https://llvm.org/bugs/show_bug.cgi?id=26491
Bug ID: 26491
Summary: _mm_movehl_ps pesimised to movaps/shufpd instead of
movhlps
Product: clang
Version: 3.7
Hardware: PC
OS: Linux
Status: NEW
Keywords: performance
Severity: normal
Priority: P
Component: LLVM Codegen
Assignee: unassignedclangb...@nondot.org
Reporter: pe...@cordes.ca
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
clang sometimes uses a different instruction that the "normal" one for the
intrinsic. Usually it gets an improvement, but not in this case where it ends
up wasting a movaps and using a bigger instruction.
float hsum_ps(__m128 v) {
__m128 tmp = _mm_movehdup_ps(v); // dest of MOVSHDUP is write-only
__m128 sums = _mm_add_ps(v, tmp);
tmp = _mm_movehl_ps(tmp, sums);
sums = _mm_add_ss(sums, tmp);
return _mm_cvtss_f32(sums);
}
compiles to: clang 3.7.1 -O3 -march=core2
movshdup xmm1, xmm0 # xmm1 = xmm0[1,1,3,3]
addps xmm1, xmm0
movaps xmm0, xmm1
shufpd xmm0, xmm0, 1 # xmm0 = xmm0[1,0]
addss xmm0, xmm1
ret
movhlps is very compact: OF 12 /r, vs. shufpd being a 5 byte insn. At least
it doesn't use shufps, which is slow on Core2. A similar sequence after a loop
produced a movaps/unpckhpd instead of the movhlps.
What I was hoping for (and which gcc produces) is the obvious literal
translation from intrinsics to instructions:
movshdup xmm1, xmm0
addps xmm0, xmm1
movhlps xmm1, xmm0
addss xmm0, xmm1
ret
I used movshdup first because its destination is write-only. I can avoid a
movaps without needing the compiler to keep some other variable live so I can
movhlps into it. (And so I could put it in a function that only takes one
parameter.) I can't use _mm_undefined_ps() without risk of denormals or NaN
slowdowns.
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
