Home | History | Annotate | Download | only in X86 
      1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
      2 ; RUN: llc < %s -mtriple=i686-unknown-unknown | FileCheck %s --check-prefixes=X86,X86-NOBMI
      3 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+bmi2 | FileCheck %s --check-prefixes=X86,X86-BMI
      4 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefixes=X64,X64-NOBMI
      5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+bmi2 | FileCheck %s --check-prefixes=X64,X64-BMI
      6 
      7 ; PR1198
      8 
      9 define i64 @foo(i64 %x, i64 %y) nounwind {
     10 ; X86-NOBMI-LABEL: foo:
     11 ; X86-NOBMI:       # %bb.0:
     12 ; X86-NOBMI-NEXT:    pushl %ebp
     13 ; X86-NOBMI-NEXT:    pushl %ebx
     14 ; X86-NOBMI-NEXT:    pushl %edi
     15 ; X86-NOBMI-NEXT:    pushl %esi
     16 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %ecx
     17 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %esi
     18 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %ebp
     19 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %edi
     20 ; X86-NOBMI-NEXT:    movl %ecx, %eax
     21 ; X86-NOBMI-NEXT:    mull %ebp
     22 ; X86-NOBMI-NEXT:    movl %edx, %ebx
     23 ; X86-NOBMI-NEXT:    movl %esi, %eax
     24 ; X86-NOBMI-NEXT:    mull %ebp
     25 ; X86-NOBMI-NEXT:    movl %edx, %ebp
     26 ; X86-NOBMI-NEXT:    movl %eax, %esi
     27 ; X86-NOBMI-NEXT:    addl %ebx, %esi
     28 ; X86-NOBMI-NEXT:    adcl $0, %ebp
     29 ; X86-NOBMI-NEXT:    movl %ecx, %eax
     30 ; X86-NOBMI-NEXT:    mull %edi
     31 ; X86-NOBMI-NEXT:    movl %edx, %ebx
     32 ; X86-NOBMI-NEXT:    addl %esi, %eax
     33 ; X86-NOBMI-NEXT:    adcl %ebp, %ebx
     34 ; X86-NOBMI-NEXT:    setb %al
     35 ; X86-NOBMI-NEXT:    movzbl %al, %ecx
     36 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
     37 ; X86-NOBMI-NEXT:    mull %edi
     38 ; X86-NOBMI-NEXT:    movl %edx, %esi
     39 ; X86-NOBMI-NEXT:    movl %eax, %ebp
     40 ; X86-NOBMI-NEXT:    addl %ebx, %ebp
     41 ; X86-NOBMI-NEXT:    adcl %ecx, %esi
     42 ; X86-NOBMI-NEXT:    xorl %ecx, %ecx
     43 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
     44 ; X86-NOBMI-NEXT:    mull %ecx
     45 ; X86-NOBMI-NEXT:    movl %edx, %edi
     46 ; X86-NOBMI-NEXT:    movl %eax, %ebx
     47 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
     48 ; X86-NOBMI-NEXT:    mull %ecx
     49 ; X86-NOBMI-NEXT:    addl %ebx, %eax
     50 ; X86-NOBMI-NEXT:    adcl %edi, %edx
     51 ; X86-NOBMI-NEXT:    addl %ebp, %eax
     52 ; X86-NOBMI-NEXT:    adcl %esi, %edx
     53 ; X86-NOBMI-NEXT:    popl %esi
     54 ; X86-NOBMI-NEXT:    popl %edi
     55 ; X86-NOBMI-NEXT:    popl %ebx
