llvm/test/CodeGen/X86/fma.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
   2 ; RUN: llc < %s -mtriple=i386-apple-darwin10  -mattr=+avx,+fma,-fma4 -show-mc-encoding | FileCheck %s --check-prefix=FMA32
   3 ; RUN: llc < %s -mtriple=i386-apple-darwin10  -mattr=+avx,-fma,-fma4 -show-mc-encoding | FileCheck %s --check-prefix=FMACALL32
   4 ; RUN: llc < %s -mtriple=x86_64-apple-darwin10 -mattr=+fma,-fma4 -show-mc-encoding | FileCheck %s --check-prefix=FMA64
   5 ; RUN: llc < %s -mtriple=x86_64-apple-darwin10  -mattr=-fma,-fma4 -show-mc-encoding | FileCheck %s --check-prefix=FMACALL64
   6 ; RUN: llc < %s -mtriple=x86_64-apple-darwin10  -mattr=+avx512f,-fma4 -show-mc-encoding | FileCheck %s --check-prefix=AVX512
   7 ; RUN: llc < %s -mtriple=x86_64-apple-darwin10  -mattr=+avx512vl,-fma4 -show-mc-encoding | FileCheck %s --check-prefix=AVX512VL
   8 ; RUN: llc < %s -mtriple=i386-apple-darwin10 -mcpu=bdver2 -mattr=-fma4 -show-mc-encoding | FileCheck %s --check-prefix=FMA32
   9 ; RUN: llc < %s -mtriple=i386-apple-darwin10 -mcpu=bdver2 -mattr=-fma,-fma4 -show-mc-encoding | FileCheck %s --check-prefixes=FMACALL32,FMACALL32_BDVER2
  10
  11 define float @test_f32(float %a, float %b, float %c) #0 {
  12 ; FMA32-LABEL: test_f32:
  13 ; FMA32:       ## %bb.0:
  14 ; FMA32-NEXT:    pushl %eax ## encoding: [0x50]
  15 ; FMA32-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x44,0x24,0x08]
  16 ; FMA32-NEXT:    ## xmm0 = mem[0],zero,zero,zero
  17 ; FMA32-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfa,0x10,0x4c,0x24,0x0c]
  18 ; FMA32-NEXT:    ## xmm1 = mem[0],zero,zero,zero
  19 ; FMA32-NEXT:    vfmadd213ss {{[0-9]+}}(%esp), %xmm0, %xmm1 ## encoding: [0xc4,0xe2,0x79,0xa9,0x4c,0x24,0x10]
  20 ; FMA32-NEXT:    ## xmm1 = (xmm0 * xmm1) + mem
  21 ; FMA32-NEXT:    vmovss %xmm1, (%esp) ## encoding: [0xc5,0xfa,0x11,0x0c,0x24]
  22 ; FMA32-NEXT:    flds (%esp) ## encoding: [0xd9,0x04,0x24]
  23 ; FMA32-NEXT:    popl %eax ## encoding: [0x58]
  24 ; FMA32-NEXT:    retl ## encoding: [0xc3]
  25 ;
  26 ; FMACALL32-LABEL: test_f32:
  27 ; FMACALL32:       ## %bb.0:
  28 ; FMACALL32-NEXT:    jmp _fmaf ## TAILCALL
  29 ; FMACALL32-NEXT:    ## encoding: [0xeb,A]
  30 ; FMACALL32-NEXT:    ## fixup A - offset: 1, value: _fmaf-1, kind: FK_PCRel_1
  31 ;
  32 ; FMA64-LABEL: test_f32:
  33 ; FMA64:       ## %bb.0:
  34 ; FMA64-NEXT:    vfmadd213ss %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0x71,0xa9,0xc2]
  35 ; FMA64-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
  36 ; FMA64-NEXT:    retq ## encoding: [0xc3]
  37 ;
  38 ; FMACALL64-LABEL: test_f32:
  39 ; FMACALL64:       ## %bb.0:
  40 ; FMACALL64-NEXT:    jmp _fmaf ## TAILCALL
  41 ; FMACALL64-NEXT:    ## encoding: [0xeb,A]
  42 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-1, kind: FK_PCRel_1
  43 ;
  44 ; AVX512-LABEL: test_f32:
  45 ; AVX512:       ## %bb.0:
  46 ; AVX512-NEXT:    vfmadd213ss %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa9,0xc2]
  47 ; AVX512-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
  48 ; AVX512-NEXT:    retq ## encoding: [0xc3]
  49 ;
  50 ; AVX512VL-LABEL: test_f32:
  51 ; AVX512VL:       ## %bb.0:
  52 ; AVX512VL-NEXT:    vfmadd213ss %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa9,0xc2]
  53 ; AVX512VL-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
  54 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
  55   %call = call float @llvm.fma.f32(float %a, float %b, float %c)
  56   ret float %call
  57 }
  58
  59 define float @test_f32_reassoc(float %a, float %b, float %c) #0 {
  60 ; FMA32-LABEL: test_f32_reassoc:
  61 ; FMA32:       ## %bb.0:
  62 ; FMA32-NEXT:    pushl %eax ## encoding: [0x50]
  63 ; FMA32-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x44,0x24,0x08]
  64 ; FMA32-NEXT:    ## xmm0 = mem[0],zero,zero,zero
  65 ; FMA32-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfa,0x10,0x4c,0x24,0x0c]
  66 ; FMA32-NEXT:    ## xmm1 = mem[0],zero,zero,zero
  67 ; FMA32-NEXT:    vfmadd213ss {{[0-9]+}}(%esp), %xmm0, %xmm1 ## encoding: [0xc4,0xe2,0x79,0xa9,0x4c,0x24,0x10]
  68 ; FMA32-NEXT:    ## xmm1 = (xmm0 * xmm1) + mem
  69 ; FMA32-NEXT:    vmovss %xmm1, (%esp) ## encoding: [0xc5,0xfa,0x11,0x0c,0x24]
  70 ; FMA32-NEXT:    flds (%esp) ## encoding: [0xd9,0x04,0x24]
  71 ; FMA32-NEXT:    popl %eax ## encoding: [0x58]
  72 ; FMA32-NEXT:    retl ## encoding: [0xc3]
  73 ;
  74 ; FMACALL32-LABEL: test_f32_reassoc:
  75 ; FMACALL32:       ## %bb.0:
  76 ; FMACALL32-NEXT:    pushl %eax ## encoding: [0x50]
  77 ; FMACALL32-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x44,0x24,0x08]
  78 ; FMACALL32-NEXT:    ## xmm0 = mem[0],zero,zero,zero
  79 ; FMACALL32-NEXT:    vmulss {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc5,0xfa,0x59,0x44,0x24,0x0c]
  80 ; FMACALL32-NEXT:    vaddss {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc5,0xfa,0x58,0x44,0x24,0x10]
  81 ; FMACALL32-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
  82 ; FMACALL32-NEXT:    flds (%esp) ## encoding: [0xd9,0x04,0x24]
  83 ; FMACALL32-NEXT:    popl %eax ## encoding: [0x58]
  84 ; FMACALL32-NEXT:    retl ## encoding: [0xc3]
  85 ;
  86 ; FMA64-LABEL: test_f32_reassoc:
  87 ; FMA64:       ## %bb.0:
  88 ; FMA64-NEXT:    vfmadd213ss %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0x71,0xa9,0xc2]
  89 ; FMA64-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
  90 ; FMA64-NEXT:    retq ## encoding: [0xc3]
  91 ;
  92 ; FMACALL64-LABEL: test_f32_reassoc:
  93 ; FMACALL64:       ## %bb.0:
  94 ; FMACALL64-NEXT:    mulss %xmm1, %xmm0 ## encoding: [0xf3,0x0f,0x59,0xc1]
  95 ; FMACALL64-NEXT:    addss %xmm2, %xmm0 ## encoding: [0xf3,0x0f,0x58,0xc2]
  96 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
  97 ;
  98 ; AVX512-LABEL: test_f32_reassoc:
  99 ; AVX512:       ## %bb.0:
 100 ; AVX512-NEXT:    vfmadd213ss %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa9,0xc2]
 101 ; AVX512-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 102 ; AVX512-NEXT:    retq ## encoding: [0xc3]
 103 ;
 104 ; AVX512VL-LABEL: test_f32_reassoc:
 105 ; AVX512VL:       ## %bb.0:
 106 ; AVX512VL-NEXT:    vfmadd213ss %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa9,0xc2]
 107 ; AVX512VL-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 108 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 109   %call = call reassoc float @llvm.fma.f32(float %a, float %b, float %c)
 110   ret float %call
 111 }
 112
 113 define double @test_f64(double %a, double %b, double %c) #0 {
 114 ; FMA32-LABEL: test_f64:
 115 ; FMA32:       ## %bb.0: ## %entry
 116 ; FMA32-NEXT:    subl $12, %esp ## encoding: [0x83,0xec,0x0c]
 117 ; FMA32-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x44,0x24,0x10]
 118 ; FMA32-NEXT:    ## xmm0 = mem[0],zero
 119 ; FMA32-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfb,0x10,0x4c,0x24,0x18]
 120 ; FMA32-NEXT:    ## xmm1 = mem[0],zero
 121 ; FMA32-NEXT:    vfmadd213sd {{[0-9]+}}(%esp), %xmm0, %xmm1 ## encoding: [0xc4,0xe2,0xf9,0xa9,0x4c,0x24,0x20]
 122 ; FMA32-NEXT:    ## xmm1 = (xmm0 * xmm1) + mem
 123 ; FMA32-NEXT:    vmovsd %xmm1, (%esp) ## encoding: [0xc5,0xfb,0x11,0x0c,0x24]
 124 ; FMA32-NEXT:    fldl (%esp) ## encoding: [0xdd,0x04,0x24]
 125 ; FMA32-NEXT:    addl $12, %esp ## encoding: [0x83,0xc4,0x0c]
 126 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 127 ;
 128 ; FMACALL32-LABEL: test_f64:
 129 ; FMACALL32:       ## %bb.0: ## %entry
 130 ; FMACALL32-NEXT:    jmp _fma ## TAILCALL
 131 ; FMACALL32-NEXT:    ## encoding: [0xeb,A]
 132 ; FMACALL32-NEXT:    ## fixup A - offset: 1, value: _fma-1, kind: FK_PCRel_1
 133 ;
 134 ; FMA64-LABEL: test_f64:
 135 ; FMA64:       ## %bb.0: ## %entry
 136 ; FMA64-NEXT:    vfmadd213sd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa9,0xc2]
 137 ; FMA64-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 138 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 139 ;
 140 ; FMACALL64-LABEL: test_f64:
 141 ; FMACALL64:       ## %bb.0: ## %entry
 142 ; FMACALL64-NEXT:    jmp _fma ## TAILCALL
 143 ; FMACALL64-NEXT:    ## encoding: [0xeb,A]
 144 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-1, kind: FK_PCRel_1
 145 ;
 146 ; AVX512-LABEL: test_f64:
 147 ; AVX512:       ## %bb.0: ## %entry
 148 ; AVX512-NEXT:    vfmadd213sd %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa9,0xc2]
 149 ; AVX512-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 150 ; AVX512-NEXT:    retq ## encoding: [0xc3]
 151 ;
 152 ; AVX512VL-LABEL: test_f64:
 153 ; AVX512VL:       ## %bb.0: ## %entry
 154 ; AVX512VL-NEXT:    vfmadd213sd %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa9,0xc2]
 155 ; AVX512VL-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 156 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 157 entry:
 158   %call = call double @llvm.fma.f64(double %a, double %b, double %c)
 159   ret double %call
 160 }
 161
 162 define x86_fp80 @test_f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c) #0 {
 163 ; FMA32-LABEL: test_f80:
 164 ; FMA32:       ## %bb.0: ## %entry
 165 ; FMA32-NEXT:    subl $60, %esp ## encoding: [0x83,0xec,0x3c]
 166 ; FMA32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x40]
 167 ; FMA32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x50]
 168 ; FMA32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x60]
 169 ; FMA32-NEXT:    fstpt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x7c,0x24,0x20]
 170 ; FMA32-NEXT:    fstpt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x7c,0x24,0x10]
 171 ; FMA32-NEXT:    fstpt (%esp) ## encoding: [0xdb,0x3c,0x24]
 172 ; FMA32-NEXT:    calll _fmal ## encoding: [0xe8,A,A,A,A]
 173 ; FMA32-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: FK_PCRel_4
 174 ; FMA32-NEXT:    addl $60, %esp ## encoding: [0x83,0xc4,0x3c]
 175 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 176 ;
 177 ; FMACALL32-LABEL: test_f80:
 178 ; FMACALL32:       ## %bb.0: ## %entry
 179 ; FMACALL32-NEXT:    subl $60, %esp ## encoding: [0x83,0xec,0x3c]
 180 ; FMACALL32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x40]
 181 ; FMACALL32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x50]
 182 ; FMACALL32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x60]
 183 ; FMACALL32-NEXT:    fstpt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x7c,0x24,0x20]
 184 ; FMACALL32-NEXT:    fstpt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x7c,0x24,0x10]
 185 ; FMACALL32-NEXT:    fstpt (%esp) ## encoding: [0xdb,0x3c,0x24]
 186 ; FMACALL32-NEXT:    calll _fmal ## encoding: [0xe8,A,A,A,A]
 187 ; FMACALL32-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: FK_PCRel_4
 188 ; FMACALL32-NEXT:    addl $60, %esp ## encoding: [0x83,0xc4,0x3c]
 189 ; FMACALL32-NEXT:    retl ## encoding: [0xc3]
 190 ;
 191 ; FMA64-LABEL: test_f80:
 192 ; FMA64:       ## %bb.0: ## %entry
 193 ; FMA64-NEXT:    subq $56, %rsp ## encoding: [0x48,0x83,0xec,0x38]
 194 ; FMA64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x40]
 195 ; FMA64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x50]
 196 ; FMA64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x60]
 197 ; FMA64-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x20]
 198 ; FMA64-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x10]
 199 ; FMA64-NEXT:    fstpt (%rsp) ## encoding: [0xdb,0x3c,0x24]
 200 ; FMA64-NEXT:    callq _fmal ## encoding: [0xe8,A,A,A,A]
 201 ; FMA64-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: reloc_branch_4byte_pcrel
 202 ; FMA64-NEXT:    addq $56, %rsp ## encoding: [0x48,0x83,0xc4,0x38]
 203 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 204 ;
 205 ; FMACALL64-LABEL: test_f80:
 206 ; FMACALL64:       ## %bb.0: ## %entry
 207 ; FMACALL64-NEXT:    subq $56, %rsp ## encoding: [0x48,0x83,0xec,0x38]
 208 ; FMACALL64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x40]
 209 ; FMACALL64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x50]
 210 ; FMACALL64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x60]
 211 ; FMACALL64-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x20]
 212 ; FMACALL64-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x10]
 213 ; FMACALL64-NEXT:    fstpt (%rsp) ## encoding: [0xdb,0x3c,0x24]
 214 ; FMACALL64-NEXT:    callq _fmal ## encoding: [0xe8,A,A,A,A]
 215 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: reloc_branch_4byte_pcrel
 216 ; FMACALL64-NEXT:    addq $56, %rsp ## encoding: [0x48,0x83,0xc4,0x38]
 217 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
 218 ;
 219 ; AVX512-LABEL: test_f80:
 220 ; AVX512:       ## %bb.0: ## %entry
 221 ; AVX512-NEXT:    subq $56, %rsp ## encoding: [0x48,0x83,0xec,0x38]
 222 ; AVX512-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x40]
 223 ; AVX512-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x50]
 224 ; AVX512-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x60]
 225 ; AVX512-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x20]
 226 ; AVX512-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x10]
 227 ; AVX512-NEXT:    fstpt (%rsp) ## encoding: [0xdb,0x3c,0x24]
 228 ; AVX512-NEXT:    callq _fmal ## encoding: [0xe8,A,A,A,A]
 229 ; AVX512-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: reloc_branch_4byte_pcrel
 230 ; AVX512-NEXT:    addq $56, %rsp ## encoding: [0x48,0x83,0xc4,0x38]
 231 ; AVX512-NEXT:    retq ## encoding: [0xc3]
 232 ;
 233 ; AVX512VL-LABEL: test_f80:
 234 ; AVX512VL:       ## %bb.0: ## %entry
 235 ; AVX512VL-NEXT:    subq $56, %rsp ## encoding: [0x48,0x83,0xec,0x38]
 236 ; AVX512VL-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x40]
 237 ; AVX512VL-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x50]
 238 ; AVX512VL-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x60]
 239 ; AVX512VL-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x20]
 240 ; AVX512VL-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x10]
 241 ; AVX512VL-NEXT:    fstpt (%rsp) ## encoding: [0xdb,0x3c,0x24]
 242 ; AVX512VL-NEXT:    callq _fmal ## encoding: [0xe8,A,A,A,A]
 243 ; AVX512VL-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: reloc_branch_4byte_pcrel
 244 ; AVX512VL-NEXT:    addq $56, %rsp ## encoding: [0x48,0x83,0xc4,0x38]
 245 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 246 entry:
 247   %call = call x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
 248   ret x86_fp80 %call
 249 }
 250
 251 define float @test_f32_cst() #0 {
 252 ; FMA32-LABEL: test_f32_cst:
 253 ; FMA32:       ## %bb.0: ## %entry
 254 ; FMA32-NEXT:    flds {{\.?LCPI[0-9]+_[0-9]+}} ## encoding: [0xd9,0x05,A,A,A,A]
 255 ; FMA32-NEXT:    ## fixup A - offset: 2, value: {{\.?LCPI[0-9]+_[0-9]+}}, kind: FK_Data_4
 256 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 257 ;
 258 ; FMACALL32-LABEL: test_f32_cst:
 259 ; FMACALL32:       ## %bb.0: ## %entry
 260 ; FMACALL32-NEXT:    flds {{\.?LCPI[0-9]+_[0-9]+}} ## encoding: [0xd9,0x05,A,A,A,A]
 261 ; FMACALL32-NEXT:    ## fixup A - offset: 2, value: {{\.?LCPI[0-9]+_[0-9]+}}, kind: FK_Data_4
 262 ; FMACALL32-NEXT:    retl ## encoding: [0xc3]
 263 ;
 264 ; FMA64-LABEL: test_f32_cst:
 265 ; FMA64:       ## %bb.0: ## %entry
 266 ; FMA64-NEXT:    vmovss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x05,A,A,A,A]
 267 ; FMA64-NEXT:    ## fixup A - offset: 4, value: {{\.?LCPI[0-9]+_[0-9]+}}-4, kind: reloc_riprel_4byte
 268 ; FMA64-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 269 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 270 ;
 271 ; FMACALL64-LABEL: test_f32_cst:
 272 ; FMACALL64:       ## %bb.0: ## %entry
 273 ; FMACALL64-NEXT:    movss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 ## encoding: [0xf3,0x0f,0x10,0x05,A,A,A,A]
 274 ; FMACALL64-NEXT:    ## fixup A - offset: 4, value: {{\.?LCPI[0-9]+_[0-9]+}}-4, kind: reloc_riprel_4byte
 275 ; FMACALL64-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 276 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
 277 ;
 278 ; AVX512-LABEL: test_f32_cst:
 279 ; AVX512:       ## %bb.0: ## %entry
 280 ; AVX512-NEXT:    vmovss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfa,0x10,0x05,A,A,A,A]
 281 ; AVX512-NEXT:    ## fixup A - offset: 4, value: {{\.?LCPI[0-9]+_[0-9]+}}-4, kind: reloc_riprel_4byte
 282 ; AVX512-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 283 ; AVX512-NEXT:    retq ## encoding: [0xc3]
 284 ;
 285 ; AVX512VL-LABEL: test_f32_cst:
 286 ; AVX512VL:       ## %bb.0: ## %entry
 287 ; AVX512VL-NEXT:    vmovss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfa,0x10,0x05,A,A,A,A]
 288 ; AVX512VL-NEXT:    ## fixup A - offset: 4, value: {{\.?LCPI[0-9]+_[0-9]+}}-4, kind: reloc_riprel_4byte
