1; RUN: llc < %s -mcpu=cyclone -verify-machineinstrs -aarch64-ccmp -aarch64-stress-ccmp | FileCheck %s 2target triple = "arm64-apple-ios" 3 4; CHECK: single_same 5; CHECK: cmp w0, #5 6; CHECK-NEXT: ccmp w1, #17, #4, ne 7; CHECK-NEXT: b.ne 8; CHECK: %if.then 9; CHECK: bl _foo 10; CHECK: %if.end 11define i32 @single_same(i32 %a, i32 %b) nounwind ssp { 12entry: 13 %cmp = icmp eq i32 %a, 5 14 %cmp1 = icmp eq i32 %b, 17 15 %or.cond = or i1 %cmp, %cmp1 16 br i1 %or.cond, label %if.then, label %if.end 17 18if.then: 19 %call = tail call i32 @foo() nounwind 20 br label %if.end 21 22if.end: 23 ret i32 7 24} 25 26; Different condition codes for the two compares. 27; CHECK: single_different 28; CHECK: cmp w0, #6 29; CHECK-NEXT: ccmp w1, #17, #0, ge 30; CHECK-NEXT: b.eq 31; CHECK: %if.then 32; CHECK: bl _foo 33; CHECK: %if.end 34define i32 @single_different(i32 %a, i32 %b) nounwind ssp { 35entry: 36 %cmp = icmp sle i32 %a, 5 37 %cmp1 = icmp ne i32 %b, 17 38 %or.cond = or i1 %cmp, %cmp1 39 br i1 %or.cond, label %if.then, label %if.end 40 41if.then: 42 %call = tail call i32 @foo() nounwind 43 br label %if.end 44 45if.end: 46 ret i32 7 47} 48 49; Second block clobbers the flags, can't convert (easily). 50; CHECK: single_flagclobber 51; CHECK: cmp 52; CHECK: b.eq 53; CHECK: cmp 54; CHECK: b.gt 55define i32 @single_flagclobber(i32 %a, i32 %b) nounwind ssp { 56entry: 57 %cmp = icmp eq i32 %a, 5 58 br i1 %cmp, label %if.then, label %lor.lhs.false 59 60lor.lhs.false: ; preds = %entry 61 %cmp1 = icmp slt i32 %b, 7 62 %mul = shl nsw i32 %b, 1 63 %add = add nsw i32 %b, 1 64 %cond = select i1 %cmp1, i32 %mul, i32 %add 65 %cmp2 = icmp slt i32 %cond, 17 66 br i1 %cmp2, label %if.then, label %if.end 67 68if.then: ; preds = %lor.lhs.false, %entry 69 %call = tail call i32 @foo() nounwind 70 br label %if.end 71 72if.end: ; preds = %if.then, %lor.lhs.false 73 ret i32 7 74} 75 76; Second block clobbers the flags and ends with a tbz terminator. 77; CHECK: single_flagclobber_tbz 78; CHECK: cmp 79; CHECK: b.eq 80; CHECK: cmp 81; CHECK: tbz 82define i32 @single_flagclobber_tbz(i32 %a, i32 %b) nounwind ssp { 83entry: 84 %cmp = icmp eq i32 %a, 5 85 br i1 %cmp, label %if.then, label %lor.lhs.false 86 87lor.lhs.false: ; preds = %entry 88 %cmp1 = icmp slt i32 %b, 7 89 %mul = shl nsw i32 %b, 1 90 %add = add nsw i32 %b, 1 91 %cond = select i1 %cmp1, i32 %mul, i32 %add 92 %and = and i32 %cond, 8 93 %cmp2 = icmp ne i32 %and, 0 94 br i1 %cmp2, label %if.then, label %if.end 95 96if.then: ; preds = %lor.lhs.false, %entry 97 %call = tail call i32 @foo() nounwind 98 br label %if.end 99 100if.end: ; preds = %if.then, %lor.lhs.false 101 ret i32 7 102} 103 104; Speculatively execute division by zero. 105; The sdiv/udiv instructions do not trap when the divisor is zero, so they are 106; safe to speculate. 107; CHECK: speculate_division 108; CHECK-NOT: cmp 109; CHECK: sdiv 110; CHECK: cmp 111; CHECK-NEXT: ccmp 112define i32 @speculate_division(i32 %a, i32 %b) nounwind ssp { 113entry: 114 %cmp = icmp sgt i32 %a, 0 115 br i1 %cmp, label %land.lhs.true, label %if.end 116 117land.lhs.true: 118 %div = sdiv i32 %b, %a 119 %cmp1 = icmp slt i32 %div, 17 120 br i1 %cmp1, label %if.then, label %if.end 121 122if.then: 123 %call = tail call i32 @foo() nounwind 124 br label %if.end 125 126if.end: 127 ret i32 7 128} 129 130; Floating point compare. 131; CHECK: single_fcmp 132; CHECK: cmp 133; CHECK-NOT: b. 134; CHECK: fccmp {{.*}}, #8, ge 135; CHECK: b.lt 136define i32 @single_fcmp(i32 %a, float %b) nounwind ssp { 137entry: 138 %cmp = icmp sgt i32 %a, 0 139 br i1 %cmp, label %land.lhs.true, label %if.end 140 141land.lhs.true: 142 %conv = sitofp i32 %a to float 143 %div = fdiv float %b, %conv 144 %cmp1 = fcmp oge float %div, 1.700000e+01 145 br i1 %cmp1, label %if.then, label %if.end 146 147if.then: 148 %call = tail call i32 @foo() nounwind 149 br label %if.end 150 151if.end: 152 ret i32 7 153} 154 155; Chain multiple compares. 