llvm/test/CodeGen/X86/x86-interrupt_cc.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
   2 ; RUN: llc -verify-machineinstrs -mtriple=x86_64-apple-macosx -show-mc-encoding -mcpu=knl < %s | FileCheck %s -check-prefix=CHECK64-KNL
   3 ; RUN: llc -verify-machineinstrs -mtriple=x86_64-apple-macosx -show-mc-encoding -mcpu=skx < %s | FileCheck %s -check-prefix=CHECK64-SKX
   4 ; RUN: llc -verify-machineinstrs -mtriple=i386-apple-macosx -show-mc-encoding -mcpu=knl < %s | FileCheck %s -check-prefix=CHECK32-KNL
   5 ; RUN: llc -verify-machineinstrs -mtriple=i386-apple-macosx -show-mc-encoding -mcpu=skx < %s | FileCheck %s -check-prefix=CHECK32-SKX
   6
   7 ; Make sure we spill the high numbered zmm registers and K registers with the right encoding.
   8
   9 define x86_intrcc void @foo(ptr byval(i8) %frame) {
  10 ; CHECK64-KNL-LABEL: foo:
  11 ; CHECK64-KNL:       ## %bb.0:
  12 ; CHECK64-KNL-NEXT:    pushq %rax ## encoding: [0x50]
  13 ; CHECK64-KNL-NEXT:    .cfi_def_cfa_offset 16
  14 ; CHECK64-KNL-NEXT:    pushq %r11 ## encoding: [0x41,0x53]
  15 ; CHECK64-KNL-NEXT:    .cfi_def_cfa_offset 24
  16 ; CHECK64-KNL-NEXT:    pushq %r10 ## encoding: [0x41,0x52]
  17 ; CHECK64-KNL-NEXT:    .cfi_def_cfa_offset 32
  18 ; CHECK64-KNL-NEXT:    pushq %r9 ## encoding: [0x41,0x51]
  19 ; CHECK64-KNL-NEXT:    .cfi_def_cfa_offset 40
  20 ; CHECK64-KNL-NEXT:    pushq %r8 ## encoding: [0x41,0x50]
  21 ; CHECK64-KNL-NEXT:    .cfi_def_cfa_offset 48
  22 ; CHECK64-KNL-NEXT:    pushq %rdi ## encoding: [0x57]
  23 ; CHECK64-KNL-NEXT:    .cfi_def_cfa_offset 56
  24 ; CHECK64-KNL-NEXT:    pushq %rsi ## encoding: [0x56]
  25 ; CHECK64-KNL-NEXT:    .cfi_def_cfa_offset 64
  26 ; CHECK64-KNL-NEXT:    pushq %rdx ## encoding: [0x52]
  27 ; CHECK64-KNL-NEXT:    .cfi_def_cfa_offset 72
  28 ; CHECK64-KNL-NEXT:    pushq %rcx ## encoding: [0x51]
  29 ; CHECK64-KNL-NEXT:    .cfi_def_cfa_offset 80
  30 ; CHECK64-KNL-NEXT:    subq $2096, %rsp ## encoding: [0x48,0x81,0xec,0x30,0x08,0x00,0x00]
  31 ; CHECK64-KNL-NEXT:    ## imm = 0x830
  32 ; CHECK64-KNL-NEXT:    kmovw %k7, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
  33 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0xbc,0x24,0x2e,0x08,0x00,0x00]
  34 ; CHECK64-KNL-NEXT:    kmovw %k6, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
  35 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0xb4,0x24,0x2c,0x08,0x00,0x00]
  36 ; CHECK64-KNL-NEXT:    kmovw %k5, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
  37 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0xac,0x24,0x2a,0x08,0x00,0x00]
  38 ; CHECK64-KNL-NEXT:    kmovw %k4, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
  39 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0xa4,0x24,0x28,0x08,0x00,0x00]
  40 ; CHECK64-KNL-NEXT:    kmovw %k3, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
  41 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x9c,0x24,0x26,0x08,0x00,0x00]
  42 ; CHECK64-KNL-NEXT:    kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
  43 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x94,0x24,0x24,0x08,0x00,0x00]
  44 ; CHECK64-KNL-NEXT:    kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
  45 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x8c,0x24,0x22,0x08,0x00,0x00]
  46 ; CHECK64-KNL-NEXT:    kmovw %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
  47 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x84,0x24,0x20,0x08,0x00,0x00]
  48 ; CHECK64-KNL-NEXT:    vmovups %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  49 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x7c,0x24,0x1f]
  50 ; CHECK64-KNL-NEXT:    vmovups %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  51 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x74,0x24,0x1e]
  52 ; CHECK64-KNL-NEXT:    vmovups %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  53 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x6c,0x24,0x1d]
  54 ; CHECK64-KNL-NEXT:    vmovups %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  55 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x64,0x24,0x1c]
  56 ; CHECK64-KNL-NEXT:    vmovups %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  57 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x5c,0x24,0x1b]
  58 ; CHECK64-KNL-NEXT:    vmovups %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  59 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x54,0x24,0x1a]
  60 ; CHECK64-KNL-NEXT:    vmovups %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  61 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x4c,0x24,0x19]
  62 ; CHECK64-KNL-NEXT:    vmovups %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  63 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x44,0x24,0x18]
  64 ; CHECK64-KNL-NEXT:    vmovups %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  65 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x7c,0x24,0x17]
  66 ; CHECK64-KNL-NEXT:    vmovups %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  67 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x74,0x24,0x16]
  68 ; CHECK64-KNL-NEXT:    vmovups %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  69 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x6c,0x24,0x15]
  70 ; CHECK64-KNL-NEXT:    vmovups %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  71 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x64,0x24,0x14]
  72 ; CHECK64-KNL-NEXT:    vmovups %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  73 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x5c,0x24,0x13]
  74 ; CHECK64-KNL-NEXT:    vmovups %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  75 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x54,0x24,0x12]
  76 ; CHECK64-KNL-NEXT:    vmovups %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  77 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x4c,0x24,0x11]
  78 ; CHECK64-KNL-NEXT:    vmovups %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  79 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x44,0x24,0x10]
  80 ; CHECK64-KNL-NEXT:    vmovups %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  81 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x7c,0x24,0x0f]
  82 ; CHECK64-KNL-NEXT:    vmovups %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  83 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x74,0x24,0x0e]
  84 ; CHECK64-KNL-NEXT:    vmovups %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  85 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x6c,0x24,0x0d]
