1==========================================
2The LLVM Target-Independent Code Generator
3==========================================
4
5.. role:: raw-html(raw)
6   :format: html
7
8.. raw:: html
9
10  <style>
11    .unknown { background-color: #C0C0C0; text-align: center; }
12    .unknown:before { content: "?" }
13    .no { background-color: #C11B17 }
14    .no:before { content: "N" }
15    .partial { background-color: #F88017 }
16    .yes { background-color: #0F0; }
17    .yes:before { content: "Y" }
18    .na { background-color: #6666FF; }
19    .na:before { content: "N/A" }
20  </style>
21
22.. contents::
23   :local:
24
25.. warning::
26  This is a work in progress.
27
28Introduction
29============
30
31The LLVM target-independent code generator is a framework that provides a suite
32of reusable components for translating the LLVM internal representation to the
33machine code for a specified target---either in assembly form (suitable for a
34static compiler) or in binary machine code format (usable for a JIT
35compiler). The LLVM target-independent code generator consists of six main
36components:
37
381. `Abstract target description`_ interfaces which capture important properties
39   about various aspects of the machine, independently of how they will be used.
40   These interfaces are defined in ``include/llvm/Target/``.
41
422. Classes used to represent the `code being generated`_ for a target.  These
43   classes are intended to be abstract enough to represent the machine code for
44   *any* target machine.  These classes are defined in
45   ``include/llvm/CodeGen/``. At this level, concepts like "constant pool
46   entries" and "jump tables" are explicitly exposed.
47
483. Classes and algorithms used to represent code as the object file level, the
49   `MC Layer`_.  These classes represent assembly level constructs like labels,
50   sections, and instructions.  At this level, concepts like "constant pool
51   entries" and "jump tables" don't exist.
52
534. `Target-independent algorithms`_ used to implement various phases of native
54   code generation (register allocation, scheduling, stack frame representation,
55   etc).  This code lives in ``lib/CodeGen/``.
56
575. `Implementations of the abstract target description interfaces`_ for
58   particular targets.  These machine descriptions make use of the components
59   provided by LLVM, and can optionally provide custom target-specific passes,
60   to build complete code generators for a specific target.  Target descriptions
61   live in ``lib/Target/``.
62
636. The target-independent JIT components.  The LLVM JIT is completely target
64   independent (it uses the ``TargetJITInfo`` structure to interface for
65   target-specific issues.  The code for the target-independent JIT lives in
66   ``lib/ExecutionEngine/JIT``.
67
68Depending on which part of the code generator you are interested in working on,
69different pieces of this will be useful to you.  In any case, you should be
70familiar with the `target description`_ and `machine code representation`_
71classes.  If you want to add a backend for a new target, you will need to
72`implement the target description`_ classes for your new target and understand
73the :doc:`LLVM code representation <LangRef>`.  If you are interested in
74implementing a new `code generation algorithm`_, it should only depend on the
75target-description and machine code representation classes, ensuring that it is
76portable.
77
78Required components in the code generator
79-----------------------------------------
80
81The two pieces of the LLVM code generator are the high-level interface to the
82code generator and the set of reusable components that can be used to build
83target-specific backends.  The two most important interfaces (:raw-html:`<tt>`
84`TargetMachine`_ :raw-html:`</tt>` and :raw-html:`<tt>` `DataLayout`_
85:raw-html:`</tt>`) are the only ones that are required to be defined for a
86backend to fit into the LLVM system, but the others must be defined if the
87reusable code generator components are going to be used.
88
89This design has two important implications.  The first is that LLVM can support
90completely non-traditional code generation targets.  For example, the C backend
91does not require register allocation, instruction selection, or any of the other
92standard components provided by the system.  As such, it only implements these
93two interfaces, and does its own thing. Note that C backend was removed from the
94trunk since LLVM 3.1 release. Another example of a code generator like this is a
95(purely hypothetical) backend that converts LLVM to the GCC RTL form and uses
96GCC to emit machine code for a target.
97
98This design also implies that it is possible to design and implement radically
99different code generators in the LLVM system that do not make use of any of the
100built-in components.  Doing so is not recommended at all, but could be required
101for radically different targets that do not fit into the LLVM machine
102description model: FPGAs for example.
103
104.. _high-level design of the code generator:
105
106The high-level design of the code generator
107-------------------------------------------
108
109The LLVM target-independent code generator is designed to support efficient and
110quality code generation for standard register-based microprocessors.  Code
111generation in this model is divided into the following stages:
112
1131. `Instruction Selection`_ --- This phase determines an efficient way to
114   express the input LLVM code in the target instruction set.  This stage
115   produces the initial code for the program in the target instruction set, then
116   makes use of virtual registers in SSA form and physical registers that
117   represent any required register assignments due to target constraints or
118   calling conventions.  This step turns the LLVM code into a DAG of target
119   instructions.
120
1212. `Scheduling and Formation`_ --- This phase takes the DAG of target
122   instructions produced by the instruction selection phase, determines an
123   ordering of the instructions, then emits the instructions as :raw-html:`<tt>`
124   `MachineInstr`_\s :raw-html:`</tt>` with that ordering.  Note that we
125   describe this in the `instruction selection section`_ because it operates on
126   a `SelectionDAG`_.
127
1283. `SSA-based Machine Code Optimizations`_ --- This optional stage consists of a
129   series of machine-code optimizations that operate on the SSA-form produced by
130   the instruction selector.  Optimizations like modulo-scheduling or peephole
131   optimization work here.
132
1334. `Register Allocation`_ --- The target code is transformed from an infinite
134   virtual register file in SSA form to the concrete register file used by the
135   target.  This phase introduces spill code and eliminates all virtual register
136   references from the program.
137
1385. `Prolog/Epilog Code Insertion`_ --- Once the machine code has been generated
139   for the function and the amount of stack space required is known (used for
140   LLVM alloca's and spill slots), the prolog and epilog code for the function
141   can be inserted and "abstract stack location references" can be eliminated.
142   This stage is responsible for implementing optimizations like frame-pointer
143   elimination and stack packing.
144
1456. `Late Machine Code Optimizations`_ --- Optimizations that operate on "final"
146   machine code can go here, such as spill code scheduling and peephole
147   optimizations.
148
1497. `Code Emission`_ --- The final stage actually puts out the code for the
150   current function, either in the target assembler format or in machine
151   code.
152
153The code generator is based on the assumption that the instruction selector will
154use an optimal pattern matching selector to create high-quality sequences of
155native instructions.  Alternative code generator designs based on pattern
156expansion and aggressive iterative peephole optimization are much slower.  This
157design permits efficient compilation (important for JIT environments) and
158aggressive optimization (used when generating code offline) by allowing
159components of varying levels of sophistication to be used for any step of
160compilation.
161
162In addition to these stages, target implementations can insert arbitrary
163target-specific passes into the flow.  For example, the X86 target uses a
164special pass to handle the 80x87 floating point stack architecture.  Other
165targets with unusual requirements can be supported with custom passes as needed.
166
167Using TableGen for target description
168-------------------------------------
169
170The target description classes require a detailed description of the target
171architecture.  These target descriptions often have a large amount of common
172information (e.g., an ``add`` instruction is almost identical to a ``sub``
173instruction).  In order to allow the maximum amount of commonality to be
174factored out, the LLVM code generator uses the
175:doc:`TableGen/index` tool to describe big chunks of the
176target machine, which allows the use of domain-specific and target-specific
177abstractions to reduce the amount of repetition.
178
179As LLVM continues to be developed and refined, we plan to move more and more of
180the target description to the ``.td`` form.  Doing so gives us a number of
181advantages.  The most important is that it makes it easier to port LLVM because
182it reduces the amount of C++ code that has to be written, and the surface area
183of the code generator that needs to be understood before someone can get
184something working.  Second, it makes it easier to change things. In particular,
185if tables and other things are all emitted by ``tblgen``, we only need a change
186in one place (``tblgen``) to update all of the targets to a new interface.
187
188.. _Abstract target description:
189.. _target description:
190
191Target description classes
192==========================
193
194The LLVM target description classes (located in the ``include/llvm/Target``
195directory) provide an abstract description of the target machine independent of
196any particular client.  These classes are designed to capture the *abstract*
197properties of the target (such as the instructions and registers it has), and do
198not incorporate any particular pieces of code generation algorithms.
199
200All of the target description classes (except the :raw-html:`<tt>` `DataLayout`_
201:raw-html:`</tt>` class) are designed to be subclassed by the concrete target
202implementation, and have virtual methods implemented.  To get to these
203implementations, the :raw-html:`<tt>` `TargetMachine`_ :raw-html:`</tt>` class
204provides accessors that should be implemented by the target.
205
206.. _TargetMachine:
207
208The ``TargetMachine`` class
209---------------------------
210
211The ``TargetMachine`` class provides virtual methods that are used to access the
212target-specific implementations of the various target description classes via
213the ``get*Info`` methods (``getInstrInfo``, ``getRegisterInfo``,
214``getFrameInfo``, etc.).  This class is designed to be specialized by a concrete
215target implementation (e.g., ``X86TargetMachine``) which implements the various
216virtual methods.  The only required target description class is the
217:raw-html:`<tt>` `DataLayout`_ :raw-html:`</tt>` class, but if the code
218generator components are to be used, the other interfaces should be implemented
219as well.
220
221.. _DataLayout:
222
223The ``DataLayout`` class
224------------------------
225
226The ``DataLayout`` class is the only required target description class, and it
227is the only class that is not extensible (you cannot derive a new class from
228it).  ``DataLayout`` specifies information about how the target lays out memory
229for structures, the alignment requirements for various data types, the size of
230pointers in the target, and whether the target is little-endian or
231big-endian.
232
233.. _TargetLowering:
234
235The ``TargetLowering`` class
236----------------------------
237
238The ``TargetLowering`` class is used by SelectionDAG based instruction selectors
239primarily to describe how LLVM code should be lowered to SelectionDAG
240operations.  Among other things, this class indicates:
241
242* an initial register class to use for various ``ValueType``\s,
243
244* which operations are natively supported by the target machine,
245
246* the return type of ``setcc`` operations,
247
248* the type to use for shift amounts, and
249
250* various high-level characteristics, like whether it is profitable to turn
251  division by a constant into a multiplication sequence.
252
253.. _TargetRegisterInfo:
254
255The ``TargetRegisterInfo`` class
256--------------------------------
257
258The ``TargetRegisterInfo`` class is used to describe the register file of the
259target and any interactions between the registers.
260
261Registers are represented in the code generator by unsigned integers.  Physical
262registers (those that actually exist in the target description) are unique
263small numbers, and virtual registers are generally large.  Note that
264register ``#0`` is reserved as a flag value.
265
266Each register in the processor description has an associated
267``TargetRegisterDesc`` entry, which provides a textual name for the register
268(used for assembly output and debugging dumps) and a set of aliases (used to
269indicate whether one register overlaps with another).
270
271In addition to the per-register description, the ``TargetRegisterInfo`` class
272exposes a set of processor specific register classes (instances of the
273``TargetRegisterClass`` class).  Each register class contains sets of registers
274that have the same properties (for example, they are all 32-bit integer
275registers).  Each SSA virtual register created by the instruction selector has
276an associated register class.  When the register allocator runs, it replaces
277virtual registers with a physical register in the set.
278
279The target-specific implementations of these classes is auto-generated from a
280:doc:`TableGen/index` description of the register file.
281
282.. _TargetInstrInfo:
283
284The ``TargetInstrInfo`` class
285-----------------------------
286
287The ``TargetInstrInfo`` class is used to describe the machine instructions
288supported by the target.  Descriptions define things like the mnemonic for
289the opcode, the number of operands, the list of implicit register uses and defs,
290whether the instruction has certain target-independent properties (accesses
291memory, is commutable, etc), and holds any target-specific flags.
292
293The ``TargetFrameLowering`` class
294---------------------------------
295
296The ``TargetFrameLowering`` class is used to provide information about the stack
297frame layout of the target. It holds the direction of stack growth, the known
298stack alignment on entry to each function, and the offset to the local area.
299The offset to the local area is the offset from the stack pointer on function
300entry to the first location where function data (local variables, spill
301locations) can be stored.
302
303The ``TargetSubtarget`` class
304-----------------------------
305
306The ``TargetSubtarget`` class is used to provide information about the specific
307chip set being targeted.  A sub-target informs code generation of which
308instructions are supported, instruction latencies and instruction execution
309itinerary; i.e., which processing units are used, in what order, and for how
310long.
311
312The ``TargetJITInfo`` class
313---------------------------
314
315The ``TargetJITInfo`` class exposes an abstract interface used by the
316Just-In-Time code generator to perform target-specific activities, such as
317emitting stubs.  If a ``TargetMachine`` supports JIT code generation, it should
318provide one of these objects through the ``getJITInfo`` method.
319
320.. _code being generated:
321.. _machine code representation:
322
323Machine code description classes
324================================
325
326At the high-level, LLVM code is translated to a machine specific representation
327formed out of :raw-html:`<tt>` `MachineFunction`_ :raw-html:`</tt>`,
328:raw-html:`<tt>` `MachineBasicBlock`_ :raw-html:`</tt>`, and :raw-html:`<tt>`
329`MachineInstr`_ :raw-html:`</tt>` instances (defined in
330``include/llvm/CodeGen``).  This representation is completely target agnostic,
331representing instructions in their most abstract form: an opcode and a series of
332operands.  This representation is designed to support both an SSA representation
333for machine code, as well as a register allocated, non-SSA form.
334
335.. _MachineInstr:
336
337The ``MachineInstr`` class
338--------------------------
339
340Target machine instructions are represented as instances of the ``MachineInstr``
341class.  This class is an extremely abstract way of representing machine
342instructions.  In particular, it only keeps track of an opcode number and a set
343of operands.
344
345The opcode number is a simple unsigned integer that only has meaning to a
346specific backend.  All of the instructions for a target should be defined in the
347``*InstrInfo.td`` file for the target. The opcode enum values are auto-generated
348from this description.  The ``MachineInstr`` class does not have any information
349about how to interpret the instruction (i.e., what the semantics of the
350instruction are); for that you must refer to the :raw-html:`<tt>`
351`TargetInstrInfo`_ :raw-html:`</tt>` class.
352
353The operands of a machine instruction can be of several different types: a
354register reference, a constant integer, a basic block reference, etc.  In
355addition, a machine operand should be marked as a def or a use of the value
356(though only registers are allowed to be defs).
357
358By convention, the LLVM code generator orders instruction operands so that all
359register definitions come before the register uses, even on architectures that
360are normally printed in other orders.  For example, the SPARC add instruction:
361"``add %i1, %i2, %i3``" adds the "%i1", and "%i2" registers and stores the
362result into the "%i3" register.  In the LLVM code generator, the operands should
363be stored as "``%i3, %i1, %i2``": with the destination first.
