llvm/test/CodeGen/X86/vec-strict-inttofp-256-fp16.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
   2 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=avx512fp16,avx512vl -O3 | FileCheck %s --check-prefixes=CHECK,X86
   3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512fp16,avx512vl -O3 | FileCheck %s --check-prefixes=CHECK,X64
   4
   5 declare <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i1(<16 x i1>, metadata, metadata)
   6 declare <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i1(<16 x i1>, metadata, metadata)
   7 declare <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i8(<16 x i8>, metadata, metadata)
   8 declare <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i8(<16 x i8>, metadata, metadata)
   9 declare <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i16(<16 x i16>, metadata, metadata)
  10 declare <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i16(<16 x i16>, metadata, metadata)
  11 declare <8 x half> @llvm.experimental.constrained.sitofp.v8f16.v8i32(<8 x i32>, metadata, metadata)
  12 declare <8 x half> @llvm.experimental.constrained.uitofp.v8f16.v8i32(<8 x i32>, metadata, metadata)
  13 declare <4 x half> @llvm.experimental.constrained.sitofp.v4f16.v4i64(<4 x i64>, metadata, metadata)
  14 declare <4 x half> @llvm.experimental.constrained.uitofp.v4f16.v4i64(<4 x i64>, metadata, metadata)
  15 declare <8 x half> @llvm.experimental.constrained.sitofp.v8f16.v8i64(<8 x i64>, metadata, metadata)
  16 declare <8 x half> @llvm.experimental.constrained.uitofp.v8f16.v8i64(<8 x i64>, metadata, metadata)
  17
  18 define <16 x half> @sitofp_v16i1_v16f16(<16 x i1> %x) #0 {
  19 ; CHECK-LABEL: sitofp_v16i1_v16f16:
  20 ; CHECK:       # %bb.0:
  21 ; CHECK-NEXT:    vpmovzxbw {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero
  22 ; CHECK-NEXT:    vpsllw $15, %ymm0, %ymm0
  23 ; CHECK-NEXT:    vpsraw $15, %ymm0, %ymm0
  24 ; CHECK-NEXT:    vcvtw2ph %ymm0, %ymm0
  25 ; CHECK-NEXT:    ret{{[l|q]}}
  26  %result = call <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i1(<16 x i1> %x,
  27                                                               metadata !"round.dynamic",
  28                                                               metadata !"fpexcept.strict") #0
  29   ret <16 x half> %result
  30 }
  31
  32 define <16 x half> @uitofp_v16i1_v16f16(<16 x i1> %x) #0 {
  33 ; X86-LABEL: uitofp_v16i1_v16f16:
  34 ; X86:       # %bb.0:
  35 ; X86-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}{1to4}, %xmm0, %xmm0
  36 ; X86-NEXT:    vpmovzxbw {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero
  37 ; X86-NEXT:    vcvtw2ph %ymm0, %ymm0
  38 ; X86-NEXT:    retl
  39 ;
  40 ; X64-LABEL: uitofp_v16i1_v16f16:
  41 ; X64:       # %bb.0:
  42 ; X64-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
  43 ; X64-NEXT:    vpmovzxbw {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero
  44 ; X64-NEXT:    vcvtw2ph %ymm0, %ymm0
  45 ; X64-NEXT:    retq
  46  %result = call <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i1(<16 x i1> %x,
  47                                                               metadata !"round.dynamic",
  48                                                               metadata !"fpexcept.strict") #0
  49   ret <16 x half> %result
  50 }
  51
  52 define <16 x half> @sitofp_v16i8_v16f16(<16 x i8> %x) #0 {
  53 ; CHECK-LABEL: sitofp_v16i8_v16f16:
  54 ; CHECK:       # %bb.0:
  55 ; CHECK-NEXT:    vpmovsxbw %xmm0, %ymm0
  56 ; CHECK-NEXT:    vcvtw2ph %ymm0, %ymm0
  57 ; CHECK-NEXT:    ret{{[l|q]}}
  58  %result = call <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i8(<16 x i8> %x,
  59                                                               metadata !"round.dynamic",
  60                                                               metadata !"fpexcept.strict") #0
  61   ret <16 x half> %result
  62 }
  63
  64 define <16 x half> @uitofp_v16i8_v16f16(<16 x i8> %x) #0 {
  65 ; CHECK-LABEL: uitofp_v16i8_v16f16:
  66 ; CHECK:       # %bb.0:
  67 ; CHECK-NEXT:    vpmovzxbw {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero
  68 ; CHECK-NEXT:    vcvtw2ph %ymm0, %ymm0
  69 ; CHECK-NEXT:    ret{{[l|q]}}
  70  %result = call <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i8(<16 x i8> %x,
  71                                                               metadata !"round.dynamic",
  72                                                               metadata !"fpexcept.strict") #0
  73   ret <16 x half> %result
  74 }
  75
  76 define <16 x half> @sitofp_v16i16_v16f16(<16 x i16> %x) #0 {
  77 ; CHECK-LABEL: sitofp_v16i16_v16f16:
  78 ; CHECK:       # %bb.0:
  79 ; CHECK-NEXT:    vcvtw2ph %ymm0, %ymm0
