https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63737
Bug ID: 63737
Summary: Missed optimization: -ffixed-reg and unelided copies
Product: gcc
Version: 4.9.1
Status: UNCONFIRMED
Severity: trivial
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: tstache1 at binghamton dot edu
Excuse me if I'm misunderstanding the semantics of -ffixed-reg and pinned
registers, but take this trivial example:
#include <emmintrin.h>
register __m128i fixed_reg_1 __asm__ ("xmm6");
register __m128i fixed_reg_2 __asm__ ("xmm7");
__m128i xmm_add(__m128i a, __m128i b) {
__m128i dest;
fixed_reg_1 = a + b;
fixed_reg_2 = fixed_reg_1 + b;
dest = fixed_reg_1 + fixed_reg_2;
return dest;
}
Compiling it (-c -O2 -ffixed-xmm6 -ffixed-reg-xmm7 -mavx) seems to produce the
following, non-optimal code (the compiler does not elide the copies):
0000000000000000 <xmm_add>:
0: c5 f9 d4 c1 vpaddq %xmm1,%xmm0,%xmm0
4: c5 f9 6f f0 vmovdqa %xmm0,%xmm6
8: c5 f1 d4 c0 vpaddq %xmm0,%xmm1,%xmm0
c: c5 f9 6f f8 vmovdqa %xmm0,%xmm7
10: c5 f9 d4 c6 vpaddq %xmm6,%xmm0,%xmm0
14: c3 retq
Ideally, I would think that the compiler should generate code like so, no?
0000000000000000 <xmm_add_opt>:
0: c5 f9 d4 f1 vpaddq %xmm1,%xmm0,%xmm6
4: c5 f1 d4 fe vpaddq %xmm6,%xmm1,%xmm7
8: c5 c1 d4 c6 vpaddq %xmm6,%xmm7,%xmm0
c: c3 retq
