArrayS32FloatLevel";
595   case NVPTXISD::TexUnified2DArrayS32FloatGrad:
596     return "NVPTXISD::TexUnified2DArrayS32FloatGrad";
597   case NVPTXISD::TexUnified2DArrayU32S32:
598     return "NVPTXISD::TexUnified2DArrayU32S32";
599   case NVPTXISD::TexUnified2DArrayU32Float:
600     return "NVPTXISD::TexUnified2DArrayU32Float";
601   case NVPTXISD::TexUnified2DArrayU32FloatLevel:
602     return "NVPTXISD::TexUnified2DArrayU32FloatLevel";
603   case NVPTXISD::TexUnified2DArrayU32FloatGrad:
604     return "NVPTXISD::TexUnified2DArrayU32FloatGrad";
605   case NVPTXISD::TexUnified3DFloatS32:
606     return "NVPTXISD::TexUnified3DFloatS32";
607   case NVPTXISD::TexUnified3DFloatFloat:
608     return "NVPTXISD::TexUnified3DFloatFloat";
609   case NVPTXISD::TexUnified3DFloatFloatLevel:
610     return "NVPTXISD::TexUnified3DFloatFloatLevel";
611   case NVPTXISD::TexUnified3DFloatFloatGrad:
612     return "NVPTXISD::TexUnified3DFloatFloatGrad";
613   case NVPTXISD::TexUnified3DS32S32:
614     return "NVPTXISD::TexUnified3DS32S32";
615   case NVPTXISD::TexUnified3DS32Float:
616     return "NVPTXISD::TexUnified3DS32Float";
617   case NVPTXISD::TexUnified3DS32FloatLevel:
618     return "NVPTXISD::TexUnified3DS32FloatLevel";
619   case NVPTXISD::TexUnified3DS32FloatGrad:
620     return "NVPTXISD::TexUnified3DS32FloatGrad";
621   case NVPTXISD::TexUnified3DU32S32:
622     return "NVPTXISD::TexUnified3DU32S32";
623   case NVPTXISD::TexUnified3DU32Float:
624     return "NVPTXISD::TexUnified3DU32Float";
625   case NVPTXISD::TexUnified3DU32FloatLevel:
626     return "NVPTXISD::TexUnified3DU32FloatLevel";
627   case NVPTXISD::TexUnified3DU32FloatGrad:
628     return "NVPTXISD::TexUnified3DU32FloatGrad";
629   case NVPTXISD::TexUnifiedCubeFloatFloat:
630     return "NVPTXISD::TexUnifiedCubeFloatFloat";
631   case NVPTXISD::TexUnifiedCubeFloatFloatLevel:
632     return "NVPTXISD::TexUnifiedCubeFloatFloatLevel";
633   case NVPTXISD::TexUnifiedCubeS32Float:
634     return "NVPTXISD::TexUnifiedCubeS32Float";
635   case NVPTXISD::TexUnifiedCubeS32FloatLevel:
636     return "NVPTXISD::TexUnifiedCubeS32FloatLevel";
637   case NVPTXISD::TexUnifiedCubeU32Float:
638     return "NVPTXISD::TexUnifiedCubeU32Float";
639   case NVPTXISD::TexUnifiedCubeU32FloatLevel:
640     return "NVPTXISD::TexUnifiedCubeU32FloatLevel";
641   case NVPTXISD::TexUnifiedCubeArrayFloatFloat:
642     return "NVPTXISD::TexUnifiedCubeArrayFloatFloat";
643   case NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel:
644     return "NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel";
645   case NVPTXISD::TexUnifiedCubeArrayS32Float:
646     return "NVPTXISD::TexUnifiedCubeArrayS32Float";
647   case NVPTXISD::TexUnifiedCubeArrayS32FloatLevel:
648     return "NVPTXISD::TexUnifiedCubeArrayS32FloatLevel";
649   case NVPTXISD::TexUnifiedCubeArrayU32Float:
650     return "NVPTXISD::TexUnifiedCubeArrayU32Float";
651   case NVPTXISD::TexUnifiedCubeArrayU32FloatLevel:
652     return "NVPTXISD::TexUnifiedCubeArrayU32FloatLevel";
653   case NVPTXISD::Tld4UnifiedR2DFloatFloat:
654     return "NVPTXISD::Tld4UnifiedR2DFloatFloat";
655   case NVPTXISD::Tld4UnifiedG2DFloatFloat:
656     return "NVPTXISD::Tld4UnifiedG2DFloatFloat";
657   case NVPTXISD::Tld4UnifiedB2DFloatFloat:
658     return "NVPTXISD::Tld4UnifiedB2DFloatFloat";
659   case NVPTXISD::Tld4UnifiedA2DFloatFloat:
660     return "NVPTXISD::Tld4UnifiedA2DFloatFloat";
661   case NVPTXISD::Tld4UnifiedR2DS64Float:
662     return "NVPTXISD::Tld4UnifiedR2DS64Float";
663   case NVPTXISD::Tld4UnifiedG2DS64Float:
664     return "NVPTXISD::Tld4UnifiedG2DS64Float";
665   case NVPTXISD::Tld4UnifiedB2DS64Float:
666     return "NVPTXISD::Tld4UnifiedB2DS64Float";
667   case NVPTXISD::Tld4UnifiedA2DS64Float:
668     return "NVPTXISD::Tld4UnifiedA2DS64Float";
669   case NVPTXISD::Tld4UnifiedR2DU64Float:
670     return "NVPTXISD::Tld4UnifiedR2DU64Float";
671   case NVPTXISD::Tld4UnifiedG2DU64Float:
672     return "NVPTXISD::Tld4UnifiedG2DU64Float";
673   case NVPTXISD::Tld4UnifiedB2DU64Float:
674     return "NVPTXISD::Tld4UnifiedB2DU64Float";
675   case NVPTXISD::Tld4UnifiedA2DU64Float:
676     return "NVPTXISD::Tld4UnifiedA2DU64Float";
677 
678   case NVPTXISD::Suld1DI8Clamp:          return "NVPTXISD::Suld1DI8Clamp";
679   case NVPTXISD::Suld1DI16Clamp:         return "NVPTXISD::Suld1DI16Clamp";
680   case NVPTXISD::Suld1DI32Clamp:         return "NVPTXISD::Suld1DI32Clamp";
681   case NVPTXISD::Suld1DI64Clamp:         return "NVPTXISD::Suld1DI64Clamp";
682   case NVPTXISD::Suld1DV2I8Clamp:        return "NVPTXISD::Suld1DV2I8Clamp";
683   case NVPTXISD::Suld1DV2I16Clamp:       return "NVPTXISD::Suld1DV2I16Clamp";
684   case NVPTXISD::Suld1DV2I32Clamp:       return "NVPTXISD::Suld1DV2I32Clamp";
685   case NVPTXISD::Suld1DV2I64Clamp:       return "NVPTXISD::Suld1DV2I64Clamp";
686   case NVPTXISD::Suld1DV4I8Clamp:        return "NVPTXISD::Suld1DV4I8Clamp";
687   case NVPTXISD::Suld1DV4I16Clamp:       return "NVPTXISD::Suld1DV4I16Clamp";
688   case NVPTXISD::Suld1DV4I32Clamp:       return "NVPTXISD::Suld1DV4I32Clamp";
689 
690   case NVPTXISD::Suld1DArrayI8Clamp:   return "NVPTXISD::Suld1DArrayI8Clamp";
691   case NVPTXISD::Suld1DArrayI16Clamp:  return "NVPTXISD::Suld1DArrayI16Clamp";
692   case NVPTXISD::Suld1DArrayI32Clamp:  return "NVPTXISD::Suld1DArrayI32Clamp";
693   case NVPTXISD::Suld1DArrayI64Clamp:  return "NVPTXISD::Suld1DArrayI64Clamp";
694   case NVPTXISD::Suld1DArrayV2I8Clamp: return "NVPTXISD::Suld1DArrayV2I8Clamp";
695   case NVPTXISD::Suld1DArrayV2I16Clamp:return "NVPTXISD::Suld1DArrayV2I16Clamp";
696   case NVPTXISD::Suld1DArrayV2I32Clamp:return "NVPTXISD::Suld1DArrayV2I32Clamp";
697   case NVPTXISD::Suld1DArrayV2I64Clamp:return "NVPTXISD::Suld1DArrayV2I64Clamp";
698   case NVPTXISD::Suld1DArrayV4I8Clamp: return "NVPTXISD::Suld1DArrayV4I8Clamp";
699   case NVPTXISD::Suld1DArrayV4I16Clamp:return "NVPTXISD::Suld1DArrayV4I16Clamp";
700   case NVPTXISD::Suld1DArrayV4I32Clamp:return "NVPTXISD::Suld1DArrayV4I32Clamp";
701 
702   case NVPTXISD::Suld2DI8Clamp:          return "NVPTXISD::Suld2DI8Clamp";
703   case NVPTXISD::Suld2DI16Clamp:         return "NVPTXISD::Suld2DI16Clamp";
704   case NVPTXISD::Suld2DI32Clamp:         return "NVPTXISD::Suld2DI32Clamp";
705   case NVPTXISD::Suld2DI64Clamp:         return "NVPTXISD::Suld2DI64Clamp";
706   case NVPTXISD::Suld2DV2I8Clamp:        return "NVPTXISD::Suld2DV2I8Clamp";
707   case NVPTXISD::Suld2DV2I16Clamp:       return "NVPTXISD::Suld2DV2I16Clamp";
708   case NVPTXISD::Suld2DV2I32Clamp:       return "NVPTXISD::Suld2DV2I32Clamp";
709   case NVPTXISD::Suld2DV2I64Clamp:       return "NVPTXISD::Suld2DV2I64Clamp";
710   case NVPTXISD::Suld2DV4I8Clamp:        return "NVPTXISD::Suld2DV4I8Clamp";
711   case NVPTXISD::Suld2DV4I16Clamp:       return "NVPTXISD::Suld2DV4I16Clamp";
712   case NVPTXISD::Suld2DV4I32Clamp:       return "NVPTXISD::Suld2DV4I32Clamp";
713 
714   case NVPTXISD::Suld2DArrayI8Clamp:   return "NVPTXISD::Suld2DArrayI8Clamp";
715   case NVPTXISD::Suld2DArrayI16Clamp:  return "NVPTXISD::Suld2DArrayI16Clamp";
716   case NVPTXISD::Suld2DArrayI32Clamp:  return "NVPTXISD::Suld2DArrayI32Clamp";
717   case NVPTXISD::Suld2DArrayI64Clamp:  return "NVPTXISD::Suld2DArrayI64Clamp";
718   case NVPTXISD::Suld2DArrayV2I8Clamp: return "NVPTXISD::Suld2DArrayV2I8Clamp";
719   case NVPTXISD::Suld2DArrayV2I16Clamp:return "NVPTXISD::Suld2DArrayV2I16Clamp";
720   case NVPTXISD::Suld2DArrayV2I32Clamp:return "NVPTXISD::Suld2DArrayV2I32Clamp";
721   case NVPTXISD::Suld2DArrayV2I64Clamp:return "NVPTXISD::Suld2DArrayV2I64Clamp";
722   case NVPTXISD::Suld2DArrayV4I8Clamp: return "NVPTXISD::Suld2DArrayV4I8Clamp";
723   case NVPTXISD::Suld2DArrayV4I16Clamp:return "NVPTXISD::Suld2DArrayV4I16Clamp";
724   case NVPTXISD::Suld2DArrayV4I32Clamp:return "NVPTXISD::Suld2DArrayV4I32Clamp";
725 
726   case NVPTXISD::Suld3DI8Clamp:          return "NVPTXISD::Suld3DI8Clamp";
727   case NVPTXISD::Suld3DI16Clamp:         return "NVPTXISD::Suld3DI16Clamp";
728   case NVPTXISD::Suld3DI32Clamp:         return "NVPTXISD::Suld3DI32Clamp";
729   case NVPTXISD::Suld3DI64Clamp:         return "NVPTXISD::Suld3DI64Clamp";
730   case NVPTXISD::Suld3DV2I8Clamp:        return "NVPTXISD::Suld3DV2I8Clamp";
731   case NVPTXISD::Suld3DV2I16Clamp:       return "NVPTXISD::Suld3DV2I16Clamp";
732   case NVPTXISD::Suld3DV2I32Clamp:       return "NVPTXISD::Suld3DV2I32Clamp";
733   case NVPTXISD::Suld3DV2I64Clamp:       return "NVPTXISD::Suld3DV2I64Clamp";
734   case NVPTXISD::Suld3DV4I8Clamp:        return "NVPTXISD::Suld3DV4I8Clamp";
735   case NVPTXISD::Suld3DV4I16Clamp:       return "NVPTXISD::Suld3DV4I16Clamp";
736   case NVPTXISD::Suld3DV4I32Clamp:       return "NVPTXISD::Suld3DV4I32Clamp";
737 
738   case NVPTXISD::Suld1DI8Trap:          return "NVPTXISD::Suld1DI8Trap";
739   case NVPTXISD::Suld1DI16Trap:         return "NVPTXISD::Suld1DI16Trap";
740   case NVPTXISD::Suld1DI32Trap:         return "NVPTXISD::Suld1DI32Trap";
741   case NVPTXISD::Suld1DI64Trap:         return "NVPTXISD::Suld1DI64Trap";
742   case NVPTXISD::Suld1DV2I8Trap:        return "NVPTXISD::Suld1DV2I8Trap";
743   case NVPTXISD::Suld1DV2I16Trap:       return "NVPTXISD::Suld1DV2I16Trap";
744   case NVPTXISD::Suld1DV2I32Trap:       return "NVPTXISD::Suld1DV2I32Trap";
745   case NVPTXISD::Suld1DV2I64Trap:       return "NVPTXISD::Suld1DV2I64Trap";
746   case NVPTXISD::Suld1DV4I8Trap:        return "NVPTXISD::Suld1DV4I8Trap";
747   case NVPTXISD::Suld1DV4I16Trap:       return "NVPTXISD::Suld1DV4I16Trap";
748   case NVPTXISD::Suld1DV4I32Trap:       return "NVPTXISD::Suld1DV4I32Trap";
749 
750   case NVPTXISD::Suld1DArrayI8Trap:     return "NVPTXISD::Suld1DArrayI8Trap";
751   case NVPTXISD::Suld1DArrayI16Trap:    return "NVPTXISD::Suld1DArrayI16Trap";
752   case NVPTXISD::Suld1DArrayI32Trap:    return "NVPTXISD::Suld1DArrayI32Trap";
753   case NVPTXISD::Suld1DArrayI64Trap:    return "NVPTXISD::Suld1DArrayI64Trap";
754   case NVPTXISD::Suld1DArrayV2I8Trap:   return "NVPTXISD::Suld1DArrayV2I8Trap";
755   case NVPTXISD::Suld1DArrayV2I16Trap:  return "NVPTXISD::Suld1DArrayV2I16Trap";
756   case NVPTXISD::Suld1DArrayV2I32Trap:  return "NVPTXISD::Suld1DArrayV2I32Trap";
757   case NVPTXISD::Suld1DArrayV2I64Trap:  return "NVPTXISD::Suld1DArrayV2I64Trap";
758   case NVPTXISD::Suld1DArrayV4I8Trap:   return "NVPTXISD::Suld1DArrayV4I8Trap";
759   case NVPTXISD::Suld1DArrayV4I16Trap:  return "NVPTXISD::Suld1DArrayV4I16Trap";
760   case NVPTXISD::Suld1DArrayV4I32Trap:  return "NVPTXISD::Suld1DArrayV4I32Trap";
761 
762   case NVPTXISD::Suld2DI8Trap:          return "NVPTXISD::Suld2DI8Trap";
763   case NVPTXISD::Suld2DI16Trap:         return "NVPTXISD::Suld2DI16Trap";
764   case NVPTXISD::Suld2DI32Trap:         return "NVPTXISD::Suld2DI32Trap";
765   case NVPTXISD::Suld2DI64Trap:         return "NVPTXISD::Suld2DI64Trap";
766   case NVPTXISD::Suld2DV2I8Trap:        return "NVPTXISD::Suld2DV2I8Trap";
767   case NVPTXISD::Suld2DV2I16Trap:       return "NVPTXISD::Suld2DV2I16Trap";
768   case NVPTXISD::Suld2DV2I32Trap:       return "NVPTXISD::Suld2DV2I32Trap";
769   case NVPTXISD::Suld2DV2I64Trap:       return "NVPTXISD::Suld2DV2I64Trap";
770   case NVPTXISD::Suld2DV4I8Trap:        return "NVPTXISD::Suld2DV4I8Trap";
771   case NVPTXISD::Suld2DV4I16Trap:       return "NVPTXISD::Suld2DV4I16Trap";
772   case NVPTXISD::Suld2DV4I32Trap:       return "NVPTXISD::Suld2DV4I32Trap";
773 
774   case NVPTXISD::Suld2DArrayI8Trap:     return "NVPTXISD::Suld2DArrayI8Trap";
775   case NVPTXISD::Suld2DArrayI16Trap:    return "NVPTXISD::Suld2DArrayI16Trap";
776   case NVPTXISD::Suld2DArrayI32Trap:    return "NVPTXISD::Suld2DArrayI32Trap";
777   case NVPTXISD::Suld2DArrayI64Trap:    return "NVPTXISD::Suld2DArrayI64Trap";
778   case NVPTXISD::Suld2DArrayV2I8Trap:   return "NVPTXISD::Suld2DArrayV2I8Trap";
779   case NVPTXISD::Suld2DArrayV2I16Trap:  return "NVPTXISD::Suld2DArrayV2I16Trap";
780   case NVPTXISD::Suld2DArrayV2I32Trap:  return "NVPTXISD::Suld2DArrayV2I32Trap";
781   case NVPTXISD::Suld2DArrayV2I64Trap:  return "NVPTXISD::Suld2DArrayV2I64Trap";
782   case NVPTXISD::Suld2DArrayV4I8Trap:   return "NVPTXISD::Suld2DArrayV4I8Trap";
783   case NVPTXISD::Suld2DArrayV4I16Trap:  return "NVPTXISD::Suld2DArrayV4I16Trap";
784   case NVPTXISD::Suld2DArrayV4I32Trap:  return "NVPTXISD::Suld2DArrayV4I32Trap";
785 
786   case NVPTXISD::Suld3DI8Trap:          return "NVPTXISD::Suld3DI8Trap";
787   case NVPTXISD::Suld3DI16Trap:         return "NVPTXISD::Suld3DI16Trap";
788   case NVPTXISD::Suld3DI32Trap:         return "NVPTXISD::Suld3DI32Trap";
789   case NVPTXISD::Suld3DI64Trap:         return "NVPTXISD::Suld3DI64Trap";
790   case NVPTXISD::Suld3DV2I8Trap:        return "NVPTXISD::Suld3DV2I8Trap";
791   case NVPTXISD::Suld3DV2I16Trap:       return "NVPTXISD::Suld3DV2I16Trap";
792   case NVPTXISD::Suld3DV2I32Trap:       return "NVPTXISD::Suld3DV2I32Trap";
793   case NVPTXISD::Suld3DV2I64Trap:       return "NVPTXISD::Suld3DV2I64Trap";
794   case NVPTXISD::Suld3DV4I8Trap:        return "NVPTXISD::Suld3DV4I8Trap";
795   case NVPTXISD::Suld3DV4I16Trap:       return "NVPTXISD::Suld3DV4I16Trap";
796   case NVPTXISD::Suld3DV4I32Trap:       return "NVPTXISD::Suld3DV4I32Trap";
797 
798   case NVPTXISD::Suld1DI8Zero:          return "NVPTXISD::Suld1DI8Zero";
799   case NVPTXISD::Suld1DI16Zero:         return "NVPTXISD::Suld1DI16Zero";
800   case NVPTXISD::Suld1DI32Zero:         return "NVPTXISD::Suld1DI32Zero";
801   case NVPTXISD::Suld1DI64Zero:         return "NVPTXISD::Suld1DI64Zero";
802   case NVPTXISD::Suld1DV2I8Zero:        return "NVPTXISD::Suld1DV2I8Zero";
803   case NVPTXISD::Suld1DV2I16Zero:       return "NVPTXISD::Suld1DV2I16Zero";
804   case NVPTXISD::Suld1DV2I32Zero:       return "NVPTXISD::Suld1DV2I32Zero";
805   case NVPTXISD::Suld1DV2I64Zero:       return "NVPTXISD::Suld1DV2I64Zero";
806   case NVPTXISD::Suld1DV4I8Zero:        return "NVPTXISD::Suld1DV4I8Zero";
807   case NVPTXISD::Suld1DV4I16Zero:       return "NVPTXISD::Suld1DV4I16Zero";
808   case NVPTXISD::Suld1DV4I32Zero:       return "NVPTXISD::Suld1DV4I32Zero";
809 
810   case NVPTXISD::Suld1DArrayI8Zero:     return "NVPTXISD::Suld1DArrayI8Zero";
811   case NVPTXISD::Suld1DArrayI16Zero:    return "NVPTXISD::Suld1DArrayI16Zero";
812   case NVPTXISD::Suld1DArrayI32Zero:    return "NVPTXISD::Suld1DArrayI32Zero";
813   case NVPTXISD::Suld1DArrayI64Zero:    return "NVPTXISD::Suld1DArrayI64Zero";
814   case NVPTXISD::Suld1DArrayV2I8Zero:   return "NVPTXISD::Suld1DArrayV2I8Zero";
815   case NVPTXISD::Suld1DArrayV2I16Zero:  return "NVPTXISD::Suld1DArrayV2I16Zero";
816   case NVPTXISD::Suld1DArrayV2I32Zero:  return "NVPTXISD::Suld1DArrayV2I32Zero";
817   case NVPTXISD::Suld1DArrayV2I64Zero:  return "NVPTXISD::Suld1DArrayV2I64Zero";
818   case NVPTXISD::Suld1DArrayV4I8Zero:   return "NVPTXISD::Suld1DArrayV4I8Zero";
819   case NVPTXISD::Suld1DArrayV4I16Zero:  return "NVPTXISD::Suld1DArrayV4I16Zero";
820   case NVPTXISD::Suld1DArrayV4I32Zero:  return "NVPTXISD::Suld1DArrayV4I32Zero";
821 
822   case NVPTXISD::Suld2DI8Zero:          return "NVPTXISD::Suld2DI8Zero";
823   case NVPTXISD::Suld2DI16Zero:         return "NVPTXISD::Suld2DI16Zero";
824   case NVPTXISD::Suld2DI32Zero:         return "NVPTXISD::Suld2DI32Zero";
825   case NVPTXISD::Suld2DI64Zero:         return "NVPTXISD::Suld2DI64Zero";
826   case NVPTXISD::Suld2DV2I8Zero:        return "NVPTXISD::Suld2DV2I8Zero";
827   case NVPTXISD::Suld2DV2I16Zero:       return "NVPTXISD::Suld2DV2I16Zero";
828   case NVPTXISD::Suld2DV2I32Zero:       return "NVPTXISD::Suld2DV2I32Zero";
829   case NVPTXISD::Suld2DV2I64Zero:       return "NVPTXISD::Suld2DV2I64Zero";
830   case NVPTXISD::Suld2DV4I8Zero:        return "NVPTXISD::Suld2DV4I8Zero";
831   case NVPTXISD::Suld2DV4I16Zero:       return "NVPTXISD::Suld2DV4I16Zero";
832   case NVPTXISD::Suld2DV4I32Zero:       return "NVPTXISD::Suld2DV4I32Zero";
833 
834   case NVPTXISD::Suld2DArrayI8Zero:     return "NVPTXISD::Suld2DArrayI8Zero";
835   case NVPTXISD::Suld2DArrayI16Zero:    return "NVPTXISD::Suld2DArrayI16Zero";
836   case NVPTXISD::Suld2DArrayI32Zero:    return "NVPTXISD::Suld2DArrayI32Zero";
837   case NVPTXISD::Suld2DArrayI64Zero:    return "NVPTXISD::Suld2DArrayI64Zero";
838   case NVPTXISD::Suld2DArrayV2I8Zero:   return "NVPTXISD::Suld2DArrayV2I8Zero";
839   case NVPTXISD::Suld2DArrayV2I16Zero:  return "NVPTXISD::Suld2DArrayV2I16Zero";
840   case NVPTXISD::Suld2DArrayV2I32Zero:  return "NVPTXISD::Suld2DArrayV2I32Zero";
841   case NVPTXISD::Suld2DArrayV2I64Zero:  return "NVPTXISD::Suld2DArrayV2I64Zero";
842   case NVPTXISD::Suld2DArrayV4I8Zero:   return "NVPTXISD::Suld2DArrayV4I8Zero";
843   case NVPTXISD::Suld2DArrayV4I16Zero:  return "NVPTXISD::Suld2DArrayV4I16Zero";
844   case NVPTXISD::Suld2DArrayV4I32Zero:  return "NVPTXISD::Suld2DArrayV4I32Zero";
845 
846   case NVPTXISD::Suld3DI8Zero:          return "NVPTXISD::Suld3DI8Zero";
847   case NVPTXISD::Suld3DI16Zero:         return "NVPTXISD::Suld3DI16Zero";
848   case NVPTXISD::Suld3DI32Zero:         return "NVPTXISD::Suld3DI32Zero";
849   case NVPTXISD::Suld3DI64Zero:         return "NVPTXISD::Suld3DI64Zero";
850   case NVPTXISD::Suld3DV2I8Zero:        return "NVPTXISD::Suld3DV2I8Zero";
851   case NVPTXISD::Suld3DV2I16Zero:       return "NVPTXISD::Suld3DV2I16Zero";
852   case NVPTXISD::Suld3DV2I32Zero:       return "NVPTXISD::Suld3DV2I32Zero";
853   case NVPTXISD::Suld3DV2I64Zero:       return "NVPTXISD::Suld3DV2I64Zero";
854   case NVPTXISD::Suld3DV4I8Zero:        return "NVPTXISD::Suld3DV4I8Zero";
855   case NVPTXISD::Suld3DV4I16Zero:       return "NVPTXISD::Suld3DV4I16Zero";
856   case NVPTXISD::Suld3DV4I32Zero:       return "NVPTXISD::Suld3DV4I32Zero";
857   }
858 }
859 
860 TargetLoweringBase::LegalizeTypeAction
getPreferredVectorAction(EVT VT) const861 NVPTXTargetLowering::getPreferredVectorAction(EVT VT) const {
862   if (VT.getVectorNumElements() != 1 && VT.getScalarType() == MVT::i1)
863     return TypeSplitVector;
864 
865   return TargetLoweringBase::getPreferredVectorAction(VT);
866 }
867 
868 SDValue
LowerGlobalAddress(SDValue Op,SelectionDAG & DAG) const869 NVPTXTargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
870   SDLoc dl(Op);
871   const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
872   Op = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
873   return DAG.getNode(NVPTXISD::Wrapper, dl, getPointerTy(), Op);
874 }
875 
876 std::string
getPrototype(Type * retTy,const ArgListTy & Args,const SmallVectorImpl<ISD::OutputArg> & Outs,unsigned retAlignment,const ImmutableCallSite * CS) const877 NVPTXTargetLowering::getPrototype(Type *retTy, const ArgListTy &Args,
878                                   const SmallVectorImpl<ISD::OutputArg> &Outs,
879                                   unsigned retAlignment,
880                                   const ImmutableCallSite *CS) const {
881 
882   bool isABI = (STI.getSmVersion() >= 20);
883   assert(isABI && "Non-ABI compilation is not supported");
884   if (!isABI)
885     return "";
886 
887   std::stringstream O;
888   O << "prototype_" << uniqueCallSite << " : .callprototype ";
889 
890   if (retTy->getTypeID() == Type::VoidTyID) {
891     O << "()";
892   } else {
893     O << "(";
894     if (retTy->isFloatingPointTy() || retTy->isIntegerTy()) {
895       unsigned size = 0;
896       if (const IntegerType *ITy = dyn_cast<IntegerType>(retTy)) {
897         size = ITy->getBitWidth();
898         if (size < 32)
899           size = 32;
900       } else {
901         assert(retTy->isFloatingPointTy() &&
902                "Floating point type expected here");
903         size = retTy->getPrimitiveSizeInBits();
904       }
905 
906       O << ".param .b" << size << " _";
907     } else if (isa<PointerType>(retTy)) {
908       O << ".param .b" << getPointerTy().getSizeInBits() << " _";
909     } else if ((retTy->getTypeID() == Type::StructTyID) ||
910                isa<VectorType>(retTy)) {
911       O << ".param .align "
912         << retAlignment
913         << " .b8 _["