  80 ; CHECK-NEXT:    ret{{[l|q]}}
  81  %result = call <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i16(<16 x i16> %x,
  82                                                               metadata !"round.dynamic",
  83                                                               metadata !"fpexcept.strict") #0
  84   ret <16 x half> %result
  85 }
  86
  87 define <16 x half> @uitofp_v16i16_v16f16(<16 x i16> %x) #0 {
  88 ; CHECK-LABEL: uitofp_v16i16_v16f16:
  89 ; CHECK:       # %bb.0:
  90 ; CHECK-NEXT:    vcvtuw2ph %ymm0, %ymm0
  91 ; CHECK-NEXT:    ret{{[l|q]}}
  92  %result = call <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i16(<16 x i16> %x,
  93                                                               metadata !"round.dynamic",
  94                                                               metadata !"fpexcept.strict") #0
  95   ret <16 x half> %result
  96 }
  97
  98 define <8 x half> @sitofp_v8i32_v8f16(<8 x i32> %x) #0 {
  99 ; CHECK-LABEL: sitofp_v8i32_v8f16:
 100 ; CHECK:       # %bb.0:
 101 ; CHECK-NEXT:    vcvtdq2ph %ymm0, %xmm0
 102 ; CHECK-NEXT:    vzeroupper
 103 ; CHECK-NEXT:    ret{{[l|q]}}
 104  %result = call <8 x half> @llvm.experimental.constrained.sitofp.v8f16.v8i32(<8 x i32> %x,
 105                                                               metadata !"round.dynamic",
 106                                                               metadata !"fpexcept.strict") #0
 107   ret <8 x half> %result
 108 }
 109
 110 define <8 x half> @uitofp_v8i32_v8f16(<8 x i32> %x) #0 {
 111 ; CHECK-LABEL: uitofp_v8i32_v8f16:
 112 ; CHECK:       # %bb.0:
 113 ; CHECK-NEXT:    vcvtudq2ph %ymm0, %xmm0
 114 ; CHECK-NEXT:    vzeroupper
 115 ; CHECK-NEXT:    ret{{[l|q]}}
 116  %result = call <8 x half> @llvm.experimental.constrained.uitofp.v8f16.v8i32(<8 x i32> %x,
 117                                                               metadata !"round.dynamic",
 118                                                               metadata !"fpexcept.strict") #0
 119   ret <8 x half> %result
 120 }
 121
 122 define <4 x half> @sitofp_v4i64_v4f16(<4 x i64> %x) #0 {
 123 ; CHECK-LABEL: sitofp_v4i64_v4f16:
 124 ; CHECK:       # %bb.0:
 125 ; CHECK-NEXT:    vcvtqq2ph %ymm0, %xmm0
 126 ; CHECK-NEXT:    vzeroupper
 127 ; CHECK-NEXT:    ret{{[l|q]}}
 128  %result = call <4 x half> @llvm.experimental.constrained.sitofp.v4f16.v4i64(<4 x i64> %x,
 129                                                               metadata !"round.dynamic",
 130                                                               metadata !"fpexcept.strict") #0
 131   ret <4 x half> %result
 132 }
 133
 134 define <4 x half> @uitofp_v4i64_v4f16(<4 x i64> %x) #0 {
 135 ; CHECK-LABEL: uitofp_v4i64_v4f16:
 136 ; CHECK:       # %bb.0:
 137 ; CHECK-NEXT:    vcvtuqq2ph %ymm0, %xmm0
 138 ; CHECK-NEXT:    vzeroupper
 139 ; CHECK-NEXT:    ret{{[l|q]}}
 140  %result = call <4 x half> @llvm.experimental.constrained.uitofp.v4f16.v4i64(<4 x i64> %x,
 141                                                               metadata !"round.dynamic",
 142                                                               metadata !"fpexcept.strict") #0
 143   ret <4 x half> %result
 144 }
 145
 146 define <8 x half> @sitofp_v8i64_v8f16(<8 x i64> %x) #1 {
 147 ; CHECK-LABEL: sitofp_v8i64_v8f16:
 148 ; CHECK:       # %bb.0:
 149 ; CHECK-NEXT:    vcvtqq2ph %ymm1, %xmm1
 150 ; CHECK-NEXT:    vcvtqq2ph %ymm0, %xmm0
 151 ; CHECK-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 152 ; CHECK-NEXT:    vzeroupper
 153 ; CHECK-NEXT:    ret{{[l|q]}}
 154  %result = call <8 x half> @llvm.experimental.constrained.sitofp.v8f16.v8i64(<8 x i64> %x,
 155                                                               metadata !"round.dynamic",
 156                                                               metadata !"fpexcept.strict") #0
 157   ret <8 x half> %result
 158 }
 159
 160 define <8 x half> @uitofp_v8i64_v8f16(<8 x i64> %x) #1 {
 161 ; CHECK-LABEL: uitofp_v8i64_v8f16:
 162 ; CHECK:       # %bb.0:
 163 ; CHECK-NEXT:    vcvtuqq2ph %ymm1, %xmm1
 164 ; CHECK-NEXT:    vcvtuqq2ph %ymm0, %xmm0
 165 ; CHECK-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 166 ; CHECK-NEXT:    vzeroupper
 167 ; CHECK-NEXT:    ret{{[l|q]}}
 168  %result = call <8 x half> @llvm.experimental.constrained.uitofp.v8f16.v8i64(<8 x i64> %x,
 169                                                               metadata !"round.dynamic",
 170                                                               metadata !"fpexcept.strict") #0
 171   ret <8 x half> %result
 172 }
 173
 174 attributes #0 = { strictfp }
 175 attributes #1 = { strictfp "min-legal-vector-width"="256" "prefer-vector-width"="256" }