 289 ; AVX512VL-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 290 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 291 entry:
 292   %call = call float @llvm.fma.f32(float 3.0, float 3.0, float 3.0)
 293   ret float %call
 294 }
 295
 296 define <4 x float> @test_v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c) #0 {
 297 ; FMA32-LABEL: test_v4f32:
 298 ; FMA32:       ## %bb.0: ## %entry
 299 ; FMA32-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0x71,0xa8,0xc2]
 300 ; FMA32-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 301 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 302 ;
 303 ; FMA64-LABEL: test_v4f32:
 304 ; FMA64:       ## %bb.0: ## %entry
 305 ; FMA64-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0x71,0xa8,0xc2]
 306 ; FMA64-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 307 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 308 ;
 309 ; FMACALL64-LABEL: test_v4f32:
 310 ; FMACALL64:       ## %bb.0: ## %entry
 311 ; FMACALL64-NEXT:    subq $88, %rsp ## encoding: [0x48,0x83,0xec,0x58]
 312 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 313 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x30]
 314 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 315 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x10]
 316 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 317 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
 318 ; FMACALL64-NEXT:    shufps $255, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xff]
 319 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,3,3,3]
 320 ; FMACALL64-NEXT:    shufps $255, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xff]
 321 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,3,3,3]
 322 ; FMACALL64-NEXT:    shufps $255, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xff]
 323 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,3,3,3]
 324 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 325 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 326 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 327 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 328 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 329 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
 330 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 331 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 332 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 333 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 334 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 335 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 336 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 337 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x30]
 338 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 339 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 340 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 341 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 342 ; FMACALL64-NEXT:    unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
 343 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x04,0x24]
 344 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 345 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 346 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 347 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 348 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
 349 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 350 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 351 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 352 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x30]
 353 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 354 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 355 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 356 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x40]
 357 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 358 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
 359 ; FMACALL64-NEXT:    shufps $85, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0x55]
 360 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,1,1]
 361 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 362 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 363 ; FMACALL64-NEXT:    shufps $85, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0x55]
 364 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,1,1]
 365 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 366 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x30]
 367 ; FMACALL64-NEXT:    shufps $85, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0x55]
 368 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,1,1]
 369 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 370 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 371 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 372 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x40]
 373 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 374 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 375 ; FMACALL64-NEXT:    unpcklpd (%rsp), %xmm1 ## 16-byte Folded Reload
 376 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x0c,0x24]
 377 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 378 ; FMACALL64-NEXT:    movaps %xmm1, %xmm0 ## encoding: [0x0f,0x28,0xc1]
 379 ; FMACALL64-NEXT:    addq $88, %rsp ## encoding: [0x48,0x83,0xc4,0x58]
 380 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
 381 ;
 382 ; AVX512-LABEL: test_v4f32:
 383 ; AVX512:       ## %bb.0: ## %entry
 384 ; AVX512-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0x71,0xa8,0xc2]
 385 ; AVX512-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 386 ; AVX512-NEXT:    retq ## encoding: [0xc3]
 387 ;
 388 ; AVX512VL-LABEL: test_v4f32:
 389 ; AVX512VL:       ## %bb.0: ## %entry
 390 ; AVX512VL-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa8,0xc2]
 391 ; AVX512VL-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 392 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 393 ;
 394 ; FMACALL32_BDVER2-LABEL: test_v4f32:
 395 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
 396 ; FMACALL32_BDVER2-NEXT:    subl $108, %esp ## encoding: [0x83,0xec,0x6c]
 397 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 398 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x54,0x24,0x40]
 399 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 400 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x4c,0x24,0x30]
 401 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 402 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x44,0x24,0x20]
 403 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x08,0x02]
 404 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm1, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x4c,0x24,0x04,0x02]
 405 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x02]
 406 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 407 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 408 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 409 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x60]
 410 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 411 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x40]
 412 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
 413 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x54,0x24,0x30]
 414 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
 415 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x20]
 416 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
 417 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x01]
 418 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x01]
 419 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 420 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 421 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 422 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x54]
 423 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 424 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x40]
 425 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
 426 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 427 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x30]
 428 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
 429 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 430 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x20]
 431 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
 432 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 433 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 434 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 435 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x40]
 436 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
 437 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x54,0x24,0x30]
 438 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
 439 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x20]
 440 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
 441 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x03]
 442 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x03]
 443 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x1c]
 444 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 445 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x54]
 446 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x18]
 447 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 448 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x60]
 449 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x14]
 450 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 451 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 452 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x10]
 453 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x44,0x24,0x1c]
 454 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 455 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x18,0x10]
 456 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[2,3]
 457 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x14,0x20]
 458 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0],xmm0[3]
 459 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x10,0x30]
 460 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1,2],mem[0]
 461 ; FMACALL32_BDVER2-NEXT:    addl $108, %esp ## encoding: [0x83,0xc4,0x6c]
 462 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
 463 entry:
 464   %call = call <4 x float> @llvm.fma.v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c)
 465   ret <4 x float> %call
 466 }
 467
 468 define <8 x float> @test_v8f32(<8 x float> %a, <8 x float> %b, <8 x float> %c) #0 {
 469 ; FMA32-LABEL: test_v8f32:
 470 ; FMA32:       ## %bb.0: ## %entry
 471 ; FMA32-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0x75,0xa8,0xc2]
 472 ; FMA32-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
 473 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 474 ;
 475 ; FMA64-LABEL: test_v8f32:
 476 ; FMA64:       ## %bb.0: ## %entry
 477 ; FMA64-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0x75,0xa8,0xc2]
 478 ; FMA64-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
 479 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 480 ;
 481 ; FMACALL64-LABEL: test_v8f32:
 482 ; FMACALL64:       ## %bb.0: ## %entry
 483 ; FMACALL64-NEXT:    subq $136, %rsp ## encoding: [0x48,0x81,0xec,0x88,0x00,0x00,0x00]
 484 ; FMACALL64-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 485 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x6c,0x24,0x50]
 486 ; FMACALL64-NEXT:    movaps %xmm4, (%rsp) ## 16-byte Spill
 487 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x24,0x24]
 488 ; FMACALL64-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 489 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x5c,0x24,0x40]
 490 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 491 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x60]
 492 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 493 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x30]
 494 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 495 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
 496 ; FMACALL64-NEXT:    shufps $255, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xff]
 497 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,3,3,3]
 498 ; FMACALL64-NEXT:    movaps %xmm2, %xmm1 ## encoding: [0x0f,0x28,0xca]
 499 ; FMACALL64-NEXT:    shufps $255, %xmm2, %xmm1 ## encoding: [0x0f,0xc6,0xca,0xff]
 500 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,3],xmm2[3,3]
 501 ; FMACALL64-NEXT:    movaps %xmm4, %xmm2 ## encoding: [0x0f,0x28,0xd4]
 502 ; FMACALL64-NEXT:    shufps $255, %xmm4, %xmm2 ## encoding: [0x0f,0xc6,0xd4,0xff]
 503 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,3],xmm4[3,3]
 504 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 505 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 506 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 507 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
 508 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 509 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x10]
 510 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 511 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 512 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 513 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x60]
 514 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 515 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 516 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
 517 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
 518 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 519 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 520 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 521 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 522 ; FMACALL64-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
 523 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x44,0x24,0x20]
 524 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 525 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 526 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x70]
 527 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 528 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x10]
 529 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 530 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x60]
 531 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
 532 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
 533 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 534 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 535 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 536 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
 537 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 538 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x10]
 539 ; FMACALL64-NEXT:    shufps $85, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0x55]
 540 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,1,1]
 541 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 542 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x60]
 543 ; FMACALL64-NEXT:    shufps $85, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0x55]
 544 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,1,1]
 545 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
 546 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
 547 ; FMACALL64-NEXT:    shufps $85, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0x55]
 548 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,1,1]
 549 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 550 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 551 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 552 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
 553 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 554 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 555 ; FMACALL64-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
 556 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x4c,0x24,0x70]
 557 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 558 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 559 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x20]
 560 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 561 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 562 ; FMACALL64-NEXT:    shufps $255, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xff]
 563 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,3,3,3]
