[Bug target/81481] New: [7/8 Regression] Spills %xmm to stack in glibc strspn SSE 4.2 variant

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81481

            Bug ID: 81481
           Summary: [7/8 Regression] Spills %xmm to stack in glibc strspn
                    SSE 4.2 variant
           Product: gcc
           Version: 8.0
            Status: UNCONFIRMED
          Keywords: missed-optimization, ra
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---
            Target: i?86-*-*

Created attachment 41787
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41787&action=edit
reduced testcase

We're seeing crashes in 32bit programs violating the ABI by not properly
aligning the outgoing stack when calling strspn which eventually dispatches to
__strspn_sse42.  This is because GCC 7 and trunk compile this to

.L28:
        .cfi_restore_state
        movdqu  ___m128i_shift_right@GOTOFF(%edi,%ebx), %xmm3
        movl    %esi, %ebp
        andl    $-16, %ebp
        movl    $16, %eax
        movaps  %xmm3, (%esp)
        movdqa  0(%ebp), %xmm0
        pshufb  (%esp), %xmm0
        pcmpistri       $58, %xmm0, %xmm0

spilling %xmm3 for no good reason.  GCC 4.8 at least did better here and
avoided
spilling (and the crashes).  There are other string routines similarly
affected.

        movdqu  ___m128i_shift_right@GOTOFF(%ebx,%eax), %xmm0
        andl    $-16, %ebp
        movdqa  0(%ebp), %xmm1
        pshufb  %xmm0, %xmm1
        pcmpistri       $58, %xmm1, %xmm1

Reduced testcase attached, compile with -O2 -m32 -fPIC.