http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54703
Bug #: 54703
Summary: [miscompilation] _mm_sub_pd is incorrectly substituted
with vandnps
Classification: Unclassified
Product: gcc
Version: 4.8.0
Status: UNCONFIRMED
Severity: major
Priority: P3
Component: target
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: kr...@kde.org
The following testcase:
#include <xmmintrin.h>
__attribute__((aligned(16))) static const unsigned long long mask[2] = {
0xffffffffff000000ull, 0xffffffffff000000ull };
inline __m128d foo(__m128d v1) {
const __m128d h1 = _mm_and_pd(v1, _mm_load_pd(reinterpret_cast<const double
*>(&mask)));
const __m128d l1 = _mm_sub_pd(v1, h1);
return _mm_mul_pd(h1, l1);
}
__m128d test() {
__m128d a = _mm_set1_pd(2.);
return foo(foo(a));
}
compiles to
.cfi_startproc
vmovaps _ZL4mask(%rip), %xmm0
vandps .LC0(%rip), %xmm0, %xmm2
vandnps .LC0(%rip), %xmm0, %xmm1
vmulpd %xmm1, %xmm2, %xmm1
vandps %xmm0, %xmm1, %xmm0
vsubpd %xmm0, %xmm1, %xmm1
vmulpd %xmm1, %xmm0, %xmm0
ret
.cfi_endproc
The second foo call is correct: vandps and vsubpd are used. But the first call
uses vandps and vandnps. This pattern would be correct for integers, but is
obviously wrong for floating point numbers.