  86 ; CHECK64-KNL-NEXT:    vmovups %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  87 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x64,0x24,0x0c]
  88 ; CHECK64-KNL-NEXT:    vmovups %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  89 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x5c,0x24,0x0b]
  90 ; CHECK64-KNL-NEXT:    vmovups %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  91 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x54,0x24,0x0a]
  92 ; CHECK64-KNL-NEXT:    vmovups %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  93 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x4c,0x24,0x09]
  94 ; CHECK64-KNL-NEXT:    vmovups %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  95 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x44,0x24,0x08]
  96 ; CHECK64-KNL-NEXT:    vmovups %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  97 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x7c,0x24,0x07]
  98 ; CHECK64-KNL-NEXT:    vmovups %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
  99 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x74,0x24,0x06]
 100 ; CHECK64-KNL-NEXT:    vmovups %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 101 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x6c,0x24,0x05]
 102 ; CHECK64-KNL-NEXT:    vmovups %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 103 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x64,0x24,0x04]
 104 ; CHECK64-KNL-NEXT:    vmovups %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 105 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x5c,0x24,0x03]
 106 ; CHECK64-KNL-NEXT:    vmovups %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 107 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x54,0x24,0x02]
 108 ; CHECK64-KNL-NEXT:    vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 109 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x4c,0x24,0x01]
 110 ; CHECK64-KNL-NEXT:    vmovups %zmm0, (%rsp) ## 64-byte Spill
 111 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x04,0x24]
 112 ; CHECK64-KNL-NEXT:    .cfi_def_cfa_offset 2176
 113 ; CHECK64-KNL-NEXT:    .cfi_offset %rcx, -80
 114 ; CHECK64-KNL-NEXT:    .cfi_offset %rdx, -72
 115 ; CHECK64-KNL-NEXT:    .cfi_offset %rsi, -64
 116 ; CHECK64-KNL-NEXT:    .cfi_offset %rdi, -56
 117 ; CHECK64-KNL-NEXT:    .cfi_offset %r8, -48
 118 ; CHECK64-KNL-NEXT:    .cfi_offset %r9, -40
 119 ; CHECK64-KNL-NEXT:    .cfi_offset %r10, -32
 120 ; CHECK64-KNL-NEXT:    .cfi_offset %r11, -24
 121 ; CHECK64-KNL-NEXT:    .cfi_offset %rax, -16
 122 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm0, -2176
 123 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm1, -2112
 124 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm2, -2048
 125 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm3, -1984
 126 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm4, -1920
 127 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm5, -1856
 128 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm6, -1792
 129 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm7, -1728
 130 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm8, -1664
 131 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm9, -1600
 132 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm10, -1536
 133 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm11, -1472
 134 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm12, -1408
 135 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm13, -1344
 136 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm14, -1280
 137 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm15, -1216
 138 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm16, -1152
 139 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm17, -1088
 140 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm18, -1024
 141 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm19, -960
 142 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm20, -896
 143 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm21, -832
 144 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm22, -768
 145 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm23, -704
 146 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm24, -640
 147 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm25, -576
 148 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm26, -512
 149 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm27, -448
 150 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm28, -384
 151 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm29, -320
 152 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm30, -256
 153 ; CHECK64-KNL-NEXT:    .cfi_offset %xmm31, -192
 154 ; CHECK64-KNL-NEXT:    .cfi_offset %k0, -96
 155 ; CHECK64-KNL-NEXT:    .cfi_offset %k1, -94
 156 ; CHECK64-KNL-NEXT:    .cfi_offset %k2, -92
 157 ; CHECK64-KNL-NEXT:    .cfi_offset %k3, -90
 158 ; CHECK64-KNL-NEXT:    .cfi_offset %k4, -88
 159 ; CHECK64-KNL-NEXT:    .cfi_offset %k5, -86
 160 ; CHECK64-KNL-NEXT:    .cfi_offset %k6, -84
 161 ; CHECK64-KNL-NEXT:    .cfi_offset %k7, -82
 162 ; CHECK64-KNL-NEXT:    cld ## encoding: [0xfc]
 163 ; CHECK64-KNL-NEXT:    callq _bar ## encoding: [0xe8,A,A,A,A]
 164 ; CHECK64-KNL-NEXT:    ## fixup A - offset: 1, value: _bar-4, kind: reloc_branch_4byte_pcrel
 165 ; CHECK64-KNL-NEXT:    vmovups (%rsp), %zmm0 ## 64-byte Reload
 166 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x04,0x24]
 167 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 ## 64-byte Reload
 168 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x4c,0x24,0x01]
 169 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 ## 64-byte Reload
 170 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x54,0x24,0x02]
 171 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 ## 64-byte Reload
 172 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x5c,0x24,0x03]
 173 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 ## 64-byte Reload
 174 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x64,0x24,0x04]
 175 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 ## 64-byte Reload
 176 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x6c,0x24,0x05]
 177 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 ## 64-byte Reload
 178 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x74,0x24,0x06]
 179 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 ## 64-byte Reload