364
365Keeping destination (definition) operands at the beginning of the operand list
366has several advantages.  In particular, the debugging printer will print the
367instruction like this:
368
369.. code-block:: llvm
370
371  %r3 = add %i1, %i2
372
373Also if the first operand is a def, it is easier to `create instructions`_ whose
374only def is the first operand.
375
376.. _create instructions:
377
378Using the ``MachineInstrBuilder.h`` functions
379^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
380
381Machine instructions are created by using the ``BuildMI`` functions, located in
382the ``include/llvm/CodeGen/MachineInstrBuilder.h`` file.  The ``BuildMI``
383functions make it easy to build arbitrary machine instructions.  Usage of the
384``BuildMI`` functions look like this:
385
386.. code-block:: c++
387
388  // Create a 'DestReg = mov 42' (rendered in X86 assembly as 'mov DestReg, 42')
389  // instruction.  The '1' specifies how many operands will be added.
390  MachineInstr *MI = BuildMI(X86::MOV32ri, 1, DestReg).addImm(42);
391
392  // Create the same instr, but insert it at the end of a basic block.
393  MachineBasicBlock &MBB = ...
394  BuildMI(MBB, X86::MOV32ri, 1, DestReg).addImm(42);
395
396  // Create the same instr, but insert it before a specified iterator point.
397  MachineBasicBlock::iterator MBBI = ...
398  BuildMI(MBB, MBBI, X86::MOV32ri, 1, DestReg).addImm(42);
399
400  // Create a 'cmp Reg, 0' instruction, no destination reg.
401  MI = BuildMI(X86::CMP32ri, 2).addReg(Reg).addImm(0);
402
403  // Create an 'sahf' instruction which takes no operands and stores nothing.
404  MI = BuildMI(X86::SAHF, 0);
405
406  // Create a self looping branch instruction.
407  BuildMI(MBB, X86::JNE, 1).addMBB(&MBB);
408
409The key thing to remember with the ``BuildMI`` functions is that you have to
410specify the number of operands that the machine instruction will take.  This
411allows for efficient memory allocation.  You also need to specify if operands
412default to be uses of values, not definitions.  If you need to add a definition
413operand (other than the optional destination register), you must explicitly mark
414it as such:
415
416.. code-block:: c++
417
418  MI.addReg(Reg, RegState::Define);
419
420Fixed (preassigned) registers
421^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
422
423One important issue that the code generator needs to be aware of is the presence
424of fixed registers.  In particular, there are often places in the instruction
425stream where the register allocator *must* arrange for a particular value to be
426in a particular register.  This can occur due to limitations of the instruction
427set (e.g., the X86 can only do a 32-bit divide with the ``EAX``/``EDX``
428registers), or external factors like calling conventions.  In any case, the
429instruction selector should emit code that copies a virtual register into or out
430of a physical register when needed.
431
432For example, consider this simple LLVM example:
433
434.. code-block:: llvm
435
436  define i32 @test(i32 %X, i32 %Y) {
437    %Z = sdiv i32 %X, %Y
438    ret i32 %Z
439  }
440
441The X86 instruction selector might produce this machine code for the ``div`` and
442``ret``:
443
444.. code-block:: llvm
445
446  ;; Start of div
447  %EAX = mov %reg1024           ;; Copy X (in reg1024) into EAX
448  %reg1027 = sar %reg1024, 31
449  %EDX = mov %reg1027           ;; Sign extend X into EDX
450  idiv %reg1025                 ;; Divide by Y (in reg1025)
451  %reg1026 = mov %EAX           ;; Read the result (Z) out of EAX
452
453  ;; Start of ret
454  %EAX = mov %reg1026           ;; 32-bit return value goes in EAX
455  ret
456
457By the end of code generation, the register allocator would coalesce the
458registers and delete the resultant identity moves producing the following
459code:
460
461.. code-block:: llvm
462
463  ;; X is in EAX, Y is in ECX
464  mov %EAX, %EDX
465  sar %EDX, 31
466  idiv %ECX
467  ret
468
469This approach is extremely general (if it can handle the X86 architecture, it
470can handle anything!) and allows all of the target specific knowledge about the
471instruction stream to be isolated in the instruction selector.  Note that
472physical registers should have a short lifetime for good code generation, and
473all physical registers are assumed dead on entry to and exit from basic blocks
474(before register allocation).  Thus, if you need a value to be live across basic
475block boundaries, it *must* live in a virtual register.
476
477Call-clobbered registers
478^^^^^^^^^^^^^^^^^^^^^^^^
479
480Some machine instructions, like calls, clobber a large number of physical
481registers.  Rather than adding ``<def,dead>`` operands for all of them, it is
482possible to use an ``MO_RegisterMask`` operand instead.  The register mask
483operand holds a bit mask of preserved registers, and everything else is
484considered to be clobbered by the instruction.
485
486Machine code in SSA form
487^^^^^^^^^^^^^^^^^^^^^^^^
488
489``MachineInstr``'s are initially selected in SSA-form, and are maintained in
490SSA-form until register allocation happens.  For the most part, this is
491trivially simple since LLVM is already in SSA form; LLVM PHI nodes become
492machine code PHI nodes, and virtual registers are only allowed to have a single
493definition.
494
495After register allocation, machine code is no longer in SSA-form because there
496are no virtual registers left in the code.
497
498.. _MachineBasicBlock:
499
500The ``MachineBasicBlock`` class
501-------------------------------
502
503The ``MachineBasicBlock`` class contains a list of machine instructions
504(:raw-html:`<tt>` `MachineInstr`_ :raw-html:`</tt>` instances).  It roughly
505corresponds to the LLVM code input to the instruction selector, but there can be
506a one-to-many mapping (i.e. one LLVM basic block can map to multiple machine
507basic blocks). The ``MachineBasicBlock`` class has a "``getBasicBlock``" method,
508which returns the LLVM basic block that it comes from.
509
510.. _MachineFunction:
511
512The ``MachineFunction`` class
513-----------------------------
514
515The ``MachineFunction`` class contains a list of machine basic blocks
516(:raw-html:`<tt>` `MachineBasicBlock`_ :raw-html:`</tt>` instances).  It
517corresponds one-to-one with the LLVM function input to the instruction selector.
518In addition to a list of basic blocks, the ``MachineFunction`` contains a a
519``MachineConstantPool``, a ``MachineFrameInfo``, a ``MachineFunctionInfo``, and
520a ``MachineRegisterInfo``.  See ``include/llvm/CodeGen/MachineFunction.h`` for
521more information.
522
523``MachineInstr Bundles``
524------------------------
525
526LLVM code generator can model sequences of instructions as MachineInstr
527bundles. A MI bundle can model a VLIW group / pack which contains an arbitrary
528number of parallel instructions. It can also be used to model a sequential list
529of instructions (potentially with data dependencies) that cannot be legally
530separated (e.g. ARM Thumb2 IT blocks).
531
532Conceptually a MI bundle is a MI with a number of other MIs nested within:
533
534::
535
536  --------------
537  |   Bundle   | ---------
538  --------------          \
539         |           ----------------
540         |           |      MI      |
541         |           ----------------
542         |                   |
543         |           ----------------
544         |           |      MI      |
545         |           ----------------
546         |                   |
547         |           ----------------
548         |           |      MI      |
549         |           ----------------
550         |
551  --------------
552  |   Bundle   | --------
553  --------------         \
554         |           ----------------
555         |           |      MI      |
556         |           ----------------
557         |                   |
558         |           ----------------
559         |           |      MI      |
560         |           ----------------
561         |                   |
562         |                  ...
563         |
564  --------------
565  |   Bundle   | --------
566  --------------         \
567         |
568        ...
569
570MI bundle support does not change the physical representations of
571MachineBasicBlock and MachineInstr. All the MIs (including top level and nested
572ones) are stored as sequential list of MIs. The "bundled" MIs are marked with
573the 'InsideBundle' flag. A top level MI with the special BUNDLE opcode is used
574to represent the start of a bundle. It's legal to mix BUNDLE MIs with indiviual
575MIs that are not inside bundles nor represent bundles.
576
577MachineInstr passes should operate on a MI bundle as a single unit. Member
578methods have been taught to correctly handle bundles and MIs inside bundles.
579The MachineBasicBlock iterator has been modified to skip over bundled MIs to
580enforce the bundle-as-a-single-unit concept. An alternative iterator
581instr_iterator has been added to MachineBasicBlock to allow passes to iterate
582over all of the MIs in a MachineBasicBlock, including those which are nested
583inside bundles. The top level BUNDLE instruction must have the correct set of
584register MachineOperand's that represent the cumulative inputs and outputs of
585the bundled MIs.
586
587Packing / bundling of MachineInstr's should be done as part of the register
588allocation super-pass. More specifically, the pass which determines what MIs
589should be bundled together must be done after code generator exits SSA form
590(i.e. after two-address pass, PHI elimination, and copy coalescing).  Bundles
591should only be finalized (i.e. adding BUNDLE MIs and input and output register
592MachineOperands) after virtual registers have been rewritten into physical
593registers. This requirement eliminates the need to add virtual register operands
594to BUNDLE instructions which would effectively double the virtual register def
595and use lists.
596
597.. _MC Layer:
598
599The "MC" Layer
600==============
601
602The MC Layer is used to represent and process code at the raw machine code
603level, devoid of "high level" information like "constant pools", "jump tables",
604"global variables" or anything like that.  At this level, LLVM handles things
605like label names, machine instructions, and sections in the object file.  The
606code in this layer is used for a number of important purposes: the tail end of
607the code generator uses it to write a .s or .o file, and it is also used by the
608llvm-mc tool to implement standalone machine code assemblers and disassemblers.
609
610This section describes some of the important classes.  There are also a number
611of important subsystems that interact at this layer, they are described later in
612this manual.
613
614.. _MCStreamer:
615
616The ``MCStreamer`` API
617----------------------
618
619MCStreamer is best thought of as an assembler API.  It is an abstract API which
620is *implemented* in different ways (e.g. to output a .s file, output an ELF .o
621file, etc) but whose API correspond directly to what you see in a .s file.
622MCStreamer has one method per directive, such as EmitLabel, EmitSymbolAttribute,
623SwitchSection, EmitValue (for .byte, .word), etc, which directly correspond to
624assembly level directives.  It also has an EmitInstruction method, which is used
625to output an MCInst to the streamer.
626
627This API is most important for two clients: the llvm-mc stand-alone assembler is
628effectively a parser that parses a line, then invokes a method on MCStreamer. In
629the code generator, the `Code Emission`_ phase of the code generator lowers
630higher level LLVM IR and Machine* constructs down to the MC layer, emitting
631directives through MCStreamer.
632
633On the implementation side of MCStreamer, there are two major implementations:
634one for writing out a .s file (MCAsmStreamer), and one for writing out a .o
635file (MCObjectStreamer).  MCAsmStreamer is a straight-forward implementation
636that prints out a directive for each method (e.g. ``EmitValue -> .byte``), but
637MCObjectStreamer implements a full assembler.
638
639For target specific directives, the MCStreamer has a MCTargetStreamer instance.
640Each target that needs it defines a class that inherits from it and is a lot
641like MCStreamer itself: It has one method per directive and two classes that
642inherit from it, a target object streamer and a target asm streamer. The target
643asm streamer just prints it (``emitFnStart -> .fnstart``), and the object
644streamer implement the assembler logic for it.
645
646To make llvm use these classes, the target initialization must call
647TargetRegistry::RegisterAsmStreamer and TargetRegistry::RegisterMCObjectStreamer
648passing callbacks that allocate the corresponding target streamer and pass it
649to createAsmStreamer or to the appropriate object streamer constructor.
650
651The ``MCContext`` class
652-----------------------
653
654The MCContext class is the owner of a variety of uniqued data structures at the
655MC layer, including symbols, sections, etc.  As such, this is the class that you
656interact with to create symbols and sections.  This class can not be subclassed.
657
658The ``MCSymbol`` class
659----------------------
660
661The MCSymbol class represents a symbol (aka label) in the assembly file.  There
662are two interesting kinds of symbols: assembler temporary symbols, and normal
663symbols.  Assembler temporary symbols are used and processed by the assembler
664but are discarded when the object file is produced.  The distinction is usually
665represented by adding a prefix to the label, for example "L" labels are
666assembler temporary labels in MachO.
667
668MCSymbols are created by MCContext and uniqued there.  This means that MCSymbols
669can be compared for pointer equivalence to find out if they are the same symbol.
670Note that pointer inequality does not guarantee the labels will end up at
671different addresses though.  It's perfectly legal to output something like this
672to the .s file:
673
674::
675
676  foo:
677  bar:
678    .byte 4
679
680In this case, both the foo and bar symbols will have the same address.
681
682The ``MCSection`` class
683-----------------------
684
685The ``MCSection`` class represents an object-file specific section. It is
686subclassed by object file specific implementations (e.g. ``MCSectionMachO``,
687``MCSectionCOFF``, ``MCSectionELF``) and these are created and uniqued by
688MCContext.  The MCStreamer has a notion of the current section, which can be
689changed with the SwitchToSection method (which corresponds to a ".section"
690directive in a .s file).
691
692.. _MCInst:
693
694The ``MCInst`` class
695--------------------
696
697The ``MCInst`` class is a target-independent representation of an instruction.
698It is a simple class (much more so than `MachineInstr`_) that holds a
699target-specific opcode and a vector of MCOperands.  MCOperand, in turn, is a
700simple discriminated union of three cases: 1) a simple immediate, 2) a target
701register ID, 3) a symbolic expression (e.g. "``Lfoo-Lbar+42``") as an MCExpr.
702
703MCInst is the common currency used to represent machine instructions at the MC
704layer.  It is the type used by the instruction encoder, the instruction printer,
705and the type generated by the assembly parser and disassembler.
706
707.. _Target-independent algorithms:
708.. _code generation algorithm:
709
710Target-independent code generation algorithms
711=============================================
712
713This section documents the phases described in the `high-level design of the
714code generator`_.  It explains how they work and some of the rationale behind
715their design.
716
717.. _Instruction Selection:
718.. _instruction selection section:
719
720Instruction Selection
721---------------------
722
723Instruction Selection is the process of translating LLVM code presented to the
724code generator into target-specific machine instructions.  There are several
725well-known ways to do this in the literature.  LLVM uses a SelectionDAG based
726instruction selector.
727
728Portions of the DAG instruction selector are generated from the target
729description (``*.td``) files.  Our goal is for the entire instruction selector
730to be generated from these ``.td`` files, though currently there are still
731things that require custom C++ code.
