https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90594
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization
Target| |x86_64-*-*, aarch64
Status|UNCONFIRMED |NEW
Last reconfirmed| |2019-05-23
CC| |rguenth at gcc dot gnu.org
Target Milestone|--- |9.2
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed. x86_64:
bad_popcount:
.LFB0:
.cfi_startproc
testl %edi, %edi
je .L4
movq %rsi, %rcx
movl %edi, %edx
.p2align 4,,10
.p2align 3
.L3:
xorl %r8d, %r8d
addq $4, %rcx
tzcntl %edx, %r8d
blsr %edx, %edx
movl %r8d, -4(%rcx)
jne .L3
popcntl %edi, %edi
leal -1(%rdi), %eax
leaq 4(%rsi,%rax,4), %rax
ret
this is SCCP doing final value replacement for 'p' and a more general
issue of SCCP eliding out-of-loop IV uses just because it can rather
than because it will save anything (for example it may elide an
induction variable).
I've long had the opinion that this separate pass should go away and
that we instead should perform this transform only when it enables
sth else.
So not really a regression, we just got the ability to compute the
final value of p in generic infrastructure...
