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 --check-prefix=AVX51264
   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-prefixes=FMA32,FMA32_BDVER2
   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: ## %entry
  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: ## %entry
  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: ## %entry
  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: ## %entry
  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: ## %entry
  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: ## %entry
  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 entry:
  56   %call = call float @llvm.fma.f32(float %a, float %b, float %c)
  57   ret float %call
  58 }
  59
  60 define double @test_f64(double %a, double %b, double %c) #0 {
  61 ; FMA32-LABEL: test_f64:
  62 ; FMA32:       ## %bb.0: ## %entry
  63 ; FMA32-NEXT:    subl $12, %esp ## encoding: [0x83,0xec,0x0c]
  64 ; FMA32-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x44,0x24,0x10]
  65 ; FMA32-NEXT:    ## xmm0 = mem[0],zero
  66 ; FMA32-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfb,0x10,0x4c,0x24,0x18]
  67 ; FMA32-NEXT:    ## xmm1 = mem[0],zero
  68 ; FMA32-NEXT:    vfmadd213sd {{[0-9]+}}(%esp), %xmm0, %xmm1 ## encoding: [0xc4,0xe2,0xf9,0xa9,0x4c,0x24,0x20]
  69 ; FMA32-NEXT:    ## xmm1 = (xmm0 * xmm1) + mem
  70 ; FMA32-NEXT:    vmovsd %xmm1, (%esp) ## encoding: [0xc5,0xfb,0x11,0x0c,0x24]
  71 ; FMA32-NEXT:    fldl (%esp) ## encoding: [0xdd,0x04,0x24]
  72 ; FMA32-NEXT:    addl $12, %esp ## encoding: [0x83,0xc4,0x0c]
  73 ; FMA32-NEXT:    retl ## encoding: [0xc3]
  74 ;
  75 ; FMACALL32-LABEL: test_f64:
  76 ; FMACALL32:       ## %bb.0: ## %entry
  77 ; FMACALL32-NEXT:    jmp _fma ## TAILCALL
  78 ; FMACALL32-NEXT:    ## encoding: [0xeb,A]
  79 ; FMACALL32-NEXT:    ## fixup A - offset: 1, value: _fma-1, kind: FK_PCRel_1
  80 ;
  81 ; FMA64-LABEL: test_f64:
  82 ; FMA64:       ## %bb.0: ## %entry
  83 ; FMA64-NEXT:    vfmadd213sd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa9,0xc2]
  84 ; FMA64-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
  85 ; FMA64-NEXT:    retq ## encoding: [0xc3]
  86 ;
  87 ; FMACALL64-LABEL: test_f64:
  88 ; FMACALL64:       ## %bb.0: ## %entry
  89 ; FMACALL64-NEXT:    jmp _fma ## TAILCALL
  90 ; FMACALL64-NEXT:    ## encoding: [0xeb,A]
  91 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-1, kind: FK_PCRel_1
  92 ;
  93 ; AVX512-LABEL: test_f64:
  94 ; AVX512:       ## %bb.0: ## %entry
  95 ; AVX512-NEXT:    vfmadd213sd %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa9,0xc2]
  96 ; AVX512-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
  97 ; AVX512-NEXT:    retq ## encoding: [0xc3]
  98 ;
  99 ; AVX512VL-LABEL: test_f64:
 100 ; AVX512VL:       ## %bb.0: ## %entry
 101 ; AVX512VL-NEXT:    vfmadd213sd %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa9,0xc2]
 102 ; AVX512VL-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 103 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 104 entry:
 105   %call = call double @llvm.fma.f64(double %a, double %b, double %c)
 106   ret double %call
 107 }
 108
 109 define x86_fp80 @test_f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c) #0 {
 110 ; FMA32-LABEL: test_f80:
 111 ; FMA32:       ## %bb.0: ## %entry
 112 ; FMA32-NEXT:    subl $60, %esp ## encoding: [0x83,0xec,0x3c]
 113 ; FMA32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x40]
 114 ; FMA32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x50]
 115 ; FMA32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x60]
 116 ; FMA32-NEXT:    fstpt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x7c,0x24,0x20]
 117 ; FMA32-NEXT:    fstpt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x7c,0x24,0x10]
 118 ; FMA32-NEXT:    fstpt (%esp) ## encoding: [0xdb,0x3c,0x24]
 119 ; FMA32-NEXT:    calll _fmal ## encoding: [0xe8,A,A,A,A]
 120 ; FMA32-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: FK_PCRel_4
 121 ; FMA32-NEXT:    addl $60, %esp ## encoding: [0x83,0xc4,0x3c]
 122 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 123 ;
 124 ; FMACALL32-LABEL: test_f80:
 125 ; FMACALL32:       ## %bb.0: ## %entry
 126 ; FMACALL32-NEXT:    subl $60, %esp ## encoding: [0x83,0xec,0x3c]
 127 ; FMACALL32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x40]
 128 ; FMACALL32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x50]
 129 ; FMACALL32-NEXT:    fldt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x6c,0x24,0x60]
 130 ; FMACALL32-NEXT:    fstpt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x7c,0x24,0x20]
 131 ; FMACALL32-NEXT:    fstpt {{[0-9]+}}(%esp) ## encoding: [0xdb,0x7c,0x24,0x10]
 132 ; FMACALL32-NEXT:    fstpt (%esp) ## encoding: [0xdb,0x3c,0x24]
 133 ; FMACALL32-NEXT:    calll _fmal ## encoding: [0xe8,A,A,A,A]
 134 ; FMACALL32-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: FK_PCRel_4
 135 ; FMACALL32-NEXT:    addl $60, %esp ## encoding: [0x83,0xc4,0x3c]
 136 ; FMACALL32-NEXT:    retl ## encoding: [0xc3]
 137 ;
 138 ; FMA64-LABEL: test_f80:
 139 ; FMA64:       ## %bb.0: ## %entry
 140 ; FMA64-NEXT:    subq $56, %rsp ## encoding: [0x48,0x83,0xec,0x38]
 141 ; FMA64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x40]
 142 ; FMA64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x50]
 143 ; FMA64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x60]
 144 ; FMA64-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x20]
 145 ; FMA64-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x10]
 146 ; FMA64-NEXT:    fstpt (%rsp) ## encoding: [0xdb,0x3c,0x24]
 147 ; FMA64-NEXT:    callq _fmal ## encoding: [0xe8,A,A,A,A]
 148 ; FMA64-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: reloc_branch_4byte_pcrel
 149 ; FMA64-NEXT:    addq $56, %rsp ## encoding: [0x48,0x83,0xc4,0x38]
 150 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 151 ;
 152 ; FMACALL64-LABEL: test_f80:
 153 ; FMACALL64:       ## %bb.0: ## %entry
 154 ; FMACALL64-NEXT:    subq $56, %rsp ## encoding: [0x48,0x83,0xec,0x38]
 155 ; FMACALL64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x40]
 156 ; FMACALL64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x50]
 157 ; FMACALL64-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x60]
 158 ; FMACALL64-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x20]
 159 ; FMACALL64-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x10]
 160 ; FMACALL64-NEXT:    fstpt (%rsp) ## encoding: [0xdb,0x3c,0x24]
 161 ; FMACALL64-NEXT:    callq _fmal ## encoding: [0xe8,A,A,A,A]
 162 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: reloc_branch_4byte_pcrel
 163 ; FMACALL64-NEXT:    addq $56, %rsp ## encoding: [0x48,0x83,0xc4,0x38]
 164 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
 165 ;
 166 ; AVX512-LABEL: test_f80:
 167 ; AVX512:       ## %bb.0: ## %entry
 168 ; AVX512-NEXT:    subq $56, %rsp ## encoding: [0x48,0x83,0xec,0x38]
 169 ; AVX512-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x40]
 170 ; AVX512-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x50]
 171 ; AVX512-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x60]
 172 ; AVX512-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x20]
 173 ; AVX512-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x10]
 174 ; AVX512-NEXT:    fstpt (%rsp) ## encoding: [0xdb,0x3c,0x24]
 175 ; AVX512-NEXT:    callq _fmal ## encoding: [0xe8,A,A,A,A]
 176 ; AVX512-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: reloc_branch_4byte_pcrel
 177 ; AVX512-NEXT:    addq $56, %rsp ## encoding: [0x48,0x83,0xc4,0x38]
 178 ; AVX512-NEXT:    retq ## encoding: [0xc3]
 179 ;
 180 ; AVX512VL-LABEL: test_f80:
 181 ; AVX512VL:       ## %bb.0: ## %entry
 182 ; AVX512VL-NEXT:    subq $56, %rsp ## encoding: [0x48,0x83,0xec,0x38]
 183 ; AVX512VL-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x40]
 184 ; AVX512VL-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x50]
 185 ; AVX512VL-NEXT:    fldt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x6c,0x24,0x60]
 186 ; AVX512VL-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x20]
 187 ; AVX512VL-NEXT:    fstpt {{[0-9]+}}(%rsp) ## encoding: [0xdb,0x7c,0x24,0x10]
 188 ; AVX512VL-NEXT:    fstpt (%rsp) ## encoding: [0xdb,0x3c,0x24]
 189 ; AVX512VL-NEXT:    callq _fmal ## encoding: [0xe8,A,A,A,A]
 190 ; AVX512VL-NEXT:    ## fixup A - offset: 1, value: _fmal-4, kind: reloc_branch_4byte_pcrel
 191 ; AVX512VL-NEXT:    addq $56, %rsp ## encoding: [0x48,0x83,0xc4,0x38]
 192 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 193 entry:
 194   %call = call x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
 195   ret x86_fp80 %call
 196 }
 197
 198 define float @test_f32_cst() #0 {
 199 ; FMA32-LABEL: test_f32_cst:
 200 ; FMA32:       ## %bb.0: ## %entry
 201 ; FMA32-NEXT:    flds LCPI3_0 ## encoding: [0xd9,0x05,A,A,A,A]
 202 ; FMA32-NEXT:    ## fixup A - offset: 2, value: LCPI3_0, kind: FK_Data_4
 203 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 204 ;
 205 ; FMACALL32-LABEL: test_f32_cst:
 206 ; FMACALL32:       ## %bb.0: ## %entry
 207 ; FMACALL32-NEXT:    flds LCPI3_0 ## encoding: [0xd9,0x05,A,A,A,A]
 208 ; FMACALL32-NEXT:    ## fixup A - offset: 2, value: LCPI3_0, kind: FK_Data_4
 209 ; FMACALL32-NEXT:    retl ## encoding: [0xc3]
 210 ;
 211 ; FMA64-LABEL: test_f32_cst:
 212 ; FMA64:       ## %bb.0: ## %entry
 213 ; FMA64-NEXT:    vmovss {{.*}}(%rip), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x05,A,A,A,A]
 214 ; FMA64-NEXT:    ## fixup A - offset: 4, value: LCPI3_0-4, kind: reloc_riprel_4byte
 215 ; FMA64-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 216 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 217 ;
 218 ; FMACALL64-LABEL: test_f32_cst:
 219 ; FMACALL64:       ## %bb.0: ## %entry
 220 ; FMACALL64-NEXT:    movss {{.*}}(%rip), %xmm0 ## encoding: [0xf3,0x0f,0x10,0x05,A,A,A,A]
 221 ; FMACALL64-NEXT:    ## fixup A - offset: 4, value: LCPI3_0-4, kind: reloc_riprel_4byte
 222 ; FMACALL64-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 223 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
 224 ;
 225 ; AVX512-LABEL: test_f32_cst:
 226 ; AVX512:       ## %bb.0: ## %entry
 227 ; AVX512-NEXT:    vmovss {{.*}}(%rip), %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfa,0x10,0x05,A,A,A,A]
 228 ; AVX512-NEXT:    ## fixup A - offset: 4, value: LCPI3_0-4, kind: reloc_riprel_4byte
 229 ; AVX512-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 230 ; AVX512-NEXT:    retq ## encoding: [0xc3]
 231 ;
 232 ; AVX512VL-LABEL: test_f32_cst:
 233 ; AVX512VL:       ## %bb.0: ## %entry
 234 ; AVX512VL-NEXT:    vmovss {{.*}}(%rip), %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfa,0x10,0x05,A,A,A,A]
 235 ; AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI3_0-4, kind: reloc_riprel_4byte
 236 ; AVX512VL-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 237 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 238 entry:
 239   %call = call float @llvm.fma.f32(float 3.0, float 3.0, float 3.0)
 240   ret float %call
 241 }
 242
 243 define <4 x float> @test_v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c) #0 {
 244 ; FMA32-LABEL: test_v4f32:
 245 ; FMA32:       ## %bb.0: ## %entry
 246 ; FMA32-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0x71,0xa8,0xc2]
 247 ; FMA32-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 248 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 249 ;
 250 ; FMA64-LABEL: test_v4f32:
 251 ; FMA64:       ## %bb.0: ## %entry
 252 ; FMA64-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0x71,0xa8,0xc2]
 253 ; FMA64-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 254 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 255 ;
 256 ; FMACALL64-LABEL: test_v4f32:
 257 ; FMACALL64:       ## %bb.0: ## %entry
 258 ; FMACALL64-NEXT:    subq $88, %rsp ## encoding: [0x48,0x83,0xec,0x58]
 259 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 260 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x30]
 261 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 262 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x10]
 263 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 264 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
 265 ; FMACALL64-NEXT:    shufps $231, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe7]
 266 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,1,2,3]
 267 ; FMACALL64-NEXT:    shufps $231, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe7]
 268 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,1,2,3]
 269 ; FMACALL64-NEXT:    shufps $231, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe7]
 270 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,1,2,3]
 271 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 272 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 273 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 274 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 275 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 276 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
 277 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 278 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 279 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 280 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 281 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 282 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 283 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 284 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x30]
 285 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 286 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 287 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 288 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 289 ; FMACALL64-NEXT:    unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
 290 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x04,0x24]
 291 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 292 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 293 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 294 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 295 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
 296 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 297 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 298 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 299 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x30]
 300 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 301 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 302 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 303 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x40]
 304 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 305 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
 306 ; FMACALL64-NEXT:    shufps $229, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe5]
 307 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,2,3]
 308 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 309 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 310 ; FMACALL64-NEXT:    shufps $229, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe5]
 311 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,2,3]
 312 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 313 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x30]
 314 ; FMACALL64-NEXT:    shufps $229, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe5]
 315 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,2,3]
 316 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 317 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 318 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 319 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x40]