914         << getDataLayout()->getTypeAllocSize(retTy) << "]";
915     } else {
916       llvm_unreachable("Unknown return type");
917     }
918     O << ") ";
919   }
920   O << "_ (";
921 
922   bool first = true;
923   MVT thePointerTy = getPointerTy();
924 
925   unsigned OIdx = 0;
926   for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
927     Type *Ty = Args[i].Ty;
928     if (!first) {
929       O << ", ";
930     }
931     first = false;
932 
933     if (!Outs[OIdx].Flags.isByVal()) {
934       if (Ty->isAggregateType() || Ty->isVectorTy()) {
935         unsigned align = 0;
936         const CallInst *CallI = cast<CallInst>(CS->getInstruction());
937         const DataLayout *TD = getDataLayout();
938         // +1 because index 0 is reserved for return type alignment
939         if (!llvm::getAlign(*CallI, i + 1, align))
940           align = TD->getABITypeAlignment(Ty);
941         unsigned sz = TD->getTypeAllocSize(Ty);
942         O << ".param .align " << align << " .b8 ";
943         O << "_";
944         O << "[" << sz << "]";
945         // update the index for Outs
946         SmallVector<EVT, 16> vtparts;
947         ComputeValueVTs(*this, Ty, vtparts);
948         if (unsigned len = vtparts.size())
949           OIdx += len - 1;
950         continue;
951       }
952        // i8 types in IR will be i16 types in SDAG
953       assert((getValueType(Ty) == Outs[OIdx].VT ||
954              (getValueType(Ty) == MVT::i8 && Outs[OIdx].VT == MVT::i16)) &&
955              "type mismatch between callee prototype and arguments");
956       // scalar type
957       unsigned sz = 0;
958       if (isa<IntegerType>(Ty)) {
959         sz = cast<IntegerType>(Ty)->getBitWidth();
960         if (sz < 32)
961           sz = 32;
962       } else if (isa<PointerType>(Ty))
963         sz = thePointerTy.getSizeInBits();
964       else
965         sz = Ty->getPrimitiveSizeInBits();
966       O << ".param .b" << sz << " ";
967       O << "_";
968       continue;
969     }
970     const PointerType *PTy = dyn_cast<PointerType>(Ty);
971     assert(PTy && "Param with byval attribute should be a pointer type");
972     Type *ETy = PTy->getElementType();
973 
974     unsigned align = Outs[OIdx].Flags.getByValAlign();
975     unsigned sz = getDataLayout()->getTypeAllocSize(ETy);
976     O << ".param .align " << align << " .b8 ";
977     O << "_";
978     O << "[" << sz << "]";
979   }
980   O << ");";
981   return O.str();
982 }
983 
984 unsigned
getArgumentAlignment(SDValue Callee,const ImmutableCallSite * CS,Type * Ty,unsigned Idx) const985 NVPTXTargetLowering::getArgumentAlignment(SDValue Callee,
986                                           const ImmutableCallSite *CS,
987                                           Type *Ty,
988                                           unsigned Idx) const {
989   const DataLayout *TD = getDataLayout();
990   unsigned Align = 0;
991   const Value *DirectCallee = CS->getCalledFunction();
992 
993   if (!DirectCallee) {
994     // We don't have a direct function symbol, but that may be because of
995     // constant cast instructions in the call.
996     const Instruction *CalleeI = CS->getInstruction();
997     assert(CalleeI && "Call target is not a function or derived value?");
998 
999     // With bitcast'd call targets, the instruction will be the call
1000     if (isa<CallInst>(CalleeI)) {
1001       // Check if we have call alignment metadata
1002       if (llvm::getAlign(*cast<CallInst>(CalleeI), Idx, Align))
1003         return Align;
1004 
1005       const Value *CalleeV = cast<CallInst>(CalleeI)->getCalledValue();
1006       // Ignore any bitcast instructions
1007       while(isa<ConstantExpr>(CalleeV)) {
1008         const ConstantExpr *CE = cast<ConstantExpr>(CalleeV);
1009         if (!CE->isCast())
1010           break;
1011         // Look through the bitcast
1012         CalleeV = cast<ConstantExpr>(CalleeV)->getOperand(0);
1013       }
1014 
1015       // We have now looked past all of the bitcasts.  Do we finally have a
1016       // Function?
1017       if (isa<Function>(CalleeV))
1018         DirectCallee = CalleeV;
1019     }
1020   }
1021 
1022   // Check for function alignment information if we found that the
1023   // ultimate target is a Function
1024   if (DirectCallee)
1025     if (llvm::getAlign(*cast<Function>(DirectCallee), Idx, Align))
1026       return Align;
1027 
1028   // Call is indirect or alignment information is not available, fall back to
1029   // the ABI type alignment
1030   return TD->getABITypeAlignment(Ty);
1031 }
1032 
LowerCall(TargetLowering::CallLoweringInfo & CLI,SmallVectorImpl<SDValue> & InVals) const1033 SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1034                                        SmallVectorImpl<SDValue> &InVals) const {
1035   SelectionDAG &DAG = CLI.DAG;
1036   SDLoc dl = CLI.DL;
1037   SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1038   SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1039   SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1040   SDValue Chain = CLI.Chain;
1041   SDValue Callee = CLI.Callee;
1042   bool &isTailCall = CLI.IsTailCall;
1043   ArgListTy &Args = CLI.getArgs();
1044   Type *retTy = CLI.RetTy;
1045   ImmutableCallSite *CS = CLI.CS;
1046 
1047   bool isABI = (STI.getSmVersion() >= 20);
1048   assert(isABI && "Non-ABI compilation is not supported");
1049   if (!isABI)
1050     return Chain;
1051   const DataLayout *TD = getDataLayout();
1052   MachineFunction &MF = DAG.getMachineFunction();
1053   const Function *F = MF.getFunction();
1054 
1055   SDValue tempChain = Chain;
1056   Chain =
1057       DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(uniqueCallSite, true),
1058                            dl);
1059   SDValue InFlag = Chain.getValue(1);
1060 
1061   unsigned paramCount = 0;
1062   // Args.size() and Outs.size() need not match.
1063   // Outs.size() will be larger
1064   //   * if there is an aggregate argument with multiple fields (each field
1065   //     showing up separately in Outs)
1066   //   * if there is a vector argument with more than typical vector-length
1067   //     elements (generally if more than 4) where each vector element is
1068   //     individually present in Outs.
1069   // So a different index should be used for indexing into Outs/OutVals.
1070   // See similar issue in LowerFormalArguments.
1071   unsigned OIdx = 0;
1072   // Declare the .params or .reg need to pass values
1073   // to the function
1074   for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
1075     EVT VT = Outs[OIdx].VT;
1076     Type *Ty = Args[i].Ty;
1077 
1078     if (!Outs[OIdx].Flags.isByVal()) {
1079       if (Ty->isAggregateType()) {
1080         // aggregate
1081         SmallVector<EVT, 16> vtparts;
1082         SmallVector<uint64_t, 16> Offsets;
1083         ComputePTXValueVTs(*this, Ty, vtparts, &Offsets, 0);
1084 
1085         unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
1086         // declare .param .align <align> .b8 .param<n>[<size>];
1087         unsigned sz = TD->getTypeAllocSize(Ty);
1088         SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1089         SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, MVT::i32),
1090                                       DAG.getConstant(paramCount, MVT::i32),
1091                                       DAG.getConstant(sz, MVT::i32), InFlag };
1092         Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1093                             DeclareParamOps);
1094         InFlag = Chain.getValue(1);
1095         for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
1096           EVT elemtype = vtparts[j];
1097           unsigned ArgAlign = GreatestCommonDivisor64(align, Offsets[j]);
1098           if (elemtype.isInteger() && (sz < 8))
1099             sz = 8;
1100           SDValue StVal = OutVals[OIdx];
1101           if (elemtype.getSizeInBits() < 16) {
1102             StVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, StVal);
1103           }
1104           SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1105           SDValue CopyParamOps[] = { Chain,
1106                                      DAG.getConstant(paramCount, MVT::i32),
1107                                      DAG.getConstant(Offsets[j], MVT::i32),
1108                                      StVal, InFlag };
1109           Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
1110                                           CopyParamVTs, CopyParamOps,
1111                                           elemtype, MachinePointerInfo(),
1112                                           ArgAlign);
1113           InFlag = Chain.getValue(1);
1114           ++OIdx;
1115         }
1116         if (vtparts.size() > 0)
1117           --OIdx;
1118         ++paramCount;
1119         continue;
1120       }
1121       if (Ty->isVectorTy()) {
1122         EVT ObjectVT = getValueType(Ty);
1123         unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
1124         // declare .param .align <align> .b8 .param<n>[<size>];
1125         unsigned sz = TD->getTypeAllocSize(Ty);
1126         SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1127         SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, MVT::i32),
1128                                       DAG.getConstant(paramCount, MVT::i32),
1129                                       DAG.getConstant(sz, MVT::i32), InFlag };
1130         Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1131                             DeclareParamOps);
1132         InFlag = Chain.getValue(1);
1133         unsigned NumElts = ObjectVT.getVectorNumElements();
1134         EVT EltVT = ObjectVT.getVectorElementType();
1135         EVT MemVT = EltVT;
1136         bool NeedExtend = false;
1137         if (EltVT.getSizeInBits() < 16) {
1138           NeedExtend = true;
1139           EltVT = MVT::i16;
1140         }
1141 
1142         // V1 store
1143         if (NumElts == 1) {
1144           SDValue Elt = OutVals[OIdx++];
1145           if (NeedExtend)
1146             Elt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt);
1147 
1148           SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1149           SDValue CopyParamOps[] = { Chain,
1150                                      DAG.getConstant(paramCount, MVT::i32),
1151                                      DAG.getConstant(0, MVT::i32), Elt,
1152                                      InFlag };
1153           Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
1154                                           CopyParamVTs, CopyParamOps,
1155                                           MemVT, MachinePointerInfo());
1156           InFlag = Chain.getValue(1);
1157         } else if (NumElts == 2) {
1158           SDValue Elt0 = OutVals[OIdx++];
1159           SDValue Elt1 = OutVals[OIdx++];
1160           if (NeedExtend) {
1161             Elt0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt0);
1162             Elt1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt1);
1163           }
1164 
1165           SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1166           SDValue CopyParamOps[] = { Chain,
1167                                      DAG.getConstant(paramCount, MVT::i32),
1168                                      DAG.getConstant(0, MVT::i32), Elt0, Elt1,
1169                                      InFlag };
1170           Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParamV2, dl,
1171                                           CopyParamVTs, CopyParamOps,
1172                                           MemVT, MachinePointerInfo());
1173           InFlag = Chain.getValue(1);
1174         } else {
1175           unsigned curOffset = 0;
1176           // V4 stores
1177           // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
1178           // the
1179           // vector will be expanded to a power of 2 elements, so we know we can
1180           // always round up to the next multiple of 4 when creating the vector
1181           // stores.
1182           // e.g.  4 elem => 1 st.v4
1183           //       6 elem => 2 st.v4
1184           //       8 elem => 2 st.v4
1185           //      11 elem => 3 st.v4
1186           unsigned VecSize = 4;
1187           if (EltVT.getSizeInBits() == 64)
1188             VecSize = 2;
1189 
1190           // This is potentially only part of a vector, so assume all elements
1191           // are packed together.
1192           unsigned PerStoreOffset = MemVT.getStoreSizeInBits() / 8 * VecSize;
1193 
1194           for (unsigned i = 0; i < NumElts; i += VecSize) {
1195             // Get values
1196             SDValue StoreVal;
1197             SmallVector<SDValue, 8> Ops;
1198             Ops.push_back(Chain);
1199             Ops.push_back(DAG.getConstant(paramCount, MVT::i32));
1200             Ops.push_back(DAG.getConstant(curOffset, MVT::i32));
1201 
1202             unsigned Opc = NVPTXISD::StoreParamV2;
1203 
1204             StoreVal = OutVals[OIdx++];
1205             if (NeedExtend)
1206               StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1207             Ops.push_back(StoreVal);
1208 
1209             if (i + 1 < NumElts) {
1210               StoreVal = OutVals[OIdx++];
1211               if (NeedExtend)
1212                 StoreVal =
1213                     DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1214             } else {
1215               StoreVal = DAG.getUNDEF(EltVT);
1216             }
1217             Ops.push_back(StoreVal);
1218 
1219             if (VecSize == 4) {
1220               Opc = NVPTXISD::StoreParamV4;
1221               if (i + 2 < NumElts) {
1222                 StoreVal = OutVals[OIdx++];
1223                 if (NeedExtend)
1224                   StoreVal =
1225                       DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1226               } else {
1227                 StoreVal = DAG.getUNDEF(EltVT);
1228               }
1229               Ops.push_back(StoreVal);
1230 
1231               if (i + 3 < NumElts) {
1232                 StoreVal = OutVals[OIdx++];
1233                 if (NeedExtend)
1234                   StoreVal =
1235                       DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1236               } else {
1237                 StoreVal = DAG.getUNDEF(EltVT);
1238               }
1239               Ops.push_back(StoreVal);
1240             }
1241 
1242             Ops.push_back(InFlag);
1243 
1244             SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1245             Chain = DAG.getMemIntrinsicNode(Opc, dl, CopyParamVTs, Ops,
1246                                             MemVT, MachinePointerInfo());
1247             InFlag = Chain.getValue(1);
1248             curOffset += PerStoreOffset;
1249           }
1250         }
1251         ++paramCount;
1252         --OIdx;
1253         continue;
1254       }
1255       // Plain scalar
1256       // for ABI,    declare .param .b<size> .param<n>;
1257       unsigned sz = VT.getSizeInBits();
1258       bool needExtend = false;
1259       if (VT.isInteger()) {
1260         if (sz < 16)
1261           needExtend = true;
1262         if (sz < 32)
1263           sz = 32;
1264       }
1265       SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1266       SDValue DeclareParamOps[] = { Chain,
1267                                     DAG.getConstant(paramCount, MVT::i32),
1268                                     DAG.getConstant(sz, MVT::i32),
1269                                     DAG.getConstant(0, MVT::i32), InFlag };
1270       Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs,
1271                           DeclareParamOps);
1272       InFlag = Chain.getValue(1);
1273       SDValue OutV = OutVals[OIdx];
1274       if (needExtend) {
1275         // zext/sext i1 to i16
1276         unsigned opc = ISD::ZERO_EXTEND;
1277         if (Outs[OIdx].Flags.isSExt())
1278           opc = ISD::SIGN_EXTEND;
1279         OutV = DAG.getNode(opc, dl, MVT::i16, OutV);
1280       }
1281       SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1282       SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount, MVT::i32),
1283                                  DAG.getConstant(0, MVT::i32), OutV, InFlag };
1284 
1285       unsigned opcode = NVPTXISD::StoreParam;
1286       if (Outs[OIdx].Flags.isZExt())
1287         opcode = NVPTXISD::StoreParamU32;
1288       else if (Outs[OIdx].Flags.isSExt())
1289         opcode = NVPTXISD::StoreParamS32;
1290       Chain = DAG.getMemIntrinsicNode(opcode, dl, CopyParamVTs, CopyParamOps,
1291                                       VT, MachinePointerInfo());
1292 
1293       InFlag = Chain.getValue(1);
1294       ++paramCount;
1295       continue;
1296     }
1297     // struct or vector
1298     SmallVector<EVT, 16> vtparts;
1299     SmallVector<uint64_t, 16> Offsets;
1300     const PointerType *PTy = dyn_cast<PointerType>(Args[i].Ty);
1301     assert(PTy && "Type of a byval parameter should be pointer");
1302     ComputePTXValueVTs(*this, PTy->getElementType(), vtparts, &Offsets, 0);
1303 
1304     // declare .param .align <align> .b8 .param<n>[<size>];
1305     unsigned sz = Outs[OIdx].Flags.getByValSize();
1306     SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1307     unsigned ArgAlign = Outs[OIdx].Flags.getByValAlign();
1308     // The ByValAlign in the Outs[OIdx].Flags is alway set at this point,
1309     // so we don't need to worry about natural alignment or not.
1310     // See TargetLowering::LowerCallTo().
1311     SDValue DeclareParamOps[] = {
1312       Chain, DAG.getConstant(Outs[OIdx].Flags.getByValAlign(), MVT::i32),
1313       DAG.getConstant(paramCount, MVT::i32), DAG.getConstant(sz, MVT::i32),
1314       InFlag
1315     };
1316     Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1317                         DeclareParamOps);
1318     InFlag = Chain.getValue(1);
1319     for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
1320       EVT elemtype = vtparts[j];
1321       int curOffset = Offsets[j];
1322       unsigned PartAlign = GreatestCommonDivisor64(ArgAlign, curOffset);
1323       SDValue srcAddr =
1324           DAG.getNode(ISD::ADD, dl, getPointerTy(), OutVals[OIdx],
1325                       DAG.getConstant(curOffset, getPointerTy()));
1326       SDValue theVal = DAG.getLoad(elemtype, dl, tempChain, srcAddr,
1327                                    MachinePointerInfo(), false, false, false,
1328                                    PartAlign);
1329       if (elemtype.getSizeInBits() < 16) {
1330         theVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, theVal);
1331       }
1332       SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1333       SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount, MVT::i32),
1334                                  DAG.getConstant(curOffset, MVT::i32), theVal,
1335                                  InFlag };
1336       Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl, CopyParamVTs,
1337                                       CopyParamOps, elemtype,
1338                                       MachinePointerInfo());
1339 
1340       InFlag = Chain.getValue(1);
1341     }
1342     ++paramCount;
1343   }
1344 
1345   GlobalAddressSDNode *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode());
1346   unsigned retAlignment = 0;
1347 
1348   // Handle Result
1349   if (Ins.size() > 0) {
1350     SmallVector<EVT, 16> resvtparts;
1351     ComputeValueVTs(*this, retTy, resvtparts);
1352 
1353     // Declare
1354     //  .param .align 16 .b8 retval0[<size-in-bytes>], or
1355     //  .param .b<size-in-bits> retval0
1356     unsigned resultsz = TD->getTypeAllocSizeInBits(retTy);
1357     // Emit ".param .b<size-in-bits> retval0" instead of byte arrays only for
1358     // these three types to match the logic in
1359     // NVPTXAsmPrinter::printReturnValStr and NVPTXTargetLowering::getPrototype.
