https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70721
Bug ID: 70721
Summary: Suboptimal code generated when using _mm_min_sd
Product: gcc
Version: 7.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: hjl.tools at gmail dot com
CC: kirill.yukhin at intel dot com
Target Milestone: ---
_mm_min_sd is implemented as
(define_insn "<sse>_vm<code><mode>3<round_saeonly_name>"
[(set (match_operand:VF_128 0 "register_operand" "=x,v")
(vec_merge:VF_128
(smaxmin:VF_128
(match_operand:VF_128 1 "register_operand" "0,v")
(match_operand:VF_128 2 "vector_operand"
"xBm,<round_saeonly_constraint>"))
(match_dup 1)
(const_int 1)))]
The problem is smaxmin is applied to the full 128-bit operand.
Can we change it to apply only to the first 64-bit of operand
so that we can remove 2 xmm moves in
---
#include <emmintrin.h>
double
__attribute ((noinline, noclone))
foo (double a, double b)
{
__m128d x = _mm_set_sd(a);
__m128d y = _mm_set_sd(b);
return _mm_cvtsd_f64(_mm_min_sd(x, y));
}
---