 320 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 321 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 322 ; FMACALL64-NEXT:    unpcklpd (%rsp), %xmm1 ## 16-byte Folded Reload
 323 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x0c,0x24]
 324 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 325 ; FMACALL64-NEXT:    movaps %xmm1, %xmm0 ## encoding: [0x0f,0x28,0xc1]
 326 ; FMACALL64-NEXT:    addq $88, %rsp ## encoding: [0x48,0x83,0xc4,0x58]
 327 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
 328 ;
 329 ; AVX512-LABEL: test_v4f32:
 330 ; AVX512:       ## %bb.0: ## %entry
 331 ; AVX512-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0x71,0xa8,0xc2]
 332 ; AVX512-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 333 ; AVX512-NEXT:    retq ## encoding: [0xc3]
 334 ;
 335 ; AVX512VL-LABEL: test_v4f32:
 336 ; AVX512VL:       ## %bb.0: ## %entry
 337 ; AVX512VL-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa8,0xc2]
 338 ; AVX512VL-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
 339 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 340 ;
 341 ; FMACALL32_BDVER2-LABEL: test_v4f32:
 342 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
 343 ; FMACALL32_BDVER2-NEXT:    subl $108, %esp ## encoding: [0x83,0xec,0x6c]
 344 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 345 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x54,0x24,0x40]
 346 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 347 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x4c,0x24,0x30]
 348 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 349 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x44,0x24,0x20]
 350 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x08,0x02]
 351 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm1, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x4c,0x24,0x04,0x02]
 352 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x02]
 353 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 354 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 355 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 356 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x60]
 357 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 358 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x40]
 359 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
 360 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x54,0x24,0x30]
 361 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
 362 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x20]
 363 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
 364 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x01]
 365 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x01]
 366 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 367 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 368 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 369 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x54]
 370 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 371 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x40]
 372 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
 373 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 374 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x30]
 375 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
 376 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 377 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x20]
 378 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
 379 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 380 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 381 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 382 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x40]
 383 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
 384 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x54,0x24,0x30]
 385 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
 386 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x20]
 387 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
 388 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x03]
 389 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x03]
 390 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x1c]
 391 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 392 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x54]
 393 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x18]
 394 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 395 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x60]
 396 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x14]
 397 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 398 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 399 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x10]
 400 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x44,0x24,0x1c]
 401 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 402 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x18,0x10]
 403 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[2,3]
 404 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x14,0x20]
 405 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0],xmm0[3]
 406 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x10,0x30]
 407 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1,2],mem[0]
 408 ; FMACALL32_BDVER2-NEXT:    addl $108, %esp ## encoding: [0x83,0xc4,0x6c]
 409 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
 410 entry:
 411   %call = call <4 x float> @llvm.fma.v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c)
 412   ret <4 x float> %call
 413 }
 414
 415 define <8 x float> @test_v8f32(<8 x float> %a, <8 x float> %b, <8 x float> %c) #0 {
 416 ; FMA32-LABEL: test_v8f32:
 417 ; FMA32:       ## %bb.0: ## %entry
 418 ; FMA32-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0x75,0xa8,0xc2]
 419 ; FMA32-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
 420 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 421 ;
 422 ; FMA64-LABEL: test_v8f32:
 423 ; FMA64:       ## %bb.0: ## %entry
 424 ; FMA64-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0x75,0xa8,0xc2]
 425 ; FMA64-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
 426 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 427 ;
 428 ; FMACALL64-LABEL: test_v8f32:
 429 ; FMACALL64:       ## %bb.0: ## %entry
 430 ; FMACALL64-NEXT:    subq $136, %rsp ## encoding: [0x48,0x81,0xec,0x88,0x00,0x00,0x00]
 431 ; FMACALL64-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 432 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x6c,0x24,0x50]
 433 ; FMACALL64-NEXT:    movaps %xmm4, (%rsp) ## 16-byte Spill
 434 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x24,0x24]
 435 ; FMACALL64-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 436 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x5c,0x24,0x40]
 437 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 438 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x60]
 439 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 440 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x30]
 441 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 442 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
 443 ; FMACALL64-NEXT:    shufps $231, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe7]
 444 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,1,2,3]
 445 ; FMACALL64-NEXT:    movaps %xmm2, %xmm1 ## encoding: [0x0f,0x28,0xca]
 446 ; FMACALL64-NEXT:    shufps $231, %xmm2, %xmm1 ## encoding: [0x0f,0xc6,0xca,0xe7]
 447 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,1],xmm2[2,3]
 448 ; FMACALL64-NEXT:    movaps %xmm4, %xmm2 ## encoding: [0x0f,0x28,0xd4]
 449 ; FMACALL64-NEXT:    shufps $231, %xmm4, %xmm2 ## encoding: [0x0f,0xc6,0xd4,0xe7]
 450 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,1],xmm4[2,3]
 451 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 452 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 453 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 454 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
 455 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 456 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x10]
 457 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 458 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 459 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 460 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x60]
 461 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 462 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 463 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
 464 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
 465 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 466 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 467 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 468 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 469 ; FMACALL64-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
 470 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x44,0x24,0x20]
 471 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 472 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 473 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x70]
 474 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 475 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x10]
 476 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 477 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x60]
 478 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
 479 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
 480 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 481 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 482 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 483 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
 484 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 485 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x10]
 486 ; FMACALL64-NEXT:    shufps $229, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe5]
 487 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,2,3]
 488 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 489 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x60]
 490 ; FMACALL64-NEXT:    shufps $229, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe5]
 491 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,2,3]
 492 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
 493 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
 494 ; FMACALL64-NEXT:    shufps $229, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe5]
 495 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,2,3]
 496 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 497 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 498 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 499 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
 500 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 501 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 502 ; FMACALL64-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
 503 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x4c,0x24,0x70]
 504 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 505 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 506 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x20]
 507 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 508 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 509 ; FMACALL64-NEXT:    shufps $231, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe7]
 510 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,1,2,3]
 511 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 512 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x40]
 513 ; FMACALL64-NEXT:    shufps $231, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe7]
 514 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,1,2,3]
 515 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 516 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x50]
 517 ; FMACALL64-NEXT:    shufps $231, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe7]
 518 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,1,2,3]
 519 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 520 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 521 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 522 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 523 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 524 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 525 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 526 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 527 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 528 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x40]
 529 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 530 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 531 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 532 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x50]
 533 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 534 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 535 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 536 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 537 ; FMACALL64-NEXT:    unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
 538 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x04,0x24]
 539 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 540 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 541 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 542 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 543 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 544 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 545 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x40]
 546 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 547 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x50]
 548 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 549 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 550 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 551 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
 552 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 553 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 554 ; FMACALL64-NEXT:    shufps $229, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe5]
 555 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,2,3]
 556 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 557 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x40]
 558 ; FMACALL64-NEXT:    shufps $229, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe5]
 559 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,2,3]
 560 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
 561 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x50]
 562 ; FMACALL64-NEXT:    shufps $229, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe5]
 563 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,2,3]
 564 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 565 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 566 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 567 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 568 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 569 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 570 ; FMACALL64-NEXT:    unpcklpd (%rsp), %xmm1 ## 16-byte Folded Reload
 571 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x0c,0x24]
 572 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 573 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 574 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
 575 ; FMACALL64-NEXT:    addq $136, %rsp ## encoding: [0x48,0x81,0xc4,0x88,0x00,0x00,0x00]
