[Bug tree-optimization/127745] New: SRA scalarizes aggregate copy using x87 flds/fsts

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127745

            Bug ID: 127745
           Summary: SRA scalarizes aggregate copy using x87 flds/fsts
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---
            Target: i?86-*-*

struct F { float a, b, c, d; };
void copy (struct F *dst, const struct F *src)
{
  struct F s = *src;
  *dst = s;
}

is altering the values of src->{a,b,c,d}.  On x86 with -O2 -mno-sse -m32 you
get

copy:
.LFB0:
        .cfi_startproc
        movl    8(%esp), %edx
        movl    4(%esp), %eax
        flds    (%edx)
        flds    4(%edx)
        flds    8(%edx)
        flds    12(%edx)
        fxch    %st(3)
        fstps   (%eax)
        fxch    %st(1)
        fstps   4(%eax)
        fstps   8(%eax)
        fstps   12(%eax)
        ret

while w/ -fno-tree-sra we correctly use integer registers for the transfer:

        movl    32(%esp), %edx
        movl    28(%esp), %eax
        movl    (%edx), %esi
        movl    4(%edx), %ebx
        movl    8(%edx), %ecx
        movl    12(%edx), %edx
        movl    %esi, (%eax)
        movl    %ebx, 4(%eax)
        movl    %ecx, 8(%eax)
        movl    %edx, 12(%eax)


There is partial mitigation against this issue with calls of
types_risk_mangled_binary_repr_p in tree-sra.cc but it seems incomplete.