1; RUN: opt -S < %s -instcombine | FileCheck %s 2 3target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64" 4target triple = "x86_64-apple-macosx10.7.0" 5 6; Check transforms involving atomic operations 7 8define i32 @test1(i32* %p) { 9; CHECK-LABEL: define i32 @test1( 10; CHECK: %x = load atomic i32, i32* %p seq_cst, align 4 11; CHECK: shl i32 %x, 1 12 %x = load atomic i32, i32* %p seq_cst, align 4 13 %y = load i32, i32* %p, align 4 14 %z = add i32 %x, %y 15 ret i32 %z 16} 17 18define i32 @test2(i32* %p) { 19; CHECK-LABEL: define i32 @test2( 20; CHECK: %x = load volatile i32, i32* %p, align 4 21; CHECK: %y = load volatile i32, i32* %p, align 4 22 %x = load volatile i32, i32* %p, align 4 23 %y = load volatile i32, i32* %p, align 4 24 %z = add i32 %x, %y 25 ret i32 %z 26} 27 28; The exact semantics of mixing volatile and non-volatile on the same 29; memory location are a bit unclear, but conservatively, we know we don't 30; want to remove the volatile. 31define i32 @test3(i32* %p) { 32; CHECK-LABEL: define i32 @test3( 33; CHECK: %x = load volatile i32, i32* %p, align 4 34 %x = load volatile i32, i32* %p, align 4 35 %y = load i32, i32* %p, align 4 36 %z = add i32 %x, %y 37 ret i32 %z 38} 39 40; Forwarding from a stronger ordered atomic is fine 41define i32 @test4(i32* %p) { 42; CHECK-LABEL: define i32 @test4( 43; CHECK: %x = load atomic i32, i32* %p seq_cst, align 4 44; CHECK: shl i32 %x, 1 45 %x = load atomic i32, i32* %p seq_cst, align 4 46 %y = load atomic i32, i32* %p unordered, align 4 47 %z = add i32 %x, %y 48 ret i32 %z 49} 50 51; Forwarding from a non-atomic is not. (The earlier load 52; could in priciple be promoted to atomic and then forwarded, 53; but we can't just drop the atomic from the load.) 54define i32 @test5(i32* %p) { 55; CHECK-LABEL: define i32 @test5( 56; CHECK: %x = load atomic i32, i32* %p unordered, align 4 57 %x = load atomic i32, i32* %p unordered, align 4 58 %y = load i32, i32* %p, align 4 59 %z = add i32 %x, %y 60 ret i32 %z 61} 62 63; Forwarding atomic to atomic is fine 64define i32 @test6(i32* %p) { 65; CHECK-LABEL: define i32 @test6( 66; CHECK: %x = load atomic i32, i32* %p unordered, align 4 67; CHECK: shl i32 %x, 1 68 %x = load atomic i32, i32* %p unordered, align 4 69 %y = load atomic i32, i32* %p unordered, align 4 70 %z = add i32 %x, %y 71 ret i32 %z 72} 73 74; FIXME: we currently don't do anything for monotonic 75define i32 @test7(i32* %p) { 76; CHECK-LABEL: define i32 @test7( 77; CHECK: %x = load atomic i32, i32* %p seq_cst, align 4 78; CHECK: %y = load atomic i32, i32* %p monotonic, align 4 79 %x = load atomic i32, i32* %p seq_cst, align 4 80 %y = load atomic i32, i32* %p monotonic, align 4 81 %z = add i32 %x, %y 82 ret i32 %z 83} 84 85; FIXME: We could forward in racy code 86define i32 @test8(i32* %p) { 87; CHECK-LABEL: define i32 @test8( 88; CHECK: %x = load atomic i32, i32* %p seq_cst, align 4 89; CHECK: %y = load atomic i32, i32* %p acquire, align 4 90 %x = load atomic i32, i32* %p seq_cst, align 4 91 %y = load atomic i32, i32* %p acquire, align 4 92 %z = add i32 %x, %y 93 ret i32 %z 94} 95 96; An unordered access to null is still unreachable. There's no 97; ordering imposed. 