[Bug rtl-optimization/101366] New: x86 memset codegen for constant sized is suboptimal

[Bug rtl-optimization/101366] New: x86 memset codegen for constant sized is suboptimal

[kutdanila at yandex dot ru]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101366

            Bug ID: 101366
           Summary: x86 memset codegen for constant sized is suboptimal
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: kutdanila at yandex dot ru
  Target Milestone: ---

https://gcc.godbolt.org/z/hP99MYMEG

void Test(char* dst) {
    __m128i pattern = _mm_set1_epi8(dst[-1]);
    for (int i = 0; i < 4; i++) {
        _mm_loadu_si128(reinterpret_cast<__m128i*>(dst + 16 * i), pattern);
    }
}

vs

void TestStd(char* s) {
    memset(s, s[-1], 64);
}


-O3 -msse4.2

Test(char*):
        movzbl  -1(%rdi), %eax
        pxor    %xmm1, %xmm1
        movd    %eax, %xmm0
        pshufb  %xmm1, %xmm0
        movups  %xmm0, (%rdi)
        movups  %xmm0, 16(%rdi)
        movups  %xmm0, 32(%rdi)
        movups  %xmm0, 48(%rdi)
        ret
TestStd(char*):
        movabsq $72340172838076673, %rdx
        movzbl  -1(%rdi), %eax
        movq    %rax, %rcx
        imulq   %rdx, %rcx
        mulq    %rdx
        addq    %rcx, %rdx
        movq    %rax, (%rdi)
        movq    %rdx, 8(%rdi)
        movq    %rax, 16(%rdi)
        movq    %rdx, 24(%rdi)
        movq    %rax, 32(%rdi)
        movq    %rdx, 40(%rdi)
        movq    %rax, 48(%rdi)
        movq    %rdx, 56(%rdi)
        ret

[Bug target/101366] memset codegen for constant sized does not use SSE instructions

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101366

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Target|                            |x86_64-linux-gnu
   Last reconfirmed|                            |2021-07-07
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
            Summary|x86 memset codegen for      |memset codegen for constant
                   |constant sized is           |sized does not use SSE
                   |suboptimal                  |instructions
           Severity|normal                      |enhancement

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
AARCH64 is fine:
TestStd(char*):
        ldr     b0, [x0, -1]
        dup     v0.16b, v0.b[0]
        stp     q0, q0, [x0]
        stp     q0, q0, [x0, 32]
        ret

[Bug target/101366] memset codegen for constant sized does not use SSE instructions

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101366

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hjl.tools at gmail dot com

--- Comment #2 from H.J. Lu <hjl.tools at gmail dot com> ---
Should be fixed by

https://gcc.gnu.org/pipermail/gcc-patches/2021-July/574249.html

which generates:

        movsbl  -1(%rdi), %eax
        pxor    %xmm0, %xmm0
        movd    %eax, %xmm15
        pshufb  %xmm0, %xmm15
        movups  %xmm15, (%rdi)
        movups  %xmm15, 16(%rdi)
        movups  %xmm15, 32(%rdi)
        movups  %xmm15, 48(%rdi)
        ret