 564 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 565 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x40]
 566 ; FMACALL64-NEXT:    shufps $255, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xff]
 567 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,3,3,3]
 568 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 569 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x50]
 570 ; FMACALL64-NEXT:    shufps $255, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xff]
 571 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,3,3,3]
 572 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 573 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 574 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 575 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 576 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 577 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 578 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 579 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 580 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 581 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x40]
 582 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 583 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 584 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 585 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x50]
 586 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 587 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 588 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 589 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 590 ; FMACALL64-NEXT:    unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
 591 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x04,0x24]
 592 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 593 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 594 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 595 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 596 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 597 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 598 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x40]
 599 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 600 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x50]
 601 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 602 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 603 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 604 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
 605 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 606 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 607 ; FMACALL64-NEXT:    shufps $85, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0x55]
 608 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,1,1]
 609 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 610 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x40]
 611 ; FMACALL64-NEXT:    shufps $85, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0x55]
 612 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,1,1]
 613 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 614 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x50]
 615 ; FMACALL64-NEXT:    shufps $85, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0x55]
 616 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,1,1]
 617 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 618 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 619 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 620 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 621 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 622 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 623 ; FMACALL64-NEXT:    unpcklpd (%rsp), %xmm1 ## 16-byte Folded Reload
 624 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x0c,0x24]
 625 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 626 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 627 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
 628 ; FMACALL64-NEXT:    addq $136, %rsp ## encoding: [0x48,0x81,0xc4,0x88,0x00,0x00,0x00]
 629 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
 630 ;
 631 ; AVX512-LABEL: test_v8f32:
 632 ; AVX512:       ## %bb.0: ## %entry
 633 ; AVX512-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0x75,0xa8,0xc2]
 634 ; AVX512-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
 635 ; AVX512-NEXT:    retq ## encoding: [0xc3]
 636 ;
 637 ; AVX512VL-LABEL: test_v8f32:
 638 ; AVX512VL:       ## %bb.0: ## %entry
 639 ; AVX512VL-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x75,0xa8,0xc2]
 640 ; AVX512VL-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
 641 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 642 ;
 643 ; FMACALL32_BDVER2-LABEL: test_v8f32:
 644 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
 645 ; FMACALL32_BDVER2-NEXT:    subl $284, %esp ## encoding: [0x81,0xec,0x1c,0x01,0x00,0x00]
 646 ; FMACALL32_BDVER2-NEXT:    ## imm = 0x11C
 647 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm2, %xmm3 ## encoding: [0xc4,0xe3,0x7d,0x19,0xd3,0x01]
 648 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
 649 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x94,0x24,0xf0,0x00,0x00,0x00]
 650 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm1, %xmm2 ## encoding: [0xc4,0xe3,0x7d,0x19,0xca,0x01]
 651 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
 652 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x8c,0x24,0xd0,0x00,0x00,0x00]
 653 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm1 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc1,0x01]
 654 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
 655 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x84,0x24,0xb0,0x00,0x00,0x00]
 656 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 657 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x5c,0x24,0x50]
 658 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm3, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x5c,0x24,0x08,0x02]
 659 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 660 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x54,0x24,0x40]
 661 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x02]
 662 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 663 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x4c,0x24,0x30]
 664 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x02]
 665 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 666 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 667 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 668 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 669 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xa4,0x00,0x00,0x00]
 670 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 671 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x50]
 672 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
 673 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x54,0x24,0x40]
 674 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
 675 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x30]
 676 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
 677 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x01]
 678 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x01]
 679 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 680 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 681 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 682 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x98,0x00,0x00,0x00]
 683 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 684 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x50]
 685 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
 686 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 687 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x40]
 688 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
 689 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 690 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x30]
 691 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
 692 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 693 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 694 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 695 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x8c,0x00,0x00,0x00]
 696 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 697 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xf0,0x00,0x00,0x00]
 698 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
 699 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 700 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xd0,0x00,0x00,0x00]
 701 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x03]
 702 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 703 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xb0,0x00,0x00,0x00]
 704 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x03]
 705 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 706 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 707 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 708 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 709 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x80,0x00,0x00,0x00]
 710 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 711 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xf0,0x00,0x00,0x00]
 712 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x02]
 713 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 714 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xd0,0x00,0x00,0x00]
 715 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x02]
 716 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 717 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xb0,0x00,0x00,0x00]
 718 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x02]
 719 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 720 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 721 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 722 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 723 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x74]
 724 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 725 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xf0,0x00,0x00,0x00]
 726 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
 727 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 728 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xd0,0x00,0x00,0x00]
 729 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x01]
 730 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 731 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xb0,0x00,0x00,0x00]
 732 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x01]
 733 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 734 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 735 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 736 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 737 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x68]
 738 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 739 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xf0,0x00,0x00,0x00]
 740 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
 741 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 742 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xd0,0x00,0x00,0x00]
 743 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
 744 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 745 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xb0,0x00,0x00,0x00]
 746 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
 747 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 748 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 749 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 750 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 751 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x50]
 752 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
 753 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x54,0x24,0x40]
 754 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
 755 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x30]
 756 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
 757 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x03]
 758 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x03]
 759 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x2c]
 760 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 761 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x68]
 762 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x28]
 763 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 764 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x74]
 765 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x24]
 766 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 767 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x80,0x00,0x00,0x00]
 768 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x20]
 769 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 770 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x8c,0x00,0x00,0x00]
 771 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x1c]
 772 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 773 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x98,0x00,0x00,0x00]
 774 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x18]
 775 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 776 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xa4,0x00,0x00,0x00]
 777 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x14]
 778 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 779 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 780 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x10]
 781 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x44,0x24,0x2c]
 782 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 783 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x28,0x10]
 784 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[2,3]
 785 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfa,0x10,0x4c,0x24,0x1c]
 786 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero,zero,zero
 787 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x18,0x10]
 788 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0],mem[0],xmm1[2,3]
 789 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x24,0x20]
 790 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0],xmm0[3]
 791 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x14,0x20]
 792 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0],xmm1[3]
 793 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x20,0x30]
 794 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1,2],mem[0]
 795 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x10,0x30]
 796 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1,2],mem[0]
 797 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 ## encoding: [0xc4,0xe3,0x7d,0x18,0xc1,0x01]
 798 ; FMACALL32_BDVER2-NEXT:    addl $284, %esp ## encoding: [0x81,0xc4,0x1c,0x01,0x00,0x00]
 799 ; FMACALL32_BDVER2-NEXT:    ## imm = 0x11C
 800 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
 801 entry:
 802   %call = call <8 x float> @llvm.fma.v8f32(<8 x float> %a, <8 x float> %b, <8 x float> %c)
 803   ret <8 x float> %call
 804 }
 805
 806 define <16 x float> @test_v16f32(<16 x float> %a, <16 x float> %b, <16 x float> %c) #0 {
 807 ; FMA32-LABEL: test_v16f32:
 808 ; FMA32:       ## %bb.0: ## %entry
 809 ; FMA32-NEXT:    pushl %ebp ## encoding: [0x55]
 810 ; FMA32-NEXT:    movl %esp, %ebp ## encoding: [0x89,0xe5]
 811 ; FMA32-NEXT:    andl $-32, %esp ## encoding: [0x83,0xe4,0xe0]
 812 ; FMA32-NEXT:    subl $32, %esp ## encoding: [0x83,0xec,0x20]
 813 ; FMA32-NEXT:    vfmadd213ps 8(%ebp), %ymm2, %ymm0 ## encoding: [0xc4,0xe2,0x6d,0xa8,0x45,0x08]
 814 ; FMA32-NEXT:    ## ymm0 = (ymm2 * ymm0) + mem
 815 ; FMA32-NEXT:    vfmadd213ps 40(%ebp), %ymm3, %ymm1 ## encoding: [0xc4,0xe2,0x65,0xa8,0x4d,0x28]
 816 ; FMA32-NEXT:    ## ymm1 = (ymm3 * ymm1) + mem
 817 ; FMA32-NEXT:    movl %ebp, %esp ## encoding: [0x89,0xec]
 818 ; FMA32-NEXT:    popl %ebp ## encoding: [0x5d]
 819 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 820 ;
 821 ; FMA64-LABEL: test_v16f32:
 822 ; FMA64:       ## %bb.0: ## %entry
 823 ; FMA64-NEXT:    vfmadd213ps %ymm4, %ymm2, %ymm0 ## encoding: [0xc4,0xe2,0x6d,0xa8,0xc4]
 824 ; FMA64-NEXT:    ## ymm0 = (ymm2 * ymm0) + ymm4
 825 ; FMA64-NEXT:    vfmadd213ps %ymm5, %ymm3, %ymm1 ## encoding: [0xc4,0xe2,0x65,0xa8,0xcd]
