i've experiencing a problem with gcc-4.6.1 using avx intrinsics to write a sign
function. when compiling the following code with -ffast-math, some wrong code is
generated.
the code is the following:
essentially it seems that the sign_mask bitmask is initialized as 0 instead of
0x80000000 and therefore one `and' and one `or' instruction are optimized away.
template <typename VecType>
inline VecType vec_sign(VecType const & arg)
{
typedef VecType vec;
const vec zero = vec::gen_zero();
const vec one = vec::gen_one();
const vec sign_mask = vec::gen_sign_mask(); // wrongly initialized
const vec nonzero = mask_neq(arg, zero);
const vec sign = arg & sign_mask;
const vec abs_ret = nonzero & one;
const vec ret = sign | abs_ret;
return ret;
}
&, | and mask_neq are implemented with _mm256_and_ps, _mm256_or_ps and
_mm256_cmp_ps. sign_mask is generated via:
template <>
struct vec<float>
{
[...]
__m256 data_;
static inline __m256 gen_sign_mask(void)
{
return set_bitmask(0x80000000);
}
static inline __m256 set_bitmask(unsigned int mask)
{
union {
unsigned int i;
float f;
} u;
u.i = mask;
return _mm256_set1_ps(u.f);
}
[...]
};
the generated code looks is essentially the following:
00000000000056f0 <sign_nova(UnaryOpUGen*, int)>:
56f0: 55 push %rbp
56f1: c1 ee 04 shr $0x4,%esi
56f4: c5 f8 57 c0 vxorps %xmm0,%xmm0,%xmm0
56f8: 48 89 e5 mov %rsp,%rbp
56fb: 48 83 e4 e0 and $0xffffffffffffffe0,%rsp
56ff: 83 ee 01 sub $0x1,%esi
5702: 48 83 c4 10 add $0x10,%rsp
5706: 48 c1 e6 06 shl $0x6,%rsi
570a: 48 8b 47 50 mov 0x50(%rdi),%rax
570e: c5 fc 28 0d ca 88 00 vmovaps 0x88ca(%rip),%ymm1 # dfe0
<_fini+0x498>
5715: 00
5716: 48 8b 08 mov (%rax),%rcx
5719: 48 8b 47 48 mov 0x48(%rdi),%rax
571d: 48 8d 51 40 lea 0x40(%rcx),%rdx
5721: 48 8b 00 mov (%rax),%rax
5724: 48 01 d6 add %rdx,%rsi
5727: eb 0b jmp 5734 <sign_nova(UnaryOpUGen*,
int)+0x44>
5729: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
5730: 48 83 c2 40 add $0x40,%rdx
5734: c5 fc 28 10 vmovaps (%rax),%ymm2
5738: c5 ec c2 d0 04 vcmpneqps %ymm0,%ymm2,%ymm2
573d: c5 ec 54 d1 vandps %ymm1,%ymm2,%ymm2
5741: c5 fc 29 11 vmovaps %ymm2,(%rcx)
5745: c5 fc 28 50 20 vmovaps 0x20(%rax),%ymm2
574a: 48 83 c0 40 add $0x40,%rax
574e: 48 39 f2 cmp %rsi,%rdx
5751: c5 ec c2 d0 04 vcmpneqps %ymm0,%ymm2,%ymm2
5756: c5 ec 54 d1 vandps %ymm1,%ymm2,%ymm2
575a: c5 fc 29 51 20 vmovaps %ymm2,0x20(%rcx)
575f: 48 89 d1 mov %rdx,%rcx
5762: 75 cc jne 5730 <sign_nova(UnaryOpUGen*,
int)+0x40>
5764: c9 leaveq
5765: c5 f8 77 vzeroupper
5768: c3 retq
5769: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
without -ffast-math the code is correct:
0000000000001350 <sign_nova(UnaryOpUGen*, int)>:
1350: 55 push %rbp
1351: c1 ee 04 shr $0x4,%esi
1354: c5 f0 57 c9 vxorps %xmm1,%xmm1,%xmm1
1358: 48 89 e5 mov %rsp,%rbp
135b: 48 83 e4 e0 and $0xffffffffffffffe0,%rsp
135f: 83 ee 01 sub $0x1,%esi
1362: 48 83 c4 10 add $0x10,%rsp
1366: 48 c1 e6 06 shl $0x6,%rsi
136a: 48 8b 47 50 mov 0x50(%rdi),%rax
136e: c5 fc 28 1d 00 00 00 vmovaps 0x0(%rip),%ymm3 # 1376
<sign_nova(UnaryOpUGen*, int)+0x26>
1375: 00
1376: c5 fc 28 15 00 00 00 vmovaps 0x0(%rip),%ymm2 # 137e
<sign_nova(UnaryOpUGen*, int)+0x2e>
137d: 00
137e: 48 8b 08 mov (%rax),%rcx
1381: 48 8b 47 48 mov 0x48(%rdi),%rax
1385: 48 8d 51 40 lea 0x40(%rcx),%rdx
1389: 48 8b 00 mov (%rax),%rax
138c: 48 01 d6 add %rdx,%rsi
138f: eb 0b jmp 139c <sign_nova(UnaryOpUGen*,
int)+0x4c>
1391: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
1398: 48 83 c2 40 add $0x40,%rdx
139c: c5 fc 28 00 vmovaps (%rax),%ymm0
13a0: c5 fc c2 e1 04 vcmpneqps %ymm1,%ymm0,%ymm4
13a5: c5 fc 54 c3 vandps %ymm3,%ymm0,%ymm0
13a9: c5 dc 54 e2 vandps %ymm2,%ymm4,%ymm4
13ad: c5 fc 56 c4 vorps %ymm4,%ymm0,%ymm0
13b1: c5 fc 29 01 vmovaps %ymm0,(%rcx)
13b5: c5 fc 28 40 20 vmovaps 0x20(%rax),%ymm0
13ba: 48 83 c0 40 add $0x40,%rax
13be: 48 39 f2 cmp %rsi,%rdx
13c1: c5 fc c2 e1 04 vcmpneqps %ymm1,%ymm0,%ymm4
13c6: c5 fc 54 c3 vandps %ymm3,%ymm0,%ymm0
13ca: c5 dc 54 e2 vandps %ymm2,%ymm4,%ymm4
13ce: c5 fc 56 c4 vorps %ymm4,%ymm0,%ymm0
13d2: c5 fc 29 41 20 vmovaps %ymm0,0x20(%rcx)
13d7: 48 89 d1 mov %rdx,%rcx
13da: 75 bc jne 1398 <sign_nova(UnaryOpUGen*,
int)+0x48>
13dc: c9 leaveq
13dd: c5 f8 77 vzeroupper
13e0: c3 retq
13e1: 66 66 66 66 66 66 2e data32 data32 data32 data32 data32 nopw
%cs:0x0(%rax,%rax,1)
13e8: 0f 1f 84 00 00 00 00
13ef: 00
is this a bug in the compiler or have i merely been lucky that my sse-specific
code was never hit by this optimization?
thanks, tim