98define i32 @test9() { 99; CHECK-LABEL: define i32 @test9( 100; CHECK: store i32 undef, i32* null 101 %x = load atomic i32, i32* null unordered, align 4 102 ret i32 %x 103} 104 105define i32 @test9_no_null_opt() #0 { 106; CHECK-LABEL: define i32 @test9_no_null_opt( 107; CHECK: load atomic i32, i32* null unordered 108 %x = load atomic i32, i32* null unordered, align 4 109 ret i32 %x 110} 111 112; FIXME: Could also fold 113define i32 @test10() { 114; CHECK-LABEL: define i32 @test10( 115; CHECK: load atomic i32, i32* null monotonic 116 %x = load atomic i32, i32* null monotonic, align 4 117 ret i32 %x 118} 119 120define i32 @test10_no_null_opt() #0 { 121; CHECK-LABEL: define i32 @test10_no_null_opt( 122; CHECK: load atomic i32, i32* null monotonic 123 %x = load atomic i32, i32* null monotonic, align 4 124 ret i32 %x 125} 126 127; Would this be legal to fold? Probably? 128define i32 @test11() { 129; CHECK-LABEL: define i32 @test11( 130; CHECK: load atomic i32, i32* null seq_cst 131 %x = load atomic i32, i32* null seq_cst, align 4 132 ret i32 %x 133} 134 135define i32 @test11_no_null_opt() #0 { 136; CHECK-LABEL: define i32 @test11_no_null_opt( 137; CHECK: load atomic i32, i32* null seq_cst 138 %x = load atomic i32, i32* null seq_cst, align 4 139 ret i32 %x 140} 141 142; An unordered access to null is still unreachable. There's no 143; ordering imposed. 144define i32 @test12() { 145; CHECK-LABEL: define i32 @test12( 146; CHECK: store atomic i32 undef, i32* null 147 store atomic i32 0, i32* null unordered, align 4 148 ret i32 0 149} 150 151define i32 @test12_no_null_opt() #0 { 152; CHECK-LABEL: define i32 @test12_no_null_opt( 153; CHECK: store atomic i32 0, i32* null unordered 154 store atomic i32 0, i32* null unordered, align 4 155 ret i32 0 156} 157 158; FIXME: Could also fold 159define i32 @test13() { 160; CHECK-LABEL: define i32 @test13( 161; CHECK: store atomic i32 0, i32* null monotonic 162 store atomic i32 0, i32* null monotonic, align 4 163 ret i32 0 164} 165 166define i32 @test13_no_null_opt() #0 { 167; CHECK-LABEL: define i32 @test13_no_null_opt( 168; CHECK: store atomic i32 0, i32* null monotonic 169 store atomic i32 0, i32* null monotonic, align 4 170 ret i32 0 171} 172 173; Would this be legal to fold? Probably? 174define i32 @test14() { 175; CHECK-LABEL: define i32 @test14( 176; CHECK: store atomic i32 0, i32* null seq_cst 177 store atomic i32 0, i32* null seq_cst, align 4 178 ret i32 0 179} 180 181define i32 @test14_no_null_opt() #0 { 182; CHECK-LABEL: define i32 @test14_no_null_opt( 183; CHECK: store atomic i32 0, i32* null seq_cst 184 store atomic i32 0, i32* null seq_cst, align 4 185 ret i32 0 186} 187 188@a = external global i32 189@b = external global i32 190 191define i32 @test15(i1 %cnd) { 192; CHECK-LABEL: define i32 @test15( 193; CHECK: load atomic i32, i32* @a unordered, align 4 194; CHECK: load atomic i32, i32* @b unordered, align 4 195 %addr = select i1 %cnd, i32* @a, i32* @b 196 %x = load atomic i32, i32* %addr unordered, align 4 197 ret i32 %x 198} 199 200; FIXME: This would be legal to transform 201define i32 @test16(i1 %cnd) { 202; CHECK-LABEL: define i32 @test16( 203; CHECK: load atomic i32, i32* %addr monotonic, align 4 204 %addr = select i1 %cnd, i32* @a, i32* @b 205 %x = load atomic i32, i32* %addr monotonic, align 4 206 ret i32 %x 207} 208 209; FIXME: This would be legal to transform 210define i32 @test17(i1 %cnd) { 211; CHECK-LABEL: define i32 @test17( 212; CHECK: load atomic i32, i32* %addr seq_cst, align 4 213 %addr = select i1 %cnd, i32* @a, i32* @b 214 %x = load atomic i32, i32* %addr seq_cst, align 4 215 ret i32 %x 216} 217 218define i32 @test22(i1 %cnd) { 219; CHECK-LABEL: define i32 @test22( 220; CHECK: [[PHI:%.