156; CHECK: multi_different 157; CHECK: cmp 158; CHECK: ccmp 159; CHECK: ccmp 160; CHECK: b. 161define void @multi_different(i32 %a, i32 %b, i32 %c) nounwind ssp { 162entry: 163 %cmp = icmp sgt i32 %a, %b 164 br i1 %cmp, label %land.lhs.true, label %if.end 165 166land.lhs.true: 167 %div = sdiv i32 %b, %a 168 %cmp1 = icmp eq i32 %div, 5 169 %cmp4 = icmp sgt i32 %div, %c 170 %or.cond = and i1 %cmp1, %cmp4 171 br i1 %or.cond, label %if.then, label %if.end 172 173if.then: 174 %call = tail call i32 @foo() nounwind 175 br label %if.end 176 177if.end: 178 ret void 179} 180 181; Convert a cbz in the head block. 182; CHECK: cbz_head 183; CHECK: cmp w0, #0 184; CHECK: ccmp 185define i32 @cbz_head(i32 %a, i32 %b) nounwind ssp { 186entry: 187 %cmp = icmp eq i32 %a, 0 188 %cmp1 = icmp ne i32 %b, 17 189 %or.cond = or i1 %cmp, %cmp1 190 br i1 %or.cond, label %if.then, label %if.end 191 192if.then: 193 %call = tail call i32 @foo() nounwind 194 br label %if.end 195 196if.end: 197 ret i32 7 198} 199 200; Check that the immediate operand is in range. The ccmp instruction encodes a 201; smaller range of immediates than subs/adds. 202; The ccmp immediates must be in the range 0-31. 203; CHECK: immediate_range 204; CHECK-NOT: ccmp 205define i32 @immediate_range(i32 %a, i32 %b) nounwind ssp { 206entry: 207 %cmp = icmp eq i32 %a, 5 208 %cmp1 = icmp eq i32 %b, 32 209 %or.cond = or i1 %cmp, %cmp1 210 br i1 %or.cond, label %if.then, label %if.end 211 212if.then: 213 %call = tail call i32 @foo() nounwind 214 br label %if.end 215 216if.end: 217 ret i32 7 218} 219 220; Convert a cbz in the second block. 221; CHECK: cbz_second 222; CHECK: cmp w0, #0 223; CHECK: ccmp w1, #0, #0, ne 224; CHECK: b.eq 225define i32 @cbz_second(i32 %a, i32 %b) nounwind ssp { 226entry: 227 %cmp = icmp eq i32 %a, 0 228 %cmp1 = icmp ne i32 %b, 0 229 %or.cond = or i1 %cmp, %cmp1 230 br i1 %or.cond, label %if.then, label %if.end 231 232if.then: 233 %call = tail call i32 @foo() nounwind 234 br label %if.end 235 236if.end: 237 ret i32 7 238} 239 240; Convert a cbnz in the second block. 241; CHECK: cbnz_second 242; CHECK: cmp w0, #0 243; CHECK: ccmp w1, #0, #4, ne 244; CHECK: b.ne 245define i32 @cbnz_second(i32 %a, i32 %b) nounwind ssp { 246entry: 247 %cmp = icmp eq i32 %a, 0 248 %cmp1 = icmp eq i32 %b, 0 249 %or.cond = or i1 %cmp, %cmp1 250 br i1 %or.cond, label %if.then, label %if.end 251 252if.then: 253 %call = tail call i32 @foo() nounwind 254 br label %if.end 255 256if.end: 257 ret i32 7 258} 259declare i32 @foo() 260 261%str1 = type { %str2 } 262%str2 = type { [24 x i8], i8*, i32, %str1*, i32, [4 x i8], %str1*, %str1*, %str1*, %str1*, %str1*, %str1*, %str1*, %str1*, %str1*, i8*, i8, i8*, %str1*, i8* } 263 264; Test case distilled from 126.gcc. 265; The phi in sw.bb.i.i gets multiple operands for the %entry predecessor. 266; CHECK: build_modify_expr 267define void @build_modify_expr() nounwind ssp { 268entry: 269 switch i32 undef, label %sw.bb.i.i [ 270 i32 69, label %if.end85 271 i32 70, label %if.end85 272 i32 71, label %if.end85 273 i32 72, label %if.end85 274 i32 73, label %if.end85 275 i32 105, label %if.end85 276 i32 106, label %if.end85 277 ] 278 279if.end85: 280 ret void 281 282sw.bb.i.i: 283 %ref.tr.i.i = phi %str1* [ %0, %sw.bb.i.i ], [ undef, %entry ] 284 %operands.i.i = getelementptr inbounds %str1, %str1* %ref.tr.i.i, i64 0, i32 0, i32 2 285 %arrayidx.i.i = bitcast i32* %operands.i.i to %str1** 286 %0 = load %str1*, %str1** %arrayidx.i.i, align 8 287 %code1.i.i.phi.trans.insert = getelementptr inbounds %str1, %str1* %0, i64 0, i32 0, i32 0, i64 16 288 br label %sw.bb.i.i 289} 290