1360     // Plus, this behavior is consistent with nvcc's.
1361     if (retTy->isFloatingPointTy() || retTy->isIntegerTy() ||
1362         retTy->isPointerTy()) {
1363       // Scalar needs to be at least 32bit wide
1364       if (resultsz < 32)
1365         resultsz = 32;
1366       SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1367       SDValue DeclareRetOps[] = { Chain, DAG.getConstant(1, MVT::i32),
1368                                   DAG.getConstant(resultsz, MVT::i32),
1369                                   DAG.getConstant(0, MVT::i32), InFlag };
1370       Chain = DAG.getNode(NVPTXISD::DeclareRet, dl, DeclareRetVTs,
1371                           DeclareRetOps);
1372       InFlag = Chain.getValue(1);
1373     } else {
1374       retAlignment = getArgumentAlignment(Callee, CS, retTy, 0);
1375       SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1376       SDValue DeclareRetOps[] = { Chain,
1377                                   DAG.getConstant(retAlignment, MVT::i32),
1378                                   DAG.getConstant(resultsz / 8, MVT::i32),
1379                                   DAG.getConstant(0, MVT::i32), InFlag };
1380       Chain = DAG.getNode(NVPTXISD::DeclareRetParam, dl, DeclareRetVTs,
1381                           DeclareRetOps);
1382       InFlag = Chain.getValue(1);
1383     }
1384   }
1385 
1386   if (!Func) {
1387     // This is indirect function call case : PTX requires a prototype of the
1388     // form
1389     // proto_0 : .callprototype(.param .b32 _) _ (.param .b32 _);
1390     // to be emitted, and the label has to used as the last arg of call
1391     // instruction.
1392     // The prototype is embedded in a string and put as the operand for a
1393     // CallPrototype SDNode which will print out to the value of the string.
1394     SDVTList ProtoVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1395     std::string Proto = getPrototype(retTy, Args, Outs, retAlignment, CS);
1396     const char *ProtoStr =
1397       nvTM->getManagedStrPool()->getManagedString(Proto.c_str())->c_str();
1398     SDValue ProtoOps[] = {
1399       Chain, DAG.getTargetExternalSymbol(ProtoStr, MVT::i32), InFlag,
1400     };
1401     Chain = DAG.getNode(NVPTXISD::CallPrototype, dl, ProtoVTs, ProtoOps);
1402     InFlag = Chain.getValue(1);
1403   }
1404   // Op to just print "call"
1405   SDVTList PrintCallVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1406   SDValue PrintCallOps[] = {
1407     Chain, DAG.getConstant((Ins.size() == 0) ? 0 : 1, MVT::i32), InFlag
1408   };
1409   Chain = DAG.getNode(Func ? (NVPTXISD::PrintCallUni) : (NVPTXISD::PrintCall),
1410                       dl, PrintCallVTs, PrintCallOps);
1411   InFlag = Chain.getValue(1);
1412 
1413   // Ops to print out the function name
1414   SDVTList CallVoidVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1415   SDValue CallVoidOps[] = { Chain, Callee, InFlag };
1416   Chain = DAG.getNode(NVPTXISD::CallVoid, dl, CallVoidVTs, CallVoidOps);
1417   InFlag = Chain.getValue(1);
1418 
1419   // Ops to print out the param list
1420   SDVTList CallArgBeginVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1421   SDValue CallArgBeginOps[] = { Chain, InFlag };
1422   Chain = DAG.getNode(NVPTXISD::CallArgBegin, dl, CallArgBeginVTs,
1423                       CallArgBeginOps);
1424   InFlag = Chain.getValue(1);
1425 
1426   for (unsigned i = 0, e = paramCount; i != e; ++i) {
1427     unsigned opcode;
1428     if (i == (e - 1))
1429       opcode = NVPTXISD::LastCallArg;
1430     else
1431       opcode = NVPTXISD::CallArg;
1432     SDVTList CallArgVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1433     SDValue CallArgOps[] = { Chain, DAG.getConstant(1, MVT::i32),
1434                              DAG.getConstant(i, MVT::i32), InFlag };
1435     Chain = DAG.getNode(opcode, dl, CallArgVTs, CallArgOps);
1436     InFlag = Chain.getValue(1);
1437   }
1438   SDVTList CallArgEndVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1439   SDValue CallArgEndOps[] = { Chain, DAG.getConstant(Func ? 1 : 0, MVT::i32),
1440                               InFlag };
1441   Chain = DAG.getNode(NVPTXISD::CallArgEnd, dl, CallArgEndVTs, CallArgEndOps);
1442   InFlag = Chain.getValue(1);
1443 
1444   if (!Func) {
1445     SDVTList PrototypeVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1446     SDValue PrototypeOps[] = { Chain, DAG.getConstant(uniqueCallSite, MVT::i32),
1447                                InFlag };
1448     Chain = DAG.getNode(NVPTXISD::Prototype, dl, PrototypeVTs, PrototypeOps);
1449     InFlag = Chain.getValue(1);
1450   }
1451 
1452   // Generate loads from param memory/moves from registers for result
1453   if (Ins.size() > 0) {
1454     if (retTy && retTy->isVectorTy()) {
1455       EVT ObjectVT = getValueType(retTy);
1456       unsigned NumElts = ObjectVT.getVectorNumElements();
1457       EVT EltVT = ObjectVT.getVectorElementType();
1458       assert(STI.getTargetLowering()->getNumRegisters(F->getContext(),
1459                                                       ObjectVT) == NumElts &&
1460              "Vector was not scalarized");
1461       unsigned sz = EltVT.getSizeInBits();
1462       bool needTruncate = sz < 8;
1463 
1464       if (NumElts == 1) {
1465         // Just a simple load
1466         SmallVector<EVT, 4> LoadRetVTs;
1467         if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1468           // If loading i1/i8 result, generate
1469           //   load.b8 i16
1470           //   if i1
1471           //   trunc i16 to i1
1472           LoadRetVTs.push_back(MVT::i16);
1473         } else
1474           LoadRetVTs.push_back(EltVT);
1475         LoadRetVTs.push_back(MVT::Other);
1476         LoadRetVTs.push_back(MVT::Glue);
1477         SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, MVT::i32),
1478                                 DAG.getConstant(0, MVT::i32), InFlag};
1479         SDValue retval = DAG.getMemIntrinsicNode(
1480             NVPTXISD::LoadParam, dl,
1481             DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
1482         Chain = retval.getValue(1);
1483         InFlag = retval.getValue(2);
1484         SDValue Ret0 = retval;
1485         if (needTruncate)
1486           Ret0 = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Ret0);
1487         InVals.push_back(Ret0);
1488       } else if (NumElts == 2) {
1489         // LoadV2
1490         SmallVector<EVT, 4> LoadRetVTs;
1491         if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1492           // If loading i1/i8 result, generate
1493           //   load.b8 i16
1494           //   if i1
1495           //   trunc i16 to i1
1496           LoadRetVTs.push_back(MVT::i16);
1497           LoadRetVTs.push_back(MVT::i16);
1498         } else {
1499           LoadRetVTs.push_back(EltVT);
1500           LoadRetVTs.push_back(EltVT);
1501         }
1502         LoadRetVTs.push_back(MVT::Other);
1503         LoadRetVTs.push_back(MVT::Glue);
1504         SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, MVT::i32),
1505                                 DAG.getConstant(0, MVT::i32), InFlag};
1506         SDValue retval = DAG.getMemIntrinsicNode(
1507             NVPTXISD::LoadParamV2, dl,
1508             DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
1509         Chain = retval.getValue(2);
1510         InFlag = retval.getValue(3);
1511         SDValue Ret0 = retval.getValue(0);
1512         SDValue Ret1 = retval.getValue(1);
1513         if (needTruncate) {
1514           Ret0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret0);
1515           InVals.push_back(Ret0);
1516           Ret1 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret1);
1517           InVals.push_back(Ret1);
1518         } else {
1519           InVals.push_back(Ret0);
1520           InVals.push_back(Ret1);
1521         }
1522       } else {
1523         // Split into N LoadV4
1524         unsigned Ofst = 0;
1525         unsigned VecSize = 4;
1526         unsigned Opc = NVPTXISD::LoadParamV4;
1527         if (EltVT.getSizeInBits() == 64) {
1528           VecSize = 2;
1529           Opc = NVPTXISD::LoadParamV2;
1530         }
1531         EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
1532         for (unsigned i = 0; i < NumElts; i += VecSize) {
1533           SmallVector<EVT, 8> LoadRetVTs;
1534           if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1535             // If loading i1/i8 result, generate
1536             //   load.b8 i16
1537             //   if i1
1538             //   trunc i16 to i1
1539             for (unsigned j = 0; j < VecSize; ++j)
1540               LoadRetVTs.push_back(MVT::i16);
1541           } else {
1542             for (unsigned j = 0; j < VecSize; ++j)
1543               LoadRetVTs.push_back(EltVT);
1544           }
1545           LoadRetVTs.push_back(MVT::Other);
1546           LoadRetVTs.push_back(MVT::Glue);
1547           SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, MVT::i32),
1548                                   DAG.getConstant(Ofst, MVT::i32), InFlag};
1549           SDValue retval = DAG.getMemIntrinsicNode(
1550               Opc, dl, DAG.getVTList(LoadRetVTs),
1551               LoadRetOps, EltVT, MachinePointerInfo());
1552           if (VecSize == 2) {
1553             Chain = retval.getValue(2);
1554             InFlag = retval.getValue(3);
1555           } else {
1556             Chain = retval.getValue(4);
1557             InFlag = retval.getValue(5);
1558           }
1559 
1560           for (unsigned j = 0; j < VecSize; ++j) {
1561             if (i + j >= NumElts)
1562               break;
1563             SDValue Elt = retval.getValue(j);
1564             if (needTruncate)
1565               Elt = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
1566             InVals.push_back(Elt);
1567           }
1568           Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
1569         }
1570       }
1571     } else {
1572       SmallVector<EVT, 16> VTs;
1573       SmallVector<uint64_t, 16> Offsets;
1574       ComputePTXValueVTs(*this, retTy, VTs, &Offsets, 0);
1575       assert(VTs.size() == Ins.size() && "Bad value decomposition");
1576       unsigned RetAlign = getArgumentAlignment(Callee, CS, retTy, 0);
1577       for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
1578         unsigned sz = VTs[i].getSizeInBits();
1579         unsigned AlignI = GreatestCommonDivisor64(RetAlign, Offsets[i]);
1580         bool needTruncate = sz < 8;
1581         if (VTs[i].isInteger() && (sz < 8))
1582           sz = 8;
1583 
1584         SmallVector<EVT, 4> LoadRetVTs;
1585         EVT TheLoadType = VTs[i];
1586         if (retTy->isIntegerTy() &&
1587             TD->getTypeAllocSizeInBits(retTy) < 32) {
1588           // This is for integer types only, and specifically not for
1589           // aggregates.
1590           LoadRetVTs.push_back(MVT::i32);
1591           TheLoadType = MVT::i32;
1592         } else if (sz < 16) {
1593           // If loading i1/i8 result, generate
1594           //   load i8 (-> i16)
1595           //   trunc i16 to i1/i8
1596           LoadRetVTs.push_back(MVT::i16);
1597         } else
1598           LoadRetVTs.push_back(Ins[i].VT);
1599         LoadRetVTs.push_back(MVT::Other);
1600         LoadRetVTs.push_back(MVT::Glue);
1601 
1602         SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, MVT::i32),
1603                                 DAG.getConstant(Offsets[i], MVT::i32), InFlag};
1604         SDValue retval = DAG.getMemIntrinsicNode(
1605             NVPTXISD::LoadParam, dl,
1606             DAG.getVTList(LoadRetVTs), LoadRetOps,
1607             TheLoadType, MachinePointerInfo(), AlignI);
1608         Chain = retval.getValue(1);
1609         InFlag = retval.getValue(2);
1610         SDValue Ret0 = retval.getValue(0);
1611         if (needTruncate)
1612           Ret0 = DAG.getNode(ISD::TRUNCATE, dl, Ins[i].VT, Ret0);
1613         InVals.push_back(Ret0);
1614       }
1615     }
1616   }
1617 
1618   Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(uniqueCallSite, true),
1619                              DAG.getIntPtrConstant(uniqueCallSite + 1, true),
1620                              InFlag, dl);
1621   uniqueCallSite++;
1622 
1623   // set isTailCall to false for now, until we figure out how to express
1624   // tail call optimization in PTX
1625   isTailCall = false;
1626   return Chain;
1627 }
1628 
1629 // By default CONCAT_VECTORS is lowered by ExpandVectorBuildThroughStack()
1630 // (see LegalizeDAG.cpp). This is slow and uses local memory.
1631 // We use extract/insert/build vector just as what LegalizeOp() does in llvm 2.5
1632 SDValue
LowerCONCAT_VECTORS(SDValue Op,SelectionDAG & DAG) const1633 NVPTXTargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
1634   SDNode *Node = Op.getNode();
1635   SDLoc dl(Node);
1636   SmallVector<SDValue, 8> Ops;
1637   unsigned NumOperands = Node->getNumOperands();
1638   for (unsigned i = 0; i < NumOperands; ++i) {
1639     SDValue SubOp = Node->getOperand(i);
1640     EVT VVT = SubOp.getNode()->getValueType(0);
1641     EVT EltVT = VVT.getVectorElementType();
1642     unsigned NumSubElem = VVT.getVectorNumElements();
1643     for (unsigned j = 0; j < NumSubElem; ++j) {
1644       Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, SubOp,
1645                                 DAG.getIntPtrConstant(j)));
1646     }
1647   }
1648   return DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Ops);
1649 }
1650 
1651 /// LowerShiftRightParts - Lower SRL_PARTS, SRA_PARTS, which
1652 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
1653 ///    amount, or
1654 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
1655 ///    amount.
LowerShiftRightParts(SDValue Op,SelectionDAG & DAG) const1656 SDValue NVPTXTargetLowering::LowerShiftRightParts(SDValue Op,
1657                                                   SelectionDAG &DAG) const {
1658   assert(Op.getNumOperands() == 3 && "Not a double-shift!");
1659   assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
1660 
1661   EVT VT = Op.getValueType();
1662   unsigned VTBits = VT.getSizeInBits();
1663   SDLoc dl(Op);
1664   SDValue ShOpLo = Op.getOperand(0);
1665   SDValue ShOpHi = Op.getOperand(1);
1666   SDValue ShAmt  = Op.getOperand(2);
1667   unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
1668 
1669   if (VTBits == 32 && STI.getSmVersion() >= 35) {
1670 
1671     // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
1672     // {dHi, dLo} = {aHi, aLo} >> Amt
1673     //   dHi = aHi >> Amt
1674     //   dLo = shf.r.clamp aLo, aHi, Amt
1675 
1676     SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
1677     SDValue Lo = DAG.getNode(NVPTXISD::FUN_SHFR_CLAMP, dl, VT, ShOpLo, ShOpHi,
1678                              ShAmt);
1679 
1680     SDValue Ops[2] = { Lo, Hi };
1681     return DAG.getMergeValues(Ops, dl);
1682   }
1683   else {
1684 
1685     // {dHi, dLo} = {aHi, aLo} >> Amt
1686     // - if (Amt>=size) then
1687     //      dLo = aHi >> (Amt-size)
1688     //      dHi = aHi >> Amt (this is either all 0 or all 1)
1689     //   else
1690     //      dLo = (aLo >>logic Amt) | (aHi << (size-Amt))
1691     //      dHi = aHi >> Amt
1692 
1693     SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
1694                                    DAG.getConstant(VTBits, MVT::i32), ShAmt);
1695     SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
1696     SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
1697                                      DAG.getConstant(VTBits, MVT::i32));
1698     SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
1699     SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
1700     SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
1701 
1702     SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
1703                                DAG.getConstant(VTBits, MVT::i32), ISD::SETGE);
1704     SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
1705     SDValue Lo = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
1706 
1707     SDValue Ops[2] = { Lo, Hi };
1708     return DAG.getMergeValues(Ops, dl);
1709   }
1710 }
1711 
1712 /// LowerShiftLeftParts - Lower SHL_PARTS, which
1713 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
1714 ///    amount, or
1715 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
1716 ///    amount.
LowerShiftLeftParts(SDValue Op,SelectionDAG & DAG) const1717 SDValue NVPTXTargetLowering::LowerShiftLeftParts(SDValue Op,
1718                                                  SelectionDAG &DAG) const {
1719   assert(Op.getNumOperands() == 3 && "Not a double-shift!");
1720   assert(Op.getOpcode() == ISD::SHL_PARTS);
1721 
1722   EVT VT = Op.getValueType();
1723   unsigned VTBits = VT.getSizeInBits();
1724   SDLoc dl(Op);
1725   SDValue ShOpLo = Op.getOperand(0);
1726   SDValue ShOpHi = Op.getOperand(1);
1727   SDValue ShAmt  = Op.getOperand(2);
1728 
1729   if (VTBits == 32 && STI.getSmVersion() >= 35) {
1730 
1731     // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
1732     // {dHi, dLo} = {aHi, aLo} << Amt
1733     //   dHi = shf.l.clamp aLo, aHi, Amt
1734     //   dLo = aLo << Amt
1735 
1736     SDValue Hi = DAG.getNode(NVPTXISD::FUN_SHFL_CLAMP, dl, VT, ShOpLo, ShOpHi,
1737                              ShAmt);
1738     SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
1739 
1740     SDValue Ops[2] = { Lo, Hi };
1741     return DAG.getMergeValues(Ops, dl);
1742   }
1743   else {
1744 
1745     // {dHi, dLo} = {aHi, aLo} << Amt
1746     // - if (Amt>=size) then
1747     //      dLo = aLo << Amt (all 0)
1748     //      dLo = aLo << (Amt-size)
1749     //   else
1750     //      dLo = aLo << Amt
1751     //      dHi = (aHi << Amt) | (aLo >> (size-Amt))
1752 
1753     SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
1754                                    DAG.getConstant(VTBits, MVT::i32), ShAmt);
1755     SDValue Tmp1 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
1756     SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
1757                                      DAG.getConstant(VTBits, MVT::i32));
1758     SDValue Tmp2 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
1759     SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
1760     SDValue TrueVal = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
1761 
1762     SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
1763                                DAG.getConstant(VTBits, MVT::i32), ISD::SETGE);
1764     SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
1765     SDValue Hi = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
1766 
1767     SDValue Ops[2] = { Lo, Hi };
1768     return DAG.getMergeValues(Ops, dl);
1769   }
1770 }
1771 
1772 SDValue
LowerOperation(SDValue Op,SelectionDAG & DAG) const1773 NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
1774   switch (Op.getOpcode()) {
1775   case ISD::RETURNADDR:
1776     return SDValue();
1777   case ISD::FRAMEADDR:
1778     return SDValue();
1779   case ISD::GlobalAddress:
1780     return LowerGlobalAddress(Op, DAG);
1781   case ISD::INTRINSIC_W_CHAIN:
1782     return Op;
1783   case ISD::BUILD_VECTOR:
1784   case ISD::EXTRACT_SUBVECTOR:
1785     return Op;
1786   case ISD::CONCAT_VECTORS:
1787     return LowerCONCAT_VECTORS(Op, DAG);
1788   case ISD::STORE:
1789     return LowerSTORE(Op, DAG);
1790   case ISD::LOAD:
1791     return LowerLOAD(Op, DAG);
1792   case ISD::SHL_PARTS:
1793     return LowerShiftLeftParts(Op, DAG);
1794   case ISD::SRA_PARTS:
1795   case ISD::SRL_PARTS:
1796     return LowerShiftRightParts(Op, DAG);
1797   case ISD::SELECT:
1798     return LowerSelect(Op, DAG);
1799   default:
1800     llvm_unreachable("Custom lowering not defined for operation");
1801   }
1802 }
1803 
LowerSelect(SDValue Op,SelectionDAG & DAG) const1804 SDValue NVPTXTargetLowering::LowerSelect(SDValue Op, SelectionDAG &DAG) const {
1805   SDValue Op0 = Op->getOperand(0);
1806   SDValue Op1 = Op->getOperand(1);
1807   SDValue Op2 = Op->getOperand(2);
1808   SDLoc DL(Op.getNode());
1809 
1810   assert(Op.getValueType() == MVT::i1 && "Custom lowering enabled only for i1");
1811 
1812   Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op1);
1813   Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op2);
1814   SDValue Select = DAG.getNode(ISD::SELECT, DL, MVT::i32, Op0, Op1, Op2);
1815   SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Select);
1816 
1817   return Trunc;
1818 }
1819 
LowerLOAD(SDValue Op,SelectionDAG & DAG) const1820 SDValue NVPTXTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1821   if (Op.getValueType() == MVT::i1)
1822     return LowerLOADi1(Op, DAG);
1823   else
1824     return SDValue();
1825 }
1826 
1827 // v = ld i1* addr
1828 //   =>
1829 // v1 = ld i8* addr (-> i16)
1830 // v = trunc i16 to i1
LowerLOADi1(SDValue Op,SelectionDAG & DAG) const1831 SDValue NVPTXTargetLowering::LowerLOADi1(SDValue Op, SelectionDAG &DAG) const {
1832   SDNode *Node = Op.getNode();
1833   LoadSDNode *LD = cast<LoadSDNode>(Node);
1834   SDLoc dl(Node);
1835   assert(LD->getExtensionType() == ISD::NON_EXTLOAD);
1836   assert(Node->getValueType(0) == MVT::i1 &&
1837          "Custom lowering for i1 load only");
1838   SDValue newLD =
1839       DAG.getLoad(MVT::i16, dl, LD->getChain(), LD->getBasePtr(),
1840                   LD->getPointerInfo(), LD->isVolatile(), LD->isNonTemporal(),
1841                   LD->isInvariant(), LD->getAlignment());
1842   SDValue result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, newLD);
1843   // The legalizer (the caller) is expecting two values from the legalized
1844   // load, so we build a MergeValues node for it. See ExpandUnalignedLoad()
1845   // in LegalizeDAG.cpp which also uses MergeValues.