 576 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
 577 ;
 578 ; AVX512-LABEL: test_v8f32:
 579 ; AVX512:       ## %bb.0: ## %entry
 580 ; AVX512-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0x75,0xa8,0xc2]
 581 ; AVX512-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
 582 ; AVX512-NEXT:    retq ## encoding: [0xc3]
 583 ;
 584 ; AVX512VL-LABEL: test_v8f32:
 585 ; AVX512VL:       ## %bb.0: ## %entry
 586 ; AVX512VL-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x75,0xa8,0xc2]
 587 ; AVX512VL-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
 588 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
 589 ;
 590 ; FMACALL32_BDVER2-LABEL: test_v8f32:
 591 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
 592 ; FMACALL32_BDVER2-NEXT:    subl $316, %esp ## encoding: [0x81,0xec,0x3c,0x01,0x00,0x00]
 593 ; FMACALL32_BDVER2-NEXT:    ## imm = 0x13C
 594 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm2, %xmm3 ## encoding: [0xc4,0xe3,0x7d,0x19,0xd3,0x01]
 595 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
 596 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x94,0x24,0x00,0x01,0x00,0x00]
 597 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm1, %xmm2 ## encoding: [0xc4,0xe3,0x7d,0x19,0xca,0x01]
 598 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
 599 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x8c,0x24,0xe0,0x00,0x00,0x00]
 600 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm1 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc1,0x01]
 601 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
 602 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x84,0x24,0xc0,0x00,0x00,0x00]
 603 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 604 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x5c,0x24,0x60]
 605 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm3, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x5c,0x24,0x08,0x02]
 606 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 607 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x54,0x24,0x50]
 608 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x02]
 609 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
 610 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x4c,0x24,0x40]
 611 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x02]
 612 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 613 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 614 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 615 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 616 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xb4,0x00,0x00,0x00]
 617 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 618 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x60]
 619 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
 620 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x54,0x24,0x50]
 621 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
 622 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x40]
 623 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
 624 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x01]
 625 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x01]
 626 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 627 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 628 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 629 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xa8,0x00,0x00,0x00]
 630 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 631 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x60]
 632 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
 633 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 634 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x50]
 635 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
 636 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 637 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x40]
 638 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
 639 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 640 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 641 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 642 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x9c,0x00,0x00,0x00]
 643 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 644 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0x00,0x01,0x00,0x00]
 645 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
 646 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 647 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xe0,0x00,0x00,0x00]
 648 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x03]
 649 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 650 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xc0,0x00,0x00,0x00]
 651 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x03]
 652 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 653 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 654 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 655 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 656 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x90,0x00,0x00,0x00]
 657 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 658 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0x00,0x01,0x00,0x00]
 659 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x02]
 660 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 661 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xe0,0x00,0x00,0x00]
 662 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x02]
 663 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 664 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xc0,0x00,0x00,0x00]
 665 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x02]
 666 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 667 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 668 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 669 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 670 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x84,0x00,0x00,0x00]
 671 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 672 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0x00,0x01,0x00,0x00]
 673 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
 674 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 675 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xe0,0x00,0x00,0x00]
 676 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x01]
 677 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 678 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xc0,0x00,0x00,0x00]
 679 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x01]
 680 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 681 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 682 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 683 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
 684 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x78]
 685 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 686 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0x00,0x01,0x00,0x00]
 687 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
 688 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 689 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xe0,0x00,0x00,0x00]
 690 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
 691 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
 692 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xc0,0x00,0x00,0x00]
 693 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
 694 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 695 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 696 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 697 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
 698 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x60]
 699 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
 700 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x54,0x24,0x50]
 701 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
 702 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x40]
 703 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
 704 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x03]
 705 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x03]
 706 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x3c]
 707 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 708 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x78]
 709 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x38]
 710 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 711 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x84,0x00,0x00,0x00]
 712 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x34]
 713 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 714 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x90,0x00,0x00,0x00]
 715 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x30]
 716 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 717 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x9c,0x00,0x00,0x00]
 718 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x2c]
 719 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 720 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xa8,0x00,0x00,0x00]
 721 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x28]
 722 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
 723 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xb4,0x00,0x00,0x00]
 724 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x24]
 725 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
 726 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
 727 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x20]
 728 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x44,0x24,0x3c]
 729 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero,zero,zero
 730 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x38,0x10]
 731 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[2,3]
 732 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfa,0x10,0x4c,0x24,0x2c]
 733 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero,zero,zero
 734 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x28,0x10]
 735 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0],mem[0],xmm1[2,3]
 736 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x34,0x20]
 737 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0],xmm0[3]
 738 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x24,0x20]
 739 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0],xmm1[3]
 740 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x30,0x30]
 741 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1,2],mem[0]
 742 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x20,0x30]
 743 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1,2],mem[0]
 744 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 ## encoding: [0xc4,0xe3,0x7d,0x18,0xc1,0x01]
 745 ; FMACALL32_BDVER2-NEXT:    addl $316, %esp ## encoding: [0x81,0xc4,0x3c,0x01,0x00,0x00]
 746 ; FMACALL32_BDVER2-NEXT:    ## imm = 0x13C
 747 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
 748 entry:
 749   %call = call <8 x float> @llvm.fma.v8f32(<8 x float> %a, <8 x float> %b, <8 x float> %c)
 750   ret <8 x float> %call
 751 }
 752
 753 define <16 x float> @test_v16f32(<16 x float> %a, <16 x float> %b, <16 x float> %c) #0 {
 754 ; FMA32-LABEL: test_v16f32:
 755 ; FMA32:       ## %bb.0: ## %entry
 756 ; FMA32-NEXT:    pushl %ebp ## encoding: [0x55]
 757 ; FMA32-NEXT:    movl %esp, %ebp ## encoding: [0x89,0xe5]
 758 ; FMA32-NEXT:    andl $-32, %esp ## encoding: [0x83,0xe4,0xe0]
 759 ; FMA32-NEXT:    subl $32, %esp ## encoding: [0x83,0xec,0x20]
 760 ; FMA32-NEXT:    vfmadd213ps 8(%ebp), %ymm2, %ymm0 ## encoding: [0xc4,0xe2,0x6d,0xa8,0x45,0x08]
 761 ; FMA32-NEXT:    ## ymm0 = (ymm2 * ymm0) + mem
 762 ; FMA32-NEXT:    vfmadd213ps 40(%ebp), %ymm3, %ymm1 ## encoding: [0xc4,0xe2,0x65,0xa8,0x4d,0x28]
 763 ; FMA32-NEXT:    ## ymm1 = (ymm3 * ymm1) + mem
 764 ; FMA32-NEXT:    movl %ebp, %esp ## encoding: [0x89,0xec]
 765 ; FMA32-NEXT:    popl %ebp ## encoding: [0x5d]
 766 ; FMA32-NEXT:    retl ## encoding: [0xc3]
 767 ;
 768 ; FMA64-LABEL: test_v16f32:
 769 ; FMA64:       ## %bb.0: ## %entry
 770 ; FMA64-NEXT:    vfmadd213ps %ymm4, %ymm2, %ymm0 ## encoding: [0xc4,0xe2,0x6d,0xa8,0xc4]
 771 ; FMA64-NEXT:    ## ymm0 = (ymm2 * ymm0) + ymm4
 772 ; FMA64-NEXT:    vfmadd213ps %ymm5, %ymm3, %ymm1 ## encoding: [0xc4,0xe2,0x65,0xa8,0xcd]
 773 ; FMA64-NEXT:    ## ymm1 = (ymm3 * ymm1) + ymm5
 774 ; FMA64-NEXT:    retq ## encoding: [0xc3]
 775 ;
 776 ; FMACALL64-LABEL: test_v16f32:
 777 ; FMACALL64:       ## %bb.0: ## %entry
 778 ; FMACALL64-NEXT:    subq $168, %rsp ## encoding: [0x48,0x81,0xec,0xa8,0x00,0x00,0x00]
 779 ; FMACALL64-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 780 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0xbc,0x24,0x80,0x00,0x00,0x00]
 781 ; FMACALL64-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 782 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x74,0x24,0x20]
 783 ; FMACALL64-NEXT:    movaps %xmm5, (%rsp) ## 16-byte Spill
 784 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x2c,0x24]
 785 ; FMACALL64-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 786 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x64,0x24,0x10]
 787 ; FMACALL64-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 788 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x5c,0x24,0x70]
 789 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 790 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x40]
 791 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 792 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x30]
 793 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 794 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x50]
 795 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
 796 ; FMACALL64-NEXT:    shufps $231, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe7]
 797 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,1,2,3]
 798 ; FMACALL64-NEXT:    movaps %xmm4, %xmm1 ## encoding: [0x0f,0x28,0xcc]
 799 ; FMACALL64-NEXT:    shufps $231, %xmm4, %xmm1 ## encoding: [0x0f,0xc6,0xcc,0xe7]
 800 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,1],xmm4[2,3]
 801 ; FMACALL64-NEXT:    shufps $231, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe7]
 802 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,1,2,3]
 803 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 804 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 805 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 806 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x60]
 807 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 808 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
 809 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 810 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 811 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 812 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 813 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 814 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 815 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
 816 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 817 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 818 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 819 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 820 ; FMACALL64-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
 821 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x44,0x24,0x60]
 822 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 823 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 824 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x84,0x24,0x90,0x00,0x00,0x00]
