https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70359
--- Comment #23 from Aldy Hernandez <aldyh at gcc dot gnu.org> ---
For the curious, on x86 with -ftree-forwprop we get an additional jump:
inttostr:
.LFB0:
.cfi_startproc
movl %edi, %eax
movslq %edx, %rdx
movl $-858993459, %r9d
sarl $31, %eax
leaq -1(%rsi,%rdx), %rsi
movl %eax, %ecx
movb $0, (%rsi)
xorl %edi, %ecx
subl %eax, %ecx
jmp .L2 ;; boo!
.p2align 4,,10
.p2align 3
.L4:
movq %r8, %rsi
movl %edx, %ecx
.L2:
...
...
...
movb $45, -1(%r8)
leaq -2(%rsi), %r8
Not to mention that the last two instructions are slower (ok, larger) than
without forwprop, probably because rsi encodes better than r8.
movb $45, -1(%rsi)
subq $1, %rsi
Also, the inequality comparison on x86 is shorter than comparing with > 9. But
that's probably irrelevant.
All in all:
x86 -ftree-forwprop: 139 bytes
x86 -fno-tree-forwprop: 126 bytes
