http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49358
Summary: optimization regression in 4.6.1 from 4.5.4
Product: gcc
Version: 4.6.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: v.hais...@sh.cvut.cz
The code:
--------8<----------
double f (double y)
{
return y*y*y*y;
}
--------8<----------
This compiles using version 4.5.4 20110602 (prerelease) (FreeBSD Ports
Collection) to this:
--------8<----------
_Z1fd:
.LFB87:
.cfi_startproc
movapd %xmm0, %xmm1 # y, y
mulsd %xmm1, %xmm0 # y, tmp63
mulsd %xmm1, %xmm0 # y, tmp63
mulsd %xmm1, %xmm0 # y, tmp63
ret
--------8<----------
But gcc version 4.6.1 20110408 (prerelease) (GCC) produces what seems to be
less optimal version:
--------8<----------
_Z1fd:
.LFB85:
.cfi_startproc
movapd %xmm0, %xmm1 # y, tmp64
mulsd %xmm0, %xmm1 # y, tmp64
mulsd %xmm0, %xmm1 # y, tmp64
mulsd %xmm0, %xmm1 # y, tmp64
movapd %xmm1, %xmm0 # tmp64,
ret
--------8<----------
GCC 4.5.4 seems to be smarter in register allocation avoiding the final movapd.
In both cases I have used -O3 -fverbose-asm -save-temps parameters to g++.