 825 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 826 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
 827 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 828 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 829 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
 830 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 831 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 832 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 833 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x60]
 834 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 835 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
 836 ; FMACALL64-NEXT:    shufps $229, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe5]
 837 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,2,3]
 838 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 839 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 840 ; FMACALL64-NEXT:    shufps $229, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe5]
 841 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,2,3]
 842 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
 843 ; FMACALL64-NEXT:    shufps $229, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe5]
 844 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,2,3]
 845 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 846 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 847 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 848 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x60]
 849 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 850 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 851 ; FMACALL64-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
 852 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x8c,0x24,0x90,0x00,0x00,0x00]
 853 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 854 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 855 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x60]
 856 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 857 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 858 ; FMACALL64-NEXT:    shufps $231, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe7]
 859 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,1,2,3]
 860 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
 861 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
 862 ; FMACALL64-NEXT:    shufps $231, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe7]
 863 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,1,2,3]
 864 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
 865 ; FMACALL64-NEXT:    shufps $231, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe7]
 866 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,1,2,3]
 867 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 868 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 869 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 870 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
 871 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 872 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 873 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 874 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 875 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
 876 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
 877 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 878 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 879 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
 880 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 881 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 882 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 883 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 884 ; FMACALL64-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
 885 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x44,0x24,0x10]
 886 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 887 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 888 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x50]
 889 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 890 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 891 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
 892 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
 893 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
 894 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 895 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 896 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 897 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
 898 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 899 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
 900 ; FMACALL64-NEXT:    shufps $229, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe5]
 901 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,2,3]
 902 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
 903 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
 904 ; FMACALL64-NEXT:    shufps $229, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe5]
 905 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,2,3]
 906 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
 907 ; FMACALL64-NEXT:    shufps $229, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe5]
 908 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,2,3]
 909 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 910 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 911 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 912 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
 913 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 914 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 915 ; FMACALL64-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
 916 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x4c,0x24,0x50]
 917 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 918 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 919 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x10]
 920 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 921 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
 922 ; FMACALL64-NEXT:    shufps $231, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe7]
 923 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,1,2,3]
 924 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 925 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
 926 ; FMACALL64-NEXT:    shufps $231, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe7]
 927 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,1,2,3]
 928 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
 929 ; FMACALL64-NEXT:    shufps $231, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe7]
 930 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,1,2,3]
 931 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 932 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 933 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 934 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 935 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 936 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
 937 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
 938 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
 939 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 940 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
 941 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
 942 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
 943 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
 944 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
 945 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
 946 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 947 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 948 ; FMACALL64-NEXT:    unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
 949 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x04,0x24]
 950 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 951 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 952 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x30]
 953 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 954 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
 955 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 956 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
 957 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
 958 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 959 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 960 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
 961 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
 962 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 963 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
 964 ; FMACALL64-NEXT:    shufps $229, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe5]
 965 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,2,3]
 966 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 967 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
 968 ; FMACALL64-NEXT:    shufps $229, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe5]
 969 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,2,3]
 970 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
 971 ; FMACALL64-NEXT:    shufps $229, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe5]
 972 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,2,3]
 973 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 974 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 975 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
 976 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
 977 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm1 ## encoding: [0x0f,0x14,0xc8]
 978 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 979 ; FMACALL64-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
 980 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x4c,0x24,0x30]
 981 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],mem[0]
 982 ; FMACALL64-NEXT:    movaps %xmm1, (%rsp) ## 16-byte Spill
 983 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x0c,0x24]
 984 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
 985 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x70]
 986 ; FMACALL64-NEXT:    shufps $231, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe7]
 987 ; FMACALL64-NEXT:    ## xmm0 = xmm0[3,1,2,3]
 988 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
 989 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x8c,0x24,0x80,0x00,0x00,0x00]
 990 ; FMACALL64-NEXT:    shufps $231, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe7]
 991 ; FMACALL64-NEXT:    ## xmm1 = xmm1[3,1,2,3]
 992 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xe0,0x00,0x00,0x00]
 993 ; FMACALL64-NEXT:    shufps $231, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe7]
 994 ; FMACALL64-NEXT:    ## xmm2 = xmm2[3,1,2,3]
 995 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
 996 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
 997 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
 998 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
 999 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1000 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x70]
1001 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1002 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1003 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1004 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x8c,0x24,0x80,0x00,0x00,0x00]
1005 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1006 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1007 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xe0,0x00,0x00,0x00]
1008 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1009 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1010 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
1011 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
1012 ; FMACALL64-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
1013 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x14,0x44,0x24,0x20]
1014 ; FMACALL64-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
1015 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1016 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
1017 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1018 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x70]
1019 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1020 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x8c,0x24,0x80,0x00,0x00,0x00]
1021 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xe0,0x00,0x00,0x00]
1022 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
1023 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
1024 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1025 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x40]
1026 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1027 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x70]
1028 ; FMACALL64-NEXT:    shufps $229, %xmm0, %xmm0 ## encoding: [0x0f,0xc6,0xc0,0xe5]
1029 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1,2,3]
1030 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1031 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x8c,0x24,0x80,0x00,0x00,0x00]
1032 ; FMACALL64-NEXT:    shufps $229, %xmm1, %xmm1 ## encoding: [0x0f,0xc6,0xc9,0xe5]
1033 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1,2,3]
1034 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xe0,0x00,0x00,0x00]
1035 ; FMACALL64-NEXT:    shufps $229, %xmm2, %xmm2 ## encoding: [0x0f,0xc6,0xd2,0xe5]
1036 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1,2,3]
1037 ; FMACALL64-NEXT:    callq _fmaf ## encoding: [0xe8,A,A,A,A]
1038 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: reloc_branch_4byte_pcrel
1039 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Reload
1040 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x5c,0x24,0x40]
1041 ; FMACALL64-NEXT:    unpcklps %xmm0, %xmm3 ## encoding: [0x0f,0x14,0xd8]
1042 ; FMACALL64-NEXT:    ## xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
1043 ; FMACALL64-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Folded Reload
1044 ; FMACALL64-NEXT:    ## encoding: [0x66,0x0f,0x14,0x5c,0x24,0x20]
1045 ; FMACALL64-NEXT:    ## xmm3 = xmm3[0],mem[0]
1046 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1047 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x60]
1048 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1049 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1050 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
1051 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
1052 ; FMACALL64-NEXT:    addq $168, %rsp ## encoding: [0x48,0x81,0xc4,0xa8,0x00,0x00,0x00]
1053 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
1054 ;
1055 ; AVX512-LABEL: test_v16f32:
1056 ; AVX512:       ## %bb.0: ## %entry
1057 ; AVX512-NEXT:    vfmadd213ps %zmm2, %zmm1, %zmm0 ## encoding: [0x62,0xf2,0x75,0x48,0xa8,0xc2]
1058 ; AVX512-NEXT:    ## zmm0 = (zmm1 * zmm0) + zmm2
1059 ; AVX512-NEXT:    retq ## encoding: [0xc3]
1060 ;
1061 ; AVX512VL-LABEL: test_v16f32:
1062 ; AVX512VL:       ## %bb.0: ## %entry
1063 ; AVX512VL-NEXT:    vfmadd213ps %zmm2, %zmm1, %zmm0 ## encoding: [0x62,0xf2,0x75,0x48,0xa8,0xc2]
1064 ; AVX512VL-NEXT:    ## zmm0 = (zmm1 * zmm0) + zmm2
1065 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
1066 ;
1067 ; FMACALL32_BDVER2-LABEL: test_v16f32:
1068 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
1069 ; FMACALL32_BDVER2-NEXT:    pushl %ebp ## encoding: [0x55]