1846   SDValue Ops[] = { result, LD->getChain() };
1847   return DAG.getMergeValues(Ops, dl);
1848 }
1849 
LowerSTORE(SDValue Op,SelectionDAG & DAG) const1850 SDValue NVPTXTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1851   EVT ValVT = Op.getOperand(1).getValueType();
1852   if (ValVT == MVT::i1)
1853     return LowerSTOREi1(Op, DAG);
1854   else if (ValVT.isVector())
1855     return LowerSTOREVector(Op, DAG);
1856   else
1857     return SDValue();
1858 }
1859 
1860 SDValue
LowerSTOREVector(SDValue Op,SelectionDAG & DAG) const1861 NVPTXTargetLowering::LowerSTOREVector(SDValue Op, SelectionDAG &DAG) const {
1862   SDNode *N = Op.getNode();
1863   SDValue Val = N->getOperand(1);
1864   SDLoc DL(N);
1865   EVT ValVT = Val.getValueType();
1866 
1867   if (ValVT.isVector()) {
1868     // We only handle "native" vector sizes for now, e.g. <4 x double> is not
1869     // legal.  We can (and should) split that into 2 stores of <2 x double> here
1870     // but I'm leaving that as a TODO for now.
1871     if (!ValVT.isSimple())
1872       return SDValue();
1873     switch (ValVT.getSimpleVT().SimpleTy) {
1874     default:
1875       return SDValue();
1876     case MVT::v2i8:
1877     case MVT::v2i16:
1878     case MVT::v2i32:
1879     case MVT::v2i64:
1880     case MVT::v2f32:
1881     case MVT::v2f64:
1882     case MVT::v4i8:
1883     case MVT::v4i16:
1884     case MVT::v4i32:
1885     case MVT::v4f32:
1886       // This is a "native" vector type
1887       break;
1888     }
1889 
1890     MemSDNode *MemSD = cast<MemSDNode>(N);
1891     const DataLayout *TD = getDataLayout();
1892 
1893     unsigned Align = MemSD->getAlignment();
1894     unsigned PrefAlign =
1895       TD->getPrefTypeAlignment(ValVT.getTypeForEVT(*DAG.getContext()));
1896     if (Align < PrefAlign) {
1897       // This store is not sufficiently aligned, so bail out and let this vector
1898       // store be scalarized.  Note that we may still be able to emit smaller
1899       // vector stores.  For example, if we are storing a <4 x float> with an
1900       // alignment of 8, this check will fail but the legalizer will try again
1901       // with 2 x <2 x float>, which will succeed with an alignment of 8.
1902       return SDValue();
1903     }
1904 
1905     unsigned Opcode = 0;
1906     EVT EltVT = ValVT.getVectorElementType();
1907     unsigned NumElts = ValVT.getVectorNumElements();
1908 
1909     // Since StoreV2 is a target node, we cannot rely on DAG type legalization.
1910     // Therefore, we must ensure the type is legal.  For i1 and i8, we set the
1911     // stored type to i16 and propagate the "real" type as the memory type.
1912     bool NeedExt = false;
1913     if (EltVT.getSizeInBits() < 16)
1914       NeedExt = true;
1915 
1916     switch (NumElts) {
1917     default:
1918       return SDValue();
1919     case 2:
1920       Opcode = NVPTXISD::StoreV2;
1921       break;
1922     case 4: {
1923       Opcode = NVPTXISD::StoreV4;
1924       break;
1925     }
1926     }
1927 
1928     SmallVector<SDValue, 8> Ops;
1929 
1930     // First is the chain
1931     Ops.push_back(N->getOperand(0));
1932 
1933     // Then the split values
1934     for (unsigned i = 0; i < NumElts; ++i) {
1935       SDValue ExtVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Val,
1936                                    DAG.getIntPtrConstant(i));
1937       if (NeedExt)
1938         ExtVal = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i16, ExtVal);
1939       Ops.push_back(ExtVal);
1940     }
1941 
1942     // Then any remaining arguments
1943     Ops.append(N->op_begin() + 2, N->op_end());
1944 
1945     SDValue NewSt = DAG.getMemIntrinsicNode(
1946         Opcode, DL, DAG.getVTList(MVT::Other), Ops,
1947         MemSD->getMemoryVT(), MemSD->getMemOperand());
1948 
1949     //return DCI.CombineTo(N, NewSt, true);
1950     return NewSt;
1951   }
1952 
1953   return SDValue();
1954 }
1955 
1956 // st i1 v, addr
1957 //    =>
1958 // v1 = zxt v to i16
1959 // st.u8 i16, addr
LowerSTOREi1(SDValue Op,SelectionDAG & DAG) const1960 SDValue NVPTXTargetLowering::LowerSTOREi1(SDValue Op, SelectionDAG &DAG) const {
1961   SDNode *Node = Op.getNode();
1962   SDLoc dl(Node);
1963   StoreSDNode *ST = cast<StoreSDNode>(Node);
1964   SDValue Tmp1 = ST->getChain();
1965   SDValue Tmp2 = ST->getBasePtr();
1966   SDValue Tmp3 = ST->getValue();
1967   assert(Tmp3.getValueType() == MVT::i1 && "Custom lowering for i1 store only");
1968   unsigned Alignment = ST->getAlignment();
1969   bool isVolatile = ST->isVolatile();
1970   bool isNonTemporal = ST->isNonTemporal();
1971   Tmp3 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Tmp3);
1972   SDValue Result = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2,
1973                                      ST->getPointerInfo(), MVT::i8, isNonTemporal,
1974                                      isVolatile, Alignment);
1975   return Result;
1976 }
1977 
getExtSymb(SelectionDAG & DAG,const char * inname,int idx,EVT v) const1978 SDValue NVPTXTargetLowering::getExtSymb(SelectionDAG &DAG, const char *inname,
1979                                         int idx, EVT v) const {
1980   std::string *name = nvTM->getManagedStrPool()->getManagedString(inname);
1981   std::stringstream suffix;
1982   suffix << idx;
1983   *name += suffix.str();
1984   return DAG.getTargetExternalSymbol(name->c_str(), v);
1985 }
1986 
1987 SDValue
getParamSymbol(SelectionDAG & DAG,int idx,EVT v) const1988 NVPTXTargetLowering::getParamSymbol(SelectionDAG &DAG, int idx, EVT v) const {
1989   std::string ParamSym;
1990   raw_string_ostream ParamStr(ParamSym);
1991 
1992   ParamStr << DAG.getMachineFunction().getName() << "_param_" << idx;
1993   ParamStr.flush();
1994 
1995   std::string *SavedStr =
1996     nvTM->getManagedStrPool()->getManagedString(ParamSym.c_str());
1997   return DAG.getTargetExternalSymbol(SavedStr->c_str(), v);
1998 }
1999 
getParamHelpSymbol(SelectionDAG & DAG,int idx)2000 SDValue NVPTXTargetLowering::getParamHelpSymbol(SelectionDAG &DAG, int idx) {
2001   return getExtSymb(DAG, ".HLPPARAM", idx);
2002 }
2003 
2004 // Check to see if the kernel argument is image*_t or sampler_t
2005 
isImageOrSamplerVal(const Value * arg,const Module * context)2006 bool llvm::isImageOrSamplerVal(const Value *arg, const Module *context) {
2007   static const char *const specialTypes[] = { "struct._image2d_t",
2008                                               "struct._image3d_t",
2009                                               "struct._sampler_t" };
2010 
2011   const Type *Ty = arg->getType();
2012   const PointerType *PTy = dyn_cast<PointerType>(Ty);
2013 
2014   if (!PTy)
2015     return false;
2016 
2017   if (!context)
2018     return false;
2019 
2020   const StructType *STy = dyn_cast<StructType>(PTy->getElementType());
2021   const std::string TypeName = STy && !STy->isLiteral() ? STy->getName() : "";
2022 
2023   for (int i = 0, e = array_lengthof(specialTypes); i != e; ++i)
2024     if (TypeName == specialTypes[i])
2025       return true;
2026 
2027   return false;
2028 }
2029 
LowerFormalArguments(SDValue Chain,CallingConv::ID CallConv,bool isVarArg,const SmallVectorImpl<ISD::InputArg> & Ins,SDLoc dl,SelectionDAG & DAG,SmallVectorImpl<SDValue> & InVals) const2030 SDValue NVPTXTargetLowering::LowerFormalArguments(
2031     SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2032     const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2033     SmallVectorImpl<SDValue> &InVals) const {
2034   MachineFunction &MF = DAG.getMachineFunction();
2035   const DataLayout *TD = getDataLayout();
2036 
2037   const Function *F = MF.getFunction();
2038   const AttributeSet &PAL = F->getAttributes();
2039   const TargetLowering *TLI = STI.getTargetLowering();
2040 
2041   SDValue Root = DAG.getRoot();
2042   std::vector<SDValue> OutChains;
2043 
2044   bool isKernel = llvm::isKernelFunction(*F);
2045   bool isABI = (STI.getSmVersion() >= 20);
2046   assert(isABI && "Non-ABI compilation is not supported");
2047   if (!isABI)
2048     return Chain;
2049 
2050   std::vector<Type *> argTypes;
2051   std::vector<const Argument *> theArgs;
2052   for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
2053        I != E; ++I) {
2054     theArgs.push_back(I);
2055     argTypes.push_back(I->getType());
2056   }
2057   // argTypes.size() (or theArgs.size()) and Ins.size() need not match.
2058   // Ins.size() will be larger
2059   //   * if there is an aggregate argument with multiple fields (each field
2060   //     showing up separately in Ins)
2061   //   * if there is a vector argument with more than typical vector-length
2062   //     elements (generally if more than 4) where each vector element is
2063   //     individually present in Ins.
2064   // So a different index should be used for indexing into Ins.
2065   // See similar issue in LowerCall.
2066   unsigned InsIdx = 0;
2067 
2068   int idx = 0;
2069   for (unsigned i = 0, e = theArgs.size(); i != e; ++i, ++idx, ++InsIdx) {
2070     Type *Ty = argTypes[i];
2071 
2072     // If the kernel argument is image*_t or sampler_t, convert it to
2073     // a i32 constant holding the parameter position. This can later
2074     // matched in the AsmPrinter to output the correct mangled name.
2075     if (isImageOrSamplerVal(
2076             theArgs[i],
2077             (theArgs[i]->getParent() ? theArgs[i]->getParent()->getParent()
2078                                      : nullptr))) {
2079       assert(isKernel && "Only kernels can have image/sampler params");
2080       InVals.push_back(DAG.getConstant(i + 1, MVT::i32));
2081       continue;
2082     }
2083 
2084     if (theArgs[i]->use_empty()) {
2085       // argument is dead
2086       if (Ty->isAggregateType()) {
2087         SmallVector<EVT, 16> vtparts;
2088 
2089         ComputePTXValueVTs(*this, Ty, vtparts);
2090         assert(vtparts.size() > 0 && "empty aggregate type not expected");
2091         for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
2092              ++parti) {
2093           InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2094           ++InsIdx;
2095         }
2096         if (vtparts.size() > 0)
2097           --InsIdx;
2098         continue;
2099       }
2100       if (Ty->isVectorTy()) {
2101         EVT ObjectVT = getValueType(Ty);
2102         unsigned NumRegs = TLI->getNumRegisters(F->getContext(), ObjectVT);
2103         for (unsigned parti = 0; parti < NumRegs; ++parti) {
2104           InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2105           ++InsIdx;
2106         }
2107         if (NumRegs > 0)
2108           --InsIdx;
2109         continue;
2110       }
2111       InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2112       continue;
2113     }
2114 
2115     // In the following cases, assign a node order of "idx+1"
2116     // to newly created nodes. The SDNodes for params have to
2117     // appear in the same order as their order of appearance
2118     // in the original function. "idx+1" holds that order.
2119     if (!PAL.hasAttribute(i + 1, Attribute::ByVal)) {
2120       if (Ty->isAggregateType()) {
2121         SmallVector<EVT, 16> vtparts;
2122         SmallVector<uint64_t, 16> offsets;
2123 
2124         // NOTE: Here, we lose the ability to issue vector loads for vectors
2125         // that are a part of a struct.  This should be investigated in the
2126         // future.
2127         ComputePTXValueVTs(*this, Ty, vtparts, &offsets, 0);
2128         assert(vtparts.size() > 0 && "empty aggregate type not expected");
2129         bool aggregateIsPacked = false;
2130         if (StructType *STy = llvm::dyn_cast<StructType>(Ty))
2131           aggregateIsPacked = STy->isPacked();
2132 
2133         SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2134         for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
2135              ++parti) {
2136           EVT partVT = vtparts[parti];
2137           Value *srcValue = Constant::getNullValue(
2138               PointerType::get(partVT.getTypeForEVT(F->getContext()),
2139                                llvm::ADDRESS_SPACE_PARAM));
2140           SDValue srcAddr =
2141               DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
2142                           DAG.getConstant(offsets[parti], getPointerTy()));
2143           unsigned partAlign =
2144               aggregateIsPacked ? 1
2145                                 : TD->getABITypeAlignment(
2146                                       partVT.getTypeForEVT(F->getContext()));
2147           SDValue p;
2148           if (Ins[InsIdx].VT.getSizeInBits() > partVT.getSizeInBits()) {
2149             ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ?
2150                                      ISD::SEXTLOAD : ISD::ZEXTLOAD;
2151             p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, srcAddr,
2152                                MachinePointerInfo(srcValue), partVT, false,
2153                                false, false, partAlign);
2154           } else {
2155             p = DAG.getLoad(partVT, dl, Root, srcAddr,
2156                             MachinePointerInfo(srcValue), false, false, false,
2157                             partAlign);
2158           }
2159           if (p.getNode())
2160             p.getNode()->setIROrder(idx + 1);
2161           InVals.push_back(p);
2162           ++InsIdx;
2163         }
2164         if (vtparts.size() > 0)
2165           --InsIdx;
2166         continue;
2167       }
2168       if (Ty->isVectorTy()) {
2169         EVT ObjectVT = getValueType(Ty);
2170         SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2171         unsigned NumElts = ObjectVT.getVectorNumElements();
2172         assert(TLI->getNumRegisters(F->getContext(), ObjectVT) == NumElts &&
2173                "Vector was not scalarized");
2174         EVT EltVT = ObjectVT.getVectorElementType();
2175 
2176         // V1 load
2177         // f32 = load ...
2178         if (NumElts == 1) {
2179           // We only have one element, so just directly load it
2180           Value *SrcValue = Constant::getNullValue(PointerType::get(
2181               EltVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2182           SDValue P = DAG.getLoad(
2183               EltVT, dl, Root, Arg, MachinePointerInfo(SrcValue), false,
2184               false, true,
2185               TD->getABITypeAlignment(EltVT.getTypeForEVT(F->getContext())));
2186           if (P.getNode())
2187             P.getNode()->setIROrder(idx + 1);
2188 
2189           if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
2190             P = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, P);
2191           InVals.push_back(P);
2192           ++InsIdx;
2193         } else if (NumElts == 2) {
2194           // V2 load
2195           // f32,f32 = load ...
2196           EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, 2);
2197           Value *SrcValue = Constant::getNullValue(PointerType::get(
2198               VecVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2199           SDValue P = DAG.getLoad(
2200               VecVT, dl, Root, Arg, MachinePointerInfo(SrcValue), false,
2201               false, true,
2202               TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())));
2203           if (P.getNode())
2204             P.getNode()->setIROrder(idx + 1);
2205 
2206           SDValue Elt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2207                                      DAG.getIntPtrConstant(0));
2208           SDValue Elt1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2209                                      DAG.getIntPtrConstant(1));
2210 
2211           if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits()) {
2212             Elt0 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt0);
2213             Elt1 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt1);
2214           }
2215 
2216           InVals.push_back(Elt0);
2217           InVals.push_back(Elt1);
2218           InsIdx += 2;
2219         } else {
2220           // V4 loads
2221           // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
2222           // the
2223           // vector will be expanded to a power of 2 elements, so we know we can
2224           // always round up to the next multiple of 4 when creating the vector
2225           // loads.
2226           // e.g.  4 elem => 1 ld.v4
2227           //       6 elem => 2 ld.v4
2228           //       8 elem => 2 ld.v4
2229           //      11 elem => 3 ld.v4
2230           unsigned VecSize = 4;
2231           if (EltVT.getSizeInBits() == 64) {
2232             VecSize = 2;
2233           }
2234           EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
2235           unsigned Ofst = 0;
2236           for (unsigned i = 0; i < NumElts; i += VecSize) {
2237             Value *SrcValue = Constant::getNullValue(
2238                 PointerType::get(VecVT.getTypeForEVT(F->getContext()),
2239                                  llvm::ADDRESS_SPACE_PARAM));
2240             SDValue SrcAddr =
2241                 DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
2242                             DAG.getConstant(Ofst, getPointerTy()));
2243             SDValue P = DAG.getLoad(
2244                 VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false,
2245                 false, true,
2246                 TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())));
2247             if (P.getNode())
2248               P.getNode()->setIROrder(idx + 1);
2249 
2250             for (unsigned j = 0; j < VecSize; ++j) {
2251               if (i + j >= NumElts)
2252                 break;
2253               SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2254                                         DAG.getIntPtrConstant(j));
2255               if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
2256                 Elt = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt);
2257               InVals.push_back(Elt);
2258             }
2259             Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
2260           }
2261           InsIdx += NumElts;
2262         }
2263 
2264         if (NumElts > 0)
2265           --InsIdx;
2266         continue;
2267       }
2268       // A plain scalar.
2269       EVT ObjectVT = getValueType(Ty);
2270       // If ABI, load from the param symbol
2271       SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2272       Value *srcValue = Constant::getNullValue(PointerType::get(
2273           ObjectVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2274       SDValue p;
2275        if (ObjectVT.getSizeInBits() < Ins[InsIdx].VT.getSizeInBits()) {
2276         ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ?
2277                                        ISD::SEXTLOAD : ISD::ZEXTLOAD;
2278         p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, Arg,
2279                            MachinePointerInfo(srcValue), ObjectVT, false, false,
2280                            false,
2281         TD->getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
2282       } else {
2283         p = DAG.getLoad(Ins[InsIdx].VT, dl, Root, Arg,
2284                         MachinePointerInfo(srcValue), false, false, false,
2285         TD->getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
2286       }
2287       if (p.getNode())
2288         p.getNode()->setIROrder(idx + 1);
2289       InVals.push_back(p);
2290       continue;
2291     }
2292 
2293     // Param has ByVal attribute
2294     // Return MoveParam(param symbol).
2295     // Ideally, the param symbol can be returned directly,
2296     // but when SDNode builder decides to use it in a CopyToReg(),
2297     // machine instruction fails because TargetExternalSymbol
2298     // (not lowered) is target dependent, and CopyToReg assumes
2299     // the source is lowered.
2300     EVT ObjectVT = getValueType(Ty);
2301     assert(ObjectVT == Ins[InsIdx].VT &&
2302            "Ins type did not match function type");
2303     SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2304     SDValue p = DAG.getNode(NVPTXISD::MoveParam, dl, ObjectVT, Arg);
2305     if (p.getNode())
2306       p.getNode()->setIROrder(idx + 1);
2307     if (isKernel)
2308       InVals.push_back(p);
2309     else {
2310       SDValue p2 = DAG.getNode(
2311           ISD::INTRINSIC_WO_CHAIN, dl, ObjectVT,
2312           DAG.getConstant(Intrinsic::nvvm_ptr_local_to_gen, MVT::i32), p);
2313       InVals.push_back(p2);
2314     }
2315   }
2316 
2317   // Clang will check explicit VarArg and issue error if any. However, Clang
2318   // will let code with
2319   // implicit var arg like f() pass. See bug 617733.
2320   // We treat this case as if the arg list is empty.
2321   // if (F.isVarArg()) {
2322   // assert(0 && "VarArg not supported yet!");
2323   //}
2324 
2325   if (!OutChains.empty())
2326     DAG.setRoot(DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains));
2327 
2328   return Chain;
2329 }
2330 
2331 
2332 SDValue
LowerReturn(SDValue Chain,CallingConv::ID CallConv,bool isVarArg,const SmallVectorImpl<ISD::OutputArg> & Outs,const SmallVectorImpl<SDValue> & OutVals,SDLoc dl,SelectionDAG & DAG) const2333 NVPTXTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2334                                  bool isVarArg,
2335                                  const SmallVectorImpl<ISD::OutputArg> &Outs,
2336                                  const SmallVectorImpl<SDValue> &OutVals,
2337                                  SDLoc dl, SelectionDAG &DAG) const {
2338   MachineFunction &MF = DAG.getMachineFunction();
2339   const Function *F = MF.getFunction();
2340   Type *RetTy = F->getReturnType();
2341   const DataLayout *TD = getDataLayout();
2342 
2343   bool isABI = (STI.getSmVersion() >= 20);
2344   assert(isABI && "Non-ABI compilation is not supported");
2345   if (!isABI)
2346     return Chain;
2347 
2348   if (VectorType *VTy = dyn_cast<VectorType>(RetTy)) {
2349     // If we have a vector type, the OutVals array will be the scalarized
2350     // components and we have combine them into 1 or more vector stores.
2351     unsigned NumElts = VTy->getNumElements();
2352     assert(NumElts == Outs.size() && "Bad scalarization of return value");
2353 
2354     // const_cast can be removed in later LLVM versions
2355     EVT EltVT = getValueType(RetTy).getVectorElementType();
2356     bool NeedExtend = false;
2357     if (EltVT.getSizeInBits() < 16)
2358       NeedExtend = true;
2359 
2360     // V1 store
2361     if (NumElts == 1) {
2362       SDValue StoreVal = OutVals[0];
2363       // We only have one element, so just directly store it
2364       if (NeedExtend)
2365         StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
2366       SDValue Ops[] = { Chain, DAG.getConstant(0, MVT::i32), StoreVal };
2367       Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
2368                                       DAG.getVTList(MVT::Other), Ops,
2369                                       EltVT, MachinePointerInfo());
2370 
2371     } else if (NumElts == 2) {
2372       // V2 store
2373       SDValue StoreVal0 = OutVals[0];
2374       SDValue StoreVal1 = OutVals[1];
2375 
2376       if (NeedExtend) {
2377         StoreVal0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal0);
2378         StoreVal1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal1);
2379       }
2380 
2381       SDValue Ops[] = { Chain, DAG.getConstant(0, MVT::i32), StoreVal0,
2382                         StoreVal1 };
2383       Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetvalV2, dl,
2384                                       DAG.getVTList(MVT::Other), Ops,
2385                                       EltVT, MachinePointerInfo());
2386     } else {
2387       // V4 stores
2388       // We have at least 4 elements (<3 x Ty> expands to 4 elements) and the
2389       // vector will be expanded to a power of 2 elements, so we know we can
2390       // always round up to the next multiple of 4 when creating the vector
2391       // stores.