732
733.. _SelectionDAG:
734
735Introduction to SelectionDAGs
736^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
737
738The SelectionDAG provides an abstraction for code representation in a way that
739is amenable to instruction selection using automatic techniques
740(e.g. dynamic-programming based optimal pattern matching selectors). It is also
741well-suited to other phases of code generation; in particular, instruction
742scheduling (SelectionDAG's are very close to scheduling DAGs post-selection).
743Additionally, the SelectionDAG provides a host representation where a large
744variety of very-low-level (but target-independent) `optimizations`_ may be
745performed; ones which require extensive information about the instructions
746efficiently supported by the target.
747
748The SelectionDAG is a Directed-Acyclic-Graph whose nodes are instances of the
749``SDNode`` class.  The primary payload of the ``SDNode`` is its operation code
750(Opcode) that indicates what operation the node performs and the operands to the
751operation.  The various operation node types are described at the top of the
752``include/llvm/CodeGen/ISDOpcodes.h`` file.
753
754Although most operations define a single value, each node in the graph may
755define multiple values.  For example, a combined div/rem operation will define
756both the dividend and the remainder. Many other situations require multiple
757values as well.  Each node also has some number of operands, which are edges to
758the node defining the used value.  Because nodes may define multiple values,
759edges are represented by instances of the ``SDValue`` class, which is a
760``<SDNode, unsigned>`` pair, indicating the node and result value being used,
761respectively.  Each value produced by an ``SDNode`` has an associated ``MVT``
762(Machine Value Type) indicating what the type of the value is.
763
764SelectionDAGs contain two different kinds of values: those that represent data
765flow and those that represent control flow dependencies.  Data values are simple
766edges with an integer or floating point value type.  Control edges are
767represented as "chain" edges which are of type ``MVT::Other``.  These edges
768provide an ordering between nodes that have side effects (such as loads, stores,
769calls, returns, etc).  All nodes that have side effects should take a token
770chain as input and produce a new one as output.  By convention, token chain
771inputs are always operand #0, and chain results are always the last value
772produced by an operation. However, after instruction selection, the
773machine nodes have their chain after the instruction's operands, and
774may be followed by glue nodes.
775
776A SelectionDAG has designated "Entry" and "Root" nodes.  The Entry node is
777always a marker node with an Opcode of ``ISD::EntryToken``.  The Root node is
778the final side-effecting node in the token chain. For example, in a single basic
779block function it would be the return node.
780
781One important concept for SelectionDAGs is the notion of a "legal" vs.
782"illegal" DAG.  A legal DAG for a target is one that only uses supported
783operations and supported types.  On a 32-bit PowerPC, for example, a DAG with a
784value of type i1, i8, i16, or i64 would be illegal, as would a DAG that uses a
785SREM or UREM operation.  The `legalize types`_ and `legalize operations`_ phases
786are responsible for turning an illegal DAG into a legal DAG.
787
788.. _SelectionDAG-Process:
789
790SelectionDAG Instruction Selection Process
791^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
792
793SelectionDAG-based instruction selection consists of the following steps:
794
795#. `Build initial DAG`_ --- This stage performs a simple translation from the
796   input LLVM code to an illegal SelectionDAG.
797
798#. `Optimize SelectionDAG`_ --- This stage performs simple optimizations on the
799   SelectionDAG to simplify it, and recognize meta instructions (like rotates
800   and ``div``/``rem`` pairs) for targets that support these meta operations.
801   This makes the resultant code more efficient and the `select instructions
802   from DAG`_ phase (below) simpler.
803
804#. `Legalize SelectionDAG Types`_ --- This stage transforms SelectionDAG nodes
805   to eliminate any types that are unsupported on the target.
806
807#. `Optimize SelectionDAG`_ --- The SelectionDAG optimizer is run to clean up
808   redundancies exposed by type legalization.
809
810#. `Legalize SelectionDAG Ops`_ --- This stage transforms SelectionDAG nodes to
811   eliminate any operations that are unsupported on the target.
812
813#. `Optimize SelectionDAG`_ --- The SelectionDAG optimizer is run to eliminate
814   inefficiencies introduced by operation legalization.
815
816#. `Select instructions from DAG`_ --- Finally, the target instruction selector
817   matches the DAG operations to target instructions.  This process translates
818   the target-independent input DAG into another DAG of target instructions.
819
820#. `SelectionDAG Scheduling and Formation`_ --- The last phase assigns a linear
821   order to the instructions in the target-instruction DAG and emits them into
822   the MachineFunction being compiled.  This step uses traditional prepass
823   scheduling techniques.
824
825After all of these steps are complete, the SelectionDAG is destroyed and the
826rest of the code generation passes are run.
827
828One great way to visualize what is going on here is to take advantage of a few
829LLC command line options.  The following options pop up a window displaying the
830SelectionDAG at specific times (if you only get errors printed to the console
831while using this, you probably `need to configure your
832system <ProgrammersManual.html#viewing-graphs-while-debugging-code>`_ to add support for it).
833
834* ``-view-dag-combine1-dags`` displays the DAG after being built, before the
835  first optimization pass.
836
837* ``-view-legalize-dags`` displays the DAG before Legalization.
838
839* ``-view-dag-combine2-dags`` displays the DAG before the second optimization
840  pass.
841
842* ``-view-isel-dags`` displays the DAG before the Select phase.
843
844* ``-view-sched-dags`` displays the DAG before Scheduling.
845
846The ``-view-sunit-dags`` displays the Scheduler's dependency graph.  This graph
847is based on the final SelectionDAG, with nodes that must be scheduled together
848bundled into a single scheduling-unit node, and with immediate operands and
849other nodes that aren't relevant for scheduling omitted.
850
851The option ``-filter-view-dags`` allows to select the name of the basic block
852that you are interested to visualize and filters all the previous
853``view-*-dags`` options.
854
855.. _Build initial DAG:
856
857Initial SelectionDAG Construction
858^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
859
860The initial SelectionDAG is na\ :raw-html:`&iuml;`\ vely peephole expanded from
861the LLVM input by the ``SelectionDAGBuilder`` class.  The intent of this pass
862is to expose as much low-level, target-specific details to the SelectionDAG as
863possible.  This pass is mostly hard-coded (e.g. an LLVM ``add`` turns into an
864``SDNode add`` while a ``getelementptr`` is expanded into the obvious
865arithmetic). This pass requires target-specific hooks to lower calls, returns,
866varargs, etc.  For these features, the :raw-html:`<tt>` `TargetLowering`_
867:raw-html:`</tt>` interface is used.
868
869.. _legalize types:
870.. _Legalize SelectionDAG Types:
871.. _Legalize SelectionDAG Ops:
872
873SelectionDAG LegalizeTypes Phase
874^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
875
876The Legalize phase is in charge of converting a DAG to only use the types that
877are natively supported by the target.
878
879There are two main ways of converting values of unsupported scalar types to
880values of supported types: converting small types to larger types ("promoting"),
881and breaking up large integer types into smaller ones ("expanding").  For
882example, a target might require that all f32 values are promoted to f64 and that
883all i1/i8/i16 values are promoted to i32.  The same target might require that
884all i64 values be expanded into pairs of i32 values.  These changes can insert
885sign and zero extensions as needed to make sure that the final code has the same
886behavior as the input.
887
888There are two main ways of converting values of unsupported vector types to
889value of supported types: splitting vector types, multiple times if necessary,
890until a legal type is found, and extending vector types by adding elements to
891the end to round them out to legal types ("widening").  If a vector gets split
892all the way down to single-element parts with no supported vector type being
893found, the elements are converted to scalars ("scalarizing").
894
895A target implementation tells the legalizer which types are supported (and which
896register class to use for them) by calling the ``addRegisterClass`` method in
897its ``TargetLowering`` constructor.
898
899.. _legalize operations:
900.. _Legalizer:
901
902SelectionDAG Legalize Phase
903^^^^^^^^^^^^^^^^^^^^^^^^^^^
904
905The Legalize phase is in charge of converting a DAG to only use the operations
906that are natively supported by the target.
907
908Targets often have weird constraints, such as not supporting every operation on
909every supported datatype (e.g. X86 does not support byte conditional moves and
910PowerPC does not support sign-extending loads from a 16-bit memory location).
911Legalize takes care of this by open-coding another sequence of operations to
912emulate the operation ("expansion"), by promoting one type to a larger type that
913supports the operation ("promotion"), or by using a target-specific hook to
914implement the legalization ("custom").
915
916A target implementation tells the legalizer which operations are not supported
917(and which of the above three actions to take) by calling the
918``setOperationAction`` method in its ``TargetLowering`` constructor.
919
920Prior to the existence of the Legalize passes, we required that every target
921`selector`_ supported and handled every operator and type even if they are not
922natively supported.  The introduction of the Legalize phases allows all of the
923canonicalization patterns to be shared across targets, and makes it very easy to
924optimize the canonicalized code because it is still in the form of a DAG.
925
926.. _optimizations:
927.. _Optimize SelectionDAG:
928.. _selector:
929
930SelectionDAG Optimization Phase: the DAG Combiner
931^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
932
933The SelectionDAG optimization phase is run multiple times for code generation,
934immediately after the DAG is built and once after each legalization.  The first
935run of the pass allows the initial code to be cleaned up (e.g. performing
936optimizations that depend on knowing that the operators have restricted type
937inputs).  Subsequent runs of the pass clean up the messy code generated by the
938Legalize passes, which allows Legalize to be very simple (it can focus on making
939code legal instead of focusing on generating *good* and legal code).
940
941One important class of optimizations performed is optimizing inserted sign and
942zero extension instructions.  We currently use ad-hoc techniques, but could move
943to more rigorous techniques in the future.  Here are some good papers on the
944subject:
945
946"`Widening integer arithmetic <http://www.eecs.harvard.edu/~nr/pubs/widen-abstract.html>`_" :raw-html:`<br>`
947Kevin Redwine and Norman Ramsey :raw-html:`<br>`
948International Conference on Compiler Construction (CC) 2004
949
950"`Effective sign extension elimination <http://portal.acm.org/citation.cfm?doid=512529.512552>`_"  :raw-html:`<br>`
951Motohiro Kawahito, Hideaki Komatsu, and Toshio Nakatani :raw-html:`<br>`
952Proceedings of the ACM SIGPLAN 2002 Conference on Programming Language Design
953and Implementation.
954
955.. _Select instructions from DAG:
956
957SelectionDAG Select Phase
958^^^^^^^^^^^^^^^^^^^^^^^^^
959
960The Select phase is the bulk of the target-specific code for instruction
961selection.  This phase takes a legal SelectionDAG as input, pattern matches the
962instructions supported by the target to this DAG, and produces a new DAG of
963target code.  For example, consider the following LLVM fragment:
964
965.. code-block:: llvm
966
967  %t1 = fadd float %W, %X
968  %t2 = fmul float %t1, %Y
969  %t3 = fadd float %t2, %Z
970
971This LLVM code corresponds to a SelectionDAG that looks basically like this:
972
973.. code-block:: llvm
974
975  (fadd:f32 (fmul:f32 (fadd:f32 W, X), Y), Z)
976
977If a target supports floating point multiply-and-add (FMA) operations, one of
978the adds can be merged with the multiply.  On the PowerPC, for example, the
979output of the instruction selector might look like this DAG:
980
981::
982
983  (FMADDS (FADDS W, X), Y, Z)
984
985The ``FMADDS`` instruction is a ternary instruction that multiplies its first
986two operands and adds the third (as single-precision floating-point numbers).
987The ``FADDS`` instruction is a simple binary single-precision add instruction.
988To perform this pattern match, the PowerPC backend includes the following
989instruction definitions:
990
991.. code-block:: text
992  :emphasize-lines: 4-5,9
993
994  def FMADDS : AForm_1<59, 29,
995                      (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
996                      "fmadds $FRT, $FRA, $FRC, $FRB",
997                      [(set F4RC:$FRT, (fadd (fmul F4RC:$FRA, F4RC:$FRC),
998                                             F4RC:$FRB))]>;
999  def FADDS : AForm_2<59, 21,
1000                      (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
1001                      "fadds $FRT, $FRA, $FRB",
1002                      [(set F4RC:$FRT, (fadd F4RC:$FRA, F4RC:$FRB))]>;
1003
1004The highlighted portion of the instruction definitions indicates the pattern
1005used to match the instructions. The DAG operators (like ``fmul``/``fadd``)
1006are defined in the ``include/llvm/Target/TargetSelectionDAG.td`` file.
1007"``F4RC``" is the register class of the input and result values.
1008
1009The TableGen DAG instruction selector generator reads the instruction patterns
1010in the ``.td`` file and automatically builds parts of the pattern matching code
1011for your target.  It has the following strengths:
1012
1013* At compiler-compiler time, it analyzes your instruction patterns and tells you
1014  if your patterns make sense or not.
1015
1016* It can handle arbitrary constraints on operands for the pattern match.  In
1017  particular, it is straight-forward to say things like "match any immediate
1018  that is a 13-bit sign-extended value".  For examples, see the ``immSExt16``
1019  and related ``tblgen`` classes in the PowerPC backend.
1020
1021* It knows several important identities for the patterns defined.  For example,
1022  it knows that addition is commutative, so it allows the ``FMADDS`` pattern
1023  above to match "``(fadd X, (fmul Y, Z))``" as well as "``(fadd (fmul X, Y),
1024  Z)``", without the target author having to specially handle this case.
1025
1026* It has a full-featured type-inferencing system.  In particular, you should
1027  rarely have to explicitly tell the system what type parts of your patterns
1028  are.  In the ``FMADDS`` case above, we didn't have to tell ``tblgen`` that all
1029  of the nodes in the pattern are of type 'f32'.  It was able to infer and
1030  propagate this knowledge from the fact that ``F4RC`` has type 'f32'.
1031
1032* Targets can define their own (and rely on built-in) "pattern fragments".
1033  Pattern fragments are chunks of reusable patterns that get inlined into your
1034  patterns during compiler-compiler time.  For example, the integer "``(not
1035  x)``" operation is actually defined as a pattern fragment that expands as
1036  "``(xor x, -1)``", since the SelectionDAG does not have a native '``not``'
1037  operation.  Targets can define their own short-hand fragments as they see fit.
1038  See the definition of '``not``' and '``ineg``' for examples.
1039
1040* In addition to instructions, targets can specify arbitrary patterns that map
1041  to one or more instructions using the 'Pat' class.  For example, the PowerPC
1042  has no way to load an arbitrary integer immediate into a register in one
1043  instruction. To tell tblgen how to do this, it defines:
1044
1045  ::
1046
1047    // Arbitrary immediate support.  Implement in terms of LIS/ORI.