1070 ; FMACALL32_BDVER2-NEXT:    movl %esp, %ebp ## encoding: [0x89,0xe5]
1071 ; FMACALL32_BDVER2-NEXT:    andl $-32, %esp ## encoding: [0x83,0xe4,0xe0]
1072 ; FMACALL32_BDVER2-NEXT:    subl $448, %esp ## encoding: [0x81,0xec,0xc0,0x01,0x00,0x00]
1073 ; FMACALL32_BDVER2-NEXT:    ## imm = 0x1C0
1074 ; FMACALL32_BDVER2-NEXT:    vmovaps 56(%ebp), %xmm4 ## encoding: [0xc5,0xf8,0x28,0x65,0x38]
1075 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1076 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x94,0x24,0x60,0x01,0x00,0x00]
1077 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm3, %xmm2 ## encoding: [0xc4,0xe3,0x7d,0x19,0xda,0x01]
1078 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1079 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x8c,0x24,0x80,0x00,0x00,0x00]
1080 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm1, %xmm1 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc9,0x01]
1081 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm3, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1082 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x5c,0x24,0x60]
1083 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1084 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x84,0x24,0x80,0x01,0x00,0x00]
1085 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1086 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x94,0x24,0xc0,0x00,0x00,0x00]
1087 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x02]
1088 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1089 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x8c,0x24,0xb0,0x00,0x00,0x00]
1090 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x02]
1091 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm4, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x64,0x24,0x08,0x02]
1092 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1093 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1094 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1095 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1096 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x54,0x01,0x00,0x00]
1097 ; FMACALL32_BDVER2-NEXT:    vmovaps 56(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x38]
1098 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
1099 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
1100 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
1101 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x8c,0x24,0xb0,0x00,0x00,0x00]
1102 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
1103 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x01]
1104 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x01]
1105 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1106 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1107 ; FMACALL32_BDVER2-NEXT:    vmovaps 56(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x38]
1108 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1109 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x48,0x01,0x00,0x00]
1110 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
1111 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1112 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x84,0x24,0xc0,0x00,0x00,0x00]
1113 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
1114 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1115 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x84,0x24,0xb0,0x00,0x00,0x00]
1116 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
1117 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1118 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1119 ; FMACALL32_BDVER2-NEXT:    vmovaps 40(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x28]
1120 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1121 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x3c,0x01,0x00,0x00]
1122 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
1123 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1124 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x44,0x24,0x60]
1125 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x03]
1126 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1127 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1128 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x03]
1129 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1130 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1131 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1132 ; FMACALL32_BDVER2-NEXT:    vmovaps 40(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x28]
1133 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1134 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x30,0x01,0x00,0x00]
1135 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x02]
1136 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1137 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x44,0x24,0x60]
1138 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x02]
1139 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1140 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1141 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x02]
1142 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1143 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1144 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1145 ; FMACALL32_BDVER2-NEXT:    vmovaps 40(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x28]
1146 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1147 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x24,0x01,0x00,0x00]
1148 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
1149 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1150 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x44,0x24,0x60]
1151 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x01]
1152 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1153 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1154 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x01]
1155 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1156 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1157 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1158 ; FMACALL32_BDVER2-NEXT:    vmovaps 40(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x28]
1159 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1160 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x18,0x01,0x00,0x00]
1161 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
1162 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1163 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x44,0x24,0x60]
1164 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
1165 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1166 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1167 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
1168 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1169 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1170 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1171 ; FMACALL32_BDVER2-NEXT:    vmovaps 24(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x18]
1172 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1173 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x0c,0x01,0x00,0x00]
1174 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
1175 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1176 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x60,0x01,0x00,0x00]
1177 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm0 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc0,0x01]
1178 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1179 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x84,0x24,0x80,0x00,0x00,0x00]
1180 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x03]
1181 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1182 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x01,0x00,0x00]
1183 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm0 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc0,0x01]
1184 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1185 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x44,0x24,0x60]
1186 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x03]
1187 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1188 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1189 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1190 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1191 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x00,0x01,0x00,0x00]
1192 ; FMACALL32_BDVER2-NEXT:    vmovaps 24(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x18]
1193 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
1194 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x94,0x24,0x80,0x00,0x00,0x00]
1195 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
1196 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x60]
1197 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x02]
1198 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x02]
1199 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x02]
1200 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1201 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1202 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1203 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xf4,0x00,0x00,0x00]
1204 ; FMACALL32_BDVER2-NEXT:    vmovaps 24(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x18]
1205 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
1206 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x94,0x24,0x80,0x00,0x00,0x00]
1207 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
1208 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x4c,0x24,0x60]
1209 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
1210 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x01]
1211 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x01]
1212 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1213 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1214 ; FMACALL32_BDVER2-NEXT:    vmovaps 24(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x18]
1215 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1216 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xe8,0x00,0x00,0x00]
1217 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
1218 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1219 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1220 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
1221 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1222 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x60]
1223 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
1224 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1225 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1226 ; FMACALL32_BDVER2-NEXT:    vmovaps 8(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x08]
1227 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1228 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x80,0x00,0x00,0x00]
1229 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
1230 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1231 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x60,0x01,0x00,0x00]
1232 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x03]
1233 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1234 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x01,0x00,0x00]
1235 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x03]
1236 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1237 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1238 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1239 ; FMACALL32_BDVER2-NEXT:    vmovaps 8(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x08]
1240 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1241 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x60]
1242 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x02]
1243 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1244 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x60,0x01,0x00,0x00]
1245 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x02]
1246 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1247 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x01,0x00,0x00]
1248 ; FMACALL32_BDVER2-NEXT:    vextractps $2, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x02]
1249 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1250 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1251 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1252 ; FMACALL32_BDVER2-NEXT:    vmovaps 8(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x08]
1253 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1254 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xdc,0x00,0x00,0x00]
1255 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x01]
1256 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1257 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x60,0x01,0x00,0x00]
1258 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x04,0x01]
1259 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1260 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x01,0x00,0x00]
1261 ; FMACALL32_BDVER2-NEXT:    vextractps $1, %xmm0, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x04,0x24,0x01]
1262 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1263 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1264 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1265 ; FMACALL32_BDVER2-NEXT:    vmovaps 8(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x08]
1266 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1267 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xd0,0x00,0x00,0x00]
1268 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x08]
1269 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1270 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x60,0x01,0x00,0x00]
1271 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xfa,0x11,0x44,0x24,0x04]
1272 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1273 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x80,0x01,0x00,0x00]
1274 ; FMACALL32_BDVER2-NEXT:    vmovss %xmm0, (%esp) ## encoding: [0xc5,0xfa,0x11,0x04,0x24]
1275 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1276 ; FMACALL32_BDVER2-NEXT:    calll _fmaf ## encoding: [0xe8,A,A,A,A]
1277 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fmaf-4, kind: FK_PCRel_4
1278 ; FMACALL32_BDVER2-NEXT:    vmovaps 56(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x38]
1279 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 ## 16-byte Reload
1280 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
1281 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 ## 16-byte Reload
1282 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x8c,0x24,0xb0,0x00,0x00,0x00]
1283 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x44,0x24,0x08,0x03]
1284 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x54,0x24,0x04,0x03]
1285 ; FMACALL32_BDVER2-NEXT:    vextractps $3, %xmm1, (%esp) ## encoding: [0xc4,0xe3,0x79,0x17,0x0c,0x24,0x03]
1286 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x3c]
1287 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1288 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xd0,0x00,0x00,0x00]
1289 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x38]
1290 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1291 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xdc,0x00,0x00,0x00]
1292 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x34]
1293 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1294 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x60]
1295 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x30]
