https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91723
Bug ID: 91723
Summary: builtin fma is not optimized or vectorized as *+
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: nsz at gcc dot gnu.org
Target Milestone: ---
i'd expect a*b+c to generate the same code as __builtin_fmaf(a,b,c)
when hw instruction is available for fmaf, but the later generates
significantly worst code in some cases, e.g. when vectorization
is involved.
consider:
void
foo (float *restrict r, const float *restrict a,
const float *restrict b, const float *restrict c)
{
for (int i=0; i < 4; i++) {
float x;
#ifdef BUILTIN
x = __builtin_fmaf(a[i],b[i],c[i]);
x = __builtin_fmaf(a[i],b[i],x);
#else
x = a[i]*b[i]+c[i];
x = a[i]*b[i]+x;
#endif
r[i] = x;
}
}
with gcc -O3 -mfma -mavx -ffp-contract=fast -fno-math-errno i get good code:
foo:
vmovups (%rdx), %xmm0
vmovups (%rsi), %xmm1
vmovaps %xmm0, %xmm2
vfmadd213ps (%rcx), %xmm1, %xmm2
vfmadd132ps %xmm1, %xmm