     56 ; X86-NOBMI-NEXT:    popl %ebp
     57 ; X86-NOBMI-NEXT:    retl
     58 ;
     59 ; X86-BMI-LABEL: foo:
     60 ; X86-BMI:       # %bb.0:
     61 ; X86-BMI-NEXT:    pushl %ebp
     62 ; X86-BMI-NEXT:    pushl %ebx
     63 ; X86-BMI-NEXT:    pushl %edi
     64 ; X86-BMI-NEXT:    pushl %esi
     65 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %ecx
     66 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
     67 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %esi
     68 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %edi
     69 ; X86-BMI-NEXT:    movl %ecx, %edx
     70 ; X86-BMI-NEXT:    mulxl %esi, %edx, %ebx
     71 ; X86-BMI-NEXT:    movl %eax, %edx
     72 ; X86-BMI-NEXT:    mulxl %esi, %ebp, %eax
     73 ; X86-BMI-NEXT:    addl %ebx, %ebp
     74 ; X86-BMI-NEXT:    adcl $0, %eax
     75 ; X86-BMI-NEXT:    movl %ecx, %edx
     76 ; X86-BMI-NEXT:    mulxl %edi, %edx, %ebx
     77 ; X86-BMI-NEXT:    addl %ebp, %edx
     78 ; X86-BMI-NEXT:    adcl %eax, %ebx
     79 ; X86-BMI-NEXT:    setb %al
     80 ; X86-BMI-NEXT:    movzbl %al, %eax
     81 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
     82 ; X86-BMI-NEXT:    mulxl %edi, %edi, %ebp
     83 ; X86-BMI-NEXT:    addl %ebx, %edi
     84 ; X86-BMI-NEXT:    adcl %eax, %ebp
     85 ; X86-BMI-NEXT:    xorl %eax, %eax
     86 ; X86-BMI-NEXT:    movl %esi, %edx
     87 ; X86-BMI-NEXT:    mulxl %eax, %ebx, %esi
     88 ; X86-BMI-NEXT:    movl %ecx, %edx
     89 ; X86-BMI-NEXT:    mulxl %eax, %eax, %edx
     90 ; X86-BMI-NEXT:    addl %ebx, %eax
     91 ; X86-BMI-NEXT:    adcl %esi, %edx
     92 ; X86-BMI-NEXT:    addl %edi, %eax
     93 ; X86-BMI-NEXT:    adcl %ebp, %edx
     94 ; X86-BMI-NEXT:    popl %esi
     95 ; X86-BMI-NEXT:    popl %edi
     96 ; X86-BMI-NEXT:    popl %ebx
     97 ; X86-BMI-NEXT:    popl %ebp
     98 ; X86-BMI-NEXT:    retl
     99 ;
    100 ; X64-NOBMI-LABEL: foo:
    101 ; X64-NOBMI:       # %bb.0:
    102 ; X64-NOBMI-NEXT:    movq %rdi, %rax
    103 ; X64-NOBMI-NEXT:    mulq %rsi
    104 ; X64-NOBMI-NEXT:    movq %rdx, %rax
    105 ; X64-NOBMI-NEXT:    retq
    106 ;
    107 ; X64-BMI-LABEL: foo:
    108 ; X64-BMI:       # %bb.0:
    109 ; X64-BMI-NEXT:    movq %rdi, %rdx
    110 ; X64-BMI-NEXT:    mulxq %rsi, %rcx, %rax
    111 ; X64-BMI-NEXT:    retq
    112   %tmp0 = zext i64 %x to i128
    113   %tmp1 = zext i64 %y to i128
    114   %tmp2 = mul i128 %tmp0, %tmp1
    115   %tmp7 = zext i32 64 to i128
    116   %tmp3 = lshr i128 %tmp2, %tmp7
    117   %tmp4 = trunc i128 %tmp3 to i64
    118   ret i64 %tmp4
    119 }
    120 
    121 ; <rdar://problem/14096009> superfluous multiply by high part of
    122 ; zero-extended value.
