--- Short description
Use of xmm intrinsics generates useless code: mmx registers seem to be written
back to the stack systematicaly, even when this is not needed. Maybe it's
related to already reported bug about MMX.
--- GCC version (4.2.1, prebuild version for mingw)
Bugs was seen at least with 4.0.x and 4.1.x, under linux as well. I don't
remember exact build params, but as long as I remember this prevented me to use
mmx intrinsics.
-march=pentium4 option is irelevant (command line below). Use donly to enable
mmx/sse.
> g++ -v
Using built-in specs.
Target: mingw32
Configured with: ../gcc-4.2.1-2-src/configure --with-gcc --enable-libgomp
--host=mingw32 --build=mingw32 --target=mingw32 --program-suffix=-dw2
--with-arch=i486 --with-tune=generic --disable-werror --prefix=/mingw
--with-local-prefix=/mingw --enable-threads --disable-nls
--enable-languages=c,c++,fortran,objc,obj-c++,ada --disable-win32-registry
--disable-sjlj-exceptions --enable-libstdcxx-debug
--enable-cxx-flags=-fno-function-sections -fno-data-sections
--enable-version-specific-runtime-libs --disable-bootstrap
Thread model: win32
gcc version 4.2.1-dw2 (mingw32-2)
--- Source code ---
> cat bug.cc
#if defined (SSE)
#include <emmintrin.h>
#elif defined (MMX)
#include <mmintrin.h>
#endif
struct X {
#if defined (SSE)
__m128i data[2];
#elif defined (MMX)
__m64 data[2];
#else
int data[2];
#endif
};
void foo (X& a, const X& b)
{
for (int k = 0; k < 2; ++k) {
#if defined (SSE)
a.data[k] = _mm_xor_si128 (a.data[k], b.data[k]);
#elif defined (MMX)
a.data[k] = _mm_xor_si64 (a.data[k], b.data[k]);
#else
a.data[k] ^= b.data[k];
#endif
}
}
--- Assembly (SSE, OK)
> g++ -S -O3 bug.cc -DSSE -march=pentium4 -fomit-frame-pointer
.globl __Z3fooR1XRKS_
.def __Z3fooR1XRKS_; .scl 2; .type 32; .endef
__Z3fooR1XRKS_:
LFB472:
movl 4(%esp), %eax
movl 8(%esp), %edx
movdqa (%eax), %xmm0
pxor (%edx), %xmm0
movdqa %xmm0, (%eax)
movdqa 16(%eax), %xmm0
pxor 16(%edx), %xmm0
movdqa %xmm0, 16(%eax)
ret
LFE472:
--- Assembly (MMX, BAD)
> g++ -S -O3 bug.cc -DMMX -march=pentium4 -fomit-frame-pointer
.globl __Z3fooR1XRKS_
.def __Z3fooR1XRKS_; .scl 2; .type 32; .endef
__Z3fooR1XRKS_:
LFB124:
subl $20, %esp <--- Useless
LCFI0:
movl 24(%esp), %eax
movl 28(%esp), %edx
movq (%eax), %mm0
movq %mm0, 8(%esp) <--- Useless
pxor (%edx), %mm0
movq %mm0, (%eax)
movq 8(%eax), %mm0
movq %mm0, (%esp) <--- Useless
pxor 8(%edx), %mm0
movq %mm0, 8(%eax)
addl $20, %esp <--- Useless
ret
LFE124:
--
Summary: MMX bad optimization with intrinsics
Product: gcc
Version: 4.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: etjq78kl at free dot fr
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35142