 826 ; FMA64-NEXT:    ## ymm1 = (ymm3 * ymm1) + ymm5
 827 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 828 ;
 829 ; FMACALL64-LABEL: test_v16f32:
 830 ; FMACALL64:       ## %bb.0: ## %entry
 831 ; FMACALL64-NEXT:    subq $168, %rsp ## encoding: [0x48,0x81,0xec,0xa8,0x00,0x00,0x00]
 832 ; FMACALL64-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 833 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0xbc,0x24,0x80,0x00,0x00,0x00]
 834 ; FMACALL64-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 835 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x74,0x24,0x20]
 836 ; FMACALL64-NEXT:    movaps %xmm5, (%rsp) ## 16-byte Spill
 837 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x2c,0x24]
 838 ; FMACALL64-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 839 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x64,0x24,0x10]
 840 ; FMACALL64-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 841 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x5c,0x24,0x70]
 842 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 843 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x40]
 844 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 845 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x30]
 846 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 847 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x50]
 848 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
 849 ; FMACALL64-NEXT:    shufps $255, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xff]
 850 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,3,3,3]
 851 ; FMACALL64-NEXT:    movaps %xmm4, %xmm1 ## encoding: [0x0f,0x28,0xcc]
 852 ; FMACALL64-NEXT:    shufps $255, %xmm4, %xmm1 ## encoding: [0x0f,0xc6,0xcc,0xff]
 853 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,3],xmm4[3,3]
 854 ; FMACALL64-NEXT:    shufps $255, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xff]
 855 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,3,3,3]
 856 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 857 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 858 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 859 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x60]
 860 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 861 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
 862 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 863 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 864 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 865 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 866 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 867 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 868 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
 869 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 870 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 871 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 872 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 873 ; FMACALL64-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
 874 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x44,0x24,0x60]
 875 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 876 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 877 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x84,0x24,0x90,0x00,0x00,0x00]
 878 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 879 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
 880 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 881 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 882 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
 883 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 884 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 885 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 886 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x60]
 887 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 888 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
 889 ; FMACALL64-NEXT:    shufps $85, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0x55]
 890 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,1,1]
 891 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 892 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 893 ; FMACALL64-NEXT:    shufps $85, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0x55]
 894 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,1,1]
 895 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
 896 ; FMACALL64-NEXT:    shufps $85, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0x55]
 897 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,1,1]
 898 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 899 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 900 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 901 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x60]
 902 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 903 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 904 ; FMACALL64-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
 905 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x8c,0x24,0x90,0x00,0x00,0x00]
 906 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 907 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 908 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x60]
 909 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 910 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 911 ; FMACALL64-NEXT:    shufps $255, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xff]
 912 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,3,3,3]
 913 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
 914 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
 915 ; FMACALL64-NEXT:    shufps $255, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xff]
 916 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,3,3,3]
 917 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
 918 ; FMACALL64-NEXT:    shufps $255, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xff]
 919 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,3,3,3]
 920 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 921 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 922 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 923 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
 924 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 925 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 926 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 927 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 928 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
 929 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
 930 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 931 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 932 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
 933 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 934 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 935 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 936 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 937 ; FMACALL64-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
 938 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x44,0x24,0x10]
 939 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 940 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 941 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x50]
 942 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 943 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 944 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
 945 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
 946 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
 947 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 948 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 949 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 950 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
 951 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 952 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 953 ; FMACALL64-NEXT:    shufps $85, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0x55]
 954 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,1,1]
 955 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
 956 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
 957 ; FMACALL64-NEXT:    shufps $85, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0x55]
 958 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,1,1]
 959 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
 960 ; FMACALL64-NEXT:    shufps $85, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0x55]
 961 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,1,1]
 962 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 963 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 964 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 965 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 966 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 967 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 968 ; FMACALL64-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
 969 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x4c,0x24,0x50]
 970 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 971 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 972 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x10]
 973 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 974 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
 975 ; FMACALL64-NEXT:    shufps $255, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xff]
 976 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,3,3,3]
 977 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 978 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
 979 ; FMACALL64-NEXT:    shufps $255, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xff]
 980 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,3,3,3]
 981 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
 982 ; FMACALL64-NEXT:    shufps $255, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xff]
 983 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,3,3,3]
 984 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 985 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 986 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 987 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 988 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 989 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
 990 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 991 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 992 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 993 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
 994 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 995 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 996 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
 997 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 998 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 999 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
1000 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
1001 ; FMACALL64-NEXT:    unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
1002 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x04,0x24]
1003 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
1004 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1005 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x30]
1006 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1007 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
1008 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1009 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
1010 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
1011 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
1012 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
1013 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
1014 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
1015 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1016 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
1017 ; FMACALL64-NEXT:    shufps $85, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0x55]
1018 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,1,1]
1019 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1020 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
1021 ; FMACALL64-NEXT:    shufps $85, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0x55]
1022 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,1,1]
1023 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
1024 ; FMACALL64-NEXT:    shufps $85, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0x55]
1025 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,1,1]
1026 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
1027 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
1028 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
1029 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
1030 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
1031 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
1032 ; FMACALL64-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
1033 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x4c,0x24,0x30]
1034 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
1035 ; FMACALL64-NEXT:    movaps %xmm1, (%rsp) ## 16-byte Spill
1036 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x0c,0x24]
1037 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1038 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x70]
1039 ; FMACALL64-NEXT:    shufps $255, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xff]
1040 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,3,3,3]
1041 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1042 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x8c,0x24,0x80,0x00,0x00,0x00]
1043 ; FMACALL64-NEXT:    shufps $255, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xff]
1044 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,3,3,3]
1045 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xe0,0x00,0x00,0x00]
1046 ; FMACALL64-NEXT:    shufps $255, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xff]
1047 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,3,3,3]
1048 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
1049 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
1050 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1051 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
1052 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1053 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x70]
1054 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1055 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1056 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1057 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x8c,0x24,0x80,0x00,0x00,0x00]
1058 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1059 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1060 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xe0,0x00,0x00,0x00]
1061 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1062 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1063 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
1064 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
1065 ; FMACALL64-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
1066 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x44,0x24,0x20]
1067 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
1068 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1069 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
1070 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1071 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x70]
1072 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1073 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x8c,0x24,0x80,0x00,0x00,0x00]
1074 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xe0,0x00,0x00,0x00]
1075 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
1076 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
1077 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1078 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x40]
1079 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1080 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x70]
1081 ; FMACALL64-NEXT:    shufps $85, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0x55]
1082 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,1,1]
