https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93828
Bug ID: 93828
Summary: [10 Regression] incorrect shufps instruction emitted
for -march=k8
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Keywords: wrong-code
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: kretz at kde dot org
Target Milestone: ---
Target: x86_64-*-*
Test case (https://godbolt.org/z/ramAe3):
using float2 [[gnu::vector_size(8)]] = float;
using int2 [[gnu::vector_size(8)]] = int;
float2 y = {2, 2};
int main() {
const auto k = y == float2{2, 2};
if (k[1] == 0)
__builtin_abort();
const auto a = k & int2{2, 2};
return a[0] - 2;
}
Compile with `-O2 -march=k8`. The resulting instruction sequence uses:
movlps xmm0, QWORD PTR y[rip]
shufps xmm1, xmm0, 0xe5
to place y[0] in xmm0[0] and y[1] in xmm1[0]. The latter is missing a `movaps
xmm1, xmm0` to work correctly, though. Most/all other -march flags load
individual floats (using movss) from y.
