147 lines · cpp
1// RUN: %clang_cc1 -O0 -triple spir -emit-llvm %s -o - | FileCheck %s2// RUN: %clang_cc1 -O0 -triple spir64 -emit-llvm %s -o - | FileCheck %s3 4// This file tests that using the _Float16 type with the spir target will not5// use the llvm intrinsics but instead will use the half arithmetic6// instructions directly.7 8// Previously attempting to use a constant _Float16 with a comparison9// instruction when the target is spir or spir64 lead to an assert being hit.10bool fcmp_const() {11 _Float16 a = 0.0f16;12 const _Float16 b = 1.0f16;13 14 // CHECK-NOT: llvm.convert.to.fp1615 // CHECK-NOT: llvm.convert.from.fp1616 17 // CHECK: [[REG1:%.*]] = load half, ptr %a, align 218 // CHECK-NEXT: fcmp olt half [[REG1]], 0xH3C0019 20 // CHECK: [[REG2:%.*]] = load half, ptr %a, align 221 // CHECK-NEXT: fcmp olt half [[REG2]], 0xH400022 23 // CHECK: [[REG3:%.*]] = load half, ptr %a, align 224 // CHECK-NEXT: fcmp ogt half [[REG3]], 0xH3C0025 26 // CHECK: [[REG4:%.*]] = load half, ptr %a, align 227 // CHECK-NEXT: fcmp ogt half [[REG4]], 0xH420028 29 // CHECK: [[REG5:%.*]] = load half, ptr %a, align 230 // CHECK-NEXT: fcmp oeq half [[REG5]], 0xH3C0031 32 // CHECK: [[REG7:%.*]] = load half, ptr %a, align 233 // CHECK-NEXT: fcmp oeq half [[REG7]], 0xH440034 35 // CHECK: [[REG8:%.*]] = load half, ptr %a, align 236 // CHECK-NEXT: fcmp une half [[REG8]], 0xH3C0037 38 // CHECK: [[REG9:%.*]] = load half, ptr %a, align 239 // CHECK-NEXT: fcmp une half [[REG9]], 0xH450040 41 // CHECK: [[REG10:%.*]] = load half, ptr %a, align 242 // CHECK-NEXT: fcmp ole half [[REG10]], 0xH3C0043 44 // CHECK: [[REG11:%.*]] = load half, ptr %a, align 245 // CHECK-NEXT: fcmp ole half [[REG11]], 0xH460046 47 // CHECK: [[REG12:%.*]] = load half, ptr %a, align 248 // CHECK-NEXT: fcmp oge half [[REG12]], 0xH3C0049 50 // CHECK: [[REG13:%.*]] = load half, ptr %a, align 251 // CHECK-NEXT: fcmp oge half [[REG13]], 0xH470052 return a < b || a < 2.0f16 || a > b || a > 3.0f16 || a == b || a == 4.0f16 ||53 a != b || a != 5.0f16 || a <= b || a <= 6.0f16 || a >= b ||54 a >= 7.0f16;55}56 57bool fcmp() {58 _Float16 a = 0.0f16;59 _Float16 b = 1.0f16;60 61 // CHECK-NOT: llvm.convert.to.fp1662 // CHECK-NOT: llvm.convert.from.fp1663 // CHECK: [[REG1:%.*]] = load half, ptr %a, align 264 // CHECK-NEXT: [[REG2:%.*]] = load half, ptr %b, align 265 // CHECK-NEXT: fcmp olt half [[REG1]], [[REG2]]66 67 // CHECK: [[REG3:%.*]] = load half, ptr %a, align 268 // CHECK-NEXT: [[REG4:%.*]] = load half, ptr %b, align 269 // CHECK-NEXT: fcmp ogt half [[REG3]], [[REG4]]70 71 // CHECK: [[REG5:%.