1083 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1084 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x8c,0x24,0x80,0x00,0x00,0x00]
1085 ; FMACALL64-NEXT:    shufps $85, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0x55]
1086 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,1,1]
1087 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xe0,0x00,0x00,0x00]
1088 ; FMACALL64-NEXT:    shufps $85, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0x55]
1089 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,1,1]
1090 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
1091 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
1092 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Reload
1093 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x5c,0x24,0x40]
1094 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm3 ## encoding: [0x0f,0x14,0xd8]
1095 ; FMACALL64-NEXT:    ## xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
1096 ; FMACALL64-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Folded Reload
1097 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x5c,0x24,0x20]
1098 ; FMACALL64-NEXT:    ## xmm3 = xmm3[0],mem[0]
1099 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1100 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x60]
1101 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1102 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1103 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
1104 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
1105 ; FMACALL64-NEXT:    addq $168, %rsp ## encoding: [0x48,0x81,0xc4,0xa8,0x00,0x00,0x00]
1106 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
1107 ;
1108 ; AVX512-LABEL: test_v16f32:
1109 ; AVX512:       ## %bb.0: ## %entry
1110 ; AVX512-NEXT:    vfmadd213ps %zmm2, %zmm1, %zmm0 ## encoding: [0x62,0xf2,0x75,0x48,0xa8,0xc2]
1111 ; AVX512-NEXT:    ## zmm0 = (zmm1 * zmm0) + zmm2
1112 ; AVX512-NEXT:    retq ## encoding: [0xc3]
1113 ;
1114 ; AVX512VL-LABEL: test_v16f32:
1115 ; AVX512VL:       ## %bb.0: ## %entry
1116 ; AVX512VL-NEXT:    vfmadd213ps %zmm2, %zmm1, %zmm0 ## encoding: [0x62,0xf2,0x75,0x48,0xa8,0xc2]
1117 ; AVX512VL-NEXT:    ## zmm0 = (zmm1 * zmm0) + zmm2
1118 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
1119 ;
1120 ; FMACALL32_BDVER2-LABEL: test_v16f32:
1121 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
1122 ; FMACALL32_BDVER2-NEXT:    pushl %ebp ## encoding: [0x55]
1123 ; FMACALL32_BDVER2-NEXT:    movl %esp, %ebp ## encoding: [0x89,0xe5]
1124 ; FMACALL32_BDVER2-NEXT:    andl $-32, %esp ## encoding: [0x83,0xe4,0xe0]
1125 ; FMACALL32_BDVER2-NEXT:    subl $448, %esp ## encoding: [0x81,0xec,0xc0,0x01,0x00,0x00]
1126 ; FMACALL32_BDVER2-NEXT:    ## imm = 0x1C0
1127 ; FMACALL32_BDVER2-NEXT:    vmovaps 56(%ebp), %xmm4 ## encoding: [0xc5,0xf8,0x28,0x65,0x38]
1128 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1129 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x94,0x24,0x60,0x01,0x00,0x00]
1130 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm3, %xmm2 ## encoding: [0xc4,0xe3,0x7d,0x19,0xda,0x01]
1131 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1132 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x4c,0x24,0x60]
1133 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm1, %xmm1 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc9,0x01]
1134 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm3, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1135 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x9c,0x24,0x80,0x00,0x00,0x00]
1136 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1137 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x84,0x24,0x80,0x01,0x00,0x00]
1138 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1139 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x94,0x24,0xc0,0x00,0x00,0x00]
1140 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x02]
1141 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1142 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x8c,0x24,0xb0,0x00,0x00,0x00]
1143 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x02]
1144 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm4, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x64,0x24,0x08,0x02]
1145 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1146 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1147 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1148 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1149 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x54,0x01,0x00,0x00]
1150 ; FMACALL32_BDVER2-NEXT:    vmovaps 56(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x38]
1151 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
1152 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
1153 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
1154 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x8c,0x24,0xb0,0x00,0x00,0x00]
1155 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
1156 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x01]
1157 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x01]
1158 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1159 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1160 ; FMACALL32_BDVER2-NEXT:    vmovaps 56(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x38]
1161 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1162 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x48,0x01,0x00,0x00]
1163 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
1164 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1165 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x84,0x24,0xc0,0x00,0x00,0x00]
1166 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
1167 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1168 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x84,0x24,0xb0,0x00,0x00,0x00]
1169 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
1170 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1171 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1172 ; FMACALL32_BDVER2-NEXT:    vmovaps 40(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x28]
1173 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1174 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x3c,0x01,0x00,0x00]
1175 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
1176 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1177 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1178 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x03]
1179 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1180 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x44,0x24,0x60]
1181 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x03]
1182 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1183 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1184 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1185 ; FMACALL32_BDVER2-NEXT:    vmovaps 40(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x28]
1186 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1187 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x30,0x01,0x00,0x00]
1188 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x02]
1189 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1190 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1191 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x02]
1192 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1193 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x44,0x24,0x60]
1194 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x02]
1195 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1196 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1197 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1198 ; FMACALL32_BDVER2-NEXT:    vmovaps 40(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x28]
1199 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1200 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x24,0x01,0x00,0x00]
1201 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
1202 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1203 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1204 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x01]
1205 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1206 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x44,0x24,0x60]
1207 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x01]
1208 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1209 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1210 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1211 ; FMACALL32_BDVER2-NEXT:    vmovaps 40(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x28]
1212 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1213 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x18,0x01,0x00,0x00]
1214 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
1215 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1216 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1217 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
1218 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1219 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x44,0x24,0x60]
1220 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
1221 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1222 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1223 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1224 ; FMACALL32_BDVER2-NEXT:    vmovaps 24(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x18]
1225 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1226 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x0c,0x01,0x00,0x00]
1227 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
1228 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1229 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x60,0x01,0x00,0x00]
1230 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm0 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc0,0x01]
1231 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1232 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x84,0x24,0x80,0x00,0x00,0x00]
1233 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x03]
1234 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1235 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x01,0x00,0x00]
1236 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm0 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc0,0x01]
1237 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1238 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x44,0x24,0x60]
1239 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x03]
1240 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1241 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1242 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1243 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1244 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x00,0x01,0x00,0x00]
1245 ; FMACALL32_BDVER2-NEXT:    vmovaps 24(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x18]
1246 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
1247 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x94,0x24,0x80,0x00,0x00,0x00]
1248 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
1249 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x60]
1250 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x02]
1251 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x02]
1252 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x02]
1253 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1254 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1255 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1256 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xf4,0x00,0x00,0x00]
1257 ; FMACALL32_BDVER2-NEXT:    vmovaps 24(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x18]
1258 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
1259 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x94,0x24,0x80,0x00,0x00,0x00]
1260 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
1261 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x60]
1262 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
1263 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x01]
1264 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x01]
1265 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1266 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1267 ; FMACALL32_BDVER2-NEXT:    vmovaps 24(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x18]
1268 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1269 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xe8,0x00,0x00,0x00]
1270 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
1271 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1272 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1273 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
1274 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1275 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x60]
1276 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
1277 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1278 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1279 ; FMACALL32_BDVER2-NEXT:    vmovaps 8(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x08]
1280 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1281 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x80,0x00,0x00,0x00]
1282 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
1283 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1284 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x60,0x01,0x00,0x00]
1285 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x03]
1286 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1287 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x01,0x00,0x00]
1288 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x03]
1289 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1290 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1291 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1292 ; FMACALL32_BDVER2-NEXT:    vmovaps 8(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x08]
1293 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1294 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x60]
1295 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x02]
1296 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1297 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x60,0x01,0x00,0x00]
1298 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x02]
1299 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1300 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x01,0x00,0x00]
1301 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x02]
1302 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1303 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1304 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1305 ; FMACALL32_BDVER2-NEXT:    vmovaps 8(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x08]
1306 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1307 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xdc,0x00,0x00,0x00]
1308 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
1309 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1310 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x60,0x01,0x00,0x00]
1311 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x01]
1312 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1313 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x01,0x00,0x00]
1314 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x01]
1315 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1316 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1317 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1318 ; FMACALL32_BDVER2-NEXT:    vmovaps 8(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x08]
1319 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1320 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xd0,0x00,0x00,0x00]
1321 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
1322 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1323 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x60,0x01,0x00,0x00]
1324 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
1325 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1326 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x01,0x00,0x00]
1327 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
1328 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1329 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1330 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1331 ; FMACALL32_BDVER2-NEXT:    vmovaps 56(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x38]