1296 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1297 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x80,0x00,0x00,0x00]
1298 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x2c]
1299 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1300 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xe8,0x00,0x00,0x00]
1301 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x28]
1302 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1303 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xf4,0x00,0x00,0x00]
1304 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x24]
1305 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1306 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x00,0x01,0x00,0x00]
1307 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x20]
1308 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1309 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x0c,0x01,0x00,0x00]
1310 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x5c]
1311 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1312 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x18,0x01,0x00,0x00]
1313 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x58]
1314 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1315 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x24,0x01,0x00,0x00]
1316 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x54]
1317 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1318 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x30,0x01,0x00,0x00]
1319 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x50]
1320 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1321 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x3c,0x01,0x00,0x00]
1322 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x4c]
1323 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1324 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x48,0x01,0x00,0x00]
1325 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x48]
1326 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1327 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x54,0x01,0x00,0x00]
1328 ; FMACALL32_BDVER2-NEXT:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x44]
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:    fstps {{[0-9]+}}(%esp) ## encoding: [0xd9,0x5c,0x24,0x40]
1332 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfa,0x10,0x44,0x24,0x3c]
1333 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero,zero,zero
1334 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x38,0x10]
1335 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0],xmm0[2,3]
1336 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfa,0x10,0x4c,0x24,0x2c]
1337 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero,zero,zero
1338 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x28,0x10]
1339 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0],mem[0],xmm1[2,3]
1340 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm2 ## encoding: [0xc5,0xfa,0x10,0x54,0x24,0x4c]
1341 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = mem[0],zero,zero,zero
1342 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm2, %xmm2 ## encoding: [0xc4,0xe3,0x69,0x21,0x54,0x24,0x48,0x10]
1343 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = xmm2[0],mem[0],xmm2[2,3]
1344 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x34,0x20]
1345 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0],xmm0[3]
1346 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x24,0x20]
1347 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0],xmm1[3]
1348 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm2, %xmm2 ## encoding: [0xc4,0xe3,0x69,0x21,0x54,0x24,0x44,0x20]
1349 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = xmm2[0,1],mem[0],xmm2[3]
1350 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x21,0x44,0x24,0x30,0x30]
1351 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1,2],mem[0]
1352 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x20,0x30]
1353 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1,2],mem[0]
1354 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm2, %xmm2 ## encoding: [0xc4,0xe3,0x69,0x21,0x54,0x24,0x40,0x30]
1355 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = xmm2[0,1,2],mem[0]
1356 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 ## encoding: [0xc4,0xe3,0x7d,0x18,0xc1,0x01]
1357 ; FMACALL32_BDVER2-NEXT:    vmovss {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfa,0x10,0x4c,0x24,0x5c]
1358 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero,zero,zero
1359 ; FMACALL32_BDVER2-NEXT:    vinsertps $16, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x58,0x10]
1360 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0],mem[0],xmm1[2,3]
1361 ; FMACALL32_BDVER2-NEXT:    vinsertps $32, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x54,0x20]
1362 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0],xmm1[3]
1363 ; FMACALL32_BDVER2-NEXT:    vinsertps $48, {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x21,0x4c,0x24,0x50,0x30]
1364 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1,2],mem[0]
1365 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1 ## encoding: [0xc4,0xe3,0x75,0x18,0xca,0x01]
1366 ; FMACALL32_BDVER2-NEXT:    movl %ebp, %esp ## encoding: [0x89,0xec]
1367 ; FMACALL32_BDVER2-NEXT:    popl %ebp ## encoding: [0x5d]
1368 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
1369 entry:
1370   %call = call <16 x float> @llvm.fma.v16f32(<16 x float> %a, <16 x float> %b, <16 x float> %c)
1371   ret <16 x float> %call
1372 }
1373
1374 define <2 x double> @test_v2f64(<2 x double> %a, <2 x double> %b, <2 x double> %c) #0 {
1375 ; FMA32-LABEL: test_v2f64:
1376 ; FMA32:       ## %bb.0: ## %entry
1377 ; FMA32-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1378 ; FMA32-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1379 ; FMA32-NEXT:    retl ## encoding: [0xc3]
1380 ;
1381 ; FMA64-LABEL: test_v2f64:
1382 ; FMA64:       ## %bb.0: ## %entry
1383 ; FMA64-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1384 ; FMA64-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1385 ; FMA64-NEXT:    retq ## encoding: [0xc3]
1386 ;
1387 ; FMACALL64-LABEL: test_v2f64:
1388 ; FMACALL64:       ## %bb.0: ## %entry
1389 ; FMACALL64-NEXT:    subq $72, %rsp ## encoding: [0x48,0x83,0xec,0x48]
1390 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1391 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x20]
1392 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1393 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x10]
1394 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
1395 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
1396 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1397 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1398 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1399 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x30]
1400 ; FMACALL64-NEXT:    movaps (%rsp), %xmm0 ## 16-byte Reload
1401 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x04,0x24]
1402 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1403 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1404 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1405 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1406 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1407 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1408 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
1409 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x20]
1410 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1411 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1412 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1413 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1414 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1415 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x30]
1416 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1417 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1418 ; FMACALL64-NEXT:    movaps %xmm1, %xmm0 ## encoding: [0x0f,0x28,0xc1]
1419 ; FMACALL64-NEXT:    addq $72, %rsp ## encoding: [0x48,0x83,0xc4,0x48]
1420 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
1421 ;
1422 ; AVX512-LABEL: test_v2f64:
1423 ; AVX512:       ## %bb.0: ## %entry
1424 ; AVX512-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1425 ; AVX512-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1426 ; AVX512-NEXT:    retq ## encoding: [0xc3]
1427 ;
1428 ; AVX512VL-LABEL: test_v2f64:
1429 ; AVX512VL:       ## %bb.0: ## %entry
1430 ; AVX512VL-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
1431 ; AVX512VL-NEXT:    ## xmm0 = (xmm1 * xmm0) + xmm2
1432 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
1433 ;
1434 ; FMACALL32_BDVER2-LABEL: test_v2f64:
1435 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
1436 ; FMACALL32_BDVER2-NEXT:    subl $108, %esp ## encoding: [0x83,0xec,0x6c]
1437 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1438 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x44,0x24,0x50]
1439 ; FMACALL32_BDVER2-NEXT:    vmovlhps %xmm1, %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0xc1]
1440 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],xmm1[0]
1441 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1442 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x54,0x24,0x30]
1443 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1444 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x4c,0x24,0x40]
1445 ; FMACALL32_BDVER2-NEXT:    vmovlps %xmm2, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x13,0x54,0x24,0x10]
1446 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1447 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1448 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1449 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1450 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x30]
1451 ; FMACALL32_BDVER2-NEXT:    vmovhps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x17,0x44,0x24,0x10]
1452 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1453 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x40]
1454 ; FMACALL32_BDVER2-NEXT:    vmovlps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1455 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x12,0x44,0x24,0x58]
1456 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0,1],xmm0[2,3]
1457 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1458 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x28]
1459 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1460 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1461 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x20]
1462 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x44,0x24,0x28]
1463 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
1464 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0x44,0x24,0x20]
1465 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0,1]
1466 ; FMACALL32_BDVER2-NEXT:    addl $108, %esp ## encoding: [0x83,0xc4,0x6c]
1467 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
1468 entry:
1469   %call = call <2 x double> @llvm.fma.v2f64(<2 x double> %a, <2 x double> %b, <2 x double> %c)
1470   ret <2 x double> %call
1471 }
1472
1473 define <4 x double> @test_v4f64(<4 x double> %a, <4 x double> %b, <4 x double> %c) #0 {
1474 ; FMA32-LABEL: test_v4f64:
1475 ; FMA32:       ## %bb.0: ## %entry
1476 ; FMA32-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0xf5,0xa8,0xc2]
1477 ; FMA32-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
1478 ; FMA32-NEXT:    retl ## encoding: [0xc3]
1479 ;
1480 ; FMA64-LABEL: test_v4f64:
1481 ; FMA64:       ## %bb.0: ## %entry
1482 ; FMA64-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0xf5,0xa8,0xc2]
1483 ; FMA64-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
1484 ; FMA64-NEXT:    retq ## encoding: [0xc3]
1485 ;
1486 ; FMACALL64-LABEL: test_v4f64:
1487 ; FMACALL64:       ## %bb.0: ## %entry
1488 ; FMACALL64-NEXT:    subq $120, %rsp ## encoding: [0x48,0x83,0xec,0x78]
1489 ; FMACALL64-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1490 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x6c,0x24,0x40]
1491 ; FMACALL64-NEXT:    movaps %xmm4, (%rsp) ## 16-byte Spill
1492 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x24,0x24]
1493 ; FMACALL64-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1494 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x5c,0x24,0x30]
1495 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1496 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x60]
1497 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1498 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x20]
1499 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1500 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x50]
1501 ; FMACALL64-NEXT:    movaps %xmm2, %xmm1 ## encoding: [0x0f,0x28,0xca]
1502 ; FMACALL64-NEXT:    movaps %xmm4, %xmm2 ## encoding: [0x0f,0x28,0xd4]
1503 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1504 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1505 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1506 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
1507 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1508 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
1509 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1510 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1511 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1512 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x60]
1513 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1514 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1515 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
1516 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
1517 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1518 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1519 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1520 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1521 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1522 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1523 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1524 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1525 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1526 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x10]
1527 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1528 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
1529 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1530 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x30]
1531 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
1532 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x40]
1533 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1534 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1535 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
1536 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
1537 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1538 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x20]
1539 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1540 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1541 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1542 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x30]
1543 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1544 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1545 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 ## 16-byte Reload
1546 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x54,0x24,0x40]