    123 
    124 define i64 @mul1(i64 %n, i64* nocapture %z, i64* nocapture %x, i64 %y) nounwind {
    125 ; X86-NOBMI-LABEL: mul1:
    126 ; X86-NOBMI:       # %bb.0: # %entry
    127 ; X86-NOBMI-NEXT:    pushl %ebp
    128 ; X86-NOBMI-NEXT:    pushl %ebx
    129 ; X86-NOBMI-NEXT:    pushl %edi
    130 ; X86-NOBMI-NEXT:    pushl %esi
    131 ; X86-NOBMI-NEXT:    subl $28, %esp
    132 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %ecx
    133 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
    134 ; X86-NOBMI-NEXT:    orl %ecx, %eax
    135 ; X86-NOBMI-NEXT:    je .LBB1_3
    136 ; X86-NOBMI-NEXT:  # %bb.1: # %for.body.preheader
    137 ; X86-NOBMI-NEXT:    xorl %eax, %eax
    138 ; X86-NOBMI-NEXT:    xorl %edx, %edx
    139 ; X86-NOBMI-NEXT:    xorl %ebx, %ebx
    140 ; X86-NOBMI-NEXT:    movl $0, {{[0-9]+}}(%esp) # 4-byte Folded Spill
    141 ; X86-NOBMI-NEXT:    .p2align 4, 0x90
    142 ; X86-NOBMI-NEXT:  .LBB1_2: # %for.body
    143 ; X86-NOBMI-NEXT:    # =>This Inner Loop Header: Depth=1
    144 ; X86-NOBMI-NEXT:    movl %edx, {{[0-9]+}}(%esp) # 4-byte Spill
    145 ; X86-NOBMI-NEXT:    movl %eax, {{[0-9]+}}(%esp) # 4-byte Spill
    146 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
    147 ; X86-NOBMI-NEXT:    movl %eax, %ecx
    148 ; X86-NOBMI-NEXT:    movl (%eax,%ebx,8), %ebp
    149 ; X86-NOBMI-NEXT:    movl 4(%eax,%ebx,8), %esi
    150 ; X86-NOBMI-NEXT:    movl %esi, {{[0-9]+}}(%esp) # 4-byte Spill
    151 ; X86-NOBMI-NEXT:    movl %ebp, %eax
    152 ; X86-NOBMI-NEXT:    movl %ebp, {{[0-9]+}}(%esp) # 4-byte Spill
    153 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %ecx
    154 ; X86-NOBMI-NEXT:    mull %ecx
    155 ; X86-NOBMI-NEXT:    movl %edx, %edi
    156 ; X86-NOBMI-NEXT:    movl %eax, {{[0-9]+}}(%esp) # 4-byte Spill
    157 ; X86-NOBMI-NEXT:    movl %esi, %eax
    158 ; X86-NOBMI-NEXT:    mull %ecx
    159 ; X86-NOBMI-NEXT:    movl %edx, %ecx
    160 ; X86-NOBMI-NEXT:    movl %eax, %esi
    161 ; X86-NOBMI-NEXT:    addl %edi, %esi
    162 ; X86-NOBMI-NEXT:    adcl $0, %ecx
    163 ; X86-NOBMI-NEXT:    movl %ebp, %eax
    164 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
    165 ; X86-NOBMI-NEXT:    mull %edx
    166 ; X86-NOBMI-NEXT:    movl %edx, %ebp
    167 ; X86-NOBMI-NEXT:    movl %eax, %edi
    168 ; X86-NOBMI-NEXT:    addl %esi, %edi
    169 ; X86-NOBMI-NEXT:    adcl %ecx, %ebp
    170 ; X86-NOBMI-NEXT:    setb {{[0-9]+}}(%esp) # 1-byte Folded Spill
    171 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax # 4-byte Reload
    172 ; X86-NOBMI-NEXT:    mull {{[0-9]+}}(%esp)
    173 ; X86-NOBMI-NEXT:    movl %edx, %ecx
    174 ; X86-NOBMI-NEXT:    movl %eax, %esi
    175 ; X86-NOBMI-NEXT:    addl %ebp, %esi
    176 ; X86-NOBMI-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # 1-byte Folded Reload
    177 ; X86-NOBMI-NEXT:    adcl %eax, %ecx
    178 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
    179 ; X86-NOBMI-NEXT:    xorl %edx, %edx
    180 ; X86-NOBMI-NEXT:    mull %edx
    181 ; X86-NOBMI-NEXT:    movl %edx, {{[0-9]+}}(%esp) # 4-byte Spill
    182 ; X86-NOBMI-NEXT:    movl %eax, %ebp