 180 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x7c,0x24,0x07]
 181 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 ## 64-byte Reload
 182 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x44,0x24,0x08]
 183 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 ## 64-byte Reload
 184 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x4c,0x24,0x09]
 185 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 ## 64-byte Reload
 186 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x54,0x24,0x0a]
 187 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 ## 64-byte Reload
 188 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x5c,0x24,0x0b]
 189 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 ## 64-byte Reload
 190 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x64,0x24,0x0c]
 191 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 ## 64-byte Reload
 192 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x6c,0x24,0x0d]
 193 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 ## 64-byte Reload
 194 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x74,0x24,0x0e]
 195 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 ## 64-byte Reload
 196 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x7c,0x24,0x0f]
 197 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 ## 64-byte Reload
 198 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x44,0x24,0x10]
 199 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 ## 64-byte Reload
 200 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x4c,0x24,0x11]
 201 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 ## 64-byte Reload
 202 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x54,0x24,0x12]
 203 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 ## 64-byte Reload
 204 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x5c,0x24,0x13]
 205 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 ## 64-byte Reload
 206 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x64,0x24,0x14]
 207 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 ## 64-byte Reload
 208 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x6c,0x24,0x15]
 209 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 ## 64-byte Reload
 210 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x74,0x24,0x16]
 211 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 ## 64-byte Reload
 212 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x7c,0x24,0x17]
 213 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 ## 64-byte Reload
 214 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x44,0x24,0x18]
 215 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 ## 64-byte Reload
 216 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x4c,0x24,0x19]
 217 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 ## 64-byte Reload
 218 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x54,0x24,0x1a]
 219 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 ## 64-byte Reload
 220 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x5c,0x24,0x1b]
 221 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 ## 64-byte Reload
 222 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x64,0x24,0x1c]
 223 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 ## 64-byte Reload
 224 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x6c,0x24,0x1d]
 225 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 ## 64-byte Reload
 226 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x74,0x24,0x1e]
 227 ; CHECK64-KNL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 ## 64-byte Reload
 228 ; CHECK64-KNL-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x7c,0x24,0x1f]
 229 ; CHECK64-KNL-NEXT:    kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
 230 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x84,0x24,0x20,0x08,0x00,0x00]
 231 ; CHECK64-KNL-NEXT:    kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
 232 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x8c,0x24,0x22,0x08,0x00,0x00]
 233 ; CHECK64-KNL-NEXT:    kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k2 ## 2-byte Reload
 234 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x94,0x24,0x24,0x08,0x00,0x00]
 235 ; CHECK64-KNL-NEXT:    kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k3 ## 2-byte Reload
 236 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x9c,0x24,0x26,0x08,0x00,0x00]
 237 ; CHECK64-KNL-NEXT:    kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k4 ## 2-byte Reload
 238 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0xa4,0x24,0x28,0x08,0x00,0x00]
 239 ; CHECK64-KNL-NEXT:    kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k5 ## 2-byte Reload
 240 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0xac,0x24,0x2a,0x08,0x00,0x00]
 241 ; CHECK64-KNL-NEXT:    kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k6 ## 2-byte Reload
 242 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0xb4,0x24,0x2c,0x08,0x00,0x00]
 243 ; CHECK64-KNL-NEXT:    kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k7 ## 2-byte Reload
 244 ; CHECK64-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0xbc,0x24,0x2e,0x08,0x00,0x00]
 245 ; CHECK64-KNL-NEXT:    addq $2096, %rsp ## encoding: [0x48,0x81,0xc4,0x30,0x08,0x00,0x00]
 246 ; CHECK64-KNL-NEXT:    ## imm = 0x830
 247 ; CHECK64-KNL-NEXT:    popq %rcx ## encoding: [0x59]
 248 ; CHECK64-KNL-NEXT:    popq %rdx ## encoding: [0x5a]
 249 ; CHECK64-KNL-NEXT:    popq %rsi ## encoding: [0x5e]
 250 ; CHECK64-KNL-NEXT:    popq %rdi ## encoding: [0x5f]
 251 ; CHECK64-KNL-NEXT:    popq %r8 ## encoding: [0x41,0x58]
 252 ; CHECK64-KNL-NEXT:    popq %r9 ## encoding: [0x41,0x59]
 253 ; CHECK64-KNL-NEXT:    popq %r10 ## encoding: [0x41,0x5a]
 254 ; CHECK64-KNL-NEXT:    popq %r11 ## encoding: [0x41,0x5b]
 255 ; CHECK64-KNL-NEXT:    popq %rax ## encoding: [0x58]
 256 ; CHECK64-KNL-NEXT:    iretq ## encoding: [0x48,0xcf]
 257 ;
 258 ; CHECK64-SKX-LABEL: foo:
 259 ; CHECK64-SKX:       ## %bb.0:
 260 ; CHECK64-SKX-NEXT:    pushq %rax ## encoding: [0x50]
 261 ; CHECK64-SKX-NEXT:    .cfi_def_cfa_offset 16
 262 ; CHECK64-SKX-NEXT:    pushq %r11 ## encoding: [0x41,0x53]
 263 ; CHECK64-SKX-NEXT:    .cfi_def_cfa_offset 24
 264 ; CHECK64-SKX-NEXT:    pushq %r10 ## encoding: [0x41,0x52]
 265 ; CHECK64-SKX-NEXT:    .cfi_def_cfa_offset 32
 266 ; CHECK64-SKX-NEXT:    pushq %r9 ## encoding: [0x41,0x51]
 267 ; CHECK64-SKX-NEXT:    .cfi_def_cfa_offset 40
 268 ; CHECK64-SKX-NEXT:    pushq %r8 ## encoding: [0x41,0x50]
 269 ; CHECK64-SKX-NEXT:    .cfi_def_cfa_offset 48