1048    def : Pat<(i32 imm:$imm),
1049              (ORI (LIS (HI16 imm:$imm)), (LO16 imm:$imm))>;
1050
1051  If none of the single-instruction patterns for loading an immediate into a
1052  register match, this will be used.  This rule says "match an arbitrary i32
1053  immediate, turning it into an ``ORI`` ('or a 16-bit immediate') and an ``LIS``
1054  ('load 16-bit immediate, where the immediate is shifted to the left 16 bits')
1055  instruction".  To make this work, the ``LO16``/``HI16`` node transformations
1056  are used to manipulate the input immediate (in this case, take the high or low
1057  16-bits of the immediate).
1058
1059* When using the 'Pat' class to map a pattern to an instruction that has one
1060  or more complex operands (like e.g. `X86 addressing mode`_), the pattern may
1061  either specify the operand as a whole using a ``ComplexPattern``, or else it
1062  may specify the components of the complex operand separately.  The latter is
1063  done e.g. for pre-increment instructions by the PowerPC back end:
1064
1065  ::
1066
1067    def STWU  : DForm_1<37, (outs ptr_rc:$ea_res), (ins GPRC:$rS, memri:$dst),
1068                    "stwu $rS, $dst", LdStStoreUpd, []>,
1069                    RegConstraint<"$dst.reg = $ea_res">, NoEncode<"$ea_res">;
1070
1071    def : Pat<(pre_store GPRC:$rS, ptr_rc:$ptrreg, iaddroff:$ptroff),
1072              (STWU GPRC:$rS, iaddroff:$ptroff, ptr_rc:$ptrreg)>;
1073
1074  Here, the pair of ``ptroff`` and ``ptrreg`` operands is matched onto the
1075  complex operand ``dst`` of class ``memri`` in the ``STWU`` instruction.
1076
1077* While the system does automate a lot, it still allows you to write custom C++
1078  code to match special cases if there is something that is hard to
1079  express.
1080
1081While it has many strengths, the system currently has some limitations,
1082primarily because it is a work in progress and is not yet finished:
1083
1084* Overall, there is no way to define or match SelectionDAG nodes that define
1085  multiple values (e.g. ``SMUL_LOHI``, ``LOAD``, ``CALL``, etc).  This is the
1086  biggest reason that you currently still *have to* write custom C++ code
1087  for your instruction selector.
1088
1089* There is no great way to support matching complex addressing modes yet.  In
1090  the future, we will extend pattern fragments to allow them to define multiple
1091  values (e.g. the four operands of the `X86 addressing mode`_, which are
1092  currently matched with custom C++ code).  In addition, we'll extend fragments
1093  so that a fragment can match multiple different patterns.
1094
1095* We don't automatically infer flags like ``isStore``/``isLoad`` yet.
1096
1097* We don't automatically generate the set of supported registers and operations
1098  for the `Legalizer`_ yet.
1099
1100* We don't have a way of tying in custom legalized nodes yet.
1101
1102Despite these limitations, the instruction selector generator is still quite
1103useful for most of the binary and logical operations in typical instruction
1104sets.  If you run into any problems or can't figure out how to do something,
1105please let Chris know!
1106
1107.. _Scheduling and Formation:
1108.. _SelectionDAG Scheduling and Formation:
1109
1110SelectionDAG Scheduling and Formation Phase
1111^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1112
1113The scheduling phase takes the DAG of target instructions from the selection
1114phase and assigns an order.  The scheduler can pick an order depending on
1115various constraints of the machines (i.e. order for minimal register pressure or
1116try to cover instruction latencies).  Once an order is established, the DAG is
1117converted to a list of :raw-html:`<tt>` `MachineInstr`_\s :raw-html:`</tt>` and
1118the SelectionDAG is destroyed.
1119
1120Note that this phase is logically separate from the instruction selection phase,
1121but is tied to it closely in the code because it operates on SelectionDAGs.
1122
1123Future directions for the SelectionDAG
1124^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1125
1126#. Optional function-at-a-time selection.
1127
1128#. Auto-generate entire selector from ``.td`` file.
1129
1130.. _SSA-based Machine Code Optimizations:
1131
1132SSA-based Machine Code Optimizations
1133------------------------------------
1134
1135To Be Written
1136
1137Live Intervals
1138--------------
1139
1140Live Intervals are the ranges (intervals) where a variable is *live*.  They are
1141used by some `register allocator`_ passes to determine if two or more virtual
1142registers which require the same physical register are live at the same point in
1143the program (i.e., they conflict).  When this situation occurs, one virtual
1144register must be *spilled*.
1145
1146Live Variable Analysis
1147^^^^^^^^^^^^^^^^^^^^^^
1148
1149The first step in determining the live intervals of variables is to calculate
1150the set of registers that are immediately dead after the instruction (i.e., the
1151instruction calculates the value, but it is never used) and the set of registers
1152that are used by the instruction, but are never used after the instruction
1153(i.e., they are killed). Live variable information is computed for
1154each *virtual* register and *register allocatable* physical register
1155in the function.  This is done in a very efficient manner because it uses SSA to
1156sparsely compute lifetime information for virtual registers (which are in SSA
1157form) and only has to track physical registers within a block.  Before register
1158allocation, LLVM can assume that physical registers are only live within a
1159single basic block.  This allows it to do a single, local analysis to resolve
1160physical register lifetimes within each basic block. If a physical register is
1161not register allocatable (e.g., a stack pointer or condition codes), it is not
1162tracked.
1163
1164Physical registers may be live in to or out of a function. Live in values are
1165typically arguments in registers. Live out values are typically return values in
1166registers. Live in values are marked as such, and are given a dummy "defining"
1167instruction during live intervals analysis. If the last basic block of a
1168function is a ``return``, then it's marked as using all live out values in the
1169function.
1170
1171``PHI`` nodes need to be handled specially, because the calculation of the live
1172variable information from a depth first traversal of the CFG of the function
1173won't guarantee that a virtual register used by the ``PHI`` node is defined
1174before it's used. When a ``PHI`` node is encountered, only the definition is
1175handled, because the uses will be handled in other basic blocks.
1176
1177For each ``PHI`` node of the current basic block, we simulate an assignment at
1178the end of the current basic block and traverse the successor basic blocks. If a
1179successor basic block has a ``PHI`` node and one of the ``PHI`` node's operands
1180is coming from the current basic block, then the variable is marked as *alive*
1181within the current basic block and all of its predecessor basic blocks, until
1182the basic block with the defining instruction is encountered.
1183
1184Live Intervals Analysis
1185^^^^^^^^^^^^^^^^^^^^^^^
1186
1187We now have the information available to perform the live intervals analysis and
1188build the live intervals themselves.  We start off by numbering the basic blocks
1189and machine instructions.  We then handle the "live-in" values.  These are in
1190physical registers, so the physical register is assumed to be killed by the end
1191of the basic block.  Live intervals for virtual registers are computed for some
1192ordering of the machine instructions ``[1, N]``.  A live interval is an interval
1193``[i, j)``, where ``1 >= i >= j > N``, for which a variable is live.
1194
1195.. note::
1196  More to come...
1197
1198.. _Register Allocation:
1199.. _register allocator:
1200
1201Register Allocation
1202-------------------
1203
1204The *Register Allocation problem* consists in mapping a program
1205:raw-html:`<b><tt>` P\ :sub:`v`\ :raw-html:`</tt></b>`, that can use an unbounded
1206number of virtual registers, to a program :raw-html:`<b><tt>` P\ :sub:`p`\
1207:raw-html:`</tt></b>` that contains a finite (possibly small) number of physical
1208registers. Each target architecture has a different number of physical
1209registers. If the number of physical registers is not enough to accommodate all
1210the virtual registers, some of them will have to be mapped into memory. These
1211virtuals are called *spilled virtuals*.
1212
1213How registers are represented in LLVM
1214^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1215
1216In LLVM, physical registers are denoted by integer numbers that normally range
1217from 1 to 1023. To see how this numbering is defined for a particular
1218architecture, you can read the ``GenRegisterNames.inc`` file for that
1219architecture. For instance, by inspecting
1220``lib/Target/X86/X86GenRegisterInfo.inc`` we see that the 32-bit register
1221``EAX`` is denoted by 43, and the MMX register ``MM0`` is mapped to 65.
1222
1223Some architectures contain registers that share the same physical location. A
1224notable example is the X86 platform. For instance, in the X86 architecture, the
1225registers ``EAX``, ``AX`` and ``AL`` share the first eight bits. These physical
1226registers are marked as *aliased* in LLVM. Given a particular architecture, you
1227can check which registers are aliased by inspecting its ``RegisterInfo.td``
1228file. Moreover, the class ``MCRegAliasIterator`` enumerates all the physical
1229registers aliased to a register.
1230
1231Physical registers, in LLVM, are grouped in *Register Classes*.  Elements in the
1232same register class are functionally equivalent, and can be interchangeably
1233used. Each virtual register can only be mapped to physical registers of a
1234particular class. For instance, in the X86 architecture, some virtuals can only
1235be allocated to 8 bit registers.  A register class is described by
1236``TargetRegisterClass`` objects.  To discover if a virtual register is
1237compatible with a given physical, this code can be used:
1238
1239.. code-block:: c++
1240
1241  bool RegMapping_Fer::compatible_class(MachineFunction &mf,
1242                                        unsigned v_reg,
1243                                        unsigned p_reg) {
1244    assert(TargetRegisterInfo::isPhysicalRegister(p_reg) &&
1245           "Target register must be physical");
1246    const TargetRegisterClass *trc = mf.getRegInfo().getRegClass(v_reg);
1247    return trc->contains(p_reg);
1248  }
1249
1250Sometimes, mostly for debugging purposes, it is useful to change the number of
1251physical registers available in the target architecture. This must be done
1252statically, inside the ``TargetRegsterInfo.td`` file. Just ``grep`` for
1253``RegisterClass``, the last parameter of which is a list of registers. Just
1254commenting some out is one simple way to avoid them being used. A more polite
1255way is to explicitly exclude some registers from the *allocation order*. See the
1256definition of the ``GR8`` register class in
1257``lib/Target/X86/X86RegisterInfo.td`` for an example of this.
1258
1259Virtual registers are also denoted by integer numbers. Contrary to physical
1260registers, different virtual registers never share the same number. Whereas
1261physical registers are statically defined in a ``TargetRegisterInfo.td`` file
1262and cannot be created by the application developer, that is not the case with
1263virtual registers. In order to create new virtual registers, use the method
1264``MachineRegisterInfo::createVirtualRegister()``. This method will return a new
1265virtual register. Use an ``IndexedMap<Foo, VirtReg2IndexFunctor>`` to hold
1266information per virtual register. If you need to enumerate all virtual
1267registers, use the function ``TargetRegisterInfo::index2VirtReg()`` to find the
1268virtual register numbers:
1269
1270.. code-block:: c++
1271
1272    for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
1273      unsigned VirtReg = TargetRegisterInfo::index2VirtReg(i);
1274      stuff(VirtReg);
1275    }
1276
1277Before register allocation, the operands of an instruction are mostly virtual
1278registers, although physical registers may also be used. In order to check if a
1279given machine operand is a register, use the boolean function
1280``MachineOperand::isRegister()``. To obtain the integer code of a register, use
1281``MachineOperand::getReg()``. An instruction may define or use a register. For
1282instance, ``ADD reg:1026 := reg:1025 reg:1024`` defines the registers 1024, and
1283uses registers 1025 and 1026. Given a register operand, the method
1284``MachineOperand::isUse()`` informs if that register is being used by the
1285instruction. The method ``MachineOperand::isDef()`` informs if that registers is
1286being defined.
1287
1288We will call physical registers present in the LLVM bitcode before register
1289allocation *pre-colored registers*. Pre-colored registers are used in many
1290different situations, for instance, to pass parameters of functions calls, and
1291to store results of particular instructions. There are two types of pre-colored
1292registers: the ones *implicitly* defined, and those *explicitly*
1293defined. Explicitly defined registers are normal operands, and can be accessed
1294with ``MachineInstr::getOperand(int)::getReg()``.  In order to check which
1295registers are implicitly defined by an instruction, use the
1296``TargetInstrInfo::get(opcode)::ImplicitDefs``, where ``opcode`` is the opcode
1297of the target instruction. One important difference between explicit and
1298implicit physical registers is that the latter are defined statically for each
1299instruction, whereas the former may vary depending on the program being
1300compiled. For example, an instruction that represents a function call will
1301always implicitly define or use the same set of physical registers. To read the
1302registers implicitly used by an instruction, use
1303``TargetInstrInfo::get(opcode)::ImplicitUses``. Pre-colored registers impose
1304constraints on any register allocation algorithm. The register allocator must
1305make sure that none of them are overwritten by the values of virtual registers
1306while still alive.
1307
1308Mapping virtual registers to physical registers
1309^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1310
1311There are two ways to map virtual registers to physical registers (or to memory
1312slots). The first way, that we will call *direct mapping*, is based on the use
1313of methods of the classes ``TargetRegisterInfo``, and ``MachineOperand``. The
1314second way, that we will call *indirect mapping*, relies on the ``VirtRegMap``
1315class in order to insert loads and stores sending and getting values to and from
1316memory.
1317
1318The direct mapping provides more flexibility to the developer of the register
1319allocator; however, it is more error prone, and demands more implementation
1320work.  Basically, the programmer will have to specify where load and store
1321instructions should be inserted in the target function being compiled in order
1322to get and store values in memory. To assign a physical register to a virtual
1323register present in a given operand, use ``MachineOperand::setReg(p_reg)``. To
1324insert a store instruction, use ``TargetInstrInfo::storeRegToStackSlot(...)``,
1325and to insert a load instruction, use ``TargetInstrInfo::loadRegFromStackSlot``.
1326
1327The indirect mapping shields the application developer from the complexities of
1328inserting load and store instructions. In order to map a virtual register to a
1329physical one, use ``VirtRegMap::assignVirt2Phys(vreg, preg)``.  In order to map
1330a certain virtual register to memory, use
1331``VirtRegMap::assignVirt2StackSlot(vreg)``. This method will return the stack
1332slot where ``vreg``'s value will be located.  If it is necessary to map another
1333virtual register to the same stack slot, use
1334``VirtRegMap::assignVirt2StackSlot(vreg, stack_location)``. One important point
1335to consider when using the indirect mapping, is that even if a virtual register
1336is mapped to memory, it still needs to be mapped to a physical register. This
1337physical register is the location where the virtual register is supposed to be
1338found before being stored or after being reloaded.
1339
1340If the indirect strategy is used, after all the virtual registers have been
1341mapped to physical registers or stack slots, it is necessary to use a spiller
1342object to place load and store instructions in the code. Every virtual that has
1343been mapped to a stack slot will be stored to memory after being defined and will
1344be loaded before being used. The implementation of the spiller tries to recycle
1345load/store instructions, avoiding unnecessary instructions. For an example of
1346how to invoke the spiller, see ``RegAllocLinearScan::runOnMachineFunction`` in
1347``lib/CodeGen/RegAllocLinearScan.cpp``.