    183 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %eax # 4-byte Reload
    184 ; X86-NOBMI-NEXT:    xorl %edx, %edx
    185 ; X86-NOBMI-NEXT:    mull %edx
    186 ; X86-NOBMI-NEXT:    addl %ebp, %eax
    187 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %ebp
    188 ; X86-NOBMI-NEXT:    adcl {{[0-9]+}}(%esp), %edx # 4-byte Folded Reload
    189 ; X86-NOBMI-NEXT:    addl %esi, %eax
    190 ; X86-NOBMI-NEXT:    adcl %ecx, %edx
    191 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %esi # 4-byte Reload
    192 ; X86-NOBMI-NEXT:    addl {{[0-9]+}}(%esp), %esi # 4-byte Folded Reload
    193 ; X86-NOBMI-NEXT:    adcl {{[0-9]+}}(%esp), %edi # 4-byte Folded Reload
    194 ; X86-NOBMI-NEXT:    adcl $0, %eax
    195 ; X86-NOBMI-NEXT:    adcl $0, %edx
    196 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %ecx
    197 ; X86-NOBMI-NEXT:    movl %esi, (%ecx,%ebx,8)
    198 ; X86-NOBMI-NEXT:    movl %edi, 4(%ecx,%ebx,8)
    199 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %ecx
    200 ; X86-NOBMI-NEXT:    movl %ecx, %edi
    201 ; X86-NOBMI-NEXT:    addl $1, %ebx
    202 ; X86-NOBMI-NEXT:    movl {{[0-9]+}}(%esp), %esi # 4-byte Reload
    203 ; X86-NOBMI-NEXT:    adcl $0, %esi
    204 ; X86-NOBMI-NEXT:    movl %ebx, %ecx
    205 ; X86-NOBMI-NEXT:    xorl %ebp, %ecx
    206 ; X86-NOBMI-NEXT:    movl %esi, {{[0-9]+}}(%esp) # 4-byte Spill
    207 ; X86-NOBMI-NEXT:    xorl %edi, %esi
    208 ; X86-NOBMI-NEXT:    orl %ecx, %esi
    209 ; X86-NOBMI-NEXT:    jne .LBB1_2
    210 ; X86-NOBMI-NEXT:  .LBB1_3: # %for.end
    211 ; X86-NOBMI-NEXT:    xorl %eax, %eax
    212 ; X86-NOBMI-NEXT:    xorl %edx, %edx
    213 ; X86-NOBMI-NEXT:    addl $28, %esp
    214 ; X86-NOBMI-NEXT:    popl %esi
    215 ; X86-NOBMI-NEXT:    popl %edi
    216 ; X86-NOBMI-NEXT:    popl %ebx
    217 ; X86-NOBMI-NEXT:    popl %ebp
    218 ; X86-NOBMI-NEXT:    retl
    219 ;
    220 ; X86-BMI-LABEL: mul1:
    221 ; X86-BMI:       # %bb.0: # %entry
    222 ; X86-BMI-NEXT:    pushl %ebp
    223 ; X86-BMI-NEXT:    pushl %ebx
    224 ; X86-BMI-NEXT:    pushl %edi
    225 ; X86-BMI-NEXT:    pushl %esi
    226 ; X86-BMI-NEXT:    subl $20, %esp
    227 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %ecx
    228 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
    229 ; X86-BMI-NEXT:    orl %ecx, %eax
    230 ; X86-BMI-NEXT:    je .LBB1_3
    231 ; X86-BMI-NEXT:  # %bb.1: # %for.body.preheader
    232 ; X86-BMI-NEXT:    xorl %ecx, %ecx
    233 ; X86-BMI-NEXT:    xorl %edx, %edx
    234 ; X86-BMI-NEXT:    xorl %edi, %edi
    235 ; X86-BMI-NEXT:    movl $0, {{[0-9]+}}(%esp) # 4-byte Folded Spill
    236 ; X86-BMI-NEXT:    .p2align 4, 0x90
    237 ; X86-BMI-NEXT:  .LBB1_2: # %for.body
    238 ; X86-BMI-NEXT:    # =>This Inner Loop Header: Depth=1
    239 ; X86-BMI-NEXT:    movl %ecx, {{[0-9]+}}(%esp) # 4-byte Spill
    240 ; X86-BMI-NEXT:    movl %edx, {{[0-9]+}}(%esp) # 4-byte Spill
    241 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
    242 ; X86-BMI-NEXT:    movl (%eax,%edi,8), %ecx
    243 ; X86-BMI-NEXT:    movl 4(%eax,%edi,8), %ebx
    244 ; X86-BMI-NEXT:    movl %ebx, (%esp) # 4-byte Spill