1332 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
1333 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
1334 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
1335 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x8c,0x24,0xb0,0x00,0x00,0x00]
1336 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
1337 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x03]
1338 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x03]
1339 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x3c]
1340 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1341 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xd0,0x00,0x00,0x00]
1342 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x38]
1343 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1344 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xdc,0x00,0x00,0x00]
1345 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x34]
1346 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1347 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x60]
1348 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x30]
1349 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1350 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x80,0x00,0x00,0x00]
1351 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x2c]
1352 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1353 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xe8,0x00,0x00,0x00]
1354 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x28]
1355 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1356 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xf4,0x00,0x00,0x00]
1357 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x24]
1358 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1359 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x00,0x01,0x00,0x00]
1360 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x20]
1361 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1362 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x0c,0x01,0x00,0x00]
1363 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x5c]
1364 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1365 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x18,0x01,0x00,0x00]
1366 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x58]
1367 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1368 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x24,0x01,0x00,0x00]
1369 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x54]
1370 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1371 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x30,0x01,0x00,0x00]
1372 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x50]
1373 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1374 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x3c,0x01,0x00,0x00]
1375 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x4c]
1376 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1377 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x48,0x01,0x00,0x00]
1378 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x48]
1379 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1380 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x54,0x01,0x00,0x00]
1381 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x44]
1382 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1383 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1384 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x40]
1385 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x44,0x24,0x3c]
1386 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero,zero,zero
1387 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x38,0x10]
1388 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[2,3]
1389 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfa,0x10,0x4c,0x24,0x2c]
1390 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero,zero,zero
1391 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x28,0x10]
1392 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0],mem[0],xmm1[2,3]
1393 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm2 ## encoding: [0xc5,0xfa,0x10,0x54,0x24,0x4c]
1394 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = mem[0],zero,zero,zero
1395 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm2, %xmm2 ## encoding: [0xc4,0xe3,0x69,0x21,0x54,0x24,0x48,0x10]
1396 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = xmm2[0],mem[0],xmm2[2,3]
1397 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x34,0x20]
1398 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0],xmm0[3]
1399 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x24,0x20]
1400 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0],xmm1[3]
1401 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm2, %xmm2 ## encoding: [0xc4,0xe3,0x69,0x21,0x54,0x24,0x44,0x20]
1402 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = xmm2[0,1],mem[0],xmm2[3]
1403 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x30,0x30]
1404 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1,2],mem[0]
1405 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x20,0x30]
1406 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1,2],mem[0]
1407 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm2, %xmm2 ## encoding: [0xc4,0xe3,0x69,0x21,0x54,0x24,0x40,0x30]
1408 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = xmm2[0,1,2],mem[0]
1409 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 ## encoding: [0xc4,0xe3,0x7d,0x18,0xc1,0x01]
1410 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfa,0x10,0x4c,0x24,0x5c]
1411 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero,zero,zero
1412 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x58,0x10]
1413 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0],mem[0],xmm1[2,3]
1414 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x54,0x20]
1415 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0],xmm1[3]
1416 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x50,0x30]
1417 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1,2],mem[0]
1418 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1 ## encoding: [0xc4,0xe3,0x75,0x18,0xca,0x01]
1419 ; FMACALL32_BDVER2-NEXT:    movl %ebp, %esp ## encoding: [0x89,0xec]
1420 ; FMACALL32_BDVER2-NEXT:    popl %ebp ## encoding: [0x5d]
1421 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
1422 entry:
1423   %call = call <16 x float> @llvm.fma.v16f32(<16 x float> %a, <16 x float> %b, <16 x float> %c)
1424   ret <16 x float> %call
1425 }
1426
1427 define <2 x double> @test_v2f64(<2 x double> %a, <2 x double> %b, <2 x double> %c) #0 {
1428 ; FMA32-LABEL: test_v2f64:
1429 ; FMA32:       ## %bb.0:
1430 ; FMA32-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1431 ; FMA32-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1432 ; FMA32-NEXT:    retl ## encoding: [0xc3]
1433 ;
1434 ; FMA64-LABEL: test_v2f64:
1435 ; FMA64:       ## %bb.0:
1436 ; FMA64-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1437 ; FMA64-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1438 ; FMA64-NEXT:    retq ## encoding: [0xc3]
1439 ;
1440 ; FMACALL64-LABEL: test_v2f64:
1441 ; FMACALL64:       ## %bb.0:
1442 ; FMACALL64-NEXT:    subq $72, %rsp ## encoding: [0x48,0x83,0xec,0x48]
1443 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1444 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x20]
1445 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1446 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x10]
1447 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
1448 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
1449 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1450 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1451 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1452 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x30]
1453 ; FMACALL64-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
1454 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x04,0x24]
1455 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1456 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1457 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1458 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1459 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1460 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1461 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
1462 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x20]
1463 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1464 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1465 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1466 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1467 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1468 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x30]
1469 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1470 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1471 ; FMACALL64-NEXT:    movaps %xmm1, %xmm0 ## encoding: [0x0f,0x28,0xc1]
1472 ; FMACALL64-NEXT:    addq $72, %rsp ## encoding: [0x48,0x83,0xc4,0x48]
1473 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
1474 ;
1475 ; AVX512-LABEL: test_v2f64:
1476 ; AVX512:       ## %bb.0:
1477 ; AVX512-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1478 ; AVX512-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1479 ; AVX512-NEXT:    retq ## encoding: [0xc3]
1480 ;
1481 ; AVX512VL-LABEL: test_v2f64:
1482 ; AVX512VL:       ## %bb.0:
1483 ; AVX512VL-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1484 ; AVX512VL-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1485 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
1486 ;
1487 ; FMACALL32_BDVER2-LABEL: test_v2f64:
1488 ; FMACALL32_BDVER2:       ## %bb.0:
1489 ; FMACALL32_BDVER2-NEXT:    subl $108, %esp ## encoding: [0x83,0xec,0x6c]
1490 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1491 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x44,0x24,0x50]
1492 ; FMACALL32_BDVER2-NEXT:    vmovlhps %xmm1, %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0xc1]
1493 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],xmm1[0]
1494 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1495 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x54,0x24,0x30]
1496 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1497 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x4c,0x24,0x40]
1498 ; FMACALL32_BDVER2-NEXT:    vmovlps %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x13,0x54,0x24,0x10]
1499 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1500 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1501 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1502 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1503 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x30]
1504 ; FMACALL32_BDVER2-NEXT:    vmovhps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x17,0x44,0x24,0x10]
1505 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1506 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x40]
1507 ; FMACALL32_BDVER2-NEXT:    vmovlps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1508 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x12,0x44,0x24,0x58]
1509 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0,1],xmm0[2,3]
1510 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1511 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x28]
1512 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1513 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1514 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x20]
1515 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x44,0x24,0x28]
1516 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
1517 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0x44,0x24,0x20]
1518 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0,1]
1519 ; FMACALL32_BDVER2-NEXT:    addl $108, %esp ## encoding: [0x83,0xc4,0x6c]
1520 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
1521   %call = call <2 x double> @llvm.fma.v2f64(<2 x double> %a, <2 x double> %b, <2 x double> %c)
1522   ret <2 x double> %call
1523 }
1524
1525 define <2 x double> @test_v2f64_reassoc(<2 x double> %a, <2 x double> %b, <2 x double> %c) #0 {
1526 ; FMA32-LABEL: test_v2f64_reassoc:
1527 ; FMA32:       ## %bb.0:
1528 ; FMA32-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1529 ; FMA32-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1530 ; FMA32-NEXT:    retl ## encoding: [0xc3]
1531 ;
1532 ; FMACALL32-LABEL: test_v2f64_reassoc:
1533 ; FMACALL32:       ## %bb.0:
1534 ; FMACALL32-NEXT:    vmulpd %xmm1, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0x59,0xc1]
1535 ; FMACALL32-NEXT:    vaddpd %xmm2, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0x58,0xc2]
1536 ; FMACALL32-NEXT:    retl ## encoding: [0xc3]
1537 ;
1538 ; FMA64-LABEL: test_v2f64_reassoc:
1539 ; FMA64:       ## %bb.0:
1540 ; FMA64-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1541 ; FMA64-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1542 ; FMA64-NEXT:    retq ## encoding: [0xc3]
1543 ;
1544 ; FMACALL64-LABEL: test_v2f64_reassoc:
1545 ; FMACALL64:       ## %bb.0:
1546 ; FMACALL64-NEXT:    mulpd %xmm1, %xmm0 ## encoding: [0x66,0x0f,0x59,0xc1]
1547 ; FMACALL64-NEXT:    addpd %xmm2, %xmm0 ## encoding: [0x66,0x0f,0x58,0xc2]
1548 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
1549 ;
1550 ; AVX512-LABEL: test_v2f64_reassoc:
1551 ; AVX512:       ## %bb.0:
1552 ; AVX512-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1553 ; AVX512-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1554 ; AVX512-NEXT:    retq ## encoding: [0xc3]
1555 ;
1556 ; AVX512VL-LABEL: test_v2f64_reassoc:
1557 ; AVX512VL:       ## %bb.0:
1558 ; AVX512VL-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1559 ; AVX512VL-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1560 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
1561   %call = call reassoc <2 x double> @llvm.fma.v2f64(<2 x double> %a, <2 x double> %b, <2 x double> %c)
1562   ret <2 x double> %call
1563 }
1564
1565 define <4 x double> @test_v4f64(<4 x double> %a, <4 x double> %b, <4 x double> %c) #0 {
1566 ; FMA32-LABEL: test_v4f64:
1567 ; FMA32:       ## %bb.0: ## %entry
1568 ; FMA32-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0xf5,0xa8,0xc2]
1569 ; FMA32-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
1570 ; FMA32-NEXT:    retl ## encoding: [0xc3]
1571 ;
1572 ; FMA64-LABEL: test_v4f64:
1573 ; FMA64:       ## %bb.0: ## %entry
1574 ; FMA64-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0xf5,0xa8,0xc2]
1575 ; FMA64-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
1576 ; FMA64-NEXT:    retq ## encoding: [0xc3]
1577 ;
1578 ; FMACALL64-LABEL: test_v4f64:
1579 ; FMACALL64:       ## %bb.0: ## %entry
1580 ; FMACALL64-NEXT:    subq $120, %rsp ## encoding: [0x48,0x83,0xec,0x78]
1581 ; FMACALL64-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1582 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x6c,0x24,0x40]
1583 ; FMACALL64-NEXT:    movaps %xmm4, (%rsp) ## 16-byte Spill
1584 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x24,0x24]
1585 ; FMACALL64-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1586 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x5c,0x24,0x30]
1587 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1588 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x60]
1589 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1590 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x20]
1591 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1592 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x50]
1593 ; FMACALL64-NEXT:    movaps %xmm2, %xmm1 ## encoding: [0x0f,0x28,0xca]
1594 ; FMACALL64-NEXT:    movaps %xmm4, %xmm2 ## encoding: [0x0f,0x28,0xd4]
1595 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1596 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1597 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1598 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
1599 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1600 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
1601 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1602 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1603 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1604 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x60]