1348
1349Handling two address instructions
1350^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1351
1352With very rare exceptions (e.g., function calls), the LLVM machine code
1353instructions are three address instructions. That is, each instruction is
1354expected to define at most one register, and to use at most two registers.
1355However, some architectures use two address instructions. In this case, the
1356defined register is also one of the used registers. For instance, an instruction
1357such as ``ADD %EAX, %EBX``, in X86 is actually equivalent to ``%EAX = %EAX +
1358%EBX``.
1359
1360In order to produce correct code, LLVM must convert three address instructions
1361that represent two address instructions into true two address instructions. LLVM
1362provides the pass ``TwoAddressInstructionPass`` for this specific purpose. It
1363must be run before register allocation takes place. After its execution, the
1364resulting code may no longer be in SSA form. This happens, for instance, in
1365situations where an instruction such as ``%a = ADD %b %c`` is converted to two
1366instructions such as:
1367
1368::
1369
1370  %a = MOVE %b
1371  %a = ADD %a %c
1372
1373Notice that, internally, the second instruction is represented as ``ADD
1374%a[def/use] %c``. I.e., the register operand ``%a`` is both used and defined by
1375the instruction.
1376
1377The SSA deconstruction phase
1378^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1379
1380An important transformation that happens during register allocation is called
1381the *SSA Deconstruction Phase*. The SSA form simplifies many analyses that are
1382performed on the control flow graph of programs. However, traditional
1383instruction sets do not implement PHI instructions. Thus, in order to generate
1384executable code, compilers must replace PHI instructions with other instructions
1385that preserve their semantics.
1386
1387There are many ways in which PHI instructions can safely be removed from the
1388target code. The most traditional PHI deconstruction algorithm replaces PHI
1389instructions with copy instructions. That is the strategy adopted by LLVM. The
1390SSA deconstruction algorithm is implemented in
1391``lib/CodeGen/PHIElimination.cpp``. In order to invoke this pass, the identifier
1392``PHIEliminationID`` must be marked as required in the code of the register
1393allocator.
1394
1395Instruction folding
1396^^^^^^^^^^^^^^^^^^^
1397
1398*Instruction folding* is an optimization performed during register allocation
1399that removes unnecessary copy instructions. For instance, a sequence of
1400instructions such as:
1401
1402::
1403
1404  %EBX = LOAD %mem_address
1405  %EAX = COPY %EBX
1406
1407can be safely substituted by the single instruction:
1408
1409::
1410
1411  %EAX = LOAD %mem_address
1412
1413Instructions can be folded with the
1414``TargetRegisterInfo::foldMemoryOperand(...)`` method. Care must be taken when
1415folding instructions; a folded instruction can be quite different from the
1416original instruction. See ``LiveIntervals::addIntervalsForSpills`` in
1417``lib/CodeGen/LiveIntervalAnalysis.cpp`` for an example of its use.
1418
1419Built in register allocators
1420^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1421
1422The LLVM infrastructure provides the application developer with three different
1423register allocators:
1424
1425* *Fast* --- This register allocator is the default for debug builds. It
1426  allocates registers on a basic block level, attempting to keep values in
1427  registers and reusing registers as appropriate.
1428
1429* *Basic* --- This is an incremental approach to register allocation. Live
1430  ranges are assigned to registers one at a time in an order that is driven by
1431  heuristics. Since code can be rewritten on-the-fly during allocation, this
1432  framework allows interesting allocators to be developed as extensions. It is
1433  not itself a production register allocator but is a potentially useful
1434  stand-alone mode for triaging bugs and as a performance baseline.
1435
1436* *Greedy* --- *The default allocator*. This is a highly tuned implementation of
1437  the *Basic* allocator that incorporates global live range splitting. This
1438  allocator works hard to minimize the cost of spill code.
1439
1440* *PBQP* --- A Partitioned Boolean Quadratic Programming (PBQP) based register
1441  allocator. This allocator works by constructing a PBQP problem representing
1442  the register allocation problem under consideration, solving this using a PBQP
1443  solver, and mapping the solution back to a register assignment.
1444
1445The type of register allocator used in ``llc`` can be chosen with the command
1446line option ``-regalloc=...``:
1447
1448.. code-block:: bash
1449
1450  $ llc -regalloc=linearscan file.bc -o ln.s
1451  $ llc -regalloc=fast file.bc -o fa.s
1452  $ llc -regalloc=pbqp file.bc -o pbqp.s
1453
1454.. _Prolog/Epilog Code Insertion:
1455
1456Prolog/Epilog Code Insertion
1457----------------------------
1458
1459Compact Unwind
1460
1461Throwing an exception requires *unwinding* out of a function. The information on
1462how to unwind a given function is traditionally expressed in DWARF unwind
1463(a.k.a. frame) info. But that format was originally developed for debuggers to
1464backtrace, and each Frame Description Entry (FDE) requires ~20-30 bytes per
1465function. There is also the cost of mapping from an address in a function to the
1466corresponding FDE at runtime. An alternative unwind encoding is called *compact
1467unwind* and requires just 4-bytes per function.
1468
1469The compact unwind encoding is a 32-bit value, which is encoded in an
1470architecture-specific way. It specifies which registers to restore and from
1471where, and how to unwind out of the function. When the linker creates a final
1472linked image, it will create a ``__TEXT,__unwind_info`` section. This section is
1473a small and fast way for the runtime to access unwind info for any given
1474function. If we emit compact unwind info for the function, that compact unwind
1475info will be encoded in the ``__TEXT,__unwind_info`` section. If we emit DWARF
1476unwind info, the ``__TEXT,__unwind_info`` section will contain the offset of the
1477FDE in the ``__TEXT,__eh_frame`` section in the final linked image.
1478
1479For X86, there are three modes for the compact unwind encoding:
1480
1481*Function with a Frame Pointer (``EBP`` or ``RBP``)*
1482  ``EBP/RBP``-based frame, where ``EBP/RBP`` is pushed onto the stack
1483  immediately after the return address, then ``ESP/RSP`` is moved to
1484  ``EBP/RBP``. Thus to unwind, ``ESP/RSP`` is restored with the current
1485  ``EBP/RBP`` value, then ``EBP/RBP`` is restored by popping the stack, and the
1486  return is done by popping the stack once more into the PC. All non-volatile
1487  registers that need to be restored must have been saved in a small range on
1488  the stack that starts ``EBP-4`` to ``EBP-1020`` (``RBP-8`` to
1489  ``RBP-1020``). The offset (divided by 4 in 32-bit mode and 8 in 64-bit mode)
1490  is encoded in bits 16-23 (mask: ``0x00FF0000``).  The registers saved are
1491  encoded in bits 0-14 (mask: ``0x00007FFF``) as five 3-bit entries from the
1492  following table:
1493
1494    ==============  =============  ===============
1495    Compact Number  i386 Register  x86-64 Register
1496    ==============  =============  ===============
1497    1               ``EBX``        ``RBX``
1498    2               ``ECX``        ``R12``
1499    3               ``EDX``        ``R13``
1500    4               ``EDI``        ``R14``
1501    5               ``ESI``        ``R15``
1502    6               ``EBP``        ``RBP``
1503    ==============  =============  ===============
1504
1505*Frameless with a Small Constant Stack Size (``EBP`` or ``RBP`` is not used as a frame pointer)*
1506  To return, a constant (encoded in the compact unwind encoding) is added to the
1507  ``ESP/RSP``.  Then the return is done by popping the stack into the PC. All
1508  non-volatile registers that need to be restored must have been saved on the
1509  stack immediately after the return address. The stack size (divided by 4 in
1510  32-bit mode and 8 in 64-bit mode) is encoded in bits 16-23 (mask:
1511  ``0x00FF0000``). There is a maximum stack size of 1024 bytes in 32-bit mode
1512  and 2048 in 64-bit mode. The number of registers saved is encoded in bits 9-12
1513  (mask: ``0x00001C00``). Bits 0-9 (mask: ``0x000003FF``) contain which
1514  registers were saved and their order. (See the
1515  ``encodeCompactUnwindRegistersWithoutFrame()`` function in
1516  ``lib/Target/X86FrameLowering.cpp`` for the encoding algorithm.)
1517
1518*Frameless with a Large Constant Stack Size (``EBP`` or ``RBP`` is not used as a frame pointer)*
1519  This case is like the "Frameless with a Small Constant Stack Size" case, but
1520  the stack size is too large to encode in the compact unwind encoding. Instead
1521  it requires that the function contains "``subl $nnnnnn, %esp``" in its
1522  prolog. The compact encoding contains the offset to the ``$nnnnnn`` value in
1523  the function in bits 9-12 (mask: ``0x00001C00``).
1524
1525.. _Late Machine Code Optimizations:
1526
1527Late Machine Code Optimizations
1528-------------------------------
1529
1530.. note::
1531
1532  To Be Written
1533
1534.. _Code Emission:
1535
1536Code Emission
1537-------------
1538
1539The code emission step of code generation is responsible for lowering from the
1540code generator abstractions (like `MachineFunction`_, `MachineInstr`_, etc) down
1541to the abstractions used by the MC layer (`MCInst`_, `MCStreamer`_, etc).  This
1542is done with a combination of several different classes: the (misnamed)
1543target-independent AsmPrinter class, target-specific subclasses of AsmPrinter
1544(such as SparcAsmPrinter), and the TargetLoweringObjectFile class.
1545
1546Since the MC layer works at the level of abstraction of object files, it doesn't
1547have a notion of functions, global variables etc.  Instead, it thinks about
1548labels, directives, and instructions.  A key class used at this time is the
1549MCStreamer class.  This is an abstract API that is implemented in different ways
1550(e.g. to output a .s file, output an ELF .o file, etc) that is effectively an
1551"assembler API".  MCStreamer has one method per directive, such as EmitLabel,
1552EmitSymbolAttribute, SwitchSection, etc, which directly correspond to assembly
1553level directives.
1554
1555If you are interested in implementing a code generator for a target, there are
1556three important things that you have to implement for your target:
1557
1558#. First, you need a subclass of AsmPrinter for your target.  This class
1559   implements the general lowering process converting MachineFunction's into MC
1560   label constructs.  The AsmPrinter base class provides a number of useful
1561   methods and routines, and also allows you to override the lowering process in
1562   some important ways.  You should get much of the lowering for free if you are
1563   implementing an ELF, COFF, or MachO target, because the
1564   TargetLoweringObjectFile class implements much of the common logic.
1565
1566#. Second, you need to implement an instruction printer for your target.  The
1567   instruction printer takes an `MCInst`_ and renders it to a raw_ostream as
1568   text.  Most of this is automatically generated from the .td file (when you
1569   specify something like "``add $dst, $src1, $src2``" in the instructions), but
1570   you need to implement routines to print operands.
1571
1572#. Third, you need to implement code that lowers a `MachineInstr`_ to an MCInst,
1573   usually implemented in "<target>MCInstLower.cpp".  This lowering process is
1574   often target specific, and is responsible for turning jump table entries,
1575   constant pool indices, global variable addresses, etc into MCLabels as
1576   appropriate.  This translation layer is also responsible for expanding pseudo
1577   ops used by the code generator into the actual machine instructions they
1578   correspond to. The MCInsts that are generated by this are fed into the
1579   instruction printer or the encoder.
1580
1581Finally, at your choosing, you can also implement a subclass of MCCodeEmitter
1582which lowers MCInst's into machine code bytes and relocations.  This is
1583important if you want to support direct .o file emission, or would like to
1584implement an assembler for your target.
1585
1586VLIW Packetizer
1587---------------
1588
1589In a Very Long Instruction Word (VLIW) architecture, the compiler is responsible
1590for mapping instructions to functional-units available on the architecture. To
1591that end, the compiler creates groups of instructions called *packets* or
1592*bundles*. The VLIW packetizer in LLVM is a target-independent mechanism to
1593enable the packetization of machine instructions.
1594
1595Mapping from instructions to functional units
1596^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1597
1598Instructions in a VLIW target can typically be mapped to multiple functional
1599units. During the process of packetizing, the compiler must be able to reason
1600about whether an instruction can be added to a packet. This decision can be
1601complex since the compiler has to examine all possible mappings of instructions
1602to functional units. Therefore to alleviate compilation-time complexity, the
1603VLIW packetizer parses the instruction classes of a target and generates tables
1604at compiler build time. These tables can then be queried by the provided
1605machine-independent API to determine if an instruction can be accommodated in a
1606packet.
1607
1608How the packetization tables are generated and used
1609^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1610
1611The packetizer reads instruction classes from a target's itineraries and creates
1612a deterministic finite automaton (DFA) to represent the state of a packet. A DFA
1613consists of three major elements: inputs, states, and transitions. The set of
1614inputs for the generated DFA represents the instruction being added to a
1615packet. The states represent the possible consumption of functional units by
1616instructions in a packet. In the DFA, transitions from one state to another
1617occur on the addition of an instruction to an existing packet. If there is a
1618legal mapping of functional units to instructions, then the DFA contains a
1619corresponding transition. The absence of a transition indicates that a legal
1620mapping does not exist and that the instruction cannot be added to the packet.
1621
1622To generate tables for a VLIW target, add *Target*\ GenDFAPacketizer.inc as a
1623target to the Makefile in the target directory. The exported API provides three
1624functions: ``DFAPacketizer::clearResources()``,
1625``DFAPacketizer::reserveResources(MachineInstr *MI)``, and
1626``DFAPacketizer::canReserveResources(MachineInstr *MI)``. These functions allow
1627a target packetizer to add an instruction to an existing packet and to check
1628whether an instruction can be added to a packet. See
1629``llvm/CodeGen/DFAPacketizer.h`` for more information.
1630
1631Implementing a Native Assembler
1632===============================
1633
1634Though you're probably reading this because you want to write or maintain a
1635compiler backend, LLVM also fully supports building a native assembler.
1636We've tried hard to automate the generation of the assembler from the .td files
1637(in particular the instruction syntax and encodings), which means that a large
1638part of the manual and repetitive data entry can be factored and shared with the
1639compiler.
1640
1641Instruction Parsing
1642-------------------
1643
1644.. note::
1645
1646  To Be Written
1647
1648
1649Instruction Alias Processing
1650----------------------------
1651
1652Once the instruction is parsed, it enters the MatchInstructionImpl function.
1653The MatchInstructionImpl function performs alias processing and then does actual
1654matching.
