https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61299
Bug ID: 61299
Summary: [4.9 Regression] Performance regression for the SIMD
rotate operation with GCC vector extensions
Product: gcc
Version: 4.9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: siarhei.siamashka at gmail dot com
A small test:
/**************/
typedef unsigned int uint32x4 __attribute__ ((vector_size(16)));
typedef struct { uint32x4 a, b; } prng_t;
void foo(prng_t *x)
{
x->a ^= ((x->b << 17) ^ (x->b >> (32 - 17)));
}
/**************/
Gets compiled into the following slow code with GCC 4.9 (CFLAGS="-O3"):
0000000000000000 <foo>:
0: 66 0f 6f 47 10 movdqa 0x10(%rdi),%xmm0
5: 66 0f 70 c8 55 pshufd $0x55,%xmm0,%xmm1
a: 66 0f 7e c0 movd %xmm0,%eax
e: c1 c8 0f ror $0xf,%eax
11: 89 44 24 e8 mov %eax,-0x18(%rsp)
15: 66 0f 7e c8 movd %xmm1,%eax
19: 66 0f 6f c8 movdqa %xmm0,%xmm1
1d: c1 c8 0f ror $0xf,%eax
20: 66 0f 6a c8 punpckhdq %xmm0,%xmm1
24: 89 44 24 ec mov %eax,-0x14(%rsp)
28: 66 0f 70 c0 ff pshufd $0xff,%xmm0,%xmm0
2d: 66 0f 6e 5c 24 ec movd -0x14(%rsp),%xmm3
33: 66 0f 7e c8 movd %xmm1,%eax
37: c1 c8 0f ror $0xf,%eax
3a: 89 44 24 f0 mov %eax,-0x10(%rsp)
3e: 66 0f 7e c0 movd %xmm0,%eax
42: 66 0f 6e 44 24 e8 movd -0x18(%rsp),%xmm0
48: 66 0f 6e 4c 24 f0 movd -0x10(%rsp),%xmm1
4e: c1 c8 0f ror $0xf,%eax
51: 66 0f 62 c3 punpckldq %xmm3,%xmm0
55: 89 44 24 f4 mov %eax,-0xc(%rsp)
59: 66 0f 6e 54 24 f4 movd -0xc(%rsp),%xmm2
5f: 66 0f 62 ca punpckldq %xmm2,%xmm1
63: 66 0f 6c c1 punpcklqdq %xmm1,%xmm0
67: 66 0f ef 07 pxor (%rdi),%xmm0
6b: 0f 29 07 movaps %xmm0,(%rdi)
6e: c3 retq
It used to be a lot better with GCC 4.8 (CFLAGS="-O3"):
0000000000000000 <foo>:
0: 66 0f 6f 4f 10 movdqa 0x10(%rdi),%xmm1
5: 66 0f 6f c1 movdqa %xmm1,%xmm0
9: 66 0f 72 d1 0f psrld $0xf,%xmm1
e: 66 0f 72 f0 11 pslld $0x11,%xmm0
13: 66 0f ef c1 pxor %xmm1,%xmm0
17: 66 0f ef 07 pxor (%rdi),%xmm0
1b: 66 0f 7f 07 movdqa %xmm0,(%rdi)
1f: c3 retq