1605 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1606 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1607 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
1608 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
1609 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1610 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1611 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1612 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1613 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1614 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1615 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1616 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1617 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1618 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x10]
1619 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1620 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
1621 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1622 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x30]
1623 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
1624 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x40]
1625 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1626 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1627 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
1628 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
1629 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1630 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
1631 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1632 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1633 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1634 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x30]
1635 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1636 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1637 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
1638 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x40]
1639 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1640 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1641 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1642 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1643 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
1644 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
1645 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1646 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1647 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1648 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x10]
1649 ; FMACALL64-NEXT:    addq $120, %rsp ## encoding: [0x48,0x83,0xc4,0x78]
1650 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
1651 ;
1652 ; AVX512-LABEL: test_v4f64:
1653 ; AVX512:       ## %bb.0: ## %entry
1654 ; AVX512-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0xf5,0xa8,0xc2]
1655 ; AVX512-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
1656 ; AVX512-NEXT:    retq ## encoding: [0xc3]
1657 ;
1658 ; AVX512VL-LABEL: test_v4f64:
1659 ; AVX512VL:       ## %bb.0: ## %entry
1660 ; AVX512VL-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf5,0xa8,0xc2]
1661 ; AVX512VL-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
1662 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
1663 ;
1664 ; FMACALL32_BDVER2-LABEL: test_v4f64:
1665 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
1666 ; FMACALL32_BDVER2-NEXT:    subl $236, %esp ## encoding: [0x81,0xec,0xec,0x00,0x00,0x00]
1667 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm2, %xmm3 ## encoding: [0xc4,0xe3,0x7d,0x19,0xd3,0x01]
1668 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1669 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x94,0x24,0xc0,0x00,0x00,0x00]
1670 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1671 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x84,0x24,0x80,0x00,0x00,0x00]
1672 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm1, %xmm2 ## encoding: [0xc4,0xe3,0x7d,0x19,0xca,0x01]
1673 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm0 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc0,0x01]
1674 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1675 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x8c,0x24,0xa0,0x00,0x00,0x00]
1676 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1677 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x44,0x24,0x60]
1678 ; FMACALL32_BDVER2-NEXT:    vmovlhps %xmm2, %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0xc2]
1679 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],xmm2[0]
1680 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1681 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x5c,0x24,0x70]
1682 ; FMACALL32_BDVER2-NEXT:    vmovlps %xmm3, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x13,0x5c,0x24,0x10]
1683 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1684 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x54,0x24,0x50]
1685 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1686 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1687 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1688 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1689 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1690 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x44]
1691 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1692 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xc0,0x00,0x00,0x00]
1693 ; FMACALL32_BDVER2-NEXT:    vmovhps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x17,0x44,0x24,0x10]
1694 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1695 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0x80,0x00,0x00,0x00]
1696 ; FMACALL32_BDVER2-NEXT:    vunpckhpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1697 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x15,0x84,0x24,0xa0,0x00,0x00,0x00]
1698 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[1],mem[1]
1699 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1700 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1701 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1702 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1703 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1704 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x38]
1705 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1706 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xc0,0x00,0x00,0x00]
1707 ; FMACALL32_BDVER2-NEXT:    vmovlps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x13,0x44,0x24,0x10]
1708 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1709 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0x80,0x00,0x00,0x00]
1710 ; FMACALL32_BDVER2-NEXT:    vunpcklpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1711 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x14,0x84,0x24,0xa0,0x00,0x00,0x00]
1712 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0]
1713 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1714 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1715 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1716 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1717 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1718 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x70]
1719 ; FMACALL32_BDVER2-NEXT:    vmovhps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x17,0x44,0x24,0x10]
1720 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1721 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x50]
1722 ; FMACALL32_BDVER2-NEXT:    vmovlps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1723 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x12,0x44,0x24,0x68]
1724 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0,1],xmm0[2,3]
1725 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1726 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x30]
1727 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1728 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x38]
1729 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x28]
1730 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1731 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x44]
1732 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x20]
1733 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1734 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1735 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x18]
1736 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x44,0x24,0x30]
1737 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
1738 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfb,0x10,0x4c,0x24,0x20]
1739 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero
1740 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0x44,0x24,0x28]
1741 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0,1]
1742 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc5,0xf0,0x16,0x4c,0x24,0x18]
1743 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0,1]
1744 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 ## encoding: [0xc4,0xe3,0x7d,0x18,0xc1,0x01]
1745 ; FMACALL32_BDVER2-NEXT:    addl $236, %esp ## encoding: [0x81,0xc4,0xec,0x00,0x00,0x00]
1746 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
1747 entry:
1748   %call = call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> %b, <4 x double> %c)
1749   ret <4 x double> %call
1750 }
1751
1752 define <8 x double> @test_v8f64(<8 x double> %a, <8 x double> %b, <8 x double> %c) #0 {
1753 ; FMA32-LABEL: test_v8f64:
1754 ; FMA32:       ## %bb.0: ## %entry
1755 ; FMA32-NEXT:    pushl %ebp ## encoding: [0x55]
1756 ; FMA32-NEXT:    movl %esp, %ebp ## encoding: [0x89,0xe5]
1757 ; FMA32-NEXT:    andl $-32, %esp ## encoding: [0x83,0xe4,0xe0]
1758 ; FMA32-NEXT:    subl $32, %esp ## encoding: [0x83,0xec,0x20]
1759 ; FMA32-NEXT:    vfmadd213pd 8(%ebp), %ymm2, %ymm0 ## encoding: [0xc4,0xe2,0xed,0xa8,0x45,0x08]
1760 ; FMA32-NEXT:    ## ymm0 = (ymm2 * ymm0) + mem
1761 ; FMA32-NEXT:    vfmadd213pd 40(%ebp), %ymm3, %ymm1 ## encoding: [0xc4,0xe2,0xe5,0xa8,0x4d,0x28]
1762 ; FMA32-NEXT:    ## ymm1 = (ymm3 * ymm1) + mem
1763 ; FMA32-NEXT:    movl %ebp, %esp ## encoding: [0x89,0xec]
1764 ; FMA32-NEXT:    popl %ebp ## encoding: [0x5d]
1765 ; FMA32-NEXT:    retl ## encoding: [0xc3]
1766 ;
1767 ; FMA64-LABEL: test_v8f64:
1768 ; FMA64:       ## %bb.0: ## %entry
1769 ; FMA64-NEXT:    vfmadd213pd %ymm4, %ymm2, %ymm0 ## encoding: [0xc4,0xe2,0xed,0xa8,0xc4]
1770 ; FMA64-NEXT:    ## ymm0 = (ymm2 * ymm0) + ymm4
1771 ; FMA64-NEXT:    vfmadd213pd %ymm5, %ymm3, %ymm1 ## encoding: [0xc4,0xe2,0xe5,0xa8,0xcd]
1772 ; FMA64-NEXT:    ## ymm1 = (ymm3 * ymm1) + ymm5
1773 ; FMA64-NEXT:    retq ## encoding: [0xc3]
1774 ;
1775 ; FMACALL64-LABEL: test_v8f64:
1776 ; FMACALL64:       ## %bb.0: ## %entry
1777 ; FMACALL64-NEXT:    subq $152, %rsp ## encoding: [0x48,0x81,0xec,0x98,0x00,0x00,0x00]
1778 ; FMACALL64-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1779 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x7c,0x24,0x70]
1780 ; FMACALL64-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1781 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x74,0x24,0x20]
1782 ; FMACALL64-NEXT:    movaps %xmm5, (%rsp) ## 16-byte Spill
1783 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x2c,0x24]
1784 ; FMACALL64-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1785 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x64,0x24,0x10]
1786 ; FMACALL64-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1787 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x5c,0x24,0x60]
1788 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1789 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x50]
1790 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1791 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x40]
1792 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1793 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x84,0x24,0x80,0x00,0x00,0x00]
1794 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xa0,0x00,0x00,0x00]
1795 ; FMACALL64-NEXT:    movaps %xmm4, %xmm1 ## encoding: [0x0f,0x28,0xcc]
1796 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1797 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1798 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1799 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x30]
1800 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1801 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1802 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1803 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1804 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1805 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1806 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1807 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1808 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xa0,0x00,0x00,0x00]
1809 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1810 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1811 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1812 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1813 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1814 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x30]
1815 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1816 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1817 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1818 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x30]
1819 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1820 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
1821 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
1822 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
1823 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
1824 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1825 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1826 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1827 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
1828 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1829 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
1830 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1831 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1832 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
1833 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
1834 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1835 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1836 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
1837 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1838 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1839 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1840 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1841 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1842 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1843 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1844 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1845 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1846 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x10]
1847 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1848 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
1849 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1850 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
1851 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
1852 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1853 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1854 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
1855 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
1856 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1857 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
1858 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1859 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1860 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1861 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
1862 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1863 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1864 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
1865 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1866 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1867 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1868 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1869 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
1870 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
1871 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1872 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1873 ; FMACALL64-NEXT:    movaps %xmm1, (%rsp) ## 16-byte Spill