2392       // e.g.  4 elem => 1 st.v4
2393       //       6 elem => 2 st.v4
2394       //       8 elem => 2 st.v4
2395       //      11 elem => 3 st.v4
2396 
2397       unsigned VecSize = 4;
2398       if (OutVals[0].getValueType().getSizeInBits() == 64)
2399         VecSize = 2;
2400 
2401       unsigned Offset = 0;
2402 
2403       EVT VecVT =
2404           EVT::getVectorVT(F->getContext(), EltVT, VecSize);
2405       unsigned PerStoreOffset =
2406           TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
2407 
2408       for (unsigned i = 0; i < NumElts; i += VecSize) {
2409         // Get values
2410         SDValue StoreVal;
2411         SmallVector<SDValue, 8> Ops;
2412         Ops.push_back(Chain);
2413         Ops.push_back(DAG.getConstant(Offset, MVT::i32));
2414         unsigned Opc = NVPTXISD::StoreRetvalV2;
2415         EVT ExtendedVT = (NeedExtend) ? MVT::i16 : OutVals[0].getValueType();
2416 
2417         StoreVal = OutVals[i];
2418         if (NeedExtend)
2419           StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2420         Ops.push_back(StoreVal);
2421 
2422         if (i + 1 < NumElts) {
2423           StoreVal = OutVals[i + 1];
2424           if (NeedExtend)
2425             StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2426         } else {
2427           StoreVal = DAG.getUNDEF(ExtendedVT);
2428         }
2429         Ops.push_back(StoreVal);
2430 
2431         if (VecSize == 4) {
2432           Opc = NVPTXISD::StoreRetvalV4;
2433           if (i + 2 < NumElts) {
2434             StoreVal = OutVals[i + 2];
2435             if (NeedExtend)
2436               StoreVal =
2437                   DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2438           } else {
2439             StoreVal = DAG.getUNDEF(ExtendedVT);
2440           }
2441           Ops.push_back(StoreVal);
2442 
2443           if (i + 3 < NumElts) {
2444             StoreVal = OutVals[i + 3];
2445             if (NeedExtend)
2446               StoreVal =
2447                   DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2448           } else {
2449             StoreVal = DAG.getUNDEF(ExtendedVT);
2450           }
2451           Ops.push_back(StoreVal);
2452         }
2453 
2454         // Chain = DAG.getNode(Opc, dl, MVT::Other, &Ops[0], Ops.size());
2455         Chain =
2456             DAG.getMemIntrinsicNode(Opc, dl, DAG.getVTList(MVT::Other), Ops,
2457                                     EltVT, MachinePointerInfo());
2458         Offset += PerStoreOffset;
2459       }
2460     }
2461   } else {
2462     SmallVector<EVT, 16> ValVTs;
2463     SmallVector<uint64_t, 16> Offsets;
2464     ComputePTXValueVTs(*this, RetTy, ValVTs, &Offsets, 0);
2465     assert(ValVTs.size() == OutVals.size() && "Bad return value decomposition");
2466 
2467     for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
2468       SDValue theVal = OutVals[i];
2469       EVT TheValType = theVal.getValueType();
2470       unsigned numElems = 1;
2471       if (TheValType.isVector())
2472         numElems = TheValType.getVectorNumElements();
2473       for (unsigned j = 0, je = numElems; j != je; ++j) {
2474         SDValue TmpVal = theVal;
2475         if (TheValType.isVector())
2476           TmpVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
2477                                TheValType.getVectorElementType(), TmpVal,
2478                                DAG.getIntPtrConstant(j));
2479         EVT TheStoreType = ValVTs[i];
2480         if (RetTy->isIntegerTy() &&
2481             TD->getTypeAllocSizeInBits(RetTy) < 32) {
2482           // The following zero-extension is for integer types only, and
2483           // specifically not for aggregates.
2484           TmpVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, TmpVal);
2485           TheStoreType = MVT::i32;
2486         }
2487         else if (TmpVal.getValueType().getSizeInBits() < 16)
2488           TmpVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, TmpVal);
2489 
2490         SDValue Ops[] = {
2491           Chain,
2492           DAG.getConstant(Offsets[i], MVT::i32),
2493           TmpVal };
2494         Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
2495                                         DAG.getVTList(MVT::Other), Ops,
2496                                         TheStoreType,
2497                                         MachinePointerInfo());
2498       }
2499     }
2500   }
2501 
2502   return DAG.getNode(NVPTXISD::RET_FLAG, dl, MVT::Other, Chain);
2503 }
2504 
2505 
LowerAsmOperandForConstraint(SDValue Op,std::string & Constraint,std::vector<SDValue> & Ops,SelectionDAG & DAG) const2506 void NVPTXTargetLowering::LowerAsmOperandForConstraint(
2507     SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
2508     SelectionDAG &DAG) const {
2509   if (Constraint.length() > 1)
2510     return;
2511   else
2512     TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
2513 }
2514 
2515 // NVPTX suuport vector of legal types of any length in Intrinsics because the
2516 // NVPTX specific type legalizer
2517 // will legalize them to the PTX supported length.
isTypeSupportedInIntrinsic(MVT VT) const2518 bool NVPTXTargetLowering::isTypeSupportedInIntrinsic(MVT VT) const {
2519   if (isTypeLegal(VT))
2520     return true;
2521   if (VT.isVector()) {
2522     MVT eVT = VT.getVectorElementType();
2523     if (isTypeLegal(eVT))
2524       return true;
2525   }
2526   return false;
2527 }
2528 
getOpcForTextureInstr(unsigned Intrinsic)2529 static unsigned getOpcForTextureInstr(unsigned Intrinsic) {
2530   switch (Intrinsic) {
2531   default:
2532     return 0;
2533 
2534   case Intrinsic::nvvm_tex_1d_v4f32_s32:
2535     return NVPTXISD::Tex1DFloatS32;
2536   case Intrinsic::nvvm_tex_1d_v4f32_f32:
2537     return NVPTXISD::Tex1DFloatFloat;
2538   case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
2539     return NVPTXISD::Tex1DFloatFloatLevel;
2540   case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
2541     return NVPTXISD::Tex1DFloatFloatGrad;
2542   case Intrinsic::nvvm_tex_1d_v4s32_s32:
2543     return NVPTXISD::Tex1DS32S32;
2544   case Intrinsic::nvvm_tex_1d_v4s32_f32:
2545     return NVPTXISD::Tex1DS32Float;
2546   case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
2547     return NVPTXISD::Tex1DS32FloatLevel;
2548   case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
2549     return NVPTXISD::Tex1DS32FloatGrad;
2550   case Intrinsic::nvvm_tex_1d_v4u32_s32:
2551     return NVPTXISD::Tex1DU32S32;
2552   case Intrinsic::nvvm_tex_1d_v4u32_f32:
2553     return NVPTXISD::Tex1DU32Float;
2554   case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
2555     return NVPTXISD::Tex1DU32FloatLevel;
2556   case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
2557     return NVPTXISD::Tex1DU32FloatGrad;
2558 
2559   case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
2560     return NVPTXISD::Tex1DArrayFloatS32;
2561   case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
2562     return NVPTXISD::Tex1DArrayFloatFloat;
2563   case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
2564     return NVPTXISD::Tex1DArrayFloatFloatLevel;
2565   case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
2566     return NVPTXISD::Tex1DArrayFloatFloatGrad;
2567   case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
2568     return NVPTXISD::Tex1DArrayS32S32;
2569   case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
2570     return NVPTXISD::Tex1DArrayS32Float;
2571   case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
2572     return NVPTXISD::Tex1DArrayS32FloatLevel;
2573   case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
2574     return NVPTXISD::Tex1DArrayS32FloatGrad;
2575   case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
2576     return NVPTXISD::Tex1DArrayU32S32;
2577   case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
2578     return NVPTXISD::Tex1DArrayU32Float;
2579   case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
2580     return NVPTXISD::Tex1DArrayU32FloatLevel;
2581   case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
2582     return NVPTXISD::Tex1DArrayU32FloatGrad;
2583 
2584   case Intrinsic::nvvm_tex_2d_v4f32_s32:
2585     return NVPTXISD::Tex2DFloatS32;
2586   case Intrinsic::nvvm_tex_2d_v4f32_f32:
2587     return NVPTXISD::Tex2DFloatFloat;
2588   case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
2589     return NVPTXISD::Tex2DFloatFloatLevel;
2590   case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
2591     return NVPTXISD::Tex2DFloatFloatGrad;
2592   case Intrinsic::nvvm_tex_2d_v4s32_s32:
2593     return NVPTXISD::Tex2DS32S32;
2594   case Intrinsic::nvvm_tex_2d_v4s32_f32:
2595     return NVPTXISD::Tex2DS32Float;
2596   case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
2597     return NVPTXISD::Tex2DS32FloatLevel;
2598   case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
2599     return NVPTXISD::Tex2DS32FloatGrad;
2600   case Intrinsic::nvvm_tex_2d_v4u32_s32:
2601     return NVPTXISD::Tex2DU32S32;
2602   case Intrinsic::nvvm_tex_2d_v4u32_f32:
2603     return NVPTXISD::Tex2DU32Float;
2604   case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
2605     return NVPTXISD::Tex2DU32FloatLevel;
2606   case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
2607     return NVPTXISD::Tex2DU32FloatGrad;
2608 
2609   case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
2610     return NVPTXISD::Tex2DArrayFloatS32;
2611   case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
2612     return NVPTXISD::Tex2DArrayFloatFloat;
2613   case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
2614     return NVPTXISD::Tex2DArrayFloatFloatLevel;
2615   case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
2616     return NVPTXISD::Tex2DArrayFloatFloatGrad;
2617   case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
2618     return NVPTXISD::Tex2DArrayS32S32;
2619   case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
2620     return NVPTXISD::Tex2DArrayS32Float;
2621   case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
2622     return NVPTXISD::Tex2DArrayS32FloatLevel;
2623   case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
2624     return NVPTXISD::Tex2DArrayS32FloatGrad;
2625   case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
2626     return NVPTXISD::Tex2DArrayU32S32;
2627   case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
2628     return NVPTXISD::Tex2DArrayU32Float;
2629   case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
2630     return NVPTXISD::Tex2DArrayU32FloatLevel;
2631   case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
2632     return NVPTXISD::Tex2DArrayU32FloatGrad;
2633 
2634   case Intrinsic::nvvm_tex_3d_v4f32_s32:
2635     return NVPTXISD::Tex3DFloatS32;
2636   case Intrinsic::nvvm_tex_3d_v4f32_f32:
2637     return NVPTXISD::Tex3DFloatFloat;
2638   case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
2639     return NVPTXISD::Tex3DFloatFloatLevel;
2640   case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
2641     return NVPTXISD::Tex3DFloatFloatGrad;
2642   case Intrinsic::nvvm_tex_3d_v4s32_s32:
2643     return NVPTXISD::Tex3DS32S32;
2644   case Intrinsic::nvvm_tex_3d_v4s32_f32:
2645     return NVPTXISD::Tex3DS32Float;
2646   case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
2647     return NVPTXISD::Tex3DS32FloatLevel;
2648   case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
2649     return NVPTXISD::Tex3DS32FloatGrad;
2650   case Intrinsic::nvvm_tex_3d_v4u32_s32:
2651     return NVPTXISD::Tex3DU32S32;
2652   case Intrinsic::nvvm_tex_3d_v4u32_f32:
2653     return NVPTXISD::Tex3DU32Float;
2654   case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
2655     return NVPTXISD::Tex3DU32FloatLevel;
2656   case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
2657     return NVPTXISD::Tex3DU32FloatGrad;
2658 
2659   case Intrinsic::nvvm_tex_cube_v4f32_f32:
2660     return NVPTXISD::TexCubeFloatFloat;
2661   case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
2662     return NVPTXISD::TexCubeFloatFloatLevel;
2663   case Intrinsic::nvvm_tex_cube_v4s32_f32:
2664     return NVPTXISD::TexCubeS32Float;
2665   case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
2666     return NVPTXISD::TexCubeS32FloatLevel;
2667   case Intrinsic::nvvm_tex_cube_v4u32_f32:
2668     return NVPTXISD::TexCubeU32Float;
2669   case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
2670     return NVPTXISD::TexCubeU32FloatLevel;
2671 
2672   case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
2673     return NVPTXISD::TexCubeArrayFloatFloat;
2674   case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
2675     return NVPTXISD::TexCubeArrayFloatFloatLevel;
2676   case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
2677     return NVPTXISD::TexCubeArrayS32Float;
2678   case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
2679     return NVPTXISD::TexCubeArrayS32FloatLevel;
2680   case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
2681     return NVPTXISD::TexCubeArrayU32Float;
2682   case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
2683     return NVPTXISD::TexCubeArrayU32FloatLevel;
2684 
2685   case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
2686     return NVPTXISD::Tld4R2DFloatFloat;
2687   case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
2688     return NVPTXISD::Tld4G2DFloatFloat;
2689   case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
2690     return NVPTXISD::Tld4B2DFloatFloat;
2691   case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
2692     return NVPTXISD::Tld4A2DFloatFloat;
2693   case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
2694     return NVPTXISD::Tld4R2DS64Float;
2695   case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
2696     return NVPTXISD::Tld4G2DS64Float;
2697   case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
2698     return NVPTXISD::Tld4B2DS64Float;
2699   case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
2700     return NVPTXISD::Tld4A2DS64Float;
2701   case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
2702     return NVPTXISD::Tld4R2DU64Float;
2703   case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
2704     return NVPTXISD::Tld4G2DU64Float;
2705   case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
2706     return NVPTXISD::Tld4B2DU64Float;
2707   case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
2708     return NVPTXISD::Tld4A2DU64Float;
2709 
2710   case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
2711     return NVPTXISD::TexUnified1DFloatS32;
2712   case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
2713     return NVPTXISD::TexUnified1DFloatFloat;
2714   case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
2715     return NVPTXISD::TexUnified1DFloatFloatLevel;
2716   case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
2717     return NVPTXISD::TexUnified1DFloatFloatGrad;
2718   case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
2719     return NVPTXISD::TexUnified1DS32S32;
2720   case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
2721     return NVPTXISD::TexUnified1DS32Float;
2722   case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
2723     return NVPTXISD::TexUnified1DS32FloatLevel;
2724   case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
2725     return NVPTXISD::TexUnified1DS32FloatGrad;
2726   case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
2727     return NVPTXISD::TexUnified1DU32S32;
2728   case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
2729     return NVPTXISD::TexUnified1DU32Float;
2730   case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
2731     return NVPTXISD::TexUnified1DU32FloatLevel;
2732   case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
2733     return NVPTXISD::TexUnified1DU32FloatGrad;
2734 
2735   case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
2736     return NVPTXISD::TexUnified1DArrayFloatS32;
2737   case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
2738     return NVPTXISD::TexUnified1DArrayFloatFloat;
2739   case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
2740     return NVPTXISD::TexUnified1DArrayFloatFloatLevel;
2741   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
2742     return NVPTXISD::TexUnified1DArrayFloatFloatGrad;
2743   case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
2744     return NVPTXISD::TexUnified1DArrayS32S32;
2745   case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
2746     return NVPTXISD::TexUnified1DArrayS32Float;
2747   case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
2748     return NVPTXISD::TexUnified1DArrayS32FloatLevel;
2749   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
2750     return NVPTXISD::TexUnified1DArrayS32FloatGrad;
2751   case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
2752     return NVPTXISD::TexUnified1DArrayU32S32;
2753   case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
2754     return NVPTXISD::TexUnified1DArrayU32Float;
2755   case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
2756     return NVPTXISD::TexUnified1DArrayU32FloatLevel;
2757   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
2758     return NVPTXISD::TexUnified1DArrayU32FloatGrad;
2759 
2760   case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
2761     return NVPTXISD::TexUnified2DFloatS32;
2762   case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
2763     return NVPTXISD::TexUnified2DFloatFloat;
2764   case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
2765     return NVPTXISD::TexUnified2DFloatFloatLevel;
2766   case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
2767     return NVPTXISD::TexUnified2DFloatFloatGrad;
2768   case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
2769     return NVPTXISD::TexUnified2DS32S32;
2770   case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
2771     return NVPTXISD::TexUnified2DS32Float;
2772   case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
2773     return NVPTXISD::TexUnified2DS32FloatLevel;
2774   case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
2775     return NVPTXISD::TexUnified2DS32FloatGrad;
2776   case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
2777     return NVPTXISD::TexUnified2DU32S32;
2778   case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
2779     return NVPTXISD::TexUnified2DU32Float;
2780   case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
2781     return NVPTXISD::TexUnified2DU32FloatLevel;
2782   case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
2783     return NVPTXISD::TexUnified2DU32FloatGrad;
2784 
2785   case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
2786     return NVPTXISD::TexUnified2DArrayFloatS32;
2787   case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
2788     return NVPTXISD::TexUnified2DArrayFloatFloat;
2789   case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
2790     return NVPTXISD::TexUnified2DArrayFloatFloatLevel;
2791   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
2792     return NVPTXISD::TexUnified2DArrayFloatFloatGrad;
2793   case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
2794     return NVPTXISD::TexUnified2DArrayS32S32;
2795   case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
2796     return NVPTXISD::TexUnified2DArrayS32Float;
2797   case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
2798     return NVPTXISD::TexUnified2DArrayS32FloatLevel;
2799   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
2800     return NVPTXISD::TexUnified2DArrayS32FloatGrad;
2801   case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
2802     return NVPTXISD::TexUnified2DArrayU32S32;
2803   case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
2804     return NVPTXISD::TexUnified2DArrayU32Float;
2805   case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
2806     return NVPTXISD::TexUnified2DArrayU32FloatLevel;
2807   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
2808     return NVPTXISD::TexUnified2DArrayU32FloatGrad;
2809 
2810   case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
2811     return NVPTXISD::TexUnified3DFloatS32;
2812   case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
2813     return NVPTXISD::TexUnified3DFloatFloat;