    245 ; X86-BMI-NEXT:    movl %ecx, %edx
    246 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
    247 ; X86-BMI-NEXT:    movl %eax, %esi
    248 ; X86-BMI-NEXT:    mulxl %eax, %eax, %ebp
    249 ; X86-BMI-NEXT:    movl %eax, {{[0-9]+}}(%esp) # 4-byte Spill
    250 ; X86-BMI-NEXT:    movl %ebx, %edx
    251 ; X86-BMI-NEXT:    mulxl %esi, %eax, %esi
    252 ; X86-BMI-NEXT:    addl %ebp, %eax
    253 ; X86-BMI-NEXT:    adcl $0, %esi
    254 ; X86-BMI-NEXT:    movl %ecx, %edx
    255 ; X86-BMI-NEXT:    mulxl {{[0-9]+}}(%esp), %ebp, %ebx
    256 ; X86-BMI-NEXT:    addl %eax, %ebp
    257 ; X86-BMI-NEXT:    adcl %esi, %ebx
    258 ; X86-BMI-NEXT:    movl (%esp), %edx # 4-byte Reload
    259 ; X86-BMI-NEXT:    mulxl {{[0-9]+}}(%esp), %eax, %esi
    260 ; X86-BMI-NEXT:    setb %dl
    261 ; X86-BMI-NEXT:    addl %ebx, %eax
    262 ; X86-BMI-NEXT:    movzbl %dl, %edx
    263 ; X86-BMI-NEXT:    adcl %edx, %esi
    264 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %edx
    265 ; X86-BMI-NEXT:    xorl %ebx, %ebx
    266 ; X86-BMI-NEXT:    mulxl %ebx, %ebx, %edx
    267 ; X86-BMI-NEXT:    movl %edx, (%esp) # 4-byte Spill
    268 ; X86-BMI-NEXT:    movl %ecx, %edx
    269 ; X86-BMI-NEXT:    xorl %ecx, %ecx
    270 ; X86-BMI-NEXT:    mulxl %ecx, %ecx, %edx
    271 ; X86-BMI-NEXT:    addl %ebx, %ecx
    272 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %ebx
    273 ; X86-BMI-NEXT:    adcl (%esp), %edx # 4-byte Folded Reload
    274 ; X86-BMI-NEXT:    addl %eax, %ecx
    275 ; X86-BMI-NEXT:    adcl %esi, %edx
    276 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %esi # 4-byte Reload
    277 ; X86-BMI-NEXT:    addl {{[0-9]+}}(%esp), %esi # 4-byte Folded Reload
    278 ; X86-BMI-NEXT:    adcl {{[0-9]+}}(%esp), %ebp # 4-byte Folded Reload
    279 ; X86-BMI-NEXT:    adcl $0, %ecx
    280 ; X86-BMI-NEXT:    adcl $0, %edx
    281 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
    282 ; X86-BMI-NEXT:    movl %esi, (%eax,%edi,8)
    283 ; X86-BMI-NEXT:    movl %ebp, 4(%eax,%edi,8)
    284 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %eax
    285 ; X86-BMI-NEXT:    movl %eax, %esi
    286 ; X86-BMI-NEXT:    addl $1, %edi
    287 ; X86-BMI-NEXT:    movl {{[0-9]+}}(%esp), %ebp # 4-byte Reload
    288 ; X86-BMI-NEXT:    adcl $0, %ebp
    289 ; X86-BMI-NEXT:    movl %edi, %eax
    290 ; X86-BMI-NEXT:    xorl %esi, %eax
    291 ; X86-BMI-NEXT:    movl %ebp, {{[0-9]+}}(%esp) # 4-byte Spill
    292 ; X86-BMI-NEXT:    movl %ebp, %esi
    293 ; X86-BMI-NEXT:    xorl %ebx, %esi
    294 ; X86-BMI-NEXT:    orl %eax, %esi
    295 ; X86-BMI-NEXT:    jne .LBB1_2
    296 ; X86-BMI-NEXT:  .LBB1_3: # %for.end
    297 ; X86-BMI-NEXT:    xorl %eax, %eax
    298 ; X86-BMI-NEXT:    xorl %edx, %edx
    299 ; X86-BMI-NEXT:    addl $20, %esp
    300 ; X86-BMI-NEXT:    popl %esi
    301 ; X86-BMI-NEXT:    popl %edi
    302 ; X86-BMI-NEXT:    popl %ebx
    303 ; X86-BMI-NEXT:    popl %ebp
    304 ; X86-BMI-NEXT:    retl
    305 ;
    306 ; X64-NOBMI-LABEL: mul1:
    307 ; X64-NOBMI:       # %bb.0: # %entry
    308 ; X64-NOBMI-NEXT:    testq %rdi, %rdi
    309 ; X64-NOBMI-NEXT:    je .LBB1_3