1547 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1548 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1549 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1550 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1551 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
1552 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
1553 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1554 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1555 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1556 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x10]
1557 ; FMACALL64-NEXT:    addq $120, %rsp ## encoding: [0x48,0x83,0xc4,0x78]
1558 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
1559 ;
1560 ; AVX512-LABEL: test_v4f64:
1561 ; AVX512:       ## %bb.0: ## %entry
1562 ; AVX512-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 ## encoding: [0xc4,0xe2,0xf5,0xa8,0xc2]
1563 ; AVX512-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
1564 ; AVX512-NEXT:    retq ## encoding: [0xc3]
1565 ;
1566 ; AVX512VL-LABEL: test_v4f64:
1567 ; AVX512VL:       ## %bb.0: ## %entry
1568 ; AVX512VL-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf5,0xa8,0xc2]
1569 ; AVX512VL-NEXT:    ## ymm0 = (ymm1 * ymm0) + ymm2
1570 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
1571 ;
1572 ; FMACALL32_BDVER2-LABEL: test_v4f64:
1573 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
1574 ; FMACALL32_BDVER2-NEXT:    subl $252, %esp ## encoding: [0x81,0xec,0xfc,0x00,0x00,0x00]
1575 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm2, %xmm3 ## encoding: [0xc4,0xe3,0x7d,0x19,0xd3,0x01]
1576 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1577 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x94,0x24,0xc0,0x00,0x00,0x00]
1578 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1579 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x84,0x24,0x80,0x00,0x00,0x00]
1580 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm1, %xmm2 ## encoding: [0xc4,0xe3,0x7d,0x19,0xca,0x01]
1581 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm0 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc0,0x01]
1582 ; FMACALL32_BDVER2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1583 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x11,0x8c,0x24,0xa0,0x00,0x00,0x00]
1584 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1585 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x44,0x24,0x60]
1586 ; FMACALL32_BDVER2-NEXT:    vmovlhps %xmm2, %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0xc2]
1587 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],xmm2[0]
1588 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1589 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x5c,0x24,0x70]
1590 ; FMACALL32_BDVER2-NEXT:    vmovlps %xmm3, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x13,0x5c,0x24,0x10]
1591 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1592 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x54,0x24,0x50]
1593 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1594 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1595 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1596 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1597 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1598 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x44]
1599 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1600 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xc0,0x00,0x00,0x00]
1601 ; FMACALL32_BDVER2-NEXT:    vmovhps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x17,0x44,0x24,0x10]
1602 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1603 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0x80,0x00,0x00,0x00]
1604 ; FMACALL32_BDVER2-NEXT:    vunpckhpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1605 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x15,0x84,0x24,0xa0,0x00,0x00,0x00]
1606 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[1],mem[1]
1607 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1608 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1609 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1610 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1611 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1612 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x38]
1613 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1614 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0xc0,0x00,0x00,0x00]
1615 ; FMACALL32_BDVER2-NEXT:    vmovlps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x13,0x44,0x24,0x10]
1616 ; FMACALL32_BDVER2-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1617 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x10,0x84,0x24,0x80,0x00,0x00,0x00]
1618 ; FMACALL32_BDVER2-NEXT:    vunpcklpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1619 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x14,0x84,0x24,0xa0,0x00,0x00,0x00]
1620 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0]
1621 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1622 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1623 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1624 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1625 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1626 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x70]
1627 ; FMACALL32_BDVER2-NEXT:    vmovhps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x17,0x44,0x24,0x10]
1628 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1629 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x50]
1630 ; FMACALL32_BDVER2-NEXT:    vmovlps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1631 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x12,0x44,0x24,0x68]
1632 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0,1],xmm0[2,3]
1633 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1634 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x30]
1635 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1636 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x38]
1637 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x28]
1638 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1639 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x44]
1640 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x20]
1641 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1642 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1643 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x18]
1644 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x44,0x24,0x30]
1645 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
1646 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfb,0x10,0x4c,0x24,0x20]
1647 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero
1648 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0x44,0x24,0x28]
1649 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0,1]
1650 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc5,0xf0,0x16,0x4c,0x24,0x18]
1651 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0,1]
1652 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 ## encoding: [0xc4,0xe3,0x7d,0x18,0xc1,0x01]
1653 ; FMACALL32_BDVER2-NEXT:    addl $252, %esp ## encoding: [0x81,0xc4,0xfc,0x00,0x00,0x00]
1654 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
1655 entry:
1656   %call = call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> %b, <4 x double> %c)
1657   ret <4 x double> %call
1658 }
1659
1660 define <8 x double> @test_v8f64(<8 x double> %a, <8 x double> %b, <8 x double> %c) #0 {
1661 ; FMA32-LABEL: test_v8f64:
1662 ; FMA32:       ## %bb.0: ## %entry
1663 ; FMA32-NEXT:    pushl %ebp ## encoding: [0x55]
1664 ; FMA32-NEXT:    movl %esp, %ebp ## encoding: [0x89,0xe5]
1665 ; FMA32-NEXT:    andl $-32, %esp ## encoding: [0x83,0xe4,0xe0]
1666 ; FMA32-NEXT:    subl $32, %esp ## encoding: [0x83,0xec,0x20]
1667 ; FMA32-NEXT:    vfmadd213pd 8(%ebp), %ymm2, %ymm0 ## encoding: [0xc4,0xe2,0xed,0xa8,0x45,0x08]
1668 ; FMA32-NEXT:    ## ymm0 = (ymm2 * ymm0) + mem
1669 ; FMA32-NEXT:    vfmadd213pd 40(%ebp), %ymm3, %ymm1 ## encoding: [0xc4,0xe2,0xe5,0xa8,0x4d,0x28]
1670 ; FMA32-NEXT:    ## ymm1 = (ymm3 * ymm1) + mem
1671 ; FMA32-NEXT:    movl %ebp, %esp ## encoding: [0x89,0xec]
1672 ; FMA32-NEXT:    popl %ebp ## encoding: [0x5d]
1673 ; FMA32-NEXT:    retl ## encoding: [0xc3]
1674 ;
1675 ; FMA64-LABEL: test_v8f64:
1676 ; FMA64:       ## %bb.0: ## %entry
1677 ; FMA64-NEXT:    vfmadd213pd %ymm4, %ymm2, %ymm0 ## encoding: [0xc4,0xe2,0xed,0xa8,0xc4]
1678 ; FMA64-NEXT:    ## ymm0 = (ymm2 * ymm0) + ymm4
1679 ; FMA64-NEXT:    vfmadd213pd %ymm5, %ymm3, %ymm1 ## encoding: [0xc4,0xe2,0xe5,0xa8,0xcd]
1680 ; FMA64-NEXT:    ## ymm1 = (ymm3 * ymm1) + ymm5
1681 ; FMA64-NEXT:    retq ## encoding: [0xc3]
1682 ;
1683 ; FMACALL64-LABEL: test_v8f64:
1684 ; FMACALL64:       ## %bb.0: ## %entry
1685 ; FMACALL64-NEXT:    subq $152, %rsp ## encoding: [0x48,0x81,0xec,0x98,0x00,0x00,0x00]
1686 ; FMACALL64-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1687 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x7c,0x24,0x70]
1688 ; FMACALL64-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1689 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x74,0x24,0x20]
1690 ; FMACALL64-NEXT:    movaps %xmm5, (%rsp) ## 16-byte Spill
1691 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x2c,0x24]
1692 ; FMACALL64-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1693 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x64,0x24,0x10]
1694 ; FMACALL64-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1695 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x5c,0x24,0x60]
1696 ; FMACALL64-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1697 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x54,0x24,0x50]
1698 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1699 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x40]
1700 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1701 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x84,0x24,0x80,0x00,0x00,0x00]
1702 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xa0,0x00,0x00,0x00]
1703 ; FMACALL64-NEXT:    movaps %xmm4, %xmm1 ## encoding: [0x0f,0x28,0xcc]
1704 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1705 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1706 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1707 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x30]
1708 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1709 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x84,0x24,0x80,0x00,0x00,0x00]
1710 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1711 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1712 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1713 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1714 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1715 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1716 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xa0,0x00,0x00,0x00]
1717 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1718 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1719 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1720 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1721 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1722 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x30]
1723 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1724 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1725 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1726 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x30]
1727 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1728 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
1729 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
1730 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
1731 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
1732 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1733 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1734 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1735 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x10]
1736 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1737 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x40]
1738 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1739 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1740 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
1741 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
1742 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1743 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1744 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xb0,0x00,0x00,0x00]
1745 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1746 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1747 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1748 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1749 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1750 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1751 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1752 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1753 ; FMACALL64-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1754 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x4c,0x24,0x10]
1755 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1756 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
1757 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1758 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
1759 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
1760 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1761 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1762 ; FMACALL64-NEXT:    movaps %xmm0, (%rsp) ## 16-byte Spill
1763 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x04,0x24]
1764 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1765 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x50]
1766 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1767 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1768 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1769 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x20]
1770 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1771 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1772 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xc0,0x00,0x00,0x00]
1773 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1774 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1775 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1776 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1777 ; FMACALL64-NEXT:    movaps (%rsp), %xmm1 ## 16-byte Reload
1778 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x0c,0x24]
1779 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm1 ## encoding: [0x0f,0x16,0xc8]
1780 ; FMACALL64-NEXT:    ## xmm1 = xmm1[0],xmm0[0]
1781 ; FMACALL64-NEXT:    movaps %xmm1, (%rsp) ## 16-byte Spill
1782 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x0c,0x24]
1783 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1784 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x60]
1785 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1786 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x70]
1787 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
1788 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1789 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1790 ; FMACALL64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
1791 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x29,0x44,0x24,0x20]
1792 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1793 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x60]
1794 ; FMACALL64-NEXT:    movhlps %xmm0, %xmm0 ## encoding: [0x0f,0x12,0xc0]
1795 ; FMACALL64-NEXT:    ## xmm0 = xmm0[1,1]