 270 ; CHECK64-SKX-NEXT:    pushq %rdi ## encoding: [0x57]
 271 ; CHECK64-SKX-NEXT:    .cfi_def_cfa_offset 56
 272 ; CHECK64-SKX-NEXT:    pushq %rsi ## encoding: [0x56]
 273 ; CHECK64-SKX-NEXT:    .cfi_def_cfa_offset 64
 274 ; CHECK64-SKX-NEXT:    pushq %rdx ## encoding: [0x52]
 275 ; CHECK64-SKX-NEXT:    .cfi_def_cfa_offset 72
 276 ; CHECK64-SKX-NEXT:    pushq %rcx ## encoding: [0x51]
 277 ; CHECK64-SKX-NEXT:    .cfi_def_cfa_offset 80
 278 ; CHECK64-SKX-NEXT:    subq $2160, %rsp ## encoding: [0x48,0x81,0xec,0x70,0x08,0x00,0x00]
 279 ; CHECK64-SKX-NEXT:    ## imm = 0x870
 280 ; CHECK64-SKX-NEXT:    kmovq %k7, {{[-0-9]+}}(%r{{[sb]}}p) ## 8-byte Spill
 281 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0xbc,0x24,0x68,0x08,0x00,0x00]
 282 ; CHECK64-SKX-NEXT:    kmovq %k6, {{[-0-9]+}}(%r{{[sb]}}p) ## 8-byte Spill
 283 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0xb4,0x24,0x60,0x08,0x00,0x00]
 284 ; CHECK64-SKX-NEXT:    kmovq %k5, {{[-0-9]+}}(%r{{[sb]}}p) ## 8-byte Spill
 285 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0xac,0x24,0x58,0x08,0x00,0x00]
 286 ; CHECK64-SKX-NEXT:    kmovq %k4, {{[-0-9]+}}(%r{{[sb]}}p) ## 8-byte Spill
 287 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0xa4,0x24,0x50,0x08,0x00,0x00]
 288 ; CHECK64-SKX-NEXT:    kmovq %k3, {{[-0-9]+}}(%r{{[sb]}}p) ## 8-byte Spill
 289 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x9c,0x24,0x48,0x08,0x00,0x00]
 290 ; CHECK64-SKX-NEXT:    kmovq %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 8-byte Spill
 291 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x94,0x24,0x40,0x08,0x00,0x00]
 292 ; CHECK64-SKX-NEXT:    kmovq %k1, {{[-0-9]+}}(%r{{[sb]}}p) ## 8-byte Spill
 293 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x8c,0x24,0x38,0x08,0x00,0x00]
 294 ; CHECK64-SKX-NEXT:    kmovq %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 8-byte Spill
 295 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x84,0x24,0x30,0x08,0x00,0x00]
 296 ; CHECK64-SKX-NEXT:    vmovups %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 297 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x7c,0x24,0x1f]
 298 ; CHECK64-SKX-NEXT:    vmovups %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 299 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x74,0x24,0x1e]
 300 ; CHECK64-SKX-NEXT:    vmovups %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 301 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x6c,0x24,0x1d]
 302 ; CHECK64-SKX-NEXT:    vmovups %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 303 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x64,0x24,0x1c]
 304 ; CHECK64-SKX-NEXT:    vmovups %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 305 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x5c,0x24,0x1b]
 306 ; CHECK64-SKX-NEXT:    vmovups %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 307 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x54,0x24,0x1a]
 308 ; CHECK64-SKX-NEXT:    vmovups %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 309 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x4c,0x24,0x19]
 310 ; CHECK64-SKX-NEXT:    vmovups %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 311 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x11,0x44,0x24,0x18]
 312 ; CHECK64-SKX-NEXT:    vmovups %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 313 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x7c,0x24,0x17]
 314 ; CHECK64-SKX-NEXT:    vmovups %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 315 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x74,0x24,0x16]
 316 ; CHECK64-SKX-NEXT:    vmovups %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 317 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x6c,0x24,0x15]
 318 ; CHECK64-SKX-NEXT:    vmovups %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 319 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x64,0x24,0x14]
 320 ; CHECK64-SKX-NEXT:    vmovups %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 321 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x5c,0x24,0x13]
 322 ; CHECK64-SKX-NEXT:    vmovups %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 323 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x54,0x24,0x12]
 324 ; CHECK64-SKX-NEXT:    vmovups %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 325 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x4c,0x24,0x11]
 326 ; CHECK64-SKX-NEXT:    vmovups %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 327 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x11,0x44,0x24,0x10]
 328 ; CHECK64-SKX-NEXT:    vmovups %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 329 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x7c,0x24,0x0f]
 330 ; CHECK64-SKX-NEXT:    vmovups %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 331 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x74,0x24,0x0e]
 332 ; CHECK64-SKX-NEXT:    vmovups %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 333 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x6c,0x24,0x0d]
 334 ; CHECK64-SKX-NEXT:    vmovups %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 335 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x64,0x24,0x0c]
 336 ; CHECK64-SKX-NEXT:    vmovups %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 337 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x5c,0x24,0x0b]
 338 ; CHECK64-SKX-NEXT:    vmovups %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 339 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x54,0x24,0x0a]
 340 ; CHECK64-SKX-NEXT:    vmovups %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 341 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x4c,0x24,0x09]
 342 ; CHECK64-SKX-NEXT:    vmovups %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 343 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x11,0x44,0x24,0x08]
 344 ; CHECK64-SKX-NEXT:    vmovups %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 345 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x7c,0x24,0x07]
 346 ; CHECK64-SKX-NEXT:    vmovups %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 347 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x74,0x24,0x06]
 348 ; CHECK64-SKX-NEXT:    vmovups %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 349 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x6c,0x24,0x05]
 350 ; CHECK64-SKX-NEXT:    vmovups %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 351 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x64,0x24,0x04]