    310 ; X64-NOBMI-NEXT:  # %bb.1: # %for.body.preheader
    311 ; X64-NOBMI-NEXT:    movq %rcx, %r8
    312 ; X64-NOBMI-NEXT:    movq %rdx, %r9
    313 ; X64-NOBMI-NEXT:    xorl %r10d, %r10d
    314 ; X64-NOBMI-NEXT:    xorl %ecx, %ecx
    315 ; X64-NOBMI-NEXT:    .p2align 4, 0x90
    316 ; X64-NOBMI-NEXT:  .LBB1_2: # %for.body
    317 ; X64-NOBMI-NEXT:    # =>This Inner Loop Header: Depth=1
    318 ; X64-NOBMI-NEXT:    movq %r8, %rax
    319 ; X64-NOBMI-NEXT:    mulq (%r9,%rcx,8)
    320 ; X64-NOBMI-NEXT:    addq %r10, %rax
    321 ; X64-NOBMI-NEXT:    adcq $0, %rdx
    322 ; X64-NOBMI-NEXT:    movq %rax, (%rsi,%rcx,8)
    323 ; X64-NOBMI-NEXT:    incq %rcx
    324 ; X64-NOBMI-NEXT:    cmpq %rcx, %rdi
    325 ; X64-NOBMI-NEXT:    movq %rdx, %r10
    326 ; X64-NOBMI-NEXT:    jne .LBB1_2
    327 ; X64-NOBMI-NEXT:  .LBB1_3: # %for.end
    328 ; X64-NOBMI-NEXT:    xorl %eax, %eax
    329 ; X64-NOBMI-NEXT:    retq
    330 ;
    331 ; X64-BMI-LABEL: mul1:
    332 ; X64-BMI:       # %bb.0: # %entry
    333 ; X64-BMI-NEXT:    testq %rdi, %rdi
    334 ; X64-BMI-NEXT:    je .LBB1_3
    335 ; X64-BMI-NEXT:  # %bb.1: # %for.body.preheader
    336 ; X64-BMI-NEXT:    movq %rcx, %r8
    337 ; X64-BMI-NEXT:    movq %rdx, %r9
    338 ; X64-BMI-NEXT:    xorl %r10d, %r10d
    339 ; X64-BMI-NEXT:    xorl %eax, %eax
    340 ; X64-BMI-NEXT:    .p2align 4, 0x90
    341 ; X64-BMI-NEXT:  .LBB1_2: # %for.body
    342 ; X64-BMI-NEXT:    # =>This Inner Loop Header: Depth=1
    343 ; X64-BMI-NEXT:    movq %r8, %rdx
    344 ; X64-BMI-NEXT:    mulxq (%r9,%rax,8), %rcx, %rdx
    345 ; X64-BMI-NEXT:    addq %r10, %rcx
    346 ; X64-BMI-NEXT:    adcq $0, %rdx
    347 ; X64-BMI-NEXT:    movq %rcx, (%rsi,%rax,8)
    348 ; X64-BMI-NEXT:    incq %rax
    349 ; X64-BMI-NEXT:    cmpq %rax, %rdi
    350 ; X64-BMI-NEXT:    movq %rdx, %r10
    351 ; X64-BMI-NEXT:    jne .LBB1_2
    352 ; X64-BMI-NEXT:  .LBB1_3: # %for.end
    353 ; X64-BMI-NEXT:    xorl %eax, %eax
    354 ; X64-BMI-NEXT:    retq
    355 entry:
    356   %conv = zext i64 %y to i128
    357   %cmp11 = icmp eq i64 %n, 0
    358   br i1 %cmp11, label %for.end, label %for.body
    359 
    360 for.body:                                         ; preds = %entry, %for.body
    361   %carry.013 = phi i64 [ %conv6, %for.body ], [ 0, %entry ]
    362   %i.012 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
    363   %arrayidx = getelementptr inbounds i64, i64* %x, i64 %i.012
    364   %0 = load i64, i64* %arrayidx, align 8
    365   %conv2 = zext i64 %0 to i128
    366   %mul = mul i128 %conv2, %conv
    367   %conv3 = zext i64 %carry.013 to i128
    368   %add = add i128 %mul, %conv3
    369   %conv4 = trunc i128 %add to i64
    370   %arrayidx5 = getelementptr inbounds i64, i64* %z, i64 %i.012
    371   store i64 %conv4, i64* %arrayidx5, align 8
    372   %shr = lshr i128 %add, 64
    373   %conv6 = trunc i128 %shr to i64
    374   %inc = add i64 %i.012, 1
    375   %exitcond = icmp eq i64 %inc, %n
    376   br i1 %exitcond, label %for.end, label %for.body
    377 
    378 for.end:                                          ; preds = %for.body, %entry
    379   ret i64 0
    380 }
    381