[llvm-bugs] [Bug 50364] New: rob pike endianness trick not getting optimized anymore

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=50364

            Bug ID: 50364
           Summary: rob pike endianness trick not getting optimized
                    anymore
           Product: clang
           Version: 12.0
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: LLVM Codegen
          Assignee: unassignedclangbugs at nondot.org
          Reporter: soap at gentoo.org
                CC: llvm-bugs at lists.llvm.org, neeilans at live.com,
                    richard-llvm at metafoo.co.uk

I have the following repo with Rob Pike's endianness tricks:
https://github.com/SoapGentoo/portable-endianness

It seems Clang 12 broke the 64-bit store routines.

Clang 11.0.1:
store64_to_LE(unsigned long, unsigned char*):                   #
@store64_to_LE(unsigned long, unsigned char*)
        mov     qword ptr [rsi], rdi
        ret
store64_to_BE(unsigned long, unsigned char*):                   #
@store64_to_BE(unsigned long, unsigned char*)
        movbe   qword ptr [rsi], rdi
        ret

Clang 12.0.0:
.LCPI8_0:
        .quad   8                               # 0x8
        .quad   16                              # 0x10
        .quad   24                              # 0x18
        .quad   32                              # 0x20
.LCPI8_1:
        .byte   0                               # 0x0
        .byte   8                               # 0x8
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
store64_to_LE(unsigned long, unsigned char*):                   #
@store64_to_LE(unsigned long, unsigned char*)
        mov     byte ptr [rsi], dil
        vmovq   xmm0, rdi
        vpbroadcastq    ymm0, xmm0
        vpsrlvq ymm0, ymm0, ymmword ptr [rip + .LCPI8_0]
        vextracti128    xmm1, ymm0, 1
        vmovdqa xmm2, xmmword ptr [rip + .LCPI8_1] # xmm2 =
<0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
        vpshufb xmm1, xmm1, xmm2
        vpshufb xmm0, xmm0, xmm2
        vpunpcklwd      xmm0, xmm0, xmm1        # xmm0 =
xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
        vmovd   dword ptr [rsi + 1], xmm0
        mov     rax, rdi
        shr     rax, 40
        mov     byte ptr [rsi + 5], al
        mov     rax, rdi
        shr     rax, 48
        mov     byte ptr [rsi + 6], al
        shr     rdi, 56
        mov     byte ptr [rsi + 7], dil
        vzeroupper
        ret
.LCPI11_0:
        .quad   56                              # 0x38
        .quad   48                              # 0x30
        .quad   40                              # 0x28
        .quad   32                              # 0x20
.LCPI11_1:
        .byte   0                               # 0x0
        .byte   8                               # 0x8
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
        .zero   1
store64_to_BE(unsigned long, unsigned char*):                   #
@store64_to_BE(unsigned long, unsigned char*)
        mov     rax, rdi
        vmovq   xmm0, rdi
        vpbroadcastq    ymm0, xmm0
        vpsrlvq ymm0, ymm0, ymmword ptr [rip + .LCPI11_0]
        vextracti128    xmm1, ymm0, 1
        vmovdqa xmm2, xmmword ptr [rip + .LCPI11_1] # xmm2 =
<0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
        vpshufb xmm1, xmm1, xmm2
        vpshufb xmm0, xmm0, xmm2
        vpunpcklwd      xmm0, xmm0, xmm1        # xmm0 =
xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
        vmovd   dword ptr [rsi], xmm0
        mov     rcx, rdi
        shr     rcx, 24
        mov     byte ptr [rsi + 4], cl
        mov     rcx, rdi
        shr     rcx, 16
        mov     byte ptr [rsi + 5], cl
        mov     byte ptr [rsi + 6], ah
        mov     byte ptr [rsi + 7], al
        vzeroupper
        ret

I consider detecting this sort of pattern very important in order to write
performant, portable code.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20210516/2e549530/attachment.html>