2814   case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
2815     return NVPTXISD::TexUnified3DFloatFloatLevel;
2816   case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
2817     return NVPTXISD::TexUnified3DFloatFloatGrad;
2818   case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
2819     return NVPTXISD::TexUnified3DS32S32;
2820   case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
2821     return NVPTXISD::TexUnified3DS32Float;
2822   case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
2823     return NVPTXISD::TexUnified3DS32FloatLevel;
2824   case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
2825     return NVPTXISD::TexUnified3DS32FloatGrad;
2826   case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
2827     return NVPTXISD::TexUnified3DU32S32;
2828   case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
2829     return NVPTXISD::TexUnified3DU32Float;
2830   case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
2831     return NVPTXISD::TexUnified3DU32FloatLevel;
2832   case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
2833     return NVPTXISD::TexUnified3DU32FloatGrad;
2834 
2835   case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
2836     return NVPTXISD::TexUnifiedCubeFloatFloat;
2837   case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
2838     return NVPTXISD::TexUnifiedCubeFloatFloatLevel;
2839   case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
2840     return NVPTXISD::TexUnifiedCubeS32Float;
2841   case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
2842     return NVPTXISD::TexUnifiedCubeS32FloatLevel;
2843   case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
2844     return NVPTXISD::TexUnifiedCubeU32Float;
2845   case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
2846     return NVPTXISD::TexUnifiedCubeU32FloatLevel;
2847 
2848   case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
2849     return NVPTXISD::TexUnifiedCubeArrayFloatFloat;
2850   case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
2851     return NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel;
2852   case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
2853     return NVPTXISD::TexUnifiedCubeArrayS32Float;
2854   case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
2855     return NVPTXISD::TexUnifiedCubeArrayS32FloatLevel;
2856   case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
2857     return NVPTXISD::TexUnifiedCubeArrayU32Float;
2858   case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
2859     return NVPTXISD::TexUnifiedCubeArrayU32FloatLevel;
2860 
2861   case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
2862     return NVPTXISD::Tld4UnifiedR2DFloatFloat;
2863   case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
2864     return NVPTXISD::Tld4UnifiedG2DFloatFloat;
2865   case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
2866     return NVPTXISD::Tld4UnifiedB2DFloatFloat;
2867   case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32:
2868     return NVPTXISD::Tld4UnifiedA2DFloatFloat;
2869   case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
2870     return NVPTXISD::Tld4UnifiedR2DS64Float;
2871   case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
2872     return NVPTXISD::Tld4UnifiedG2DS64Float;
2873   case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
2874     return NVPTXISD::Tld4UnifiedB2DS64Float;
2875   case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
2876     return NVPTXISD::Tld4UnifiedA2DS64Float;
2877   case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
2878     return NVPTXISD::Tld4UnifiedR2DU64Float;
2879   case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
2880     return NVPTXISD::Tld4UnifiedG2DU64Float;
2881   case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
2882     return NVPTXISD::Tld4UnifiedB2DU64Float;
2883   case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32:
2884     return NVPTXISD::Tld4UnifiedA2DU64Float;
2885   }
2886 }
2887 
getOpcForSurfaceInstr(unsigned Intrinsic)2888 static unsigned getOpcForSurfaceInstr(unsigned Intrinsic) {
2889   switch (Intrinsic) {
2890   default:
2891     return 0;
2892   case Intrinsic::nvvm_suld_1d_i8_clamp:
2893     return NVPTXISD::Suld1DI8Clamp;
2894   case Intrinsic::nvvm_suld_1d_i16_clamp:
2895     return NVPTXISD::Suld1DI16Clamp;
2896   case Intrinsic::nvvm_suld_1d_i32_clamp:
2897     return NVPTXISD::Suld1DI32Clamp;
2898   case Intrinsic::nvvm_suld_1d_i64_clamp:
2899     return NVPTXISD::Suld1DI64Clamp;
2900   case Intrinsic::nvvm_suld_1d_v2i8_clamp:
2901     return NVPTXISD::Suld1DV2I8Clamp;
2902   case Intrinsic::nvvm_suld_1d_v2i16_clamp:
2903     return NVPTXISD::Suld1DV2I16Clamp;
2904   case Intrinsic::nvvm_suld_1d_v2i32_clamp:
2905     return NVPTXISD::Suld1DV2I32Clamp;
2906   case Intrinsic::nvvm_suld_1d_v2i64_clamp:
2907     return NVPTXISD::Suld1DV2I64Clamp;
2908   case Intrinsic::nvvm_suld_1d_v4i8_clamp:
2909     return NVPTXISD::Suld1DV4I8Clamp;
2910   case Intrinsic::nvvm_suld_1d_v4i16_clamp:
2911     return NVPTXISD::Suld1DV4I16Clamp;
2912   case Intrinsic::nvvm_suld_1d_v4i32_clamp:
2913     return NVPTXISD::Suld1DV4I32Clamp;
2914   case Intrinsic::nvvm_suld_1d_array_i8_clamp:
2915     return NVPTXISD::Suld1DArrayI8Clamp;
2916   case Intrinsic::nvvm_suld_1d_array_i16_clamp:
2917     return NVPTXISD::Suld1DArrayI16Clamp;
2918   case Intrinsic::nvvm_suld_1d_array_i32_clamp:
2919     return NVPTXISD::Suld1DArrayI32Clamp;
2920   case Intrinsic::nvvm_suld_1d_array_i64_clamp:
2921     return NVPTXISD::Suld1DArrayI64Clamp;
2922   case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
2923     return NVPTXISD::Suld1DArrayV2I8Clamp;
2924   case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
2925     return NVPTXISD::Suld1DArrayV2I16Clamp;
2926   case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
2927     return NVPTXISD::Suld1DArrayV2I32Clamp;
2928   case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
2929     return NVPTXISD::Suld1DArrayV2I64Clamp;
2930   case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
2931     return NVPTXISD::Suld1DArrayV4I8Clamp;
2932   case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
2933     return NVPTXISD::Suld1DArrayV4I16Clamp;
2934   case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
2935     return NVPTXISD::Suld1DArrayV4I32Clamp;
2936   case Intrinsic::nvvm_suld_2d_i8_clamp:
2937     return NVPTXISD::Suld2DI8Clamp;
2938   case Intrinsic::nvvm_suld_2d_i16_clamp:
2939     return NVPTXISD::Suld2DI16Clamp;
2940   case Intrinsic::nvvm_suld_2d_i32_clamp:
2941     return NVPTXISD::Suld2DI32Clamp;
2942   case Intrinsic::nvvm_suld_2d_i64_clamp:
2943     return NVPTXISD::Suld2DI64Clamp;
2944   case Intrinsic::nvvm_suld_2d_v2i8_clamp:
2945     return NVPTXISD::Suld2DV2I8Clamp;
2946   case Intrinsic::nvvm_suld_2d_v2i16_clamp:
2947     return NVPTXISD::Suld2DV2I16Clamp;
2948   case Intrinsic::nvvm_suld_2d_v2i32_clamp:
2949     return NVPTXISD::Suld2DV2I32Clamp;
2950   case Intrinsic::nvvm_suld_2d_v2i64_clamp:
2951     return NVPTXISD::Suld2DV2I64Clamp;
2952   case Intrinsic::nvvm_suld_2d_v4i8_clamp:
2953     return NVPTXISD::Suld2DV4I8Clamp;
2954   case Intrinsic::nvvm_suld_2d_v4i16_clamp:
2955     return NVPTXISD::Suld2DV4I16Clamp;
2956   case Intrinsic::nvvm_suld_2d_v4i32_clamp:
2957     return NVPTXISD::Suld2DV4I32Clamp;
2958   case Intrinsic::nvvm_suld_2d_array_i8_clamp:
2959     return NVPTXISD::Suld2DArrayI8Clamp;
2960   case Intrinsic::nvvm_suld_2d_array_i16_clamp:
2961     return NVPTXISD::Suld2DArrayI16Clamp;
2962   case Intrinsic::nvvm_suld_2d_array_i32_clamp:
2963     return NVPTXISD::Suld2DArrayI32Clamp;
2964   case Intrinsic::nvvm_suld_2d_array_i64_clamp:
2965     return NVPTXISD::Suld2DArrayI64Clamp;
2966   case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
2967     return NVPTXISD::Suld2DArrayV2I8Clamp;
2968   case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
2969     return NVPTXISD::Suld2DArrayV2I16Clamp;
2970   case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
2971     return NVPTXISD::Suld2DArrayV2I32Clamp;
2972   case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
2973     return NVPTXISD::Suld2DArrayV2I64Clamp;
2974   case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
2975     return NVPTXISD::Suld2DArrayV4I8Clamp;
2976   case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
2977     return NVPTXISD::Suld2DArrayV4I16Clamp;
2978   case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
2979     return NVPTXISD::Suld2DArrayV4I32Clamp;
2980   case Intrinsic::nvvm_suld_3d_i8_clamp:
2981     return NVPTXISD::Suld3DI8Clamp;
2982   case Intrinsic::nvvm_suld_3d_i16_clamp:
2983     return NVPTXISD::Suld3DI16Clamp;
2984   case Intrinsic::nvvm_suld_3d_i32_clamp:
2985     return NVPTXISD::Suld3DI32Clamp;
2986   case Intrinsic::nvvm_suld_3d_i64_clamp:
2987     return NVPTXISD::Suld3DI64Clamp;
2988   case Intrinsic::nvvm_suld_3d_v2i8_clamp:
2989     return NVPTXISD::Suld3DV2I8Clamp;
2990   case Intrinsic::nvvm_suld_3d_v2i16_clamp:
2991     return NVPTXISD::Suld3DV2I16Clamp;
2992   case Intrinsic::nvvm_suld_3d_v2i32_clamp:
2993     return NVPTXISD::Suld3DV2I32Clamp;
2994   case Intrinsic::nvvm_suld_3d_v2i64_clamp:
2995     return NVPTXISD::Suld3DV2I64Clamp;
2996   case Intrinsic::nvvm_suld_3d_v4i8_clamp:
2997     return NVPTXISD::Suld3DV4I8Clamp;
2998   case Intrinsic::nvvm_suld_3d_v4i16_clamp:
2999     return NVPTXISD::Suld3DV4I16Clamp;
3000   case Intrinsic::nvvm_suld_3d_v4i32_clamp:
3001     return NVPTXISD::Suld3DV4I32Clamp;
3002   case Intrinsic::nvvm_suld_1d_i8_trap:
3003     return NVPTXISD::Suld1DI8Trap;
3004   case Intrinsic::nvvm_suld_1d_i16_trap:
3005     return NVPTXISD::Suld1DI16Trap;
3006   case Intrinsic::nvvm_suld_1d_i32_trap:
3007     return NVPTXISD::Suld1DI32Trap;
3008   case Intrinsic::nvvm_suld_1d_i64_trap:
3009     return NVPTXISD::Suld1DI64Trap;
3010   case Intrinsic::nvvm_suld_1d_v2i8_trap:
3011     return NVPTXISD::Suld1DV2I8Trap;
3012   case Intrinsic::nvvm_suld_1d_v2i16_trap:
3013     return NVPTXISD::Suld1DV2I16Trap;
3014   case Intrinsic::nvvm_suld_1d_v2i32_trap:
3015     return NVPTXISD::Suld1DV2I32Trap;
3016   case Intrinsic::nvvm_suld_1d_v2i64_trap:
3017     return NVPTXISD::Suld1DV2I64Trap;
3018   case Intrinsic::nvvm_suld_1d_v4i8_trap:
3019     return NVPTXISD::Suld1DV4I8Trap;
3020   case Intrinsic::nvvm_suld_1d_v4i16_trap:
3021     return NVPTXISD::Suld1DV4I16Trap;
3022   case Intrinsic::nvvm_suld_1d_v4i32_trap:
3023     return NVPTXISD::Suld1DV4I32Trap;
3024   case Intrinsic::nvvm_suld_1d_array_i8_trap:
3025     return NVPTXISD::Suld1DArrayI8Trap;
3026   case Intrinsic::nvvm_suld_1d_array_i16_trap:
3027     return NVPTXISD::Suld1DArrayI16Trap;
3028   case Intrinsic::nvvm_suld_1d_array_i32_trap:
3029     return NVPTXISD::Suld1DArrayI32Trap;
3030   case Intrinsic::nvvm_suld_1d_array_i64_trap:
3031     return NVPTXISD::Suld1DArrayI64Trap;
3032   case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
3033     return NVPTXISD::Suld1DArrayV2I8Trap;
3034   case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
3035     return NVPTXISD::Suld1DArrayV2I16Trap;
3036   case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
3037     return NVPTXISD::Suld1DArrayV2I32Trap;
3038   case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
3039     return NVPTXISD::Suld1DArrayV2I64Trap;
3040   case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
3041     return NVPTXISD::Suld1DArrayV4I8Trap;
3042   case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
3043     return NVPTXISD::Suld1DArrayV4I16Trap;
3044   case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
3045     return NVPTXISD::Suld1DArrayV4I32Trap;
3046   case Intrinsic::nvvm_suld_2d_i8_trap:
3047     return NVPTXISD::Suld2DI8Trap;
3048   case Intrinsic::nvvm_suld_2d_i16_trap:
3049     return NVPTXISD::Suld2DI16Trap;
3050   case Intrinsic::nvvm_suld_2d_i32_trap:
3051     return NVPTXISD::Suld2DI32Trap;
3052   case Intrinsic::nvvm_suld_2d_i64_trap:
3053     return NVPTXISD::Suld2DI64Trap;
3054   case Intrinsic::nvvm_suld_2d_v2i8_trap:
3055     return NVPTXISD::Suld2DV2I8Trap;
3056   case Intrinsic::nvvm_suld_2d_v2i16_trap:
3057     return NVPTXISD::Suld2DV2I16Trap;
3058   case Intrinsic::nvvm_suld_2d_v2i32_trap:
3059     return NVPTXISD::Suld2DV2I32Trap;
3060   case Intrinsic::nvvm_suld_2d_v2i64_trap:
3061     return NVPTXISD::Suld2DV2I64Trap;
3062   case Intrinsic::nvvm_suld_2d_v4i8_trap:
3063     return NVPTXISD::Suld2DV4I8Trap;
3064   case Intrinsic::nvvm_suld_2d_v4i16_trap:
3065     return NVPTXISD::Suld2DV4I16Trap;
3066   case Intrinsic::nvvm_suld_2d_v4i32_trap:
3067     return NVPTXISD::Suld2DV4I32Trap;
3068   case Intrinsic::nvvm_suld_2d_array_i8_trap:
3069     return NVPTXISD::Suld2DArrayI8Trap;
3070   case Intrinsic::nvvm_suld_2d_array_i16_trap:
3071     return NVPTXISD::Suld2DArrayI16Trap;
3072   case Intrinsic::nvvm_suld_2d_array_i32_trap:
3073     return NVPTXISD::Suld2DArrayI32Trap;
3074   case Intrinsic::nvvm_suld_2d_array_i64_trap:
3075     return NVPTXISD::Suld2DArrayI64Trap;
3076   case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
3077     return NVPTXISD::Suld2DArrayV2I8Trap;
3078   case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
3079     return NVPTXISD::Suld2DArrayV2I16Trap;
3080   case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
3081     return NVPTXISD::Suld2DArrayV2I32Trap;
3082   case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
3083     return NVPTXISD::Suld2DArrayV2I64Trap;
3084   case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
3085     return NVPTXISD::Suld2DArrayV4I8Trap;
3086   case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
3087     return NVPTXISD::Suld2DArrayV4I16Trap;
3088   case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
3089     return NVPTXISD::Suld2DArrayV4I32Trap;
3090   case Intrinsic::nvvm_suld_3d_i8_trap:
3091     return NVPTXISD::Suld3DI8Trap;
3092   case Intrinsic::nvvm_suld_3d_i16_trap:
3093     return NVPTXISD::Suld3DI16Trap;
3094   case Intrinsic::nvvm_suld_3d_i32_trap:
3095     return NVPTXISD::Suld3DI32Trap;
3096   case Intrinsic::nvvm_suld_3d_i64_trap:
3097     return NVPTXISD::Suld3DI64Trap;
3098   case Intrinsic::nvvm_suld_3d_v2i8_trap:
3099     return NVPTXISD::Suld3DV2I8Trap;
3100   case Intrinsic::nvvm_suld_3d_v2i16_trap:
3101     return NVPTXISD::Suld3DV2I16Trap;
3102   case Intrinsic::nvvm_suld_3d_v2i32_trap:
3103     return NVPTXISD::Suld3DV2I32Trap;
3104   case Intrinsic::nvvm_suld_3d_v2i64_trap:
3105     return NVPTXISD::Suld3DV2I64Trap;
3106   case Intrinsic::nvvm_suld_3d_v4i8_trap:
3107     return NVPTXISD::Suld3DV4I8Trap;
3108   case Intrinsic::nvvm_suld_3d_v4i16_trap:
3109     return NVPTXISD::Suld3DV4I16Trap;
3110   case Intrinsic::nvvm_suld_3d_v4i32_trap:
3111     return NVPTXISD::Suld3DV4I32Trap;
3112   case Intrinsic::nvvm_suld_1d_i8_zero:
3113     return NVPTXISD::Suld1DI8Zero;
3114   case Intrinsic::nvvm_suld_1d_i16_zero:
3115     return NVPTXISD::Suld1DI16Zero;
3116   case Intrinsic::nvvm_suld_1d_i32_zero:
3117     return NVPTXISD::Suld1DI32Zero;
3118   case Intrinsic::nvvm_suld_1d_i64_zero:
3119     return NVPTXISD::Suld1DI64Zero;
3120   case Intrinsic::nvvm_suld_1d_v2i8_zero:
3121     return NVPTXISD::Suld1DV2I8Zero;
3122   case Intrinsic::nvvm_suld_1d_v2i16_zero:
3123     return NVPTXISD::Suld1DV2I16Zero;
3124   case Intrinsic::nvvm_suld_1d_v2i32_zero:
3125     return NVPTXISD::Suld1DV2I32Zero;
3126   case Intrinsic::nvvm_suld_1d_v2i64_zero:
3127     return NVPTXISD::Suld1DV2I64Zero;
3128   case Intrinsic::nvvm_suld_1d_v4i8_zero:
3129     return NVPTXISD::Suld1DV4I8Zero;
3130   case Intrinsic::nvvm_suld_1d_v4i16_zero:
3131     return NVPTXISD::Suld1DV4I16Zero;
3132   case Intrinsic::nvvm_suld_1d_v4i32_zero:
3133     return NVPTXISD::Suld1DV4I32Zero;
3134   case Intrinsic::nvvm_suld_1d_array_i8_zero:
3135     return NVPTXISD::Suld1DArrayI8Zero;
3136   case Intrinsic::nvvm_suld_1d_array_i16_zero:
3137     return NVPTXISD::Suld1DArrayI16Zero;
3138   case Intrinsic::nvvm_suld_1d_array_i32_zero:
3139     return NVPTXISD::Suld1DArrayI32Zero;
3140   case Intrinsic::nvvm_suld_1d_array_i64_zero:
3141     return NVPTXISD::Suld1DArrayI64Zero;
3142   case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
3143     return NVPTXISD::Suld1DArrayV2I8Zero;
3144   case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
3145     return NVPTXISD::Suld1DArrayV2I16Zero;
3146   case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
3147     return NVPTXISD::Suld1DArrayV2I32Zero;
3148   case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
3149     return NVPTXISD::Suld1DArrayV2I64Zero;
3150   case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
3151     return NVPTXISD::Suld1DArrayV4I8Zero;
3152   case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
3153     return NVPTXISD::Suld1DArrayV4I16Zero;
3154   case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
3155     return NVPTXISD::Suld1DArrayV4I32Zero;
3156   case Intrinsic::nvvm_suld_2d_i8_zero:
3157     return NVPTXISD::Suld2DI8Zero;
3158   case Intrinsic::nvvm_suld_2d_i16_zero:
3159     return NVPTXISD::Suld2DI16Zero;
3160   case Intrinsic::nvvm_suld_2d_i32_zero:
3161     return NVPTXISD::Suld2DI32Zero;
3162   case Intrinsic::nvvm_suld_2d_i64_zero:
3163     return NVPTXISD::Suld2DI64Zero;
3164   case Intrinsic::nvvm_suld_2d_v2i8_zero:
3165     return NVPTXISD::Suld2DV2I8Zero;
3166   case Intrinsic::nvvm_suld_2d_v2i16_zero:
3167     return NVPTXISD::Suld2DV2I16Zero;
3168   case Intrinsic::nvvm_suld_2d_v2i32_zero:
3169     return NVPTXISD::Suld2DV2I32Zero;
3170   case Intrinsic::nvvm_suld_2d_v2i64_zero:
3171     return NVPTXISD::Suld2DV2I64Zero;
3172   case Intrinsic::nvvm_suld_2d_v4i8_zero:
3173     return NVPTXISD::Suld2DV4I8Zero;
3174   case Intrinsic::nvvm_suld_2d_v4i16_zero:
3175     return NVPTXISD::Suld2DV4I16Zero;
3176   case Intrinsic::nvvm_suld_2d_v4i32_zero:
3177     return NVPTXISD::Suld2DV4I32Zero;
3178   case Intrinsic::nvvm_suld_2d_array_i8_zero:
3179     return NVPTXISD::Suld2DArrayI8Zero;
3180   case Intrinsic::nvvm_suld_2d_array_i16_zero:
3181     return NVPTXISD::Suld2DArrayI16Zero;
3182   case Intrinsic::nvvm_suld_2d_array_i32_zero:
3183     return NVPTXISD::Suld2DArrayI32Zero;
3184   case Intrinsic::nvvm_suld_2d_array_i64_zero:
3185     return NVPTXISD::Suld2DArrayI64Zero;
3186   case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
3187     return NVPTXISD::Suld2DArrayV2I8Zero;
3188   case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
3189     return NVPTXISD::Suld2DArrayV2I16Zero;
3190   case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
3191     return NVPTXISD::Suld2DArrayV2I32Zero;
3192   case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
3193     return NVPTXISD::Suld2DArrayV2I64Zero;
3194   case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
3195     return NVPTXISD::Suld2DArrayV4I8Zero;
3196   case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
3197     return NVPTXISD::Suld2DArrayV4I16Zero;
3198   case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
3199     return NVPTXISD::Suld2DArrayV4I32Zero;
3200   case Intrinsic::nvvm_suld_3d_i8_zero:
3201     return NVPTXISD::Suld3DI8Zero;
3202   case Intrinsic::nvvm_suld_3d_i16_zero:
3203     return NVPTXISD::Suld3DI16Zero;
3204   case Intrinsic::nvvm_suld_3d_i32_zero:
3205     return NVPTXISD::Suld3DI32Zero;
3206   case Intrinsic::nvvm_suld_3d_i64_zero:
3207     return NVPTXISD::Suld3DI64Zero;
3208   case Intrinsic::nvvm_suld_3d_v2i8_zero:
3209     return NVPTXISD::Suld3DV2I8Zero;
3210   case Intrinsic::nvvm_suld_3d_v2i16_zero:
3211     return NVPTXISD::Suld3DV2I16Zero;
3212   case Intrinsic::nvvm_suld_3d_v2i32_zero:
3213     return NVPTXISD::Suld3DV2I32Zero;
3214   case Intrinsic::nvvm_suld_3d_v2i64_zero:
3215     return NVPTXISD::Suld3DV2I64Zero;
3216   case Intrinsic::nvvm_suld_3d_v4i8_zero:
3217     return NVPTXISD::Suld3DV4I8Zero;
3218   case Intrinsic::nvvm_suld_3d_v4i16_zero:
3219     return NVPTXISD::Suld3DV4I16Zero;
3220   case Intrinsic::nvvm_suld_3d_v4i32_zero:
3221     return NVPTXISD::Suld3DV4I32Zero;
3222   }
3223 }
3224 
3225 // llvm.ptx.memcpy.const and llvm.ptx.memmove.const need to be modeled as
3226 // TgtMemIntrinsic
3227 // because we need the information that is only available in the "Value" type
3228 // of destination
3229 // pointer. In particular, the address space information.