1796 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1797 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x70]
1798 ; FMACALL64-NEXT:    movhlps %xmm1, %xmm1 ## encoding: [0x0f,0x12,0xc9]
1799 ; FMACALL64-NEXT:    ## xmm1 = xmm1[1,1]
1800 ; FMACALL64-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm2 ## encoding: [0x0f,0x28,0x94,0x24,0xd0,0x00,0x00,0x00]
1801 ; FMACALL64-NEXT:    movhlps %xmm2, %xmm2 ## encoding: [0x0f,0x12,0xd2]
1802 ; FMACALL64-NEXT:    ## xmm2 = xmm2[1,1]
1803 ; FMACALL64-NEXT:    callq _fma ## encoding: [0xe8,A,A,A,A]
1804 ; FMACALL64-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: reloc_branch_4byte_pcrel
1805 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Reload
1806 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x5c,0x24,0x20]
1807 ; FMACALL64-NEXT:    movlhps %xmm0, %xmm3 ## encoding: [0x0f,0x16,0xd8]
1808 ; FMACALL64-NEXT:    ## xmm3 = xmm3[0],xmm0[0]
1809 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
1810 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x44,0x24,0x30]
1811 ; FMACALL64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
1812 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x4c,0x24,0x10]
1813 ; FMACALL64-NEXT:    movaps (%rsp), %xmm2 ## 16-byte Reload
1814 ; FMACALL64-NEXT:    ## encoding: [0x0f,0x28,0x14,0x24]
1815 ; FMACALL64-NEXT:    addq $152, %rsp ## encoding: [0x48,0x81,0xc4,0x98,0x00,0x00,0x00]
1816 ; FMACALL64-NEXT:    retq ## encoding: [0xc3]
1817 ;
1818 ; AVX512-LABEL: test_v8f64:
1819 ; AVX512:       ## %bb.0: ## %entry
1820 ; AVX512-NEXT:    vfmadd213pd %zmm2, %zmm1, %zmm0 ## encoding: [0x62,0xf2,0xf5,0x48,0xa8,0xc2]
1821 ; AVX512-NEXT:    ## zmm0 = (zmm1 * zmm0) + zmm2
1822 ; AVX512-NEXT:    retq ## encoding: [0xc3]
1823 ;
1824 ; AVX512VL-LABEL: test_v8f64:
1825 ; AVX512VL:       ## %bb.0: ## %entry
1826 ; AVX512VL-NEXT:    vfmadd213pd %zmm2, %zmm1, %zmm0 ## encoding: [0x62,0xf2,0xf5,0x48,0xa8,0xc2]
1827 ; AVX512VL-NEXT:    ## zmm0 = (zmm1 * zmm0) + zmm2
1828 ; AVX512VL-NEXT:    retq ## encoding: [0xc3]
1829 ;
1830 ; FMACALL32_BDVER2-LABEL: test_v8f64:
1831 ; FMACALL32_BDVER2:       ## %bb.0: ## %entry
1832 ; FMACALL32_BDVER2-NEXT:    pushl %ebp ## encoding: [0x55]
1833 ; FMACALL32_BDVER2-NEXT:    movl %esp, %ebp ## encoding: [0x89,0xe5]
1834 ; FMACALL32_BDVER2-NEXT:    andl $-32, %esp ## encoding: [0x83,0xe4,0xe0]
1835 ; FMACALL32_BDVER2-NEXT:    subl $352, %esp ## encoding: [0x81,0xec,0x60,0x01,0x00,0x00]
1836 ; FMACALL32_BDVER2-NEXT:    ## imm = 0x160
1837 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1838 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x84,0x24,0xe0,0x00,0x00,0x00]
1839 ; FMACALL32_BDVER2-NEXT:    vmovaps 56(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x38]
1840 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1841 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x94,0x24,0x00,0x01,0x00,0x00]
1842 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm3, %xmm2 ## encoding: [0xc4,0xe3,0x7d,0x19,0xda,0x01]
1843 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm3, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1844 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x9c,0x24,0xc0,0x00,0x00,0x00]
1845 ; FMACALL32_BDVER2-NEXT:    vmovaps %ymm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 32-byte Spill
1846 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x29,0x8c,0x24,0xa0,0x00,0x00,0x00]
1847 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1848 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x94,0x24,0x30,0x01,0x00,0x00]
1849 ; FMACALL32_BDVER2-NEXT:    vmovlps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x13,0x44,0x24,0x10]
1850 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm1, %xmm0 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc8,0x01]
1851 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1852 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x84,0x24,0x40,0x01,0x00,0x00]
1853 ; FMACALL32_BDVER2-NEXT:    vmovlhps %xmm2, %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0xc2]
1854 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],xmm2[0]
1855 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1856 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1857 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1858 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1859 ; FMACALL32_BDVER2-NEXT:    vmovaps 40(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x28]
1860 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1861 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x94,0x00,0x00,0x00]
1862 ; FMACALL32_BDVER2-NEXT:    vmovhps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x17,0x44,0x24,0x10]
1863 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1864 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0xa0,0x00,0x00,0x00]
1865 ; FMACALL32_BDVER2-NEXT:    vunpckhpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1866 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x15,0x84,0x24,0xc0,0x00,0x00,0x00]
1867 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[1],mem[1]
1868 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1869 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1870 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1871 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1872 ; FMACALL32_BDVER2-NEXT:    vmovaps 40(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x28]
1873 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1874 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0x88,0x00,0x00,0x00]
1875 ; FMACALL32_BDVER2-NEXT:    vmovlps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x13,0x44,0x24,0x10]
1876 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1877 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0xa0,0x00,0x00,0x00]
1878 ; FMACALL32_BDVER2-NEXT:    vunpcklpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1879 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x14,0x84,0x24,0xc0,0x00,0x00,0x00]
1880 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0]
1881 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1882 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1883 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1884 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1885 ; FMACALL32_BDVER2-NEXT:    vmovaps 24(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x18]
1886 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1887 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xc0,0x00,0x00,0x00]
1888 ; FMACALL32_BDVER2-NEXT:    vmovhps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x17,0x44,0x24,0x10]
1889 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1890 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0x00,0x01,0x00,0x00]
1891 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm1 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc1,0x01]
1892 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1893 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0xe0,0x00,0x00,0x00]
1894 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1895 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x4c,0x24,0x20]
1896 ; FMACALL32_BDVER2-NEXT:    vextractf128 $1, %ymm0, %xmm0 ## encoding: [0xc4,0xe3,0x7d,0x19,0xc0,0x01]
1897 ; FMACALL32_BDVER2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 16-byte Spill
1898 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x29,0x44,0x24,0x30]
1899 ; FMACALL32_BDVER2-NEXT:    vunpckhpd %xmm1, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0x15,0xc1]
1900 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[1],xmm1[1]
1901 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1902 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1903 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1904 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1905 ; FMACALL32_BDVER2-NEXT:    vmovaps 24(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x18]
1906 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1907 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xbc,0x24,0xa0,0x00,0x00,0x00]
1908 ; FMACALL32_BDVER2-NEXT:    vmovlps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x13,0x44,0x24,0x10]
1909 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1910 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x44,0x24,0x30]
1911 ; FMACALL32_BDVER2-NEXT:    vunpcklpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1912 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x14,0x44,0x24,0x20]
1913 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0]
1914 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1915 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1916 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1917 ; FMACALL32_BDVER2-NEXT:    vmovaps 8(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x08]
1918 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1919 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x30]
1920 ; FMACALL32_BDVER2-NEXT:    vmovhps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x17,0x44,0x24,0x10]
1921 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1922 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0xe0,0x00,0x00,0x00]
1923 ; FMACALL32_BDVER2-NEXT:    vunpckhpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1924 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x15,0x84,0x24,0x00,0x01,0x00,0x00]
1925 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[1],mem[1]
1926 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1927 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1928 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1929 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1930 ; FMACALL32_BDVER2-NEXT:    vmovaps 8(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x08]
1931 ; FMACALL32_BDVER2-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
1932 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x7c,0x24,0x20]
1933 ; FMACALL32_BDVER2-NEXT:    vmovlps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x13,0x44,0x24,0x10]
1934 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %ymm0 ## 32-byte Reload
1935 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x84,0x24,0xe0,0x00,0x00,0x00]
1936 ; FMACALL32_BDVER2-NEXT:    vunpcklpd {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1937 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf9,0x14,0x84,0x24,0x00,0x01,0x00,0x00]
1938 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0],mem[0]
1939 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1940 ; FMACALL32_BDVER2-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
1941 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1942 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1943 ; FMACALL32_BDVER2-NEXT:    vmovaps 56(%ebp), %xmm0 ## encoding: [0xc5,0xf8,0x28,0x45,0x38]
1944 ; FMACALL32_BDVER2-NEXT:    vmovhps %xmm0, {{[0-9]+}}(%esp) ## encoding: [0xc5,0xf8,0x17,0x44,0x24,0x10]
1945 ; FMACALL32_BDVER2-NEXT:    vmovaps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 ## 16-byte Reload
1946 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x84,0x24,0x30,0x01,0x00,0x00]
1947 ; FMACALL32_BDVER2-NEXT:    vmovlps {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 16-byte Folded Reload
1948 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xc5,0xf8,0x12,0x84,0x24,0x48,0x01,0x00,0x00]
1949 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0,1],xmm0[2,3]
1950 ; FMACALL32_BDVER2-NEXT:    vmovups %xmm0, (%esp) ## encoding: [0xc5,0xf8,0x11,0x04,0x24]
1951 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x60]
1952 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1953 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x20]
1954 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x58]
1955 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1956 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0x6c,0x24,0x30]
1957 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x50]
1958 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1959 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xa0,0x00,0x00,0x00]
1960 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x48]
1961 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1962 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0xc0,0x00,0x00,0x00]
1963 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x9c,0x24,0x80,0x00,0x00,0x00]
1964 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1965 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x88,0x00,0x00,0x00]
1966 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x78]
1967 ; FMACALL32_BDVER2-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
1968 ; FMACALL32_BDVER2-NEXT:    ## encoding: [0xdb,0xac,0x24,0x94,0x00,0x00,0x00]
1969 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x70]
1970 ; FMACALL32_BDVER2-NEXT:    calll _fma ## encoding: [0xe8,A,A,A,A]
1971 ; FMACALL32_BDVER2-NEXT:    ## fixup A - offset: 1, value: _fma-4, kind: FK_PCRel_4
1972 ; FMACALL32_BDVER2-NEXT:    fstpl {{[0-9]+}}(%esp) ## encoding: [0xdd,0x5c,0x24,0x68]
1973 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm0 ## encoding: [0xc5,0xfb,0x10,0x44,0x24,0x60]
1974 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = mem[0],zero
1975 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfb,0x10,0x4c,0x24,0x50]
1976 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero
1977 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x16,0x44,0x24,0x58]
1978 ; FMACALL32_BDVER2-NEXT:    ## xmm0 = xmm0[0,1],mem[0,1]
1979 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc5,0xf0,0x16,0x4c,0x24,0x48]
1980 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0,1]
1981 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm2 ## encoding: [0xc5,0xfb,0x10,0x54,0x24,0x70]
1982 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = mem[0],zero
1983 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm2, %xmm2 ## encoding: [0xc5,0xe8,0x16,0x54,0x24,0x68]
1984 ; FMACALL32_BDVER2-NEXT:    ## xmm2 = xmm2[0,1],mem[0,1]
1985 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 ## encoding: [0xc4,0xe3,0x7d,0x18,0xc1,0x01]
1986 ; FMACALL32_BDVER2-NEXT:    vmovsd {{[0-9]+}}(%esp), %xmm1 ## encoding: [0xc5,0xfb,0x10,0x8c,0x24,0x80,0x00,0x00,0x00]
1987 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = mem[0],zero
1988 ; FMACALL32_BDVER2-NEXT:    vmovhps {{[0-9]+}}(%esp), %xmm1, %xmm1 ## encoding: [0xc5,0xf0,0x16,0x4c,0x24,0x78]
1989 ; FMACALL32_BDVER2-NEXT:    ## xmm1 = xmm1[0,1],mem[0,1]
1990 ; FMACALL32_BDVER2-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1 ## encoding: [0xc4,0xe3,0x75,0x18,0xca,0x01]
1991 ; FMACALL32_BDVER2-NEXT:    movl %ebp, %esp ## encoding: [0x89,0xec]
1992 ; FMACALL32_BDVER2-NEXT:    popl %ebp ## encoding: [0x5d]
1993 ; FMACALL32_BDVER2-NEXT:    retl ## encoding: [0xc3]
1994 entry:
1995   %call = call <8 x double> @llvm.fma.v8f64(<8 x double> %a, <8 x double> %b, <8 x double> %c)
1996   ret <8 x double> %call
1997 }
1998
1999 declare float @llvm.fma.f32(float, float, float)
2000 declare double @llvm.fma.f64(double, double, double)
2001 declare x86_fp80 @llvm.fma.f80(x86_fp80, x86_fp80, x86_fp80)
2002
2003 declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>)
2004 declare <8 x float> @llvm.fma.v8f32(<8 x float>, <8 x float>, <8 x float>)
2005 declare <16 x float> @llvm.fma.v16f32(<16 x float>, <16 x float>, <16 x float>)
2006
2007 declare <2 x double> @llvm.fma.v2f64(<2 x double>, <2 x double>, <2 x double>)
2008 declare <4 x double> @llvm.fma.v4f64(<4 x double>, <4 x double>, <4 x double>)
2009 declare <8 x double> @llvm.fma.v8f64(<8 x double>, <8 x double>, <8 x double>)
2010
2011 attributes #0 = { nounwind }