*]] = load half, ptr %a, align 272 // CHECK-NEXT: [[REG6:%.*]] = load half, ptr %b, align 273 // CHECK-NEXT: fcmp oeq half [[REG5]], [[REG6]]74 75 // CHECK: [[REG7:%.*]] = load half, ptr %a, align 276 // CHECK-NEXT: [[REG8:%.*]] = load half, ptr %b, align 277 // CHECK-NEXT: fcmp une half [[REG7]], [[REG8]]78 79 // CHECK: [[REG7:%.*]] = load half, ptr %a, align 280 // CHECK-NEXT: [[REG8:%.*]] = load half, ptr %b, align 281 // CHECK-NEXT: fcmp ole half [[REG7]], [[REG8]]82 83 // CHECK: [[REG7:%.*]] = load half, ptr %a, align 284 // CHECK-NEXT: [[REG8:%.*]] = load half, ptr %b, align 285 // CHECK-NEXT: fcmp oge half [[REG7]], [[REG8]]86 return a < b || a > b || a == b || a != b || a <= b || a >= b;87}88 89_Float16 fadd() {90 _Float16 a = 1.0f16;91 const _Float16 b = 2.0f16;92 93 // CHECK-NOT: llvm.convert.to.fp1694 // CHECK-NOT: llvm.convert.from.fp1695 96 // CHECK: [[REG1:%.*]] = load half, ptr %a, align 297 // CHECK-NEXT: [[REG2:%.*]] = fadd half [[REG1]], 0xH400098 // CHECK-NEXT: [[REG3:%.*]] = fadd half [[REG2]], 0xH420099 // CHECK-NEXT: ret half [[REG3]]100 return a + b + 3.0f16;101}102 103_Float16 fsub() {104 _Float16 a = 1.0f16;105 const _Float16 b = 2.0f16;106 107 // CHECK-NOT: llvm.convert.to.fp16108 // CHECK-NOT: llvm.convert.from.fp16109 110 // CHECK: [[REG1:%.*]] = load half, ptr %a, align 2111 // CHECK-NEXT: [[REG2:%.*]] = fsub half [[REG1]], 0xH4000112 // CHECK-NEXT: [[REG3:%.*]] = fsub half [[REG2]], 0xH4200113 // CHECK-NEXT: ret half [[REG3]]114 return a - b - 3.0f16;115}116 117// CHECK: define{{.*}} spir_func noundef half @_Z4fmulDF16_(half noundef %arg)118_Float16 fmul(_Float16 arg) {119 _Float16 a = 1.0f16;120 const _Float16 b = 2.0f16;121 122 // CHECK-NOT: llvm.convert.to.fp16123 // CHECK-NOT: llvm.convert.from.fp16124 125 // CHECK: [[REG1:%.*]] = load half, ptr %a, align 2126 // CHECK-NEXT: [[REG2:%.*]] = load half, ptr %arg.addr, align 2127 // CHECK-NEXT: [[REG3:%.*]] = fmul half [[REG1]], [[REG2]]128 // CHECK-NEXT: [[REG4:%.*]] = fmul half [[REG3]], 0xH4000129 // CHECK-NEXT: [[REG5:%.*]] = fmul half [[REG4]], 0xH4200130 // CHECK-NEXT: ret half [[REG5]]131 return a * arg * b * 3.0f16;132}133 134_Float16 fdiv() {135 _Float16 a = 1.0f16;136 const _Float16 b = 2.0f16;137 138 // CHECK-NOT: llvm.convert.to.fp16139 // CHECK-NOT: llvm.convert.from.fp16140 141 // CHECK: [[REG1:%.*]] = load half, ptr %a, align 2142 // CHECK-NEXT: [[REG2:%.*]] = fdiv half [[REG1]], 0xH4000143 // CHECK-NEXT: [[REG3:%.*]] = fdiv half [[REG2]], 0xH4200144 // CHECK-NEXT: ret half [[REG3]]145 return a / b / 3.0f16;146}147