 352 ; CHECK64-SKX-NEXT:    vmovups %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 353 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x5c,0x24,0x03]
 354 ; CHECK64-SKX-NEXT:    vmovups %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 355 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x54,0x24,0x02]
 356 ; CHECK64-SKX-NEXT:    vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 64-byte Spill
 357 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x4c,0x24,0x01]
 358 ; CHECK64-SKX-NEXT:    vmovups %zmm0, (%rsp) ## 64-byte Spill
 359 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x04,0x24]
 360 ; CHECK64-SKX-NEXT:    .cfi_def_cfa_offset 2240
 361 ; CHECK64-SKX-NEXT:    .cfi_offset %rcx, -80
 362 ; CHECK64-SKX-NEXT:    .cfi_offset %rdx, -72
 363 ; CHECK64-SKX-NEXT:    .cfi_offset %rsi, -64
 364 ; CHECK64-SKX-NEXT:    .cfi_offset %rdi, -56
 365 ; CHECK64-SKX-NEXT:    .cfi_offset %r8, -48
 366 ; CHECK64-SKX-NEXT:    .cfi_offset %r9, -40
 367 ; CHECK64-SKX-NEXT:    .cfi_offset %r10, -32
 368 ; CHECK64-SKX-NEXT:    .cfi_offset %r11, -24
 369 ; CHECK64-SKX-NEXT:    .cfi_offset %rax, -16
 370 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm0, -2240
 371 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm1, -2176
 372 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm2, -2112
 373 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm3, -2048
 374 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm4, -1984
 375 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm5, -1920
 376 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm6, -1856
 377 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm7, -1792
 378 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm8, -1728
 379 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm9, -1664
 380 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm10, -1600
 381 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm11, -1536
 382 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm12, -1472
 383 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm13, -1408
 384 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm14, -1344
 385 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm15, -1280
 386 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm16, -1216
 387 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm17, -1152
 388 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm18, -1088
 389 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm19, -1024
 390 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm20, -960
 391 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm21, -896
 392 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm22, -832
 393 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm23, -768
 394 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm24, -704
 395 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm25, -640
 396 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm26, -576
 397 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm27, -512
 398 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm28, -448
 399 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm29, -384
 400 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm30, -320
 401 ; CHECK64-SKX-NEXT:    .cfi_offset %xmm31, -256
 402 ; CHECK64-SKX-NEXT:    .cfi_offset %k0, -144
 403 ; CHECK64-SKX-NEXT:    .cfi_offset %k1, -136
 404 ; CHECK64-SKX-NEXT:    .cfi_offset %k2, -128
 405 ; CHECK64-SKX-NEXT:    .cfi_offset %k3, -120
 406 ; CHECK64-SKX-NEXT:    .cfi_offset %k4, -112
 407 ; CHECK64-SKX-NEXT:    .cfi_offset %k5, -104
 408 ; CHECK64-SKX-NEXT:    .cfi_offset %k6, -96
 409 ; CHECK64-SKX-NEXT:    .cfi_offset %k7, -88
 410 ; CHECK64-SKX-NEXT:    cld ## encoding: [0xfc]
 411 ; CHECK64-SKX-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 412 ; CHECK64-SKX-NEXT:    callq _bar ## encoding: [0xe8,A,A,A,A]
 413 ; CHECK64-SKX-NEXT:    ## fixup A - offset: 1, value: _bar-4, kind: reloc_branch_4byte_pcrel
 414 ; CHECK64-SKX-NEXT:    vmovups (%rsp), %zmm0 ## 64-byte Reload
 415 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x04,0x24]
 416 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 ## 64-byte Reload
 417 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x4c,0x24,0x01]
 418 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 ## 64-byte Reload
 419 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x54,0x24,0x02]
 420 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 ## 64-byte Reload
 421 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x5c,0x24,0x03]
 422 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 ## 64-byte Reload
 423 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x64,0x24,0x04]
 424 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 ## 64-byte Reload
 425 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x6c,0x24,0x05]
 426 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 ## 64-byte Reload
 427 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x74,0x24,0x06]
 428 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 ## 64-byte Reload
 429 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x7c,0x24,0x07]
 430 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 ## 64-byte Reload
 431 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x44,0x24,0x08]
 432 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 ## 64-byte Reload
 433 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x4c,0x24,0x09]
 434 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 ## 64-byte Reload
 435 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x54,0x24,0x0a]
 436 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 ## 64-byte Reload
 437 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x5c,0x24,0x0b]
 438 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 ## 64-byte Reload
 439 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x64,0x24,0x0c]
 440 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 ## 64-byte Reload
 441 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x6c,0x24,0x0d]
 442 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 ## 64-byte Reload
 443 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x74,0x24,0x0e]
 444 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 ## 64-byte Reload
 445 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x71,0x7c,0x48,0x10,0x7c,0x24,0x0f]
 446 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 ## 64-byte Reload
