Apologies if this is the wrong list.
I'm afraid I'm not much of an assembly programmer, but I was just
wondering if this generated code was 'correct', because from descriptions
of SSE that I've read, it looks like it's inefficient.
The C code:
float *vector_add4f(float va[4], const float vb[4])
{
va[0] += vb[0];
va[1] += vb[1];
va[2] += vb[2];
va[3] += vb[3];
return va;
}
Now unless my understanding is totally off, the processor should be able
to do those four additions in one by using the SSE extensions. The standard
code generated (without SSE) is as expected:
88: 55 push %ebp
89: 89 e5 mov %esp,%ebp
8b: 8b 45 08 mov 0x8(%ebp),%eax
8e: 8b 55 0c mov 0xc(%ebp),%edx
91: d9 00 flds (%eax)
93: d8 02 fadds (%edx)
95: d9 18 fstps (%eax)
97: d9 40 04 flds 0x4(%eax)
9a: d8 42 04 fadds 0x4(%edx)
9d: d9 58 04 fstps 0x4(%eax)
a0: d9 40 08 flds 0x8(%eax)
a3: d8 42 08 fadds 0x8(%edx)
a6: d9 58 08 fstps 0x8(%eax)
a9: d9 40 0c flds 0xc(%eax)
ac: d8 42 0c fadds 0xc(%edx)
af: d9 58 0c fstps 0xc(%eax)
b2: c9 leave
b3: c3 ret
Using -march=pentium3 -mtune=pentium3m -mfpmath=sse, the following
is generated:
140: 55 push %ebp
141: 89 e5 mov %esp,%ebp
143: 8b 4d 08 mov 0x8(%ebp),%ecx
146: 8b 45 08 mov 0x8(%ebp),%eax
149: 8b 55 0c mov 0xc(%ebp),%edx
14c: f3 0f 10 00 movss (%eax),%xmm0
150: f3 0f 58 02 addss (%edx),%xmm0
154: f3 0f 11 01 movss %xmm0,(%ecx)
158: 8b 4d 08 mov 0x8(%ebp),%ecx
15b: 83 c1 04 add $0x4,%ecx
15e: 8b 45 08 mov 0x8(%ebp),%eax
161: 83 c0 04 add $0x4,%eax
164: 8b 55 0c mov 0xc(%ebp),%edx
167: 83 c2 04 add $0x4,%edx
16a: f3 0f 10 00 movss (%eax),%xmm0
16e: f3 0f 58 02 addss (%edx),%xmm0
172: f3 0f 11 01 movss %xmm0,(%ecx)
176: 8b 4d 08 mov 0x8(%ebp),%ecx
179: 83 c1 08 add $0x8,%ecx
17c: 8b 45 08 mov 0x8(%ebp),%eax
17f: 83 c0 08 add $0x8,%eax
182: 8b 55 0c mov 0xc(%ebp),%edx
185: 83 c2 08 add $0x8,%edx
188: f3 0f 10 00 movss (%eax),%xmm0
18c: f3 0f 58 02 addss (%edx),%xmm0
190: f3 0f 11 01 movss %xmm0,(%ecx)
194: 8b 4d 08 mov 0x8(%ebp),%ecx
197: 83 c1 0c add $0xc,%ecx
19a: 8b 45 08 mov 0x8(%ebp),%eax
19d: 83 c0 0c add $0xc,%eax
1a0: 8b 55 0c mov 0xc(%ebp),%edx
1a3: 83 c2 0c add $0xc,%edx
1a6: f3 0f 10 00 movss (%eax),%xmm0
1aa: f3 0f 58 02 addss (%edx),%xmm0
1ae: f3 0f 11 01 movss %xmm0,(%ecx)
1b2: 8b 45 08 mov 0x8(%ebp),%eax
1b5: 5d pop %ebp
1b6: c3 ret
1b7: 89 f6 mov %esi,%esi
1b9: 8d bc 27 00 00 00 00 lea 0x0(%edi),%edi
Now, uh, isn't that four additions? Do I need to do something gcc-specific
to get it to use the 'add-packed-single' instruction to turn those four
additions into one?
MC