getTgtMemIntrinsic(IntrinsicInfo & Info,const CallInst & I,unsigned Intrinsic) const3230 bool NVPTXTargetLowering::getTgtMemIntrinsic(
3231     IntrinsicInfo &Info, const CallInst &I, unsigned Intrinsic) const {
3232   switch (Intrinsic) {
3233   default:
3234     return false;
3235 
3236   case Intrinsic::nvvm_atomic_load_add_f32:
3237     Info.opc = ISD::INTRINSIC_W_CHAIN;
3238     Info.memVT = MVT::f32;
3239     Info.ptrVal = I.getArgOperand(0);
3240     Info.offset = 0;
3241     Info.vol = 0;
3242     Info.readMem = true;
3243     Info.writeMem = true;
3244     Info.align = 0;
3245     return true;
3246 
3247   case Intrinsic::nvvm_atomic_load_inc_32:
3248   case Intrinsic::nvvm_atomic_load_dec_32:
3249     Info.opc = ISD::INTRINSIC_W_CHAIN;
3250     Info.memVT = MVT::i32;
3251     Info.ptrVal = I.getArgOperand(0);
3252     Info.offset = 0;
3253     Info.vol = 0;
3254     Info.readMem = true;
3255     Info.writeMem = true;
3256     Info.align = 0;
3257     return true;
3258 
3259   case Intrinsic::nvvm_ldu_global_i:
3260   case Intrinsic::nvvm_ldu_global_f:
3261   case Intrinsic::nvvm_ldu_global_p: {
3262 
3263     Info.opc = ISD::INTRINSIC_W_CHAIN;
3264     if (Intrinsic == Intrinsic::nvvm_ldu_global_i)
3265       Info.memVT = getValueType(I.getType());
3266     else if(Intrinsic == Intrinsic::nvvm_ldu_global_p)
3267       Info.memVT = getPointerTy();
3268     else
3269       Info.memVT = getValueType(I.getType());
3270     Info.ptrVal = I.getArgOperand(0);
3271     Info.offset = 0;
3272     Info.vol = 0;
3273     Info.readMem = true;
3274     Info.writeMem = false;
3275     Info.align = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
3276 
3277     return true;
3278   }
3279   case Intrinsic::nvvm_ldg_global_i:
3280   case Intrinsic::nvvm_ldg_global_f:
3281   case Intrinsic::nvvm_ldg_global_p: {
3282 
3283     Info.opc = ISD::INTRINSIC_W_CHAIN;
3284     if (Intrinsic == Intrinsic::nvvm_ldg_global_i)
3285       Info.memVT = getValueType(I.getType());
3286     else if(Intrinsic == Intrinsic::nvvm_ldg_global_p)
3287       Info.memVT = getPointerTy();
3288     else
3289       Info.memVT = getValueType(I.getType());
3290     Info.ptrVal = I.getArgOperand(0);
3291     Info.offset = 0;
3292     Info.vol = 0;
3293     Info.readMem = true;
3294     Info.writeMem = false;
3295     Info.align = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
3296 
3297     return true;
3298   }
3299 
3300   case Intrinsic::nvvm_tex_1d_v4f32_s32:
3301   case Intrinsic::nvvm_tex_1d_v4f32_f32:
3302   case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
3303   case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
3304   case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
3305   case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
3306   case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
3307   case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
3308   case Intrinsic::nvvm_tex_2d_v4f32_s32:
3309   case Intrinsic::nvvm_tex_2d_v4f32_f32:
3310   case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
3311   case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
3312   case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
3313   case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
3314   case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
3315   case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
3316   case Intrinsic::nvvm_tex_3d_v4f32_s32:
3317   case Intrinsic::nvvm_tex_3d_v4f32_f32:
3318   case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
3319   case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
3320   case Intrinsic::nvvm_tex_cube_v4f32_f32:
3321   case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
3322   case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
3323   case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
3324   case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
3325   case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
3326   case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
3327   case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
3328   case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
3329   case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
3330   case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
3331   case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
3332   case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
3333   case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
3334   case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
3335   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
3336   case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
3337   case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
3338   case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
3339   case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
3340   case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
3341   case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
3342   case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
3343   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
3344   case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
3345   case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
3346   case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
3347   case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
3348   case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
3349   case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
3350   case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
3351   case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
3352   case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
3353   case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
3354   case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
3355   case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32: {
3356     Info.opc = getOpcForTextureInstr(Intrinsic);
3357     Info.memVT = MVT::v4f32;
3358     Info.ptrVal = nullptr;
3359     Info.offset = 0;
3360     Info.vol = 0;
3361     Info.readMem = true;
3362     Info.writeMem = false;
3363     Info.align = 16;
3364     return true;
3365   }
3366   case Intrinsic::nvvm_tex_1d_v4s32_s32:
3367   case Intrinsic::nvvm_tex_1d_v4s32_f32:
3368   case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
3369   case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
3370   case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
3371   case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
3372   case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
3373   case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
3374   case Intrinsic::nvvm_tex_2d_v4s32_s32:
3375   case Intrinsic::nvvm_tex_2d_v4s32_f32:
3376   case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
3377   case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
3378   case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
3379   case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
3380   case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
3381   case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
3382   case Intrinsic::nvvm_tex_3d_v4s32_s32:
3383   case Intrinsic::nvvm_tex_3d_v4s32_f32:
3384   case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
3385   case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
3386   case Intrinsic::nvvm_tex_cube_v4s32_f32:
3387   case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
3388   case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
3389   case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
3390   case Intrinsic::nvvm_tex_cube_v4u32_f32:
3391   case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
3392   case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
3393   case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
3394   case Intrinsic::nvvm_tex_1d_v4u32_s32:
3395   case Intrinsic::nvvm_tex_1d_v4u32_f32:
3396   case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
3397   case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
3398   case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
3399   case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
3400   case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
3401   case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
3402   case Intrinsic::nvvm_tex_2d_v4u32_s32:
3403   case Intrinsic::nvvm_tex_2d_v4u32_f32:
3404   case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
3405   case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
3406   case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
3407   case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
3408   case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
3409   case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
3410   case Intrinsic::nvvm_tex_3d_v4u32_s32:
3411   case Intrinsic::nvvm_tex_3d_v4u32_f32:
3412   case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
3413   case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
3414   case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
3415   case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
3416   case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
3417   case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
3418   case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
3419   case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
3420   case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
3421   case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
3422   case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
3423   case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
3424   case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
3425   case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
3426   case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
3427   case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
3428   case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
3429   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
3430   case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
3431   case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
3432   case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
3433   case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
3434   case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
3435   case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
3436   case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
3437   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
3438   case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
3439   case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
3440   case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
3441   case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
3442   case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
3443   case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
3444   case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
3445   case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
3446   case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
3447   case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
3448   case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
3449   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
3450   case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
3451   case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
3452   case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
3453   case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
3454   case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
3455   case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
3456   case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
3457   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
3458   case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
3459   case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
3460   case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
3461   case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
3462   case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
3463   case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
3464   case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
3465   case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
3466   case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
3467   case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
3468   case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
3469   case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
3470   case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
3471   case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
3472   case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
3473   case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
3474   case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
3475   case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
3476   case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
3477   case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32: {
3478     Info.opc = getOpcForTextureInstr(Intrinsic);
3479     Info.memVT = MVT::v4i32;
3480     Info.ptrVal = nullptr;
3481     Info.offset = 0;
3482     Info.vol = 0;
3483     Info.readMem = true;
3484     Info.writeMem = false;
3485     Info.align = 16;
3486     return true;
3487   }
3488   case Intrinsic::nvvm_suld_1d_i8_clamp:
3489   case Intrinsic::nvvm_suld_1d_v2i8_clamp:
3490   case Intrinsic::nvvm_suld_1d_v4i8_clamp:
3491   case Intrinsic::nvvm_suld_1d_array_i8_clamp:
3492   case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
3493   case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
3494   case Intrinsic::nvvm_suld_2d_i8_clamp:
3495   case Intrinsic::nvvm_suld_2d_v2i8_clamp:
3496   case Intrinsic::nvvm_suld_2d_v4i8_clamp:
3497   case Intrinsic::nvvm_suld_2d_array_i8_clamp:
3498   case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
3499   case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
3500   case Intrinsic::nvvm_suld_3d_i8_clamp:
3501   case Intrinsic::nvvm_suld_3d_v2i8_clamp:
3502   case Intrinsic::nvvm_suld_3d_v4i8_clamp:
3503   case Intrinsic::nvvm_suld_1d_i8_trap:
3504   case Intrinsic::nvvm_suld_1d_v2i8_trap:
3505   case Intrinsic::nvvm_suld_1d_v4i8_trap:
3506   case Intrinsic::nvvm_suld_1d_array_i8_trap:
3507   case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
3508   case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
3509   case Intrinsic::nvvm_suld_2d_i8_trap:
3510   case Intrinsic::nvvm_suld_2d_v2i8_trap:
3511   case Intrinsic::nvvm_suld_2d_v4i8_trap:
3512   case Intrinsic::nvvm_suld_2d_array_i8_trap:
3513   case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
3514   case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
3515   case Intrinsic::nvvm_suld_3d_i8_trap:
3516   case Intrinsic::nvvm_suld_3d_v2i8_trap:
3517   case Intrinsic::nvvm_suld_3d_v4i8_trap:
3518   case Intrinsic::nvvm_suld_1d_i8_zero:
3519   case Intrinsic::nvvm_suld_1d_v2i8_zero:
3520   case Intrinsic::nvvm_suld_1d_v4i8_zero:
3521   case Intrinsic::nvvm_suld_1d_array_i8_zero:
3522   case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
3523   case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
3524   case Intrinsic::nvvm_suld_2d_i8_zero:
3525   case Intrinsic::nvvm_suld_2d_v2i8_zero:
3526   case Intrinsic::nvvm_suld_2d_v4i8_zero:
3527   case Intrinsic::nvvm_suld_2d_array_i8_zero:
3528   case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
3529   case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
3530   case Intrinsic::nvvm_suld_3d_i8_zero:
3531   case Intrinsic::nvvm_suld_3d_v2i8_zero:
3532   case Intrinsic::nvvm_suld_3d_v4i8_zero: {
3533     Info.opc = getOpcForSurfaceInstr(Intrinsic);
3534     Info.memVT = MVT::i8;
3535     Info.ptrVal = nullptr;
3536     Info.offset = 0;
3537     Info.vol = 0;
3538     Info.readMem = true;
3539     Info.writeMem = false;
3540     Info.align = 16;
3541     return true;
3542   }
3543   case Intrinsic::nvvm_suld_1d_i16_clamp:
3544   case Intrinsic::nvvm_suld_1d_v2i16_clamp:
3545   case Intrinsic::nvvm_suld_1d_v4i16_clamp:
3546   case Intrinsic::nvvm_suld_1d_array_i16_clamp:
3547   case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
3548   case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
3549   case Intrinsic::nvvm_suld_2d_i16_clamp:
3550   case Intrinsic::nvvm_suld_2d_v2i16_clamp:
3551   case Intrinsic::nvvm_suld_2d_v4i16_clamp:
3552   case Intrinsic::nvvm_suld_2d_array_i16_clamp:
3553   case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
3554   case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
3555   case Intrinsic::nvvm_suld_3d_i16_clamp:
3556   case Intrinsic::nvvm_suld_3d_v2i16_clamp:
3557   case Intrinsic::nvvm_suld_3d_v4i16_clamp:
3558   case Intrinsic::nvvm_suld_1d_i16_trap:
3559   case Intrinsic::nvvm_suld_1d_v2i16_trap:
3560   case Intrinsic::nvvm_suld_1d_v4i16_trap:
3561   case Intrinsic::nvvm_suld_1d_array_i16_trap:
3562   case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
3563   case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
3564   case Intrinsic::nvvm_suld_2d_i16_trap:
3565   case Intrinsic::nvvm_suld_2d_v2i16_trap:
3566   case Intrinsic::nvvm_suld_2d_v4i16_trap:
3567   case Intrinsic::nvvm_suld_2d_array_i16_trap:
3568   case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
3569   case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
3570   case Intrinsic::nvvm_suld_3d_i16_trap:
3571   case Intrinsic::nvvm_suld_3d_v2i16_trap:
3572   case Intrinsic::nvvm_suld_3d_v4i16_trap:
3573   case Intrinsic::nvvm_suld_1d_i16_zero:
3574   case Intrinsic::nvvm_suld_1d_v2i16_zero:
3575   case Intrinsic::nvvm_suld_1d_v4i16_zero:
3576   case Intrinsic::nvvm_suld_1d_array_i16_zero:
3577   case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
3578   case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
3579   case Intrinsic::nvvm_suld_2d_i16_zero:
3580   case Intrinsic::nvvm_suld_2d_v2i16_zero:
3581   case Intrinsic::nvvm_suld_2d_v4i16_zero:
3582   case Intrinsic::nvvm_suld_2d_array_i16_zero:
3583   case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
3584   case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
3585   case Intrinsic::nvvm_suld_3d_i16_zero:
3586   case Intrinsic::nvvm_suld_3d_v2i16_zero:
3587   case Intrinsic::nvvm_suld_3d_v4i16_zero: {
3588     Info.opc = getOpcForSurfaceInstr(Intrinsic);
3589     Info.memVT = MVT::i16;
3590     Info.ptrVal = nullptr;
3591     Info.offset = 0;
3592     Info.vol = 0;
3593     Info.readMem = true;
3594     Info.writeMem = false;
3595     Info.align = 16;
3596     return true;
3597   }
3598   case Intrinsic::nvvm_suld_1d_i32_clamp:
3599   case Intrinsic::nvvm_suld_1d_v2i32_clamp:
3600   case Intrinsic::nvvm_suld_1d_v4i32_clamp:
3601   case Intrinsic::nvvm_suld_1d_array_i32_clamp:
3602   case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
3603   case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
3604   case Intrinsic::nvvm_suld_2d_i32_clamp:
3605   case Intrinsic::nvvm_suld_2d_v2i32_clamp:
3606   case Intrinsic::nvvm_suld_2d_v4i32_clamp:
3607   case Intrinsic::nvvm_suld_2d_array_i32_clamp:
3608   case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
3609   case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
3610   case Intrinsic::nvvm_suld_3d_i32_clamp:
3611   case Intrinsic::nvvm_suld_3d_v2i32_clamp:
3612   case Intrinsic::nvvm_suld_3d_v4i32_clamp:
3613   case Intrinsic::nvvm_suld_1d_i32_trap:
3614   case Intrinsic::nvvm_suld_1d_v2i32_trap:
3615   case Intrinsic::nvvm_suld_1d_v4i32_trap:
3616   case Intrinsic::nvvm_suld_1d_array_i32_trap:
3617   case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
3618   case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
3619   case Intrinsic::nvvm_suld_2d_i32_trap:
3620   case Intrinsic::nvvm_suld_2d_v2i32_trap:
3621   case Intrinsic::nvvm_suld_2d_v4i32_trap:
3622   case Intrinsic::nvvm_suld_2d_array_i32_trap:
3623   case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
3624   case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
3625   case Intrinsic::nvvm_suld_3d_i32_trap:
3626   case Intrinsic::nvvm_suld_3d_v2i32_trap:
3627   case Intrinsic::nvvm_suld_3d_v4i32_trap:
3628   case Intrinsic::nvvm_suld_1d_i32_zero:
3629   case Intrinsic::nvvm_suld_1d_v2i32_zero:
3630   case Intrinsic::nvvm_suld_1d_v4i32_zero:
3631   case Intrinsic::nvvm_suld_1d_array_i32_zero:
3632   case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
3633   case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
3634   case Intrinsic::nvvm_suld_2d_i32_zero:
3635   case Intrinsic::nvvm_suld_2d_v2i32_zero:
3636   case Intrinsic::nvvm_suld_2d_v4i32_zero:
3637   case Intrinsic::nvvm_suld_2d_array_i32_zero:
3638   case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
3639   case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
3640   case Intrinsic::nvvm_suld_3d_i32_zero:
3641   case Intrinsic::nvvm_suld_3d_v2i32_zero:
3642   case Intrinsic::nvvm_suld_3d_v4i32_zero: {
3643     Info.opc = getOpcForSurfaceInstr(Intrinsic);
3644     Info.memVT = MVT::i32;
3645     Info.ptrVal = nullptr;
3646     Info.offset = 0;
3647     Info.vol = 0;
3648     Info.readMem = true;
3649     Info.writeMem = false;
3650     Info.align = 16;
3651     return true;
3652   }
3653   case Intrinsic::nvvm_suld_1d_i64_clamp:
3654   case Intrinsic::nvvm_suld_1d_v2i64_clamp:
3655   case Intrinsic::nvvm_suld_1d_array_i64_clamp:
3656   case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
3657   case Intrinsic::nvvm_suld_2d_i64_clamp:
3658   case Intrinsic::nvvm_suld_2d_v2i64_clamp:
3659   case Intrinsic::nvvm_suld_2d_array_i64_clamp:
3660   case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
3661   case Intrinsic::nvvm_suld_3d_i64_clamp:
3662   case Intrinsic::nvvm_suld_3d_v2i64_clamp:
3663   case Intrinsic::nvvm_suld_1d_i64_trap:
3664   case Intrinsic::nvvm_suld_1d_v2i64_trap:
3665   case Intrinsic::nvvm_suld_1d_array_i64_trap:
3666   case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
3667   case Intrinsic::nvvm_suld_2d_i64_trap:
3668   case Intrinsic::nvvm_suld_2d_v2i64_trap:
3669   case Intrinsic::nvvm_suld_2d_array_i64_trap:
3670   case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
3671   case Intrinsic::nvvm_suld_3d_i64_trap:
3672   case Intrinsic::nvvm_suld_3d_v2i64_trap:
3673   case Intrinsic::nvvm_suld_1d_i64_zero:
3674   case Intrinsic::nvvm_suld_1d_v2i64_zero:
3675   case Intrinsic::nvvm_suld_1d_array_i64_zero:
3676   case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
3677   case Intrinsic::nvvm_suld_2d_i64_zero:
3678   case Intrinsic::nvvm_suld_2d_v2i64_zero:
3679   case Intrinsic::nvvm_suld_2d_array_i64_zero:
3680   case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
3681   case Intrinsic::nvvm_suld_3d_i64_zero:
3682   case Intrinsic::nvvm_suld_3d_v2i64_zero: {
3683     Info.opc = getOpcForSurfaceInstr(Intrinsic);
3684     Info.memVT = MVT::i64;
3685     Info.ptrVal = nullptr;
3686     Info.offset = 0;
3687     Info.vol = 0;
3688     Info.readMem = true;
3689     Info.writeMem = false;
3690     Info.align = 16;
3691     return true;
3692   }
3693   }
3694   return false;
3695 }
3696 
3697 /// isLegalAddressingMode - Return true if the addressing mode represented
3698 /// by AM is legal for this target, for a load/store of the specified type.
3699 /// Used to guide target specific optimizations, like loop strength reduction
3700 /// (LoopStrengthReduce.cpp) and memory optimization for address mode
3701 /// (CodeGenPrepare.cpp)
isLegalAddressingMode(const AddrMode & AM,Type * Ty) const3702 bool NVPTXTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3703                                                 Type *Ty) const {
3704 
3705   // AddrMode - This represents an addressing mode of:
3706   //    BaseGV + BaseOffs + BaseReg + Scale*ScaleReg
3707   //
3708   // The legal address modes are
3709   // - [avar]
3710   // - [areg]
3711   // - [areg+immoff]
3712   // - [immAddr]
3713 
3714   if (AM.BaseGV) {
3715     if (AM.BaseOffs || AM.HasBaseReg || AM.Scale)
3716       return false;
3717     return true;
3718   }
3719 
3720   switch (AM.Scale) {
3721   case 0: // "r", "r+i" or "i" is allowed
3722     break;
3723   case 1:
3724     if (AM.HasBaseReg) // "r+r+i" or "r+r" is not allowed.
3725       return false;
3726     // Otherwise we have r+i.
3727     break;
3728   default:
3729     // No scale > 1 is allowed
3730     return false;
3731   }
3732   return true;
3733 }
3734 
3735 //===----------------------------------------------------------------------===//
3736 //                         NVPTX Inline Assembly Support
3737 //===----------------------------------------------------------------------===//
3738 
3739 /// getConstraintType - Given a constraint letter, return the type of
3740 /// constraint it is for this target.
3741 NVPTXTargetLowering::ConstraintType
getConstraintType(const std::string & Constraint) const3742 NVPTXTargetLowering::getConstraintType(const std::string &Constraint) const {
3743   if (Constraint.size() == 1) {
3744     switch (Constraint[0]) {
3745     default:
3746       break;
3747     case 'b':
3748     case 'r':
3749     case 'h':
3750     case 'c':
3751     case 'l':
3752     case 'f':
3753     case 'd':
3754     case '0':
3755     case 'N':
3756       return C_RegisterClass;
3757     }
3758   }
3759   return TargetLowering::getConstraintType(Constraint);
3760 }
3761 
3762 std::pair<unsigned, const TargetRegisterClass *>
getRegForInlineAsmConstraint(const TargetRegisterInfo * TRI,const std::string & Constraint,MVT VT) const3763 NVPTXTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
3764                                                   const std::string &Constraint,
3765                                                   MVT VT) const {
3766   if (Constraint.size() == 1) {
3767     switch (Constraint[0]) {
3768     case 'b':
3769       return std::make_pair(0U, &NVPTX::Int1RegsRegClass);
3770     case 'c':
3771       return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
3772     case 'h':
3773       return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
3774     case 'r':
3775       return std::make_pair(0U, &NVPTX::Int32RegsRegClass);
3776     case 'l':
3777     case 'N':
3778       return std::make_pair(0U, &NVPTX::Int64RegsRegClass);
3779     case 'f':
3780       return std::make_pair(0U, &NVPTX::Float32RegsRegClass);
3781     case 'd':
3782       return std::make_pair(0U, &NVPTX::Float64RegsRegClass);
3783     }
3784   }
3785   return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
3786 }
3787 
3788 /// getFunctionAlignment - Return the Log2 alignment of this function.
getFunctionAlignment(const Function *) const3789 unsigned NVPTXTargetLowering::getFunctionAlignment(const Function *) const {
3790   return 4;
3791 }
3792 
3793 //===----------------------------------------------------------------------===//
3794 //                         NVPTX DAG Combining
3795 //===----------------------------------------------------------------------===//
3796 
allowFMA(MachineFunction & MF,CodeGenOpt::Level OptLevel) const3797 bool NVPTXTargetLowering::allowFMA(MachineFunction &MF,
3798                                    CodeGenOpt::Level OptLevel) const {
3799   const Function *F = MF.getFunction();
3800   const TargetOptions &TO = MF.getTarget().Options;
3801 
3802   // Always honor command-line argument
3803   if (FMAContractLevelOpt.getNumOccurrences() > 0) {
3804     return FMAContractLevelOpt > 0;
3805   } else if (OptLevel == 0) {
3806     // Do not contract if we're not optimizing the code
3807     return false;
3808   } else if (TO.AllowFPOpFusion == FPOpFusion::Fast || TO.UnsafeFPMath) {
3809     // Honor TargetOptions flags that explicitly say fusion is okay
3810     return true;
3811   } else if (F->hasFnAttribute("unsafe-fp-math")) {
3812     // Check for unsafe-fp-math=true coming from Clang
3813     Attribute Attr = F->getFnAttribute("unsafe-fp-math");
3814     StringRef Val = Attr.getValueAsString();
3815     if (Val == "true")
3816       return true;
3817   }
3818 
3819   // We did not have a clear indication that fusion is allowed, so assume not
3820   return false;
3821 }
3822 
3823 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
3824 /// operands N0 and N1.  This is a helper for PerformADDCombine that is
3825 /// called with the default operands, and if that fails, with commuted
3826 /// operands.
PerformADDCombineWithOperands(SDNode * N,SDValue N0,SDValue N1,TargetLowering::DAGCombinerInfo & DCI,const NVPTXSubtarget & Subtarget,CodeGenOpt::Level OptLevel)3827 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
3828                                            TargetLowering::DAGCombinerInfo &DCI,
3829                                              const NVPTXSubtarget &Subtarget,
3830                                              CodeGenOpt::Level OptLevel) {
3831   SelectionDAG  &DAG = DCI.DAG;
3832   // Skip non-integer, non-scalar case
3833   EVT VT=N0.getValueType();
3834   if (VT.isVector())
3835     return SDValue();
3836 
3837   // fold (add (mul a, b), c) -> (mad a, b, c)
3838   //
3839   if (N0.getOpcode() == ISD::MUL) {
3840     assert (VT.isInteger());
3841     // For integer:
3842     // Since integer multiply-add costs the same as integer multiply
3843     // but is more costly than integer add, do the fusion only when
3844     // the mul is only used in the add.
3845     if (OptLevel==CodeGenOpt::None || VT != MVT::i32 ||
3846         !N0.getNode()->hasOneUse())
3847       return SDValue();
3848 
3849     // Do the folding
3850     return DAG.getNode(NVPTXISD::IMAD, SDLoc(N), VT,
3851                        N0.getOperand(0), N0.getOperand(1), N1);
3852   }
3853   else if (N0.getOpcode() == ISD::FMUL) {
3854     if (VT == MVT::f32 || VT == MVT::f64) {
3855       const auto *TLI = static_cast<const NVPTXTargetLowering *>(
3856           &DAG.getTargetLoweringInfo());
3857       if (!TLI->allowFMA(DAG.getMachineFunction(), OptLevel))
3858         return SDValue();
3859 
3860       // For floating point:
3861       // Do the fusion only when the mul has less than 5 uses and all
3862       // are add.
3863       // The heuristic is that if a use is not an add, then that use
3864       // cannot be fused into fma, therefore mul is still needed anyway.
3865       // If there are more than 4 uses, even if they are all add, fusing
3866       // them will increase register pressue.
3867       //
3868       int numUses = 0;
3869       int nonAddCount = 0;
3870       for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
3871            UE = N0.getNode()->use_end();
3872            UI != UE; ++UI) {
3873         numUses++;
3874         SDNode *User = *UI;
3875         if (User->getOpcode() != ISD::FADD)
3876           ++nonAddCount;
3877       }
3878       if (numUses >= 5)
3879         return SDValue();
3880       if (nonAddCount) {
3881         int orderNo = N->getIROrder();
3882         int orderNo2 = N0.getNode()->getIROrder();
3883         // simple heuristics here for considering potential register
3884         // pressure, the logics here is that the differnce are used
3885         // to measure the distance between def and use, the longer distance
3886         // more likely cause register pressure.
3887         if (orderNo - orderNo2 < 500)
3888           return SDValue();
3889 
3890         // Now, check if at least one of the FMUL's operands is live beyond the node N,
3891         // which guarantees that the FMA will not increase register pressure at node N.
3892         bool opIsLive = false;
3893         const SDNode *left = N0.getOperand(0).getNode();
3894         const SDNode *right = N0.getOperand(1).getNode();
3895 
3896         if (isa<ConstantSDNode>(left) || isa<ConstantSDNode>(right))
3897           opIsLive = true;
3898 
3899         if (!opIsLive)
3900           for (SDNode::use_iterator UI = left->use_begin(), UE = left->use_end(); UI != UE; ++UI) {
3901             SDNode *User = *UI;
3902             int orderNo3 = User->getIROrder();
3903             if (orderNo3 > orderNo) {
3904               opIsLive = true;
3905               break;
3906             }
3907           }
3908 
3909         if (!opIsLive)
3910           for (SDNode::use_iterator UI = right->use_begin(), UE = right->use_end(); UI != UE; ++UI) {
3911             SDNode *User = *UI;
3912             int orderNo3 = User->getIROrder();
3913             if (orderNo3 > orderNo) {
3914               opIsLive = true;
3915               break;
3916             }
3917           }
3918 
3919         if (!opIsLive)
3920           return SDValue();
3921       }
3922 
3923       return DAG.getNode(ISD::FMA, SDLoc(N), VT,
3924                          N0.getOperand(0), N0.getOperand(1), N1);
3925     }
3926   }
3927 
3928   return SDValue();
3929 }
3930 
3931 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3932 ///
PerformADDCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,const NVPTXSubtarget & Subtarget,CodeGenOpt::Level OptLevel)3933 static SDValue PerformADDCombine(SDNode *N,
3934                                  TargetLowering::DAGCombinerInfo &DCI,
3935                                  const NVPTXSubtarget &Subtarget,
3936                                  CodeGenOpt::Level OptLevel) {
3937   SDValue N0 = N->getOperand(0);
3938   SDValue N1 = N->getOperand(1);
3939 
3940   // First try with the default operand order.