1655
1656Alias processing is the phase that canonicalizes different lexical forms of the
1657same instructions down to one representation.  There are several different kinds
1658of alias that are possible to implement and they are listed below in the order
1659that they are processed (which is in order from simplest/weakest to most
1660complex/powerful).  Generally you want to use the first alias mechanism that
1661meets the needs of your instruction, because it will allow a more concise
1662description.
1663
1664Mnemonic Aliases
1665^^^^^^^^^^^^^^^^
1666
1667The first phase of alias processing is simple instruction mnemonic remapping for
1668classes of instructions which are allowed with two different mnemonics.  This
1669phase is a simple and unconditionally remapping from one input mnemonic to one
1670output mnemonic.  It isn't possible for this form of alias to look at the
1671operands at all, so the remapping must apply for all forms of a given mnemonic.
1672Mnemonic aliases are defined simply, for example X86 has:
1673
1674::
1675
1676  def : MnemonicAlias<"cbw",     "cbtw">;
1677  def : MnemonicAlias<"smovq",   "movsq">;
1678  def : MnemonicAlias<"fldcww",  "fldcw">;
1679  def : MnemonicAlias<"fucompi", "fucomip">;
1680  def : MnemonicAlias<"ud2a",    "ud2">;
1681
1682... and many others.  With a MnemonicAlias definition, the mnemonic is remapped
1683simply and directly.  Though MnemonicAlias's can't look at any aspect of the
1684instruction (such as the operands) they can depend on global modes (the same
1685ones supported by the matcher), through a Requires clause:
1686
1687::
1688
1689  def : MnemonicAlias<"pushf", "pushfq">, Requires<[In64BitMode]>;
1690  def : MnemonicAlias<"pushf", "pushfl">, Requires<[In32BitMode]>;
1691
1692In this example, the mnemonic gets mapped into a different one depending on
1693the current instruction set.
1694
1695Instruction Aliases
1696^^^^^^^^^^^^^^^^^^^
1697
1698The most general phase of alias processing occurs while matching is happening:
1699it provides new forms for the matcher to match along with a specific instruction
1700to generate.  An instruction alias has two parts: the string to match and the
1701instruction to generate.  For example:
1702
1703::
1704
1705  def : InstAlias<"movsx $src, $dst", (MOVSX16rr8W GR16:$dst, GR8  :$src)>;
1706  def : InstAlias<"movsx $src, $dst", (MOVSX16rm8W GR16:$dst, i8mem:$src)>;
1707  def : InstAlias<"movsx $src, $dst", (MOVSX32rr8  GR32:$dst, GR8  :$src)>;
1708  def : InstAlias<"movsx $src, $dst", (MOVSX32rr16 GR32:$dst, GR16 :$src)>;
1709  def : InstAlias<"movsx $src, $dst", (MOVSX64rr8  GR64:$dst, GR8  :$src)>;
1710  def : InstAlias<"movsx $src, $dst", (MOVSX64rr16 GR64:$dst, GR16 :$src)>;
1711  def : InstAlias<"movsx $src, $dst", (MOVSX64rr32 GR64:$dst, GR32 :$src)>;
1712
1713This shows a powerful example of the instruction aliases, matching the same
1714mnemonic in multiple different ways depending on what operands are present in
1715the assembly.  The result of instruction aliases can include operands in a
1716different order than the destination instruction, and can use an input multiple
1717times, for example:
1718
1719::
1720
1721  def : InstAlias<"clrb $reg", (XOR8rr  GR8 :$reg, GR8 :$reg)>;
1722  def : InstAlias<"clrw $reg", (XOR16rr GR16:$reg, GR16:$reg)>;
1723  def : InstAlias<"clrl $reg", (XOR32rr GR32:$reg, GR32:$reg)>;
1724  def : InstAlias<"clrq $reg", (XOR64rr GR64:$reg, GR64:$reg)>;
1725
1726This example also shows that tied operands are only listed once.  In the X86
1727backend, XOR8rr has two input GR8's and one output GR8 (where an input is tied
1728to the output).  InstAliases take a flattened operand list without duplicates
1729for tied operands.  The result of an instruction alias can also use immediates
1730and fixed physical registers which are added as simple immediate operands in the
1731result, for example:
1732
1733::
1734
1735  // Fixed Immediate operand.
1736  def : InstAlias<"aad", (AAD8i8 10)>;
1737
1738  // Fixed register operand.
1739  def : InstAlias<"fcomi", (COM_FIr ST1)>;
1740
1741  // Simple alias.
1742  def : InstAlias<"fcomi $reg", (COM_FIr RST:$reg)>;
1743
1744Instruction aliases can also have a Requires clause to make them subtarget
1745specific.
1746
1747If the back-end supports it, the instruction printer can automatically emit the
1748alias rather than what's being aliased. It typically leads to better, more
1749readable code. If it's better to print out what's being aliased, then pass a '0'
1750as the third parameter to the InstAlias definition.
1751
1752Instruction Matching
1753--------------------
1754
1755.. note::
1756
1757  To Be Written
1758
1759.. _Implementations of the abstract target description interfaces:
1760.. _implement the target description:
1761
1762Target-specific Implementation Notes
1763====================================
1764
1765This section of the document explains features or design decisions that are
1766specific to the code generator for a particular target.  First we start with a
1767table that summarizes what features are supported by each target.
1768
1769.. _target-feature-matrix:
1770
1771Target Feature Matrix
1772---------------------
1773
1774Note that this table does not include the C backend or Cpp backends, since they
1775do not use the target independent code generator infrastructure.  It also
1776doesn't list features that are not supported fully by any target yet.  It
1777considers a feature to be supported if at least one subtarget supports it.  A
1778feature being supported means that it is useful and works for most cases, it
1779does not indicate that there are zero known bugs in the implementation.  Here is
1780the key:
1781
1782:raw-html:`<table border="1" cellspacing="0">`
1783:raw-html:`<tr>`
1784:raw-html:`<th>Unknown</th>`
1785:raw-html:`<th>Not Applicable</th>`
1786:raw-html:`<th>No support</th>`
1787:raw-html:`<th>Partial Support</th>`
1788:raw-html:`<th>Complete Support</th>`
1789:raw-html:`</tr>`
1790:raw-html:`<tr>`
1791:raw-html:`<td class="unknown"></td>`
1792:raw-html:`<td class="na"></td>`
1793:raw-html:`<td class="no"></td>`
1794:raw-html:`<td class="partial"></td>`
1795:raw-html:`<td class="yes"></td>`
1796:raw-html:`</tr>`
1797:raw-html:`</table>`
1798
1799Here is the table:
1800
1801:raw-html:`<table width="689" border="1" cellspacing="0">`
1802:raw-html:`<tr><td></td>`
1803:raw-html:`<td colspan="13" align="center" style="background-color:#ffc">Target</td>`
1804:raw-html:`</tr>`
1805:raw-html:`<tr>`
1806:raw-html:`<th>Feature</th>`
1807:raw-html:`<th>ARM</th>`
1808:raw-html:`<th>Hexagon</th>`
1809:raw-html:`<th>MSP430</th>`
1810:raw-html:`<th>Mips</th>`
1811:raw-html:`<th>NVPTX</th>`
1812:raw-html:`<th>PowerPC</th>`
1813:raw-html:`<th>Sparc</th>`
1814:raw-html:`<th>SystemZ</th>`
1815:raw-html:`<th>X86</th>`
1816:raw-html:`<th>XCore</th>`
1817:raw-html:`<th>eBPF</th>`
1818:raw-html:`</tr>`
1819
1820:raw-html:`<tr>`
1821:raw-html:`<td><a href="#feat_reliable">is generally reliable</a></td>`
1822:raw-html:`<td class="yes"></td> <!-- ARM -->`
1823:raw-html:`<td class="yes"></td> <!-- Hexagon -->`
1824:raw-html:`<td class="unknown"></td> <!-- MSP430 -->`
1825:raw-html:`<td class="yes"></td> <!-- Mips -->`
1826:raw-html:`<td class="yes"></td> <!-- NVPTX -->`
1827:raw-html:`<td class="yes"></td> <!-- PowerPC -->`
1828:raw-html:`<td class="yes"></td> <!-- Sparc -->`
1829:raw-html:`<td class="yes"></td> <!-- SystemZ -->`
1830:raw-html:`<td class="yes"></td> <!-- X86 -->`
1831:raw-html:`<td class="yes"></td> <!-- XCore -->`
1832:raw-html:`<td class="yes"></td> <!-- eBPF -->`
1833:raw-html:`</tr>`
1834
1835:raw-html:`<tr>`
1836:raw-html:`<td><a href="#feat_asmparser">assembly parser</a></td>`
1837:raw-html:`<td class="no"></td> <!-- ARM -->`
1838:raw-html:`<td class="no"></td> <!-- Hexagon -->`
1839:raw-html:`<td class="no"></td> <!-- MSP430 -->`
1840:raw-html:`<td class="no"></td> <!-- Mips -->`
1841:raw-html:`<td class="no"></td> <!-- NVPTX -->`
1842:raw-html:`<td class="no"></td> <!-- PowerPC -->`
1843:raw-html:`<td class="no"></td> <!-- Sparc -->`
1844:raw-html:`<td class="yes"></td> <!-- SystemZ -->`
1845:raw-html:`<td class="yes"></td> <!-- X86 -->`
1846:raw-html:`<td class="no"></td> <!-- XCore -->`
1847:raw-html:`<td class="no"></td> <!-- eBPF -->`
1848:raw-html:`</tr>`
1849
1850:raw-html:`<tr>`
1851:raw-html:`<td><a href="#feat_disassembler">disassembler</a></td>`
1852:raw-html:`<td class="yes"></td> <!-- ARM -->`
1853:raw-html:`<td class="no"></td> <!-- Hexagon -->`
1854:raw-html:`<td class="no"></td> <!-- MSP430 -->`
1855:raw-html:`<td class="no"></td> <!-- Mips -->`
1856:raw-html:`<td class="na"></td> <!-- NVPTX -->`
1857:raw-html:`<td class="no"></td> <!-- PowerPC -->`
1858:raw-html:`<td class="yes"></td> <!-- SystemZ -->`
1859:raw-html:`<td class="no"></td> <!-- Sparc -->`
1860:raw-html:`<td class="yes"></td> <!-- X86 -->`
1861:raw-html:`<td class="yes"></td> <!-- XCore -->`
1862:raw-html:`<td class="yes"></td> <!-- eBPF -->`
1863:raw-html:`</tr>`
1864
1865:raw-html:`<tr>`
1866:raw-html:`<td><a href="#feat_inlineasm">inline asm</a></td>`
1867:raw-html:`<td class="yes"></td> <!-- ARM -->`
1868:raw-html:`<td class="yes"></td> <!-- Hexagon -->`
1869:raw-html:`<td class="unknown"></td> <!-- MSP430 -->`
1870:raw-html:`<td class="no"></td> <!-- Mips -->`
1871:raw-html:`<td class="yes"></td> <!-- NVPTX -->`
1872:raw-html:`<td class="yes"></td> <!-- PowerPC -->`
1873:raw-html:`<td class="unknown"></td> <!-- Sparc -->`
1874:raw-html:`<td class="yes"></td> <!-- SystemZ -->`
1875:raw-html:`<td class="yes"></td> <!-- X86 -->`
1876:raw-html:`<td class="yes"></td> <!-- XCore -->`
1877:raw-html:`<td class="no"></td> <!-- eBPF -->`
1878:raw-html:`</tr>`
1879
1880:raw-html:`<tr>`
1881:raw-html:`<td><a href="#feat_jit">jit</a></td>`
1882:raw-html:`<td class="partial"><a href="#feat_jit_arm">*</a></td> <!-- ARM -->`
1883:raw-html:`<td class="no"></td> <!-- Hexagon -->`
1884:raw-html:`<td class="unknown"></td> <!-- MSP430 -->`
1885:raw-html:`<td class="yes"></td> <!-- Mips -->`
1886:raw-html:`<td class="na"></td> <!-- NVPTX -->`
1887:raw-html:`<td class="yes"></td> <!-- PowerPC -->`
1888:raw-html:`<td class="unknown"></td> <!-- Sparc -->`
1889:raw-html:`<td class="yes"></td> <!-- SystemZ -->`
1890:raw-html:`<td class="yes"></td> <!-- X86 -->`
1891:raw-html:`<td class="no"></td> <!-- XCore -->`
1892:raw-html:`<td class="yes"></td> <!-- eBPF -->`
1893:raw-html:`</tr>`
1894
1895:raw-html:`<tr>`
1896:raw-html:`<td><a href="#feat_objectwrite">.o&nbsp;file writing</a></td>`
1897:raw-html:`<td class="no"></td> <!-- ARM -->`
1898:raw-html:`<td class="no"></td> <!-- Hexagon -->`
1899:raw-html:`<td class="no"></td> <!-- MSP430 -->`
1900:raw-html:`<td class="no"></td> <!-- Mips -->`
1901:raw-html:`<td class="na"></td> <!-- NVPTX -->`
1902:raw-html:`<td class="no"></td> <!-- PowerPC -->`
1903:raw-html:`<td class="no"></td> <!-- Sparc -->`
1904:raw-html:`<td class="yes"></td> <!-- SystemZ -->`
1905:raw-html:`<td class="yes"></td> <!-- X86 -->`
1906:raw-html:`<td class="no"></td> <!-- XCore -->`
1907:raw-html:`<td class="yes"></td> <!-- eBPF -->`
1908:raw-html:`</tr>`
1909
1910:raw-html:`<tr>`
1911:raw-html:`<td><a hr:raw-html:`ef="#feat_tailcall">tail calls</a></td>`
1912:raw-html:`<td class="yes"></td> <!-- ARM -->`
1913:raw-html:`<td class="yes"></td> <!-- Hexagon -->`
1914:raw-html:`<td class="unknown"></td> <!-- MSP430 -->`
1915:raw-html:`<td class="no"></td> <!-- Mips -->`
1916:raw-html:`<td class="no"></td> <!-- NVPTX -->`
1917:raw-html:`<td class="yes"></td> <!-- PowerPC -->`
1918:raw-html:`<td class="unknown"></td> <!-- Sparc -->`
1919:raw-html:`<td class="no"></td> <!-- SystemZ -->`
1920:raw-html:`<td class="yes"></td> <!-- X86 -->`
1921:raw-html:`<td class="no"></td> <!-- XCore -->`
1922:raw-html:`<td class="no"></td> <!-- eBPF -->`
1923:raw-html:`</tr>`
1924
1925:raw-html:`<tr>`
1926:raw-html:`<td><a href="#feat_segstacks">segmented stacks</a></td>`
1927:raw-html:`<td class="no"></td> <!-- ARM -->`
1928:raw-html:`<td class="no"></td> <!-- Hexagon -->`
1929:raw-html:`<td class="no"></td> <!-- MSP430 -->`
1930:raw-html:`<td class="no"></td> <!-- Mips -->`
1931:raw-html:`<td class="no"></td> <!-- NVPTX -->`
1932:raw-html:`<td class="no"></td> <!-- PowerPC -->`
1933:raw-html:`<td class="no"></td> <!-- Sparc -->`
1934:raw-html:`<td class="no"></td> <!-- SystemZ -->`
1935:raw-html:`<td class="partial"><a href="#feat_segstacks_x86">*</a></td> <!-- X86 -->`
1936:raw-html:`<td class="no"></td> <!-- XCore -->`
1937:raw-html:`<td class="no"></td> <!-- eBPF -->`
1938:raw-html:`</tr>`
1939
1940:raw-html:`</table>`
1941
1942.. _feat_reliable:
1943
1944Is Generally Reliable
1945^^^^^^^^^^^^^^^^^^^^^
1946
1947This box indicates whether the target is considered to be production quality.