 447 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x44,0x24,0x10]
 448 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 ## 64-byte Reload
 449 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x4c,0x24,0x11]
 450 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 ## 64-byte Reload
 451 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x54,0x24,0x12]
 452 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 ## 64-byte Reload
 453 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x5c,0x24,0x13]
 454 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 ## 64-byte Reload
 455 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x64,0x24,0x14]
 456 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 ## 64-byte Reload
 457 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x6c,0x24,0x15]
 458 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 ## 64-byte Reload
 459 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x74,0x24,0x16]
 460 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 ## 64-byte Reload
 461 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0xe1,0x7c,0x48,0x10,0x7c,0x24,0x17]
 462 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 ## 64-byte Reload
 463 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x44,0x24,0x18]
 464 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 ## 64-byte Reload
 465 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x4c,0x24,0x19]
 466 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 ## 64-byte Reload
 467 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x54,0x24,0x1a]
 468 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 ## 64-byte Reload
 469 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x5c,0x24,0x1b]
 470 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 ## 64-byte Reload
 471 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x64,0x24,0x1c]
 472 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 ## 64-byte Reload
 473 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x6c,0x24,0x1d]
 474 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 ## 64-byte Reload
 475 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x74,0x24,0x1e]
 476 ; CHECK64-SKX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 ## 64-byte Reload
 477 ; CHECK64-SKX-NEXT:    ## encoding: [0x62,0x61,0x7c,0x48,0x10,0x7c,0x24,0x1f]
 478 ; CHECK64-SKX-NEXT:    kmovq {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 8-byte Reload
 479 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x84,0x24,0x30,0x08,0x00,0x00]
 480 ; CHECK64-SKX-NEXT:    kmovq {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 8-byte Reload
 481 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x8c,0x24,0x38,0x08,0x00,0x00]
 482 ; CHECK64-SKX-NEXT:    kmovq {{[-0-9]+}}(%r{{[sb]}}p), %k2 ## 8-byte Reload
 483 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x94,0x24,0x40,0x08,0x00,0x00]
 484 ; CHECK64-SKX-NEXT:    kmovq {{[-0-9]+}}(%r{{[sb]}}p), %k3 ## 8-byte Reload
 485 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x9c,0x24,0x48,0x08,0x00,0x00]
 486 ; CHECK64-SKX-NEXT:    kmovq {{[-0-9]+}}(%r{{[sb]}}p), %k4 ## 8-byte Reload
 487 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0xa4,0x24,0x50,0x08,0x00,0x00]
 488 ; CHECK64-SKX-NEXT:    kmovq {{[-0-9]+}}(%r{{[sb]}}p), %k5 ## 8-byte Reload
 489 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0xac,0x24,0x58,0x08,0x00,0x00]
 490 ; CHECK64-SKX-NEXT:    kmovq {{[-0-9]+}}(%r{{[sb]}}p), %k6 ## 8-byte Reload
 491 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0xb4,0x24,0x60,0x08,0x00,0x00]
 492 ; CHECK64-SKX-NEXT:    kmovq {{[-0-9]+}}(%r{{[sb]}}p), %k7 ## 8-byte Reload
 493 ; CHECK64-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0xbc,0x24,0x68,0x08,0x00,0x00]
 494 ; CHECK64-SKX-NEXT:    addq $2160, %rsp ## encoding: [0x48,0x81,0xc4,0x70,0x08,0x00,0x00]
 495 ; CHECK64-SKX-NEXT:    ## imm = 0x870
 496 ; CHECK64-SKX-NEXT:    popq %rcx ## encoding: [0x59]
 497 ; CHECK64-SKX-NEXT:    popq %rdx ## encoding: [0x5a]
 498 ; CHECK64-SKX-NEXT:    popq %rsi ## encoding: [0x5e]
 499 ; CHECK64-SKX-NEXT:    popq %rdi ## encoding: [0x5f]
 500 ; CHECK64-SKX-NEXT:    popq %r8 ## encoding: [0x41,0x58]
 501 ; CHECK64-SKX-NEXT:    popq %r9 ## encoding: [0x41,0x59]
 502 ; CHECK64-SKX-NEXT:    popq %r10 ## encoding: [0x41,0x5a]
 503 ; CHECK64-SKX-NEXT:    popq %r11 ## encoding: [0x41,0x5b]
 504 ; CHECK64-SKX-NEXT:    popq %rax ## encoding: [0x58]
 505 ; CHECK64-SKX-NEXT:    iretq ## encoding: [0x48,0xcf]
 506 ;
 507 ; CHECK32-KNL-LABEL: foo:
 508 ; CHECK32-KNL:       ## %bb.0:
 509 ; CHECK32-KNL-NEXT:    pushl %ebp ## encoding: [0x55]
 510 ; CHECK32-KNL-NEXT:    .cfi_def_cfa_offset 8
 511 ; CHECK32-KNL-NEXT:    .cfi_offset %ebp, -8
 512 ; CHECK32-KNL-NEXT:    movl %esp, %ebp ## encoding: [0x89,0xe5]
 513 ; CHECK32-KNL-NEXT:    .cfi_def_cfa_register %ebp
 514 ; CHECK32-KNL-NEXT:    pushl %edx ## encoding: [0x52]
 515 ; CHECK32-KNL-NEXT:    pushl %ecx ## encoding: [0x51]
 516 ; CHECK32-KNL-NEXT:    pushl %eax ## encoding: [0x50]
 517 ; CHECK32-KNL-NEXT:    andl $-16, %esp ## encoding: [0x83,0xe4,0xf0]
 518 ; CHECK32-KNL-NEXT:    subl $560, %esp ## encoding: [0x81,0xec,0x30,0x02,0x00,0x00]
 519 ; CHECK32-KNL-NEXT:    ## imm = 0x230
 520 ; CHECK32-KNL-NEXT:    kmovw %k7, {{[-0-9]+}}(%e{{[sb]}}p) ## 2-byte Spill
 521 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x7d,0xf2]
 522 ; CHECK32-KNL-NEXT:    kmovw %k6, {{[-0-9]+}}(%e{{[sb]}}p) ## 2-byte Spill
 523 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x75,0xf0]
 524 ; CHECK32-KNL-NEXT:    kmovw %k5, {{[-0-9]+}}(%e{{[sb]}}p) ## 2-byte Spill
 525 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x6d,0xee]
 526 ; CHECK32-KNL-NEXT:    kmovw %k4, {{[-0-9]+}}(%e{{[sb]}}p) ## 2-byte Spill
 527 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x65,0xec]
 528 ; CHECK32-KNL-NEXT:    kmovw %k3, {{[-0-9]+}}(%e{{[sb]}}p) ## 2-byte Spill
 529 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x5d,0xea]
 530 ; CHECK32-KNL-NEXT:    kmovw %k2, {{[-0-9]+}}(%e{{[sb]}}p) ## 2-byte Spill
 531 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x55,0xe8]
 532 ; CHECK32-KNL-NEXT:    kmovw %k1, {{[-0-9]+}}(%e{{[sb]}}p) ## 2-byte Spill
 533 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x4d,0xe6]
 534 ; CHECK32-KNL-NEXT:    kmovw %k0, {{[-0-9]+}}(%e{{[sb]}}p) ## 2-byte Spill
 535 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x91,0x45,0xe4]
 536 ; CHECK32-KNL-NEXT:    vmovups %zmm7, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 537 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0xbd,0x88,0xff,0xff,0xff]