*]] = phi i32 221; CHECK: store atomic i32 [[PHI]], i32* @a unordered, align 4 222 br i1 %cnd, label %block1, label %block2 223 224block1: 225 store atomic i32 1, i32* @a unordered, align 4 226 br label %merge 227block2: 228 store atomic i32 2, i32* @a unordered, align 4 229 br label %merge 230 231merge: 232 ret i32 0 233} 234 235; TODO: probably also legal here 236define i32 @test23(i1 %cnd) { 237; CHECK-LABEL: define i32 @test23( 238; CHECK: br i1 %cnd, label %block1, label %block2 239 br i1 %cnd, label %block1, label %block2 240 241block1: 242 store atomic i32 1, i32* @a monotonic, align 4 243 br label %merge 244block2: 245 store atomic i32 2, i32* @a monotonic, align 4 246 br label %merge 247 248merge: 249 ret i32 0 250} 251 252declare void @clobber() 253 254define i32 @test18(float* %p) { 255; CHECK-LABEL: define i32 @test18( 256; CHECK: load atomic i32, i32* [[A:%.*]] unordered, align 4 257; CHECK: store atomic i32 [[B:%.*]], i32* [[C:%.*]] unordered, align 4 258 %x = load atomic float, float* %p unordered, align 4 259 call void @clobber() ;; keep the load around 260 store atomic float %x, float* %p unordered, align 4 261 ret i32 0 262} 263 264; TODO: probably also legal in this case 265define i32 @test19(float* %p) { 266; CHECK-LABEL: define i32 @test19( 267; CHECK: load atomic float, float* %p seq_cst, align 4 268; CHECK: store atomic float %x, float* %p seq_cst, align 4 269 %x = load atomic float, float* %p seq_cst, align 4 270 call void @clobber() ;; keep the load around 271 store atomic float %x, float* %p seq_cst, align 4 272 ret i32 0 273} 274 275define i32 @test20(i32** %p, i8* %v) { 276; CHECK-LABEL: define i32 @test20( 277; CHECK: store atomic i8* %v, i8** [[D:%.*]] unordered, align 4 278 %cast = bitcast i8* %v to i32* 279 store atomic i32* %cast, i32** %p unordered, align 4 280 ret i32 0 281} 282 283define i32 @test21(i32** %p, i8* %v) { 284; CHECK-LABEL: define i32 @test21( 285; CHECK: store atomic i32* %cast, i32** %p monotonic, align 4 286 %cast = bitcast i8* %v to i32* 287 store atomic i32* %cast, i32** %p monotonic, align 4 288 ret i32 0 289} 290 291define void @pr27490a(i8** %p1, i8** %p2) { 292; CHECK-LABEL: define void @pr27490 293; CHECK: %1 = bitcast i8** %p1 to i64* 294; CHECK: %l1 = load i64, i64* %1, align 8 295; CHECK: %2 = bitcast i8** %p2 to i64* 296; CHECK: store volatile i64 %l1, i64* %2, align 8 297 %l = load i8*, i8** %p1 298 store volatile i8* %l, i8** %p2 299 ret void 300} 301 302define void @pr27490b(i8** %p1, i8** %p2) { 303; CHECK-LABEL: define void @pr27490 304; CHECK: %1 = bitcast i8** %p1 to i64* 305; CHECK: %l1 = load i64, i64* %1, align 8 306; CHECK: %2 = bitcast i8** %p2 to i64* 307; CHECK: store atomic i64 %l1, i64* %2 seq_cst, align 8 308 %l = load i8*, i8** %p1 309 store atomic i8* %l, i8** %p2 seq_cst, align 8 310 ret void 311} 312 313;; At the moment, we can't form atomic vectors by folding since these are 314;; not representable in the IR. This was pr29121. The right long term 315;; solution is to extend the IR to handle this case. 316define <2 x float> @no_atomic_vector_load(i64* %p) { 317; CHECK-LABEL @no_atomic_vector_load 318; CHECK: load atomic i64, i64* %p unordered, align 8 319 %load = load atomic i64, i64* %p unordered, align 8 320 %.cast = bitcast i64 %load to <2 x float> 321 ret <2 x float> %.cast 322} 323 324define void @no_atomic_vector_store(<2 x float> %p, i8* %p2) { 325; CHECK-LABEL: @no_atomic_vector_store 326; CHECK: store atomic i64 %1, i64* %2 unordered, align 8 327 %1 = bitcast <2 x float> %p to i64 328 %2 = bitcast i8* %p2 to i64* 329 store atomic i64 %1, i64* %2 unordered, align 8 330 ret void 331} 332 333attributes #0 = { "null-pointer-is-valid"="true" } 334