1948This indicates that the target has been used as a static compiler to compile
1949large amounts of code by a variety of different people and is in continuous use.
1950
1951.. _feat_asmparser:
1952
1953Assembly Parser
1954^^^^^^^^^^^^^^^
1955
1956This box indicates whether the target supports parsing target specific .s files
1957by implementing the MCAsmParser interface.  This is required for llvm-mc to be
1958able to act as a native assembler and is required for inline assembly support in
1959the native .o file writer.
1960
1961.. _feat_disassembler:
1962
1963Disassembler
1964^^^^^^^^^^^^
1965
1966This box indicates whether the target supports the MCDisassembler API for
1967disassembling machine opcode bytes into MCInst's.
1968
1969.. _feat_inlineasm:
1970
1971Inline Asm
1972^^^^^^^^^^
1973
1974This box indicates whether the target supports most popular inline assembly
1975constraints and modifiers.
1976
1977.. _feat_jit:
1978
1979JIT Support
1980^^^^^^^^^^^
1981
1982This box indicates whether the target supports the JIT compiler through the
1983ExecutionEngine interface.
1984
1985.. _feat_jit_arm:
1986
1987The ARM backend has basic support for integer code in ARM codegen mode, but
1988lacks NEON and full Thumb support.
1989
1990.. _feat_objectwrite:
1991
1992.o File Writing
1993^^^^^^^^^^^^^^^
1994
1995This box indicates whether the target supports writing .o files (e.g. MachO,
1996ELF, and/or COFF) files directly from the target.  Note that the target also
1997must include an assembly parser and general inline assembly support for full
1998inline assembly support in the .o writer.
1999
2000Targets that don't support this feature can obviously still write out .o files,
2001they just rely on having an external assembler to translate from a .s file to a
2002.o file (as is the case for many C compilers).
2003
2004.. _feat_tailcall:
2005
2006Tail Calls
2007^^^^^^^^^^
2008
2009This box indicates whether the target supports guaranteed tail calls.  These are
2010calls marked "`tail <LangRef.html#i_call>`_" and use the fastcc calling
2011convention.  Please see the `tail call section`_ for more details.
2012
2013.. _feat_segstacks:
2014
2015Segmented Stacks
2016^^^^^^^^^^^^^^^^
2017
2018This box indicates whether the target supports segmented stacks. This replaces
2019the traditional large C stack with many linked segments. It is compatible with
2020the `gcc implementation <http://gcc.gnu.org/wiki/SplitStacks>`_ used by the Go
2021front end.
2022
2023.. _feat_segstacks_x86:
2024
2025Basic support exists on the X86 backend. Currently vararg doesn't work and the
2026object files are not marked the way the gold linker expects, but simple Go
2027programs can be built by dragonegg.
2028
2029.. _tail call section:
2030
2031Tail call optimization
2032----------------------
2033
2034Tail call optimization, callee reusing the stack of the caller, is currently
2035supported on x86/x86-64 and PowerPC. It is performed if:
2036
2037* Caller and callee have the calling convention ``fastcc``, ``cc 10`` (GHC
2038  calling convention) or ``cc 11`` (HiPE calling convention).
2039
2040* The call is a tail call - in tail position (ret immediately follows call and
2041  ret uses value of call or is void).
2042
2043* Option ``-tailcallopt`` is enabled.
2044
2045* Platform-specific constraints are met.
2046
2047x86/x86-64 constraints:
2048
2049* No variable argument lists are used.
2050
2051* On x86-64 when generating GOT/PIC code only module-local calls (visibility =
2052  hidden or protected) are supported.
2053
2054PowerPC constraints:
2055
2056* No variable argument lists are used.
2057
2058* No byval parameters are used.
2059
2060* On ppc32/64 GOT/PIC only module-local calls (visibility = hidden or protected)
2061  are supported.
2062
2063Example:
2064
2065Call as ``llc -tailcallopt test.ll``.
2066
2067.. code-block:: llvm
2068
2069  declare fastcc i32 @tailcallee(i32 inreg %a1, i32 inreg %a2, i32 %a3, i32 %a4)
2070
2071  define fastcc i32 @tailcaller(i32 %in1, i32 %in2) {
2072    %l1 = add i32 %in1, %in2
2073    %tmp = tail call fastcc i32 @tailcallee(i32 %in1 inreg, i32 %in2 inreg, i32 %in1, i32 %l1)
2074    ret i32 %tmp
2075  }
2076
2077Implications of ``-tailcallopt``:
2078
2079To support tail call optimization in situations where the callee has more
2080arguments than the caller a 'callee pops arguments' convention is used. This
2081currently causes each ``fastcc`` call that is not tail call optimized (because
2082one or more of above constraints are not met) to be followed by a readjustment
2083of the stack. So performance might be worse in such cases.
2084
2085Sibling call optimization
2086-------------------------
2087
2088Sibling call optimization is a restricted form of tail call optimization.
2089Unlike tail call optimization described in the previous section, it can be
2090performed automatically on any tail calls when ``-tailcallopt`` option is not
2091specified.
2092
2093Sibling call optimization is currently performed on x86/x86-64 when the
2094following constraints are met:
2095
2096* Caller and callee have the same calling convention. It can be either ``c`` or
2097  ``fastcc``.
2098
2099* The call is a tail call - in tail position (ret immediately follows call and
2100  ret uses value of call or is void).
2101
2102* Caller and callee have matching return type or the callee result is not used.
2103
2104* If any of the callee arguments are being passed in stack, they must be
2105  available in caller's own incoming argument stack and the frame offsets must
2106  be the same.
2107
2108Example:
2109
2110.. code-block:: llvm
2111
2112  declare i32 @bar(i32, i32)
2113
2114  define i32 @foo(i32 %a, i32 %b, i32 %c) {
2115  entry:
2116    %0 = tail call i32 @bar(i32 %a, i32 %b)
2117    ret i32 %0
2118  }
2119
2120The X86 backend
2121---------------
2122
2123The X86 code generator lives in the ``lib/Target/X86`` directory.  This code
2124generator is capable of targeting a variety of x86-32 and x86-64 processors, and
2125includes support for ISA extensions such as MMX and SSE.
2126
2127X86 Target Triples supported
2128^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2129
2130The following are the known target triples that are supported by the X86
2131backend.  This is not an exhaustive list, and it would be useful to add those
2132that people test.
2133
2134* **i686-pc-linux-gnu** --- Linux
2135
2136* **i386-unknown-freebsd5.3** --- FreeBSD 5.3
2137
2138* **i686-pc-cygwin** --- Cygwin on Win32
2139
2140* **i686-pc-mingw32** --- MingW on Win32
2141
2142* **i386-pc-mingw32msvc** --- MingW crosscompiler on Linux
2143
2144* **i686-apple-darwin*** --- Apple Darwin on X86
2145
2146* **x86_64-unknown-linux-gnu** --- Linux
2147
2148X86 Calling Conventions supported
2149^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2150
2151The following target-specific calling conventions are known to backend:
2152
2153* **x86_StdCall** --- stdcall calling convention seen on Microsoft Windows
2154  platform (CC ID = 64).
2155
2156* **x86_FastCall** --- fastcall calling convention seen on Microsoft Windows
2157  platform (CC ID = 65).
2158
2159* **x86_ThisCall** --- Similar to X86_StdCall. Passes first argument in ECX,
2160  others via stack. Callee is responsible for stack cleaning. This convention is
2161  used by MSVC by default for methods in its ABI (CC ID = 70).
2162
2163.. _X86 addressing mode:
2164
2165Representing X86 addressing modes in MachineInstrs
2166^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2167
2168The x86 has a very flexible way of accessing memory.  It is capable of forming
2169memory addresses of the following expression directly in integer instructions
2170(which use ModR/M addressing):
2171
2172::
2173
2174  SegmentReg: Base + [1,2,4,8] * IndexReg + Disp32
2175
2176In order to represent this, LLVM tracks no less than 5 operands for each memory
2177operand of this form.  This means that the "load" form of '``mov``' has the
2178following ``MachineOperand``\s in this order:
2179
2180::
2181
2182  Index:        0     |    1        2       3           4          5
2183  Meaning:   DestReg, | BaseReg,  Scale, IndexReg, Displacement Segment
2184  OperandTy: VirtReg, | VirtReg, UnsImm, VirtReg,   SignExtImm  PhysReg
2185
2186Stores, and all other instructions, treat the four memory operands in the same
2187way and in the same order.  If the segment register is unspecified (regno = 0),
2188then no segment override is generated.  "Lea" operations do not have a segment
2189register specified, so they only have 4 operands for their memory reference.
2190
2191X86 address spaces supported
2192^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2193
2194x86 has a feature which provides the ability to perform loads and stores to
2195different address spaces via the x86 segment registers.  A segment override
2196prefix byte on an instruction causes the instruction's memory access to go to
2197the specified segment.  LLVM address space 0 is the default address space, which
2198includes the stack, and any unqualified memory accesses in a program.  Address
2199spaces 1-255 are currently reserved for user-defined code.  The GS-segment is
2200represented by address space 256, while the FS-segment is represented by address
2201space 257. Other x86 segments have yet to be allocated address space
2202numbers.
2203
2204While these address spaces may seem similar to TLS via the ``thread_local``
2205keyword, and often use the same underlying hardware, there are some fundamental
2206differences.
2207
2208The ``thread_local`` keyword applies to global variables and specifies that they
2209are to be allocated in thread-local memory. There are no type qualifiers
2210involved, and these variables can be pointed to with normal pointers and
2211accessed with normal loads and stores.  The ``thread_local`` keyword is
2212target-independent at the LLVM IR level (though LLVM doesn't yet have
2213implementations of it for some configurations)
2214
2215Special address spaces, in contrast, apply to static types. Every load and store
2216has a particular address space in its address operand type, and this is what
2217determines which address space is accessed.  LLVM ignores these special address
2218space qualifiers on global variables, and does not provide a way to directly
2219allocate storage in them.  At the LLVM IR level, the behavior of these special
2220address spaces depends in part on the underlying OS or runtime environment, and
2221they are specific to x86 (and LLVM doesn't yet handle them correctly in some
2222cases).
2223
2224Some operating systems and runtime environments use (or may in the future use)
2225the FS/GS-segment registers for various low-level purposes, so care should be
2226taken when considering them.
2227
2228Instruction naming
2229^^^^^^^^^^^^^^^^^^
2230
2231An instruction name consists of the base name, a default operand size, and a a
2232character per operand with an optional special size. For example:
2233
2234::
2235
2236  ADD8rr      -> add, 8-bit register, 8-bit register
2237  IMUL16rmi   -> imul, 16-bit register, 16-bit memory, 16-bit immediate
2238  IMUL16rmi8  -> imul, 16-bit register, 16-bit memory, 8-bit immediate
2239  MOVSX32rm16 -> movsx, 32-bit register, 16-bit memory
2240
2241The PowerPC backend
2242-------------------
2243
2244The PowerPC code generator lives in the lib/Target/PowerPC directory.  The code
2245generation is retargetable to several variations or *subtargets* of the PowerPC
2246ISA; including ppc32, ppc64 and altivec.
2247
2248LLVM PowerPC ABI
2249^^^^^^^^^^^^^^^^
2250
2251LLVM follows the AIX PowerPC ABI, with two deviations. LLVM uses a PC relative
2252(PIC) or static addressing for accessing global values, so no TOC (r2) is
2253used. Second, r31 is used as a frame pointer to allow dynamic growth of a stack
2254frame.  LLVM takes advantage of having no TOC to provide space to save the frame
2255pointer in the PowerPC linkage area of the caller frame.  Other details of
2256PowerPC ABI can be found at `PowerPC ABI
2257<http://developer.apple.com/documentation/DeveloperTools/Conceptual/LowLevelABI/Articles/32bitPowerPC.html>`_\
2258. Note: This link describes the 32 bit ABI.  The 64 bit ABI is similar except
2259space for GPRs are 8 bytes wide (not 4) and r13 is reserved for system use.
2260
2261Frame Layout
2262^^^^^^^^^^^^
2263
2264The size of a PowerPC frame is usually fixed for the duration of a function's
2265invocation.  Since the frame is fixed size, all references into the frame can be
2266accessed via fixed offsets from the stack pointer.  The exception to this is
2267when dynamic alloca or variable sized arrays are present, then a base pointer
2268(r31) is used as a proxy for the stack pointer and stack pointer is free to grow
2269or shrink.  A base pointer is also used if llvm-gcc is not passed the
2270-fomit-frame-pointer flag. The stack pointer is always aligned to 16 bytes, so
2271that space allocated for altivec vectors will be properly aligned.
2272
2273An invocation frame is laid out as follows (low memory at top):
2274
2275:raw-html:`<table border="1" cellspacing="0">`
2276:raw-html:`<tr>`
2277:raw-html:`<td>Linkage<br><br></td>`
2278:raw-html:`</tr>`
2279:raw-html:`<tr>`
2280:raw-html:`<td>Parameter area<br><br></td>`
2281:raw-html:`</tr>`
2282:raw-html:`<tr>`
2283:raw-html:`<td>Dynamic area<br><br></td>`
2284:raw-html:`</tr>`
2285:raw-html:`<tr>`
2286:raw-html:`<td>Locals area<br><br></td>`
2287:raw-html:`</tr>`
2288:raw-html:`<tr>`
2289:raw-html:`<td>Saved registers area<br><br></td>`
2290:raw-html:`</tr>`
2291:raw-html:`<tr style="border-style: none hidden none hidden;">`
2292:raw-html:`<td><br></td>`
2293:raw-html:`</tr>`
2294:raw-html:`<tr>`
2295:raw-html:`<td>Previous Frame<br><br></td>`
2296:raw-html:`</tr>`
2297:raw-html:`</table>`
2298
2299The *linkage* area is used by a callee to save special registers prior to
2300allocating its own frame.  Only three entries are relevant to LLVM. The first
2301entry is the previous stack pointer (sp), aka link.  This allows probing tools
2302like gdb or exception handlers to quickly scan the frames in the stack.  A
2303function epilog can also use the link to pop the frame from the stack.  The
2304third entry in the linkage area is used to save the return address from the lr
2305register. Finally, as mentioned above, the last entry is used to save the
2306previous frame pointer (r31.)  The entries in the linkage area are the size of a
2307GPR, thus the linkage area is 24 bytes long in 32 bit mode and 48 bytes in 64
2308bit mode.