 538 ; CHECK32-KNL-NEXT:    vmovups %zmm6, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 539 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0xb5,0x48,0xff,0xff,0xff]
 540 ; CHECK32-KNL-NEXT:    vmovups %zmm5, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 541 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0xad,0x08,0xff,0xff,0xff]
 542 ; CHECK32-KNL-NEXT:    vmovups %zmm4, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 543 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0xa5,0xc8,0xfe,0xff,0xff]
 544 ; CHECK32-KNL-NEXT:    vmovups %zmm3, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 545 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x9d,0x88,0xfe,0xff,0xff]
 546 ; CHECK32-KNL-NEXT:    vmovups %zmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 547 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x95,0x48,0xfe,0xff,0xff]
 548 ; CHECK32-KNL-NEXT:    vmovups %zmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 549 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x8d,0x08,0xfe,0xff,0xff]
 550 ; CHECK32-KNL-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 551 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x85,0xc8,0xfd,0xff,0xff]
 552 ; CHECK32-KNL-NEXT:    .cfi_offset %eax, -20
 553 ; CHECK32-KNL-NEXT:    .cfi_offset %ecx, -16
 554 ; CHECK32-KNL-NEXT:    .cfi_offset %edx, -12
 555 ; CHECK32-KNL-NEXT:    .cfi_offset %xmm0, -576
 556 ; CHECK32-KNL-NEXT:    .cfi_offset %xmm1, -512
 557 ; CHECK32-KNL-NEXT:    .cfi_offset %xmm2, -448
 558 ; CHECK32-KNL-NEXT:    .cfi_offset %xmm3, -384
 559 ; CHECK32-KNL-NEXT:    .cfi_offset %xmm4, -320
 560 ; CHECK32-KNL-NEXT:    .cfi_offset %xmm5, -256
 561 ; CHECK32-KNL-NEXT:    .cfi_offset %xmm6, -192
 562 ; CHECK32-KNL-NEXT:    .cfi_offset %xmm7, -128
 563 ; CHECK32-KNL-NEXT:    .cfi_offset %k0, -36
 564 ; CHECK32-KNL-NEXT:    .cfi_offset %k1, -34
 565 ; CHECK32-KNL-NEXT:    .cfi_offset %k2, -32
 566 ; CHECK32-KNL-NEXT:    .cfi_offset %k3, -30
 567 ; CHECK32-KNL-NEXT:    .cfi_offset %k4, -28
 568 ; CHECK32-KNL-NEXT:    .cfi_offset %k5, -26
 569 ; CHECK32-KNL-NEXT:    .cfi_offset %k6, -24
 570 ; CHECK32-KNL-NEXT:    .cfi_offset %k7, -22
 571 ; CHECK32-KNL-NEXT:    cld ## encoding: [0xfc]
 572 ; CHECK32-KNL-NEXT:    calll _bar ## encoding: [0xe8,A,A,A,A]
 573 ; CHECK32-KNL-NEXT:    ## fixup A - offset: 1, value: _bar-4, kind: FK_PCRel_4
 574 ; CHECK32-KNL-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm0 ## 64-byte Reload
 575 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x85,0xc8,0xfd,0xff,0xff]
 576 ; CHECK32-KNL-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm1 ## 64-byte Reload
 577 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x8d,0x08,0xfe,0xff,0xff]
 578 ; CHECK32-KNL-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm2 ## 64-byte Reload
 579 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x95,0x48,0xfe,0xff,0xff]
 580 ; CHECK32-KNL-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm3 ## 64-byte Reload
 581 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x9d,0x88,0xfe,0xff,0xff]
 582 ; CHECK32-KNL-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm4 ## 64-byte Reload
 583 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0xa5,0xc8,0xfe,0xff,0xff]
 584 ; CHECK32-KNL-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm5 ## 64-byte Reload
 585 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0xad,0x08,0xff,0xff,0xff]
 586 ; CHECK32-KNL-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm6 ## 64-byte Reload
 587 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0xb5,0x48,0xff,0xff,0xff]
 588 ; CHECK32-KNL-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm7 ## 64-byte Reload
 589 ; CHECK32-KNL-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0xbd,0x88,0xff,0xff,0xff]
 590 ; CHECK32-KNL-NEXT:    kmovw {{[-0-9]+}}(%e{{[sb]}}p), %k0 ## 2-byte Reload
 591 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x45,0xe4]
 592 ; CHECK32-KNL-NEXT:    kmovw {{[-0-9]+}}(%e{{[sb]}}p), %k1 ## 2-byte Reload
 593 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x4d,0xe6]
 594 ; CHECK32-KNL-NEXT:    kmovw {{[-0-9]+}}(%e{{[sb]}}p), %k2 ## 2-byte Reload
 595 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x55,0xe8]
 596 ; CHECK32-KNL-NEXT:    kmovw {{[-0-9]+}}(%e{{[sb]}}p), %k3 ## 2-byte Reload
 597 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x5d,0xea]
 598 ; CHECK32-KNL-NEXT:    kmovw {{[-0-9]+}}(%e{{[sb]}}p), %k4 ## 2-byte Reload
 599 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x65,0xec]
 600 ; CHECK32-KNL-NEXT:    kmovw {{[-0-9]+}}(%e{{[sb]}}p), %k5 ## 2-byte Reload
 601 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x6d,0xee]
 602 ; CHECK32-KNL-NEXT:    kmovw {{[-0-9]+}}(%e{{[sb]}}p), %k6 ## 2-byte Reload
 603 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x75,0xf0]
 604 ; CHECK32-KNL-NEXT:    kmovw {{[-0-9]+}}(%e{{[sb]}}p), %k7 ## 2-byte Reload
 605 ; CHECK32-KNL-NEXT:    ## encoding: [0xc5,0xf8,0x90,0x7d,0xf2]
 606 ; CHECK32-KNL-NEXT:    leal -12(%ebp), %esp ## encoding: [0x8d,0x65,0xf4]
 607 ; CHECK32-KNL-NEXT:    popl %eax ## encoding: [0x58]
 608 ; CHECK32-KNL-NEXT:    popl %ecx ## encoding: [0x59]
 609 ; CHECK32-KNL-NEXT:    popl %edx ## encoding: [0x5a]
 610 ; CHECK32-KNL-NEXT:    popl %ebp ## encoding: [0x5d]
 611 ; CHECK32-KNL-NEXT:    iretl ## encoding: [0xcf]
 612 ;
 613 ; CHECK32-SKX-LABEL: foo:
 614 ; CHECK32-SKX:       ## %bb.0:
 615 ; CHECK32-SKX-NEXT:    pushl %ebp ## encoding: [0x55]
 616 ; CHECK32-SKX-NEXT:    .cfi_def_cfa_offset 8
 617 ; CHECK32-SKX-NEXT:    .cfi_offset %ebp, -8
 618 ; CHECK32-SKX-NEXT:    movl %esp, %ebp ## encoding: [0x89,0xe5]
 619 ; CHECK32-SKX-NEXT:    .cfi_def_cfa_register %ebp
 620 ; CHECK32-SKX-NEXT:    pushl %edx ## encoding: [0x52]
 621 ; CHECK32-SKX-NEXT:    pushl %ecx ## encoding: [0x51]
 622 ; CHECK32-SKX-NEXT:    pushl %eax ## encoding: [0x50]
 623 ; CHECK32-SKX-NEXT:    andl $-16, %esp ## encoding: [0x83,0xe4,0xf0]
 624 ; CHECK32-SKX-NEXT:    subl $624, %esp ## encoding: [0x81,0xec,0x70,0x02,0x00,0x00]
 625 ; CHECK32-SKX-NEXT:    ## imm = 0x270
 626 ; CHECK32-SKX-NEXT:    kmovq %k7, {{[-0-9]+}}(%e{{[sb]}}p) ## 8-byte Spill
 627 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x7d,0xe8]
 628 ; CHECK32-SKX-NEXT:    kmovq %k6, {{[-0-9]+}}(%e{{[sb]}}p) ## 8-byte Spill
 629 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x75,0xe0]
 630 ; CHECK32-SKX-NEXT:    kmovq %k5, {{[-0-9]+}}(%e{{[sb]}}p) ## 8-byte Spill