1874 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x0c,0x24]
1875 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1876 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x60]
1877 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1878 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x70]
1879 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
1880 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1881 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1882 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1883 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
1884 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1885 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x60]
1886 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1887 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1888 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1889 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x70]
1890 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1891 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1892 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
1893 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1894 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1895 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1896 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1897 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Reload
1898 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x5c,0x24,0x20]
1899 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm3 ## encoding: [0x0f,0x16,0xd8]
1900 ; FMACALL64-NEXT:    ## xmm3 = xmm3[0],xmm0[0]
1901 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1902 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
1903 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1904 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1905 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
1906 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
1907 ; FMACALL64-NEXT:    addq $152, %rsp ## encoding: [0x48,0x81,0xc4,0x98,0x00,0x00,0x00]
1908 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
1909 ;
1910 ; AVX512-LABEL: test_v8f64:
1911 ; AVX512:       ## %bb.0: ## %entry
1912 ; AVX512-NEXT:    vfmadd213pd %zmm2, %zmm1, %zmm0 ## encoding: [0x62,0xf2,0xf5,0x48,0xa8,0xc2]
1913 ; AVX512-NEXT:    ## zmm0 = (zmm1 * zmm0) + zmm2
1914 ; AVX512-NEXT:    retq ## encoding: [0xc3]
1915 ;
1916 ; AVX512VL-LABEL: test_v8f64:
1917 ; AVX512VL:       ## %bb.0: ## %entry
1918 ; AVX512VL-NEXT:    vfmadd213pd %zmm2, %zmm1, %zmm0 ## encoding: [0x62,0xf2,0xf5,0x48,0xa8,0xc2]
1919 ; AVX512VL-NEXT:    ## zmm0 = (zmm1 * zmm0) + zmm2
1920 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
1921 ;
1922 ; FMACALL32_BDVER2-LABEL: test_v8f64:
1923 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
1924 ; FMACALL32_BDVER2-NEXT:    pushl %ebp ## encoding: [0x55]
1925 ; FMACALL32_BDVER2-NEXT:    movl %esp, %ebp ## encoding: [0x89,0xe5]
1926 ; FMACALL32_BDVER2-NEXT:    andl $-32, %esp ## encoding: [0x83,0xe4,0xe0]
1927 ; FMACALL32_BDVER2-NEXT:    subl $352, %esp ## encoding: [0x81,0xec,0x60,0x01,0x00,0x00]
1928 ; FMACALL32_BDVER2-NEXT:    ## imm = 0x160
1929 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1930 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x84,0x24,0xe0,0x00,0x00,0x00]
1931 ; FMACALL32_BDVER2-NEXT:    vmovsd 56(%ebp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x45,0x38]
1932 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
1933 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1934 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x94,0x24,0x00,0x01,0x00,0x00]
1935 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm3, %xmm2 ## encoding: [0xc4,0xe3,0x7d,0x19,0xda,0x01]
1936 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm3, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1937 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x9c,0x24,0xc0,0x00,0x00,0x00]
1938 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1939 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x8c,0x24,0xa0,0x00,0x00,0x00]
1940 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1941 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x94,0x24,0x30,0x01,0x00,0x00]
1942 ; FMACALL32_BDVER2-NEXT:    vmovsd %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfb,0x11,0x44,0x24,0x10]
1943 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm1, %xmm0 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc8,0x01]
1944 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1945 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x84,0x24,0x40,0x01,0x00,0x00]
1946 ; FMACALL32_BDVER2-NEXT:    vmovlhps %xmm2, %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0xc2]
1947 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],xmm2[0]
1948 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1949 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1950 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1951 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1952 ; FMACALL32_BDVER2-NEXT:    vmovsd 48(%ebp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x45,0x30]
1953 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
1954 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1955 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x94,0x00,0x00,0x00]
1956 ; FMACALL32_BDVER2-NEXT:    vmovsd %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfb,0x11,0x44,0x24,0x10]
1957 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1958 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0xa0,0x00,0x00,0x00]
1959 ; FMACALL32_BDVER2-NEXT:    vunpckhpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1960 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x15,0x84,0x24,0xc0,0x00,0x00,0x00]
1961 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[1],mem[1]
1962 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1963 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1964 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1965 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1966 ; FMACALL32_BDVER2-NEXT:    vmovsd 40(%ebp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x45,0x28]
1967 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
1968 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1969 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x88,0x00,0x00,0x00]
1970 ; FMACALL32_BDVER2-NEXT:    vmovsd %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfb,0x11,0x44,0x24,0x10]
1971 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1972 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0xa0,0x00,0x00,0x00]
1973 ; FMACALL32_BDVER2-NEXT:    vunpcklpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1974 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x14,0x84,0x24,0xc0,0x00,0x00,0x00]
1975 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0]
1976 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1977 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1978 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1979 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1980 ; FMACALL32_BDVER2-NEXT:    vmovsd 32(%ebp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x45,0x20]
1981 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
1982 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1983 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xc0,0x00,0x00,0x00]
1984 ; FMACALL32_BDVER2-NEXT:    vmovsd %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfb,0x11,0x44,0x24,0x10]
1985 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1986 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x00,0x01,0x00,0x00]
1987 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm1 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc1,0x01]
1988 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1989 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0xe0,0x00,0x00,0x00]
1990 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1991 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x4c,0x24,0x20]
1992 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm0 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc0,0x01]
1993 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1994 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x44,0x24,0x30]
1995 ; FMACALL32_BDVER2-NEXT:    vunpckhpd %xmm1, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0x15,0xc1]
1996 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[1],xmm1[1]
1997 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1998 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1999 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
2000 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
2001 ; FMACALL32_BDVER2-NEXT:    vmovsd 24(%ebp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x45,0x18]
2002 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
2003 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
2004 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xa0,0x00,0x00,0x00]
2005 ; FMACALL32_BDVER2-NEXT:    vmovsd %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfb,0x11,0x44,0x24,0x10]
2006 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
2007 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x30]
2008 ; FMACALL32_BDVER2-NEXT:    vunpcklpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
2009 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x14,0x44,0x24,0x20]
2010 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0]
2011 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
2012 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
2013 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
2014 ; FMACALL32_BDVER2-NEXT:    vmovsd 16(%ebp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x45,0x10]
2015 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
2016 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
2017 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x30]
2018 ; FMACALL32_BDVER2-NEXT:    vmovsd %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfb,0x11,0x44,0x24,0x10]
2019 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
2020 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0xe0,0x00,0x00,0x00]
2021 ; FMACALL32_BDVER2-NEXT:    vunpckhpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
2022 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x15,0x84,0x24,0x00,0x01,0x00,0x00]
2023 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[1],mem[1]
2024 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
2025 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
2026 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
2027 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
2028 ; FMACALL32_BDVER2-NEXT:    vmovsd 8(%ebp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x45,0x08]
2029 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
2030 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
2031 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x20]
2032 ; FMACALL32_BDVER2-NEXT:    vmovsd %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfb,0x11,0x44,0x24,0x10]
2033 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
2034 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0xe0,0x00,0x00,0x00]
2035 ; FMACALL32_BDVER2-NEXT:    vunpcklpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
2036 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x14,0x84,0x24,0x00,0x01,0x00,0x00]
2037 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0]
2038 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
2039 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
2040 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
2041 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
2042 ; FMACALL32_BDVER2-NEXT:    vmovsd 64(%ebp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x45,0x40]
2043 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
2044 ; FMACALL32_BDVER2-NEXT:    vmovsd %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfb,0x11,0x44,0x24,0x10]
2045 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
2046 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x84,0x24,0x30,0x01,0x00,0x00]
2047 ; FMACALL32_BDVER2-NEXT:    vmovlps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
2048 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x12,0x84,0x24,0x48,0x01,0x00,0x00]
2049 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0,1],xmm0[2,3]
2050 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
2051 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x60]
2052 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
2053 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x20]
2054 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x58]
2055 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
2056 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x30]
2057 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x50]
2058 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
2059 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xa0,0x00,0x00,0x00]
2060 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x48]
2061 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
2062 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xc0,0x00,0x00,0x00]
2063 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x9c,0x24,0x80,0x00,0x00,0x00]
2064 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
2065 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x88,0x00,0x00,0x00]
2066 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x78]
2067 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
2068 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x94,0x00,0x00,0x00]
2069 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x70]
2070 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
2071 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
2072 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x68]
2073 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x44,0x24,0x60]
2074 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
2075 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfb,0x10,0x4c,0x24,0x50]
2076 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero
2077 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0x44,0x24,0x58]
2078 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0,1]
2079 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc5,0xf0,0x16,0x4c,0x24,0x48]
2080 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0,1]
2081 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm2 ## encoding: [0xc5,0xfb,0x10,0x54,0x24,0x70]
2082 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = mem[0],zero
2083 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm2, %xmm2 ## encoding: [0xc5,0xe8,0x16,0x54,0x24,0x68]
2084 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = xmm2[0,1],mem[0,1]
2085 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 ## encoding: [0xc4,0xe3,0x7d,0x18,0xc1,0x01]
2086 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfb,0x10,0x8c,0x24,0x80,0x00,0x00,0x00]
2087 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero
2088 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc5,0xf0,0x16,0x4c,0x24,0x78]
2089 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0,1]
2090 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1 ## encoding: [0xc4,0xe3,0x75,0x18,0xca,0x01]
2091 ; FMACALL32_BDVER2-NEXT:    movl %ebp, %esp ## encoding: [0x89,0xec]
2092 ; FMACALL32_BDVER2-NEXT:    popl %ebp ## encoding: [0x5d]
2093 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
2094 entry:
2095   %call = call <8 x double> @llvm.fma.v8f64(<8 x double> %a, <8 x double> %b, <8 x double> %c)
2096   ret <8 x double> %call
2097 }
2098
2099 declare float @llvm.fma.f32(float, float, float)
2100 declare double @llvm.fma.f64(double, double, double)
2101 declare x86_fp80 @llvm.fma.f80(x86_fp80, x86_fp80, x86_fp80)
2102
2103 declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>)
2104 declare <8 x float> @llvm.fma.v8f32(<8 x float>, <8 x float>, <8 x float>)
2105 declare <16 x float> @llvm.fma.v16f32(<16 x float>, <16 x float>, <16 x float>)
2106
2107 declare <2 x double> @llvm.fma.v2f64(<2 x double>, <2 x double>, <2 x double>)
2108 declare <4 x double> @llvm.fma.v4f64(<4 x double>, <4 x double>, <4 x double>)
2109 declare <8 x double> @llvm.fma.v8f64(<8 x double>, <8 x double>, <8 x double>)
2110
2111 attributes #0 = { nounwind }