2309
231032 bit linkage area:
2311
2312:raw-html:`<table  border="1" cellspacing="0">`
2313:raw-html:`<tr>`
2314:raw-html:`<td>0</td>`
2315:raw-html:`<td>Saved SP (r1)</td>`
2316:raw-html:`</tr>`
2317:raw-html:`<tr>`
2318:raw-html:`<td>4</td>`
2319:raw-html:`<td>Saved CR</td>`
2320:raw-html:`</tr>`
2321:raw-html:`<tr>`
2322:raw-html:`<td>8</td>`
2323:raw-html:`<td>Saved LR</td>`
2324:raw-html:`</tr>`
2325:raw-html:`<tr>`
2326:raw-html:`<td>12</td>`
2327:raw-html:`<td>Reserved</td>`
2328:raw-html:`</tr>`
2329:raw-html:`<tr>`
2330:raw-html:`<td>16</td>`
2331:raw-html:`<td>Reserved</td>`
2332:raw-html:`</tr>`
2333:raw-html:`<tr>`
2334:raw-html:`<td>20</td>`
2335:raw-html:`<td>Saved FP (r31)</td>`
2336:raw-html:`</tr>`
2337:raw-html:`</table>`
2338
233964 bit linkage area:
2340
2341:raw-html:`<table border="1" cellspacing="0">`
2342:raw-html:`<tr>`
2343:raw-html:`<td>0</td>`
2344:raw-html:`<td>Saved SP (r1)</td>`
2345:raw-html:`</tr>`
2346:raw-html:`<tr>`
2347:raw-html:`<td>8</td>`
2348:raw-html:`<td>Saved CR</td>`
2349:raw-html:`</tr>`
2350:raw-html:`<tr>`
2351:raw-html:`<td>16</td>`
2352:raw-html:`<td>Saved LR</td>`
2353:raw-html:`</tr>`
2354:raw-html:`<tr>`
2355:raw-html:`<td>24</td>`
2356:raw-html:`<td>Reserved</td>`
2357:raw-html:`</tr>`
2358:raw-html:`<tr>`
2359:raw-html:`<td>32</td>`
2360:raw-html:`<td>Reserved</td>`
2361:raw-html:`</tr>`
2362:raw-html:`<tr>`
2363:raw-html:`<td>40</td>`
2364:raw-html:`<td>Saved FP (r31)</td>`
2365:raw-html:`</tr>`
2366:raw-html:`</table>`
2367
2368The *parameter area* is used to store arguments being passed to a callee
2369function.  Following the PowerPC ABI, the first few arguments are actually
2370passed in registers, with the space in the parameter area unused.  However, if
2371there are not enough registers or the callee is a thunk or vararg function,
2372these register arguments can be spilled into the parameter area.  Thus, the
2373parameter area must be large enough to store all the parameters for the largest
2374call sequence made by the caller.  The size must also be minimally large enough
2375to spill registers r3-r10.  This allows callees blind to the call signature,
2376such as thunks and vararg functions, enough space to cache the argument
2377registers.  Therefore, the parameter area is minimally 32 bytes (64 bytes in 64
2378bit mode.)  Also note that since the parameter area is a fixed offset from the
2379top of the frame, that a callee can access its spilt arguments using fixed
2380offsets from the stack pointer (or base pointer.)
2381
2382Combining the information about the linkage, parameter areas and alignment. A
2383stack frame is minimally 64 bytes in 32 bit mode and 128 bytes in 64 bit mode.
2384
2385The *dynamic area* starts out as size zero.  If a function uses dynamic alloca
2386then space is added to the stack, the linkage and parameter areas are shifted to
2387top of stack, and the new space is available immediately below the linkage and
2388parameter areas.  The cost of shifting the linkage and parameter areas is minor
2389since only the link value needs to be copied.  The link value can be easily
2390fetched by adding the original frame size to the base pointer.  Note that
2391allocations in the dynamic space need to observe 16 byte alignment.
2392
2393The *locals area* is where the llvm compiler reserves space for local variables.
2394
2395The *saved registers area* is where the llvm compiler spills callee saved
2396registers on entry to the callee.
2397
2398Prolog/Epilog
2399^^^^^^^^^^^^^
2400
2401The llvm prolog and epilog are the same as described in the PowerPC ABI, with
2402the following exceptions.  Callee saved registers are spilled after the frame is
2403created.  This allows the llvm epilog/prolog support to be common with other
2404targets.  The base pointer callee saved register r31 is saved in the TOC slot of
2405linkage area.  This simplifies allocation of space for the base pointer and
2406makes it convenient to locate programatically and during debugging.
2407
2408Dynamic Allocation
2409^^^^^^^^^^^^^^^^^^
2410
2411.. note::
2412
2413  TODO - More to come.
2414
2415The NVPTX backend
2416-----------------
2417
2418The NVPTX code generator under lib/Target/NVPTX is an open-source version of
2419the NVIDIA NVPTX code generator for LLVM.  It is contributed by NVIDIA and is
2420a port of the code generator used in the CUDA compiler (nvcc).  It targets the
2421PTX 3.0/3.1 ISA and can target any compute capability greater than or equal to
24222.0 (Fermi).
2423
2424This target is of production quality and should be completely compatible with
2425the official NVIDIA toolchain.
2426
2427Code Generator Options:
2428
2429:raw-html:`<table border="1" cellspacing="0">`
2430:raw-html:`<tr>`
2431:raw-html:`<th>Option</th>`
2432:raw-html:`<th>Description</th>`
2433:raw-html:`</tr>`
2434:raw-html:`<tr>`
2435:raw-html:`<td>sm_20</td>`
2436:raw-html:`<td align="left">Set shader model/compute capability to 2.0</td>`
2437:raw-html:`</tr>`
2438:raw-html:`<tr>`
2439:raw-html:`<td>sm_21</td>`
2440:raw-html:`<td align="left">Set shader model/compute capability to 2.1</td>`
2441:raw-html:`</tr>`
2442:raw-html:`<tr>`
2443:raw-html:`<td>sm_30</td>`
2444:raw-html:`<td align="left">Set shader model/compute capability to 3.0</td>`
2445:raw-html:`</tr>`
2446:raw-html:`<tr>`
2447:raw-html:`<td>sm_35</td>`
2448:raw-html:`<td align="left">Set shader model/compute capability to 3.5</td>`
2449:raw-html:`</tr>`
2450:raw-html:`<tr>`
2451:raw-html:`<td>ptx30</td>`
2452:raw-html:`<td align="left">Target PTX 3.0</td>`
2453:raw-html:`</tr>`
2454:raw-html:`<tr>`
2455:raw-html:`<td>ptx31</td>`
2456:raw-html:`<td align="left">Target PTX 3.1</td>`
2457:raw-html:`</tr>`
2458:raw-html:`</table>`
2459
2460The extended Berkeley Packet Filter (eBPF) backend
2461--------------------------------------------------
2462
2463Extended BPF (or eBPF) is similar to the original ("classic") BPF (cBPF) used
2464to filter network packets.  The
2465`bpf() system call <http://man7.org/linux/man-pages/man2/bpf.2.html>`_
2466performs a range of operations related to eBPF.  For both cBPF and eBPF
2467programs, the Linux kernel statically analyzes the programs before loading
2468them, in order to ensure that they cannot harm the running system.  eBPF is
2469a 64-bit RISC instruction set designed for one to one mapping to 64-bit CPUs.
2470Opcodes are 8-bit encoded, and 87 instructions are defined.  There are 10
2471registers, grouped by function as outlined below.
2472
2473::
2474
2475  R0        return value from in-kernel functions; exit value for eBPF program
2476  R1 - R5   function call arguments to in-kernel functions
2477  R6 - R9   callee-saved registers preserved by in-kernel functions
2478  R10       stack frame pointer (read only)
2479
2480Instruction encoding (arithmetic and jump)
2481^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2482eBPF is reusing most of the opcode encoding from classic to simplify conversion
2483of classic BPF to eBPF.  For arithmetic and jump instructions the 8-bit 'code'
2484field is divided into three parts:
2485
2486::
2487
2488  +----------------+--------+--------------------+
2489  |   4 bits       |  1 bit |   3 bits           |
2490  | operation code | source | instruction class  |
2491  +----------------+--------+--------------------+
2492  (MSB)                                      (LSB)
2493
2494Three LSB bits store instruction class which is one of:
2495
2496::
2497
2498  BPF_LD     0x0
2499  BPF_LDX    0x1
2500  BPF_ST     0x2
2501  BPF_STX    0x3
2502  BPF_ALU    0x4
2503  BPF_JMP    0x5
2504  (unused)   0x6
2505  BPF_ALU64  0x7
2506
2507When BPF_CLASS(code) == BPF_ALU or BPF_ALU64 or BPF_JMP,
25084th bit encodes source operand
2509
2510::
2511
2512  BPF_X     0x0  use src_reg register as source operand
2513  BPF_K     0x1  use 32 bit immediate as source operand
2514
2515and four MSB bits store operation code
2516
2517::
2518
2519  BPF_ADD   0x0  add
2520  BPF_SUB   0x1  subtract
2521  BPF_MUL   0x2  multiply
2522  BPF_DIV   0x3  divide
2523  BPF_OR    0x4  bitwise logical OR
2524  BPF_AND   0x5  bitwise logical AND
2525  BPF_LSH   0x6  left shift
2526  BPF_RSH   0x7  right shift (zero extended)
2527  BPF_NEG   0x8  arithmetic negation
2528  BPF_MOD   0x9  modulo
2529  BPF_XOR   0xa  bitwise logical XOR
2530  BPF_MOV   0xb  move register to register
2531  BPF_ARSH  0xc  right shift (sign extended)
2532  BPF_END   0xd  endianness conversion
2533
2534If BPF_CLASS(code) == BPF_JMP, BPF_OP(code) is one of
2535
2536::
2537
2538  BPF_JA    0x0  unconditional jump
2539  BPF_JEQ   0x1  jump ==
2540  BPF_JGT   0x2  jump >
2541  BPF_JGE   0x3  jump >=
2542  BPF_JSET  0x4  jump if (DST & SRC)
2543  BPF_JNE   0x5  jump !=
2544  BPF_JSGT  0x6  jump signed >
2545  BPF_JSGE  0x7  jump signed >=
2546  BPF_CALL  0x8  function call
2547  BPF_EXIT  0x9  function return
2548
2549Instruction encoding (load, store)
2550^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2551For load and store instructions the 8-bit 'code' field is divided as:
2552
2553::
2554
2555  +--------+--------+-------------------+
2556  | 3 bits | 2 bits |   3 bits          |
2557  |  mode  |  size  | instruction class |
2558  +--------+--------+-------------------+
2559  (MSB)                             (LSB)
2560
2561Size modifier is one of
2562
2563::
2564
2565  BPF_W       0x0  word
2566  BPF_H       0x1  half word
2567  BPF_B       0x2  byte
2568  BPF_DW      0x3  double word
2569
2570Mode modifier is one of
2571
2572::
2573
2574  BPF_IMM     0x0  immediate
2575  BPF_ABS     0x1  used to access packet data
2576  BPF_IND     0x2  used to access packet data
2577  BPF_MEM     0x3  memory
2578  (reserved)  0x4
2579  (reserved)  0x5
2580  BPF_XADD    0x6  exclusive add
2581
2582
2583Packet data access (BPF_ABS, BPF_IND)
2584^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2585
2586Two non-generic instructions: (BPF_ABS | <size> | BPF_LD) and
2587(BPF_IND | <size> | BPF_LD) which are used to access packet data.
2588Register R6 is an implicit input that must contain pointer to sk_buff.
2589Register R0 is an implicit output which contains the data fetched
2590from the packet.  Registers R1-R5 are scratch registers and must not
2591be used to store the data across BPF_ABS | BPF_LD or BPF_IND | BPF_LD
2592instructions.  These instructions have implicit program exit condition
2593as well.  When eBPF program is trying to access the data beyond
2594the packet boundary, the interpreter will abort the execution of the program.
2595
2596BPF_IND | BPF_W | BPF_LD is equivalent to:
2597  R0 = ntohl(\*(u32 \*) (((struct sk_buff \*) R6)->data + src_reg + imm32))
2598
2599eBPF maps
2600^^^^^^^^^
2601
2602eBPF maps are provided for sharing data between kernel and user-space.
2603Currently implemented types are hash and array, with potential extension to
2604support bloom filters, radix trees, etc.  A map is defined by its type,
2605maximum number of elements, key size and value size in bytes.  eBPF syscall
2606supports create, update, find and delete functions on maps.
2607
2608Function calls
2609^^^^^^^^^^^^^^
2610
2611Function call arguments are passed using up to five registers (R1 - R5).
2612The return value is passed in a dedicated register (R0).  Four additional
2613registers (R6 - R9) are callee-saved, and the values in these registers
2614are preserved within kernel functions.  R0 - R5 are scratch registers within
2615kernel functions, and eBPF programs must therefor store/restore values in
2616these registers if needed across function calls.  The stack can be accessed
2617using the read-only frame pointer R10.  eBPF registers map 1:1 to hardware
2618registers on x86_64 and other 64-bit architectures.  For example, x86_64
2619in-kernel JIT maps them as
2620
2621::
2622
2623  R0 - rax
2624  R1 - rdi
2625  R2 - rsi
2626  R3 - rdx
2627  R4 - rcx
2628  R5 - r8
2629  R6 - rbx
2630  R7 - r13
2631  R8 - r14
2632  R9 - r15
2633  R10 - rbp
2634
2635since x86_64 ABI mandates rdi, rsi, rdx, rcx, r8, r9 for argument passing
2636and rbx, r12 - r15 are callee saved.
2637
2638Program start
2639^^^^^^^^^^^^^
2640
2641An eBPF program receives a single argument and contains
2642a single eBPF main routine; the program does not contain eBPF functions.
2643Function calls are limited to a predefined set of kernel functions.  The size
2644of a program is limited to 4K instructions:  this ensures fast termination and
2645a limited number of kernel function calls.  Prior to running an eBPF program,
2646a verifier performs static analysis to prevent loops in the code and
2647to ensure valid register usage and operand types.
2648