User account creation filtered due to spam.

Bug 19274 - temporary not eliminated in composite _mm_set_ps1
Summary: temporary not eliminated in composite _mm_set_ps1
Alias: None
Product: gcc
Classification: Unclassified
Component: target (show other bugs)
Version: 4.0.0
: P2 normal
Target Milestone: 4.0.0
Assignee: Not yet assigned to anyone
Keywords: missed-optimization, ssemmx
Depends on:
Reported: 2005-01-05 18:07 UTC by tbp
Modified: 2005-08-13 21:25 UTC (History)
1 user (show)

See Also:
Target: i?86-*-*
Known to work:
Known to fail:
Last reconfirmed: 2005-01-05 18:45:57


Note You need to log in before you can comment on or make changes to this bug.
Description tbp 2005-01-05 18:07:25 UTC
Sometime the temporary in that composite intrinsic doesn't get eliminated.

Test case: -O3 -march=k8  -fomit-frame-pointer
#include <xmmintrin.h>

__m128 not_eliminated(const float f1, const float f2) {
	const __m128
		a = _mm_set_ps1(f1),
		b = _mm_set_ps1(f2),
		c = _mm_mul_ps(a, b);
	return c;

#define broadcast(f) _mm_shuffle_ps(_mm_load_ss(&(f)),_mm_load_ss(&(f)),0)
__m128 eliminated(const float f1, const float f2) {
	const __m128
		a = broadcast(f1),
		b = broadcast(f2),
		c = _mm_mul_ps(a, b);
	return c;

int main() { return 0; }

With gcc4-20040102 + patch
(happens in older version too), i still get:
00401070 <not_eliminated(float, float)>:
  401070:       sub    $0x10,%esp
  401073:       mov    0x14(%esp),%eax
  401077:       mov    %eax,0xc(%esp)
  40107b:       mov    0x18(%esp),%eax
  40107f:       movss  0xc(%esp),%xmm0
  401085:       shufps $0x0,%xmm0,%xmm0
  401089:       mov    %eax,0xc(%esp)
  40108d:       movss  0xc(%esp),%xmm1
  401093:       add    $0x10,%esp
  401096:       shufps $0x0,%xmm1,%xmm1
  40109a:       mulps  %xmm1,%xmm0
  40109d:       ret    

004010a0 <eliminated(float, float)>:
  4010a0:       movss  0x4(%esp),%xmm0
  4010a6:       movss  0x8(%esp),%xmm1
  4010ac:       shufps $0x0,%xmm0,%xmm0
  4010b0:       shufps $0x0,%xmm1,%xmm1
  4010b4:       mulps  %xmm1,%xmm0
  4010b7:       ret    

It might happens with other intrinsics, but i haven't spotted it yet :)
Comment 1 Andrew Pinski 2005-01-05 18:45:57 UTC
Confirmed, it comes down to the defintion of _mm_set1_ps and __builtin_ia32_loadss.
Comment 2 Andrew Pinski 2005-01-12 06:07:47 UTC
This has been fixed now, I think by a patch which also fixed PR 13366.
We get the same code for both functions in 32bit mode.
But note in 64bit mode, we get better code for not_eliminated:
        shufps  $0, %xmm1, %xmm1
        shufps  $0, %xmm0, %xmm0
        mulps   %xmm1, %xmm0

        movss   %xmm0, -4(%rsp)
        movl    -4(%rsp), %eax
        movss   %xmm1, -4(%rsp)
        movd    %eax, %xmm0
        movl    -4(%rsp), %eax
        shufps  $0, %xmm0, %xmm0
        movd    %eax, %xmm1
        shufps  $0, %xmm1, %xmm1
        mulps   %xmm1, %xmm0