I'm using GCC to compile some code which uses SSE intrinsics. The code is being
compiled at -O3 -mfpmath=sse.
GCC decides to use MMX instructions for some of the operations (zeroing some
memory). There are no MMX intrinsics in the source, but an SSE _mm_setzero_ps
gets compiled into a pair of movq %mm0.
No emms instruction is emitted, possibly because gcc "knows" the FPU is not in
use because of the -mfpmath=sse switch, leaving the FPU in MMX mode.
However, the platform ABI specifies floating-point return values be returned in
FPU registers, so gcc moves return values from SSE registers to the FPU for
argument passing/returning. Since the FPU is in an invalid state because of the
lack of emms, this corrupts the floating-point values.
This behaviour seems to be very dependent on the exact version of gcc and the
exact source.
Here's a testcase:
#include <stdio.h>
#include <emmintrin.h>
// Since this is all in one file, we need to mark some functions
// noinline so that the interesting parts don't get compiled away
#define NOINLINE __attribute__((noinline))
struct OutputData
{
__m128 a, b;
};
struct InputData
{
float a, b;
};
// Something that uses an OutputData that won't be compiled away
__m128 ga, gb;
NOINLINE void doSomethingWith(const OutputData& d){
ga = d.a;
gb = d.b;
}
NOINLINE void calc(const InputData& in)
{
OutputData out;
// the next two lines are where the bug manifests
// gcc decides to use MMX instructions to write
// some zeros, but fails to clean up afterwards.
out.a = _mm_setr_ps(in.a, in.b, 0, 0);
out.b = _mm_setzero_ps();
// ensure the above is not optimised away
doSomethingWith(out);
}
NOINLINE float retFloat()
{
return 3;
}
int main()
{
InputData x = {3.4, 42};
// GCC emits MMX instructions for this function, but emits no emms
calc(x);
// This uses the FPU to return the value, which gets corrupted
printf("%f\n", retFloat());
return 0;
}
On my machine, this generates (for the function "calc"):
.globl _Z4calcRK9InputData
.type _Z4calcRK9InputData, @function
_Z4calcRK9InputData:
.LFB530:
.cfi_startproc
.cfi_personality 0x0,__gxx_personality_v0
pushl %ebp
.cfi_def_cfa_offset 8
pxor %mm0, %mm0
movl %esp, %ebp
.cfi_offset 5, -8
.cfi_def_cfa_register 5
subl $60, %esp
movl 8(%ebp), %eax
movss 4(%eax), %xmm0
movss (%eax), %xmm1
movq %mm0, -48(%ebp) ; MMX instruction
movq %mm0, -24(%ebp) ; MMX instruction
leal -40(%ebp), %eax
movq %mm0, -16(%ebp) ; MMX instruction
movl %eax, (%esp)
unpcklps %xmm0, %xmm1
movaps %xmm1, %xmm0
xorps %xmm1, %xmm1
movlhps %xmm1, %xmm0
movlps %xmm0, -40(%ebp)
movhps %xmm0, -32(%ebp)
call _Z15doSomethingWithRK10OutputData
leave
ret ; FPU stack still in MMX mode and unusable for floating point
.cfi_endproc
.LFE530:
.size _Z4calcRK9InputData, .-_Z4calcRK9InputData
System:
Ubuntu Lucid (gcc 4:4.4.3-1ubuntu1)
GCC:
Using built-in specs.
Target: i486-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Ubuntu 4.4.3-4ubuntu5'
--with-bugurl=file:///usr/share/doc/gcc-4.4/README.Bugs
--enable-languages=c,c++,fortran,objc,obj-c++ --prefix=/usr --enable-shared
--enable-multiarch --enable-linker-build-id --with-system-zlib
--libexecdir=/usr/lib --without-included-gettext --enable-threads=posix
--with-gxx-include-dir=/usr/include/c++/4.4 --program-suffix=-4.4 --enable-nls
--enable-clocale=gnu --enable-libstdcxx-debug --enable-plugin --enable-objc-gc
--enable-targets=all --disable-werror --with-arch-32=i486 --with-tune=generic
--enable-checking=release --build=i486-linux-gnu --host=i486-linux-gnu
--target=i486-linux-gnu
Thread model: posix
gcc version 4.4.3 (Ubuntu 4.4.3-4ubuntu5)
--
Summary: GCC generates MMX instructions but fails to generate
"emms"
Product: gcc
Version: 4.4.3
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: stephen dot dolan at havok dot com
GCC host triplet: i486-linux-gnu
GCC target triplet: i486-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44578
