Illegal load hoist

[pjaaskel]

I found that in some tests there is illegal code motion when running with higher optimizations. It moves a 16b load across a byte write that touches part of that word.

This is likely a gcc codegen bug, but could be something with type based alias analysis, or similar because gccbrig heavily casts pointers to different types depending on the memory access width at hand.

;; ************ Correct code optimized with -O1:

    leaq    128(%rbp,%rax), %rdx
                ;; byte 1 stored here
    movb    %r8b, (%rdx)
    movzbl  1(%rsi), %r8d
    leaq    129(%rbp,%rax), %rsi
                ;; byte 2 stored here
    movb    %r8b, (%rsi)
    movq    8(%rbx), %rax

                ;; The 16b load that accesses both bytes
    movswl  (%rdx), %r8d

;; ************ Broken code optimized with -O2:

    leaq    128(%rbp,%rax), %rdi
    movq    %rcx, %rdx
    addq    (%rbx), %rdx
    movzbl  (%rdx), %esi
                ;; byte 1 stored here
    movb    %sil, (%rdi)
    movzbl  1(%rdx), %esi

                ;; Illegal hoist above the store below!
    movswl  (%rdi), %edx

                ;; this modifies the 2nd byte:
    movb    %sil, 129(%rax,%rbp)