NetBSD has recently imported gcc 4.1.2 prerelease. I've noticed that
for sh3 there's a serious degradation in the quality of the generated
code for long long handling.
While generally the generated code is more compact, as soon as 64-bit
ints come into play, the size increases, e.g. softfloat.o explodes
from 10K to 28K just to pick one of the worst cases from our libc.
The worst case was actually bswap64 - by an oversight libc used C
version instead of hand-written asm version and it grew from a few
instructions to 100+ instructions.
A minimal self-contained test case (extracted from bswap64) that
demonstrates this is:
unsigned long long int
expand64(unsigned int x)
{
return (unsigned long long int)x << 32;
}
Compiling with -O2 -fomit-frame-pointer I get:
* gcc3
shle--netbsdelf-gcc (GCC) 3.3.6 (NetBSD nb1 20060605)
expand64:
mov r4,r5
mov #0,r4
mov r4,r0
rts
mov r5,r1
* gcc4
shle--netbsdelf-gcc (GCC) 4.1.2 20060628 (prerelease) (NetBSD nb1 20060602)
expand64:
mov #0,r5
mov r4,r1
clrt
addc r4,r1
mov r5,r2
addc r5,r2
mov r1,r3
clrt
addc r1,r3
mov r2,r4
addc r2,r4
clrt
addc r3,r3
addc r4,r4
...
/* crlt/addc/addc sequence repeated total of 32 times */
...
mov r3,r0
rts
mov r4,r1
--
Summary: SuperH: Very unoptimal code generated for 64-bit ints
Product: gcc
Version: 4.1.2
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: uwe at netbsd dot org
GCC build triplet: i386--netbsdelf
GCC host triplet: i386--netbsdelf
GCC target triplet: shle--netbsdelf
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28283