3941   SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget,
3942                                                  OptLevel);
3943   if (Result.getNode())
3944     return Result;
3945 
3946   // If that didn't work, try again with the operands commuted.
3947   return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget, OptLevel);
3948 }
3949 
PerformANDCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)3950 static SDValue PerformANDCombine(SDNode *N,
3951                                  TargetLowering::DAGCombinerInfo &DCI) {
3952   // The type legalizer turns a vector load of i8 values into a zextload to i16
3953   // registers, optionally ANY_EXTENDs it (if target type is integer),
3954   // and ANDs off the high 8 bits. Since we turn this load into a
3955   // target-specific DAG node, the DAG combiner fails to eliminate these AND
3956   // nodes. Do that here.
3957   SDValue Val = N->getOperand(0);
3958   SDValue Mask = N->getOperand(1);
3959 
3960   if (isa<ConstantSDNode>(Val)) {
3961     std::swap(Val, Mask);
3962   }
3963 
3964   SDValue AExt;
3965   // Generally, we will see zextload -> IMOV16rr -> ANY_EXTEND -> and
3966   if (Val.getOpcode() == ISD::ANY_EXTEND) {
3967     AExt = Val;
3968     Val = Val->getOperand(0);
3969   }
3970 
3971   if (Val->isMachineOpcode() && Val->getMachineOpcode() == NVPTX::IMOV16rr) {
3972     Val = Val->getOperand(0);
3973   }
3974 
3975   if (Val->getOpcode() == NVPTXISD::LoadV2 ||
3976       Val->getOpcode() == NVPTXISD::LoadV4) {
3977     ConstantSDNode *MaskCnst = dyn_cast<ConstantSDNode>(Mask);
3978     if (!MaskCnst) {
3979       // Not an AND with a constant
3980       return SDValue();
3981     }
3982 
3983     uint64_t MaskVal = MaskCnst->getZExtValue();
3984     if (MaskVal != 0xff) {
3985       // Not an AND that chops off top 8 bits
3986       return SDValue();
3987     }
3988 
3989     MemSDNode *Mem = dyn_cast<MemSDNode>(Val);
3990     if (!Mem) {
3991       // Not a MemSDNode?!?
3992       return SDValue();
3993     }
3994 
3995     EVT MemVT = Mem->getMemoryVT();
3996     if (MemVT != MVT::v2i8 && MemVT != MVT::v4i8) {
3997       // We only handle the i8 case
3998       return SDValue();
3999     }
4000 
4001     unsigned ExtType =
4002       cast<ConstantSDNode>(Val->getOperand(Val->getNumOperands()-1))->
4003         getZExtValue();
4004     if (ExtType == ISD::SEXTLOAD) {
4005       // If for some reason the load is a sextload, the and is needed to zero
4006       // out the high 8 bits
4007       return SDValue();
4008     }
4009 
4010     bool AddTo = false;
4011     if (AExt.getNode() != 0) {
4012       // Re-insert the ext as a zext.
4013       Val = DCI.DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
4014                             AExt.getValueType(), Val);
4015       AddTo = true;
4016     }
4017 
4018     // If we get here, the AND is unnecessary.  Just replace it with the load
4019     DCI.CombineTo(N, Val, AddTo);
4020   }
4021 
4022   return SDValue();
4023 }
4024 
4025 enum OperandSignedness {
4026   Signed = 0,
4027   Unsigned,
4028   Unknown
4029 };
4030 
4031 /// IsMulWideOperandDemotable - Checks if the provided DAG node is an operand
4032 /// that can be demoted to \p OptSize bits without loss of information. The
4033 /// signedness of the operand, if determinable, is placed in \p S.
IsMulWideOperandDemotable(SDValue Op,unsigned OptSize,OperandSignedness & S)4034 static bool IsMulWideOperandDemotable(SDValue Op,
4035                                       unsigned OptSize,
4036                                       OperandSignedness &S) {
4037   S = Unknown;
4038 
4039   if (Op.getOpcode() == ISD::SIGN_EXTEND ||
4040       Op.getOpcode() == ISD::SIGN_EXTEND_INREG) {
4041     EVT OrigVT = Op.getOperand(0).getValueType();
4042     if (OrigVT.getSizeInBits() <= OptSize) {
4043       S = Signed;
4044       return true;
4045     }
4046   } else if (Op.getOpcode() == ISD::ZERO_EXTEND) {
4047     EVT OrigVT = Op.getOperand(0).getValueType();
4048     if (OrigVT.getSizeInBits() <= OptSize) {
4049       S = Unsigned;
4050       return true;
4051     }
4052   }
4053 
4054   return false;
4055 }
4056 
4057 /// AreMulWideOperandsDemotable - Checks if the given LHS and RHS operands can
4058 /// be demoted to \p OptSize bits without loss of information. If the operands
4059 /// contain a constant, it should appear as the RHS operand. The signedness of
4060 /// the operands is placed in \p IsSigned.
AreMulWideOperandsDemotable(SDValue LHS,SDValue RHS,unsigned OptSize,bool & IsSigned)4061 static bool AreMulWideOperandsDemotable(SDValue LHS, SDValue RHS,
4062                                         unsigned OptSize,
4063                                         bool &IsSigned) {
4064 
4065   OperandSignedness LHSSign;
4066 
4067   // The LHS operand must be a demotable op
4068   if (!IsMulWideOperandDemotable(LHS, OptSize, LHSSign))
4069     return false;
4070 
4071   // We should have been able to determine the signedness from the LHS
4072   if (LHSSign == Unknown)
4073     return false;
4074 
4075   IsSigned = (LHSSign == Signed);
4076 
4077   // The RHS can be a demotable op or a constant
4078   if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(RHS)) {
4079     APInt Val = CI->getAPIntValue();
4080     if (LHSSign == Unsigned) {
4081       if (Val.isIntN(OptSize)) {
4082         return true;
4083       }
4084       return false;
4085     } else {
4086       if (Val.isSignedIntN(OptSize)) {
4087         return true;
4088       }
4089       return false;
4090     }
4091   } else {
4092     OperandSignedness RHSSign;
4093     if (!IsMulWideOperandDemotable(RHS, OptSize, RHSSign))
4094       return false;
4095 
4096     if (LHSSign != RHSSign)
4097       return false;
4098 
4099     return true;
4100   }
4101 }
4102 
4103 /// TryMULWIDECombine - Attempt to replace a multiply of M bits with a multiply
4104 /// of M/2 bits that produces an M-bit result (i.e. mul.wide). This transform
4105 /// works on both multiply DAG nodes and SHL DAG nodes with a constant shift
4106 /// amount.
TryMULWIDECombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)4107 static SDValue TryMULWIDECombine(SDNode *N,
4108                                  TargetLowering::DAGCombinerInfo &DCI) {
4109   EVT MulType = N->getValueType(0);
4110   if (MulType != MVT::i32 && MulType != MVT::i64) {
4111     return SDValue();
4112   }
4113 
4114   unsigned OptSize = MulType.getSizeInBits() >> 1;
4115   SDValue LHS = N->getOperand(0);
4116   SDValue RHS = N->getOperand(1);
4117 
4118   // Canonicalize the multiply so the constant (if any) is on the right
4119   if (N->getOpcode() == ISD::MUL) {
4120     if (isa<ConstantSDNode>(LHS)) {
4121       std::swap(LHS, RHS);
4122     }
4123   }
4124 
4125   // If we have a SHL, determine the actual multiply amount
4126   if (N->getOpcode() == ISD::SHL) {
4127     ConstantSDNode *ShlRHS = dyn_cast<ConstantSDNode>(RHS);
4128     if (!ShlRHS) {
4129       return SDValue();
4130     }
4131 
4132     APInt ShiftAmt = ShlRHS->getAPIntValue();
4133     unsigned BitWidth = MulType.getSizeInBits();
4134     if (ShiftAmt.sge(0) && ShiftAmt.slt(BitWidth)) {
4135       APInt MulVal = APInt(BitWidth, 1) << ShiftAmt;
4136       RHS = DCI.DAG.getConstant(MulVal, MulType);
4137     } else {
4138       return SDValue();
4139     }
4140   }
4141 
4142   bool Signed;
4143   // Verify that our operands are demotable
4144   if (!AreMulWideOperandsDemotable(LHS, RHS, OptSize, Signed)) {
4145     return SDValue();
4146   }
4147 
4148   EVT DemotedVT;
4149   if (MulType == MVT::i32) {
4150     DemotedVT = MVT::i16;
4151   } else {
4152     DemotedVT = MVT::i32;
4153   }
4154 
4155   // Truncate the operands to the correct size. Note that these are just for
4156   // type consistency and will (likely) be eliminated in later phases.
4157   SDValue TruncLHS =
4158     DCI.DAG.getNode(ISD::TRUNCATE, SDLoc(N), DemotedVT, LHS);
4159   SDValue TruncRHS =
4160     DCI.DAG.getNode(ISD::TRUNCATE, SDLoc(N), DemotedVT, RHS);
4161 
4162   unsigned Opc;
4163   if (Signed) {
4164     Opc = NVPTXISD::MUL_WIDE_SIGNED;
4165   } else {
4166     Opc = NVPTXISD::MUL_WIDE_UNSIGNED;
4167   }
4168 
4169   return DCI.DAG.getNode(Opc, SDLoc(N), MulType, TruncLHS, TruncRHS);
4170 }
4171 
4172 /// PerformMULCombine - Runs PTX-specific DAG combine patterns on MUL nodes.
PerformMULCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,CodeGenOpt::Level OptLevel)4173 static SDValue PerformMULCombine(SDNode *N,
4174                                  TargetLowering::DAGCombinerInfo &DCI,
4175                                  CodeGenOpt::Level OptLevel) {
4176   if (OptLevel > 0) {
4177     // Try mul.wide combining at OptLevel > 0
4178     SDValue Ret = TryMULWIDECombine(N, DCI);
4179     if (Ret.getNode())
4180       return Ret;
4181   }
4182 
4183   return SDValue();
4184 }
4185 
4186 /// PerformSHLCombine - Runs PTX-specific DAG combine patterns on SHL nodes.
PerformSHLCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI,CodeGenOpt::Level OptLevel)4187 static SDValue PerformSHLCombine(SDNode *N,
4188                                  TargetLowering::DAGCombinerInfo &DCI,
4189                                  CodeGenOpt::Level OptLevel) {
4190   if (OptLevel > 0) {
4191     // Try mul.wide combining at OptLevel > 0
4192     SDValue Ret = TryMULWIDECombine(N, DCI);
4193     if (Ret.getNode())
4194       return Ret;
4195   }
4196 
4197   return SDValue();
4198 }
4199 
PerformDAGCombine(SDNode * N,DAGCombinerInfo & DCI) const4200 SDValue NVPTXTargetLowering::PerformDAGCombine(SDNode *N,
4201                                                DAGCombinerInfo &DCI) const {
4202   CodeGenOpt::Level OptLevel = getTargetMachine().getOptLevel();
4203   switch (N->getOpcode()) {
4204     default: break;
4205     case ISD::ADD:
4206     case ISD::FADD:
4207       return PerformADDCombine(N, DCI, STI, OptLevel);
4208     case ISD::MUL:
4209       return PerformMULCombine(N, DCI, OptLevel);
4210     case ISD::SHL:
4211       return PerformSHLCombine(N, DCI, OptLevel);
4212     case ISD::AND:
4213       return PerformANDCombine(N, DCI);
4214   }
4215   return SDValue();
4216 }
4217 
4218 /// ReplaceVectorLoad - Convert vector loads into multi-output scalar loads.
ReplaceLoadVector(SDNode * N,SelectionDAG & DAG,const DataLayout * TD,SmallVectorImpl<SDValue> & Results)4219 static void ReplaceLoadVector(SDNode *N, SelectionDAG &DAG,
4220                               const DataLayout *TD,
4221                               SmallVectorImpl<SDValue> &Results) {
4222   EVT ResVT = N->getValueType(0);
4223   SDLoc DL(N);
4224 
4225   assert(ResVT.isVector() && "Vector load must have vector type");
4226 
4227   // We only handle "native" vector sizes for now, e.g. <4 x double> is not
4228   // legal.  We can (and should) split that into 2 loads of <2 x double> here
4229   // but I'm leaving that as a TODO for now.
4230   assert(ResVT.isSimple() && "Can only handle simple types");
4231   switch (ResVT.getSimpleVT().SimpleTy) {
4232   default:
4233     return;
4234   case MVT::v2i8:
4235   case MVT::v2i16:
4236   case MVT::v2i32:
4237   case MVT::v2i64:
4238   case MVT::v2f32:
4239   case MVT::v2f64:
4240   case MVT::v4i8:
4241   case MVT::v4i16:
4242   case MVT::v4i32:
4243   case MVT::v4f32:
4244     // This is a "native" vector type
4245     break;
4246   }
4247 
4248   LoadSDNode *LD = cast<LoadSDNode>(N);
4249 
4250   unsigned Align = LD->getAlignment();
4251   unsigned PrefAlign =
4252     TD->getPrefTypeAlignment(ResVT.getTypeForEVT(*DAG.getContext()));
4253   if (Align < PrefAlign) {
4254     // This load is not sufficiently aligned, so bail out and let this vector
4255     // load be scalarized.  Note that we may still be able to emit smaller
4256     // vector loads.  For example, if we are loading a <4 x float> with an
4257     // alignment of 8, this check will fail but the legalizer will try again
4258     // with 2 x <2 x float>, which will succeed with an alignment of 8.
4259     return;
4260   }
4261 
4262   EVT EltVT = ResVT.getVectorElementType();
4263   unsigned NumElts = ResVT.getVectorNumElements();
4264 
4265   // Since LoadV2 is a target node, we cannot rely on DAG type legalization.
4266   // Therefore, we must ensure the type is legal.  For i1 and i8, we set the
4267   // loaded type to i16 and propagate the "real" type as the memory type.
4268   bool NeedTrunc = false;
4269   if (EltVT.getSizeInBits() < 16) {
4270     EltVT = MVT::i16;
4271     NeedTrunc = true;
4272   }
4273 
4274   unsigned Opcode = 0;
4275   SDVTList LdResVTs;
4276 
4277   switch (NumElts) {
4278   default:
4279     return;
4280   case 2:
4281     Opcode = NVPTXISD::LoadV2;
4282     LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
4283     break;
4284   case 4: {
4285     Opcode = NVPTXISD::LoadV4;
4286     EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
4287     LdResVTs = DAG.getVTList(ListVTs);
4288     break;
4289   }
4290   }
4291 
4292   // Copy regular operands
4293   SmallVector<SDValue, 8> OtherOps(N->op_begin(), N->op_end());
4294 
4295   // The select routine does not have access to the LoadSDNode instance, so
4296   // pass along the extension information
4297   OtherOps.push_back(DAG.getIntPtrConstant(LD->getExtensionType()));
4298 
4299   SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
4300                                           LD->getMemoryVT(),
4301                                           LD->getMemOperand());
4302 
4303   SmallVector<SDValue, 4> ScalarRes;
4304 
4305   for (unsigned i = 0; i < NumElts; ++i) {
4306     SDValue Res = NewLD.getValue(i);
4307     if (NeedTrunc)
4308       Res = DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
4309     ScalarRes.push_back(Res);
4310   }
4311 
4312   SDValue LoadChain = NewLD.getValue(NumElts);
4313 
4314   SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, ResVT, ScalarRes);
4315 
4316   Results.push_back(BuildVec);
4317   Results.push_back(LoadChain);
4318 }
4319 
ReplaceINTRINSIC_W_CHAIN(SDNode * N,SelectionDAG & DAG,SmallVectorImpl<SDValue> & Results)4320 static void ReplaceINTRINSIC_W_CHAIN(SDNode *N, SelectionDAG &DAG,
4321                                      SmallVectorImpl<SDValue> &Results) {
4322   SDValue Chain = N->getOperand(0);
4323   SDValue Intrin = N->getOperand(1);
4324   SDLoc DL(N);
4325 
4326   // Get the intrinsic ID
4327   unsigned IntrinNo = cast<ConstantSDNode>(Intrin.getNode())->getZExtValue();
4328   switch (IntrinNo) {
4329   default:
4330     return;
4331   case Intrinsic::nvvm_ldg_global_i:
4332   case Intrinsic::nvvm_ldg_global_f:
4333   case Intrinsic::nvvm_ldg_global_p:
4334   case Intrinsic::nvvm_ldu_global_i:
4335   case Intrinsic::nvvm_ldu_global_f:
4336   case Intrinsic::nvvm_ldu_global_p: {
4337     EVT ResVT = N->getValueType(0);
4338 
4339     if (ResVT.isVector()) {
4340       // Vector LDG/LDU
4341 
4342       unsigned NumElts = ResVT.getVectorNumElements();
4343       EVT EltVT = ResVT.getVectorElementType();
4344 
4345       // Since LDU/LDG are target nodes, we cannot rely on DAG type
4346       // legalization.
4347       // Therefore, we must ensure the type is legal.  For i1 and i8, we set the
4348       // loaded type to i16 and propagate the "real" type as the memory type.
4349       bool NeedTrunc = false;
4350       if (EltVT.getSizeInBits() < 16) {
4351         EltVT = MVT::i16;
4352         NeedTrunc = true;
4353       }
4354 
4355       unsigned Opcode = 0;
4356       SDVTList LdResVTs;
4357 
4358       switch (NumElts) {
4359       default:
4360         return;
4361       case 2:
4362         switch (IntrinNo) {
4363         default:
4364           return;
4365         case Intrinsic::nvvm_ldg_global_i:
4366         case Intrinsic::nvvm_ldg_global_f:
4367         case Intrinsic::nvvm_ldg_global_p:
4368           Opcode = NVPTXISD::LDGV2;
4369           break;
4370         case Intrinsic::nvvm_ldu_global_i:
4371         case Intrinsic::nvvm_ldu_global_f:
4372         case Intrinsic::nvvm_ldu_global_p:
4373           Opcode = NVPTXISD::LDUV2;
4374           break;
4375         }
4376         LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
4377         break;
4378       case 4: {
4379         switch (IntrinNo) {
4380         default:
4381           return;
4382         case Intrinsic::nvvm_ldg_global_i:
4383         case Intrinsic::nvvm_ldg_global_f:
4384         case Intrinsic::nvvm_ldg_global_p:
4385           Opcode = NVPTXISD::LDGV4;
4386           break;
4387         case Intrinsic::nvvm_ldu_global_i:
4388         case Intrinsic::nvvm_ldu_global_f:
4389         case Intrinsic::nvvm_ldu_global_p:
4390           Opcode = NVPTXISD::LDUV4;
4391           break;
4392         }
4393         EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
4394         LdResVTs = DAG.getVTList(ListVTs);
4395         break;
4396       }
4397       }
4398 
4399       SmallVector<SDValue, 8> OtherOps;
4400 
4401       // Copy regular operands
4402 
4403       OtherOps.push_back(Chain); // Chain
4404                                  // Skip operand 1 (intrinsic ID)
4405       // Others
4406       OtherOps.append(N->op_begin() + 2, N->op_end());
4407 
4408       MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
4409 
4410       SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
4411                                               MemSD->getMemoryVT(),
4412                                               MemSD->getMemOperand());
4413 
4414       SmallVector<SDValue, 4> ScalarRes;
4415 
4416       for (unsigned i = 0; i < NumElts; ++i) {
4417         SDValue Res = NewLD.getValue(i);
4418         if (NeedTrunc)
4419           Res =
4420               DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
4421         ScalarRes.push_back(Res);
4422       }
4423 
4424       SDValue LoadChain = NewLD.getValue(NumElts);
4425 
4426       SDValue BuildVec =
4427           DAG.getNode(ISD::BUILD_VECTOR, DL, ResVT, ScalarRes);
4428 
4429       Results.push_back(BuildVec);
4430       Results.push_back(LoadChain);
4431     } else {
4432       // i8 LDG/LDU
4433       assert(ResVT.isSimple() && ResVT.getSimpleVT().SimpleTy == MVT::i8 &&
4434              "Custom handling of non-i8 ldu/ldg?");
4435 
4436       // Just copy all operands as-is
4437       SmallVector<SDValue, 4> Ops(N->op_begin(), N->op_end());
4438 
4439       // Force output to i16
4440       SDVTList LdResVTs = DAG.getVTList(MVT::i16, MVT::Other);
4441 
4442       MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
4443 
4444       // We make sure the memory type is i8, which will be used during isel
4445       // to select the proper instruction.
4446       SDValue NewLD =
4447           DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, LdResVTs, Ops,
4448                                   MVT::i8, MemSD->getMemOperand());
4449 
4450       Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
4451                                     NewLD.getValue(0)));
4452       Results.push_back(NewLD.getValue(1));
4453     }
4454   }
4455   }
4456 }
4457 
ReplaceNodeResults(SDNode * N,SmallVectorImpl<SDValue> & Results,SelectionDAG & DAG) const4458 void NVPTXTargetLowering::ReplaceNodeResults(
4459     SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
4460   switch (N->getOpcode()) {
4461   default:
4462     report_fatal_error("Unhandled custom legalization");
4463   case ISD::LOAD:
4464     ReplaceLoadVector(N, DAG, getDataLayout(), Results);
4465     return;
4466   case ISD::INTRINSIC_W_CHAIN:
4467     ReplaceINTRINSIC_W_CHAIN(N, DAG, Results);
4468     return;
4469   }
4470 }
4471 
4472 // Pin NVPTXSection's and NVPTXTargetObjectFile's vtables to this file.
anchor()4473 void NVPTXSection::anchor() {}
4474 
~NVPTXTargetObjectFile()4475 NVPTXTargetObjectFile::~NVPTXTargetObjectFile() {
4476   delete TextSection;
4477   delete DataSection;
4478   delete BSSSection;
4479   delete ReadOnlySection;
4480 
4481   delete StaticCtorSection;
4482   delete StaticDtorSection;
4483   delete LSDASection;
4484   delete EHFrameSection;
4485   delete DwarfAbbrevSection;
4486   delete DwarfInfoSection;
4487   delete DwarfLineSection;
4488   delete DwarfFrameSection;
4489   delete DwarfPubTypesSection;
4490   delete DwarfDebugInlineSection;
4491   delete DwarfStrSection;
4492   delete DwarfLocSection;
4493   delete DwarfARangesSection;
4494   delete DwarfRangesSection;
4495 }
4496 
4497 const MCSection *
SelectSectionForGlobal(const GlobalValue * GV,SectionKind Kind,Mangler & Mang,const TargetMachine & TM) const4498 NVPTXTargetObjectFile::SelectSectionForGlobal(const GlobalValue *GV,
4499                                               SectionKind Kind, Mangler &Mang,
4500                                               const TargetMachine &TM) const {
4501   return getDataSection();
4502 }
4503