 631 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x6d,0xd8]
 632 ; CHECK32-SKX-NEXT:    kmovq %k4, {{[-0-9]+}}(%e{{[sb]}}p) ## 8-byte Spill
 633 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x65,0xd0]
 634 ; CHECK32-SKX-NEXT:    kmovq %k3, {{[-0-9]+}}(%e{{[sb]}}p) ## 8-byte Spill
 635 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x5d,0xc8]
 636 ; CHECK32-SKX-NEXT:    kmovq %k2, {{[-0-9]+}}(%e{{[sb]}}p) ## 8-byte Spill
 637 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x55,0xc0]
 638 ; CHECK32-SKX-NEXT:    kmovq %k1, {{[-0-9]+}}(%e{{[sb]}}p) ## 8-byte Spill
 639 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x4d,0xb8]
 640 ; CHECK32-SKX-NEXT:    kmovq %k0, {{[-0-9]+}}(%e{{[sb]}}p) ## 8-byte Spill
 641 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x91,0x45,0xb0]
 642 ; CHECK32-SKX-NEXT:    vmovups %zmm7, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 643 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0xbd,0x48,0xff,0xff,0xff]
 644 ; CHECK32-SKX-NEXT:    vmovups %zmm6, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 645 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0xb5,0x08,0xff,0xff,0xff]
 646 ; CHECK32-SKX-NEXT:    vmovups %zmm5, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 647 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0xad,0xc8,0xfe,0xff,0xff]
 648 ; CHECK32-SKX-NEXT:    vmovups %zmm4, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 649 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0xa5,0x88,0xfe,0xff,0xff]
 650 ; CHECK32-SKX-NEXT:    vmovups %zmm3, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 651 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x9d,0x48,0xfe,0xff,0xff]
 652 ; CHECK32-SKX-NEXT:    vmovups %zmm2, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 653 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x95,0x08,0xfe,0xff,0xff]
 654 ; CHECK32-SKX-NEXT:    vmovups %zmm1, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 655 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x8d,0xc8,0xfd,0xff,0xff]
 656 ; CHECK32-SKX-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 64-byte Spill
 657 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x11,0x85,0x88,0xfd,0xff,0xff]
 658 ; CHECK32-SKX-NEXT:    .cfi_offset %eax, -20
 659 ; CHECK32-SKX-NEXT:    .cfi_offset %ecx, -16
 660 ; CHECK32-SKX-NEXT:    .cfi_offset %edx, -12
 661 ; CHECK32-SKX-NEXT:    .cfi_offset %xmm0, -640
 662 ; CHECK32-SKX-NEXT:    .cfi_offset %xmm1, -576
 663 ; CHECK32-SKX-NEXT:    .cfi_offset %xmm2, -512
 664 ; CHECK32-SKX-NEXT:    .cfi_offset %xmm3, -448
 665 ; CHECK32-SKX-NEXT:    .cfi_offset %xmm4, -384
 666 ; CHECK32-SKX-NEXT:    .cfi_offset %xmm5, -320
 667 ; CHECK32-SKX-NEXT:    .cfi_offset %xmm6, -256
 668 ; CHECK32-SKX-NEXT:    .cfi_offset %xmm7, -192
 669 ; CHECK32-SKX-NEXT:    .cfi_offset %k0, -88
 670 ; CHECK32-SKX-NEXT:    .cfi_offset %k1, -80
 671 ; CHECK32-SKX-NEXT:    .cfi_offset %k2, -72
 672 ; CHECK32-SKX-NEXT:    .cfi_offset %k3, -64
 673 ; CHECK32-SKX-NEXT:    .cfi_offset %k4, -56
 674 ; CHECK32-SKX-NEXT:    .cfi_offset %k5, -48
 675 ; CHECK32-SKX-NEXT:    .cfi_offset %k6, -40
 676 ; CHECK32-SKX-NEXT:    .cfi_offset %k7, -32
 677 ; CHECK32-SKX-NEXT:    cld ## encoding: [0xfc]
 678 ; CHECK32-SKX-NEXT:    vzeroupper ## encoding: [0xc5,0xf8,0x77]
 679 ; CHECK32-SKX-NEXT:    calll _bar ## encoding: [0xe8,A,A,A,A]
 680 ; CHECK32-SKX-NEXT:    ## fixup A - offset: 1, value: _bar-4, kind: FK_PCRel_4
 681 ; CHECK32-SKX-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm0 ## 64-byte Reload
 682 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x85,0x88,0xfd,0xff,0xff]
 683 ; CHECK32-SKX-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm1 ## 64-byte Reload
 684 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x8d,0xc8,0xfd,0xff,0xff]
 685 ; CHECK32-SKX-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm2 ## 64-byte Reload
 686 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x95,0x08,0xfe,0xff,0xff]
 687 ; CHECK32-SKX-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm3 ## 64-byte Reload
 688 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0x9d,0x48,0xfe,0xff,0xff]
 689 ; CHECK32-SKX-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm4 ## 64-byte Reload
 690 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0xa5,0x88,0xfe,0xff,0xff]
 691 ; CHECK32-SKX-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm5 ## 64-byte Reload
 692 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0xad,0xc8,0xfe,0xff,0xff]
 693 ; CHECK32-SKX-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm6 ## 64-byte Reload
 694 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0xb5,0x08,0xff,0xff,0xff]
 695 ; CHECK32-SKX-NEXT:    vmovups {{[-0-9]+}}(%e{{[sb]}}p), %zmm7 ## 64-byte Reload
 696 ; CHECK32-SKX-NEXT:    ## encoding: [0x62,0xf1,0x7c,0x48,0x10,0xbd,0x48,0xff,0xff,0xff]
 697 ; CHECK32-SKX-NEXT:    kmovq {{[-0-9]+}}(%e{{[sb]}}p), %k0 ## 8-byte Reload
 698 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x45,0xb0]
 699 ; CHECK32-SKX-NEXT:    kmovq {{[-0-9]+}}(%e{{[sb]}}p), %k1 ## 8-byte Reload
 700 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x4d,0xb8]
 701 ; CHECK32-SKX-NEXT:    kmovq {{[-0-9]+}}(%e{{[sb]}}p), %k2 ## 8-byte Reload
 702 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x55,0xc0]
 703 ; CHECK32-SKX-NEXT:    kmovq {{[-0-9]+}}(%e{{[sb]}}p), %k3 ## 8-byte Reload
 704 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x5d,0xc8]
 705 ; CHECK32-SKX-NEXT:    kmovq {{[-0-9]+}}(%e{{[sb]}}p), %k4 ## 8-byte Reload
 706 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x65,0xd0]
 707 ; CHECK32-SKX-NEXT:    kmovq {{[-0-9]+}}(%e{{[sb]}}p), %k5 ## 8-byte Reload
 708 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x6d,0xd8]
 709 ; CHECK32-SKX-NEXT:    kmovq {{[-0-9]+}}(%e{{[sb]}}p), %k6 ## 8-byte Reload
 710 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x75,0xe0]
 711 ; CHECK32-SKX-NEXT:    kmovq {{[-0-9]+}}(%e{{[sb]}}p), %k7 ## 8-byte Reload
 712 ; CHECK32-SKX-NEXT:    ## encoding: [0xc4,0xe1,0xf8,0x90,0x7d,0xe8]
 713 ; CHECK32-SKX-NEXT:    leal -12(%ebp), %esp ## encoding: [0x8d,0x65,0xf4]
 714 ; CHECK32-SKX-NEXT:    popl %eax ## encoding: [0x58]
 715 ; CHECK32-SKX-NEXT:    popl %ecx ## encoding: [0x59]
 716 ; CHECK32-SKX-NEXT:    popl %edx ## encoding: [0x5a]
 717 ; CHECK32-SKX-NEXT:    popl %ebp ## encoding: [0x5d]
 718 ; CHECK32-SKX-NEXT:    iretl ## encoding: [0xcf]
 719   call void @bar()
 720   ret void
 721 }
 722
 723 declare void @bar()
 724