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
  16 define <16 x half> @sitofp_v16i1_v16f16(<16 x i1> %x) #0 {
  17 ; CHECK-LABEL: sitofp_v16i1_v16f16:
  18 ; CHECK:       # %bb.0:
  19 ; 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
  20 ; CHECK-NEXT:    vpsllw $15, %ymm0, %ymm0
  21 ; CHECK-NEXT:    vpsraw $15, %ymm0, %ymm0
  22 ; CHECK-NEXT:    vcvtw2ph %ymm0, %ymm0
  23 ; CHECK-NEXT:    ret{{[l|q]}}
  24  %result = call <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i1(<16 x i1> %x,
  25                                                               metadata !"round.dynamic",
  26                                                               metadata !"fpexcept.strict") #0
  27   ret <16 x half> %result
  28 }
  29
  30 define <16 x half> @uitofp_v16i1_v16f16(<16 x i1> %x) #0 {
  31 ; X86-LABEL: uitofp_v16i1_v16f16:
  32 ; X86:       # %bb.0:
  33 ; X86-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
  34 ; 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
  35 ; X86-NEXT:    vcvtuw2ph %ymm0, %ymm0
  36 ; X86-NEXT:    retl
  37 ;
  38 ; X64-LABEL: uitofp_v16i1_v16f16:
  39 ; X64:       # %bb.0:
  40 ; X64-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
  41 ; 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
  42 ; X64-NEXT:    vcvtuw2ph %ymm0, %ymm0
  43 ; X64-NEXT:    retq
  44  %result = call <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i1(<16 x i1> %x,
  45                                                               metadata !"round.dynamic",
  46                                                               metadata !"fpexcept.strict") #0
  47   ret <16 x half> %result
  48 }
  49
  50 define <16 x half> @sitofp_v16i8_v16f16(<16 x i8> %x) #0 {
  51 ; CHECK-LABEL: sitofp_v16i8_v16f16:
  52 ; CHECK:       # %bb.0:
  53 ; CHECK-NEXT:    vpmovsxbw %xmm0, %ymm0
  54 ; CHECK-NEXT:    vcvtw2ph %ymm0, %ymm0
  55 ; CHECK-NEXT:    ret{{[l|q]}}
  56  %result = call <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i8(<16 x i8> %x,
  57                                                               metadata !"round.dynamic",
  58                                                               metadata !"fpexcept.strict") #0
  59   ret <16 x half> %result
  60 }
  61
  62 define <16 x half> @uitofp_v16i8_v16f16(<16 x i8> %x) #0 {
  63 ; CHECK-LABEL: uitofp_v16i8_v16f16:
  64 ; CHECK:       # %bb.0:
  65 ; 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
  66 ; CHECK-NEXT:    vcvtuw2ph %ymm0, %ymm0
  67 ; CHECK-NEXT:    ret{{[l|q]}}
  68  %result = call <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i8(<16 x i8> %x,
  69                                                               metadata !"round.dynamic",
  70                                                               metadata !"fpexcept.strict") #0
  71   ret <16 x half> %result
  72 }
  73
  74 define <16 x half> @sitofp_v16i16_v16f16(<16 x i16> %x) #0 {
  75 ; CHECK-LABEL: sitofp_v16i16_v16f16:
  76 ; CHECK:       # %bb.0:
  77 ; CHECK-NEXT:    vcvtw2ph %ymm0, %ymm0
  78 ; CHECK-NEXT:    ret{{[l|q]}}
  79  %result = call <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i16(<16 x i16> %x,
  80                                                               metadata !"round.dynamic",
  81                                                               metadata !"fpexcept.strict") #0
  82   ret <16 x half> %result
  83 }
  84
  85 define <16 x half> @uitofp_v16i16_v16f16(<16 x i16> %x) #0 {
  86 ; CHECK-LABEL: uitofp_v16i16_v16f16:
  87 ; CHECK:       # %bb.0:
  88 ; CHECK-NEXT:    vcvtuw2ph %ymm0, %ymm0
  89 ; CHECK-NEXT:    ret{{[l|q]}}
  90  %result = call <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i16(<16 x i16> %x,
  91                                                               metadata !"round.dynamic",
  92                                                               metadata !"fpexcept.strict") #0
  93   ret <16 x half> %result
  94 }
  95
  96 define <8 x half> @sitofp_v8i32_v8f16(<8 x i32> %x) #0 {
  97 ; CHECK-LABEL: sitofp_v8i32_v8f16:
  98 ; CHECK:       # %bb.0:
  99 ; CHECK-NEXT:    vcvtdq2ph %ymm0, %xmm0
 100 ; CHECK-NEXT:    vzeroupper
 101 ; CHECK-NEXT:    ret{{[l|q]}}
 102  %result = call <8 x half> @llvm.experimental.constrained.sitofp.v8f16.v8i32(<8 x i32> %x,
 103                                                               metadata !"round.dynamic",
 104                                                               metadata !"fpexcept.strict") #0
 105   ret <8 x half> %result
 106 }
 107
 108 define <8 x half> @uitofp_v8i32_v8f16(<8 x i32> %x) #0 {
 109 ; CHECK-LABEL: uitofp_v8i32_v8f16:
 110 ; CHECK:       # %bb.0:
 111 ; CHECK-NEXT:    vcvtudq2ph %ymm0, %xmm0
 112 ; CHECK-NEXT:    vzeroupper
 113 ; CHECK-NEXT:    ret{{[l|q]}}
 114  %result = call <8 x half> @llvm.experimental.constrained.uitofp.v8f16.v8i32(<8 x i32> %x,
 115                                                               metadata !"round.dynamic",
 116                                                               metadata !"fpexcept.strict") #0
 117   ret <8 x half> %result
 118 }
 119
 120 define <4 x half> @sitofp_v4i64_v4f16(<4 x i64> %x) #0 {
 121 ; CHECK-LABEL: sitofp_v4i64_v4f16:
 122 ; CHECK:       # %bb.0:
 123 ; CHECK-NEXT:    vcvtqq2ph %ymm0, %xmm0
 124 ; CHECK-NEXT:    vzeroupper
 125 ; CHECK-NEXT:    ret{{[l|q]}}
 126  %result = call <4 x half> @llvm.experimental.constrained.sitofp.v4f16.v4i64(<4 x i64> %x,
 127                                                               metadata !"round.dynamic",
 128                                                               metadata !"fpexcept.strict") #0
 129   ret <4 x half> %result
 130 }
 131
 132 define <4 x half> @uitofp_v4i64_v4f16(<4 x i64> %x) #0 {
 133 ; CHECK-LABEL: uitofp_v4i64_v4f16:
 134 ; CHECK:       # %bb.0:
 135 ; CHECK-NEXT:    vcvtuqq2ph %ymm0, %xmm0
 136 ; CHECK-NEXT:    vzeroupper
 137 ; CHECK-NEXT:    ret{{[l|q]}}
 138  %result = call <4 x half> @llvm.experimental.constrained.uitofp.v4f16.v4i64(<4 x i64> %x,
 139                                                               metadata !"round.dynamic",
 140                                                               metadata !"fpexcept.strict") #0
 141   ret <4 x half> %result
 142 }
 143
 144 attributes #0 = { strictfp }