https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71461
Bug ID: 71461
Summary: missed optimization in conditional assignment
Product: gcc
Version: 6.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: lhyatt at gmail dot com
Target Milestone: ---
=====================
int i, j;
void f1(bool b, int x) {
(b ? i : j) = x;
}
void f2(bool b, int x) {
*(b ? &i : &j) = x;
}
==================
On x86-64 at -O3, gcc outputs different code for f1() vs f2(), using a branch
for f1() and a cmove for f2(). Is it something that can/should be optimized?
gcc (version 5, 6, or current trunk) outputs:
==================
f1(bool, int):
testb %dil, %dil
jne .L6
movl %esi, j(%rip)
ret
.L6:
movl %esi, i(%rip)
ret
f2(bool, int):
testb %dil, %dil
movl $j, %edx
movl $i, %eax
cmove %rdx, %rax
movl %esi, (%rax)
ret
==================
vs clang:
==================
f1(bool, int): # @f1(bool, int)
movl $i, %eax
movl $j, %ecx
testb %dil, %dil
cmovneq %rax, %rcx
movl %esi, (%rcx)
retq
f2(bool, int): # @f2(bool, int)
movl $i, %eax
movl $j, %ecx
testb %dil, %dil
cmovneq %rax, %rcx
movl %esi, (%rcx)
retq
==================
In my tests, the branch version of f1() is about 2x-3x slower than f2() if the
branch is 50% predicted, and only about 1% faster than f2() if it is 100%
predicted.
FWIW, this version:
==================
void f3(bool b, int x) {
if(b) i = x; else j = x;
}
==================
ends up using a branch on both gcc and clang.
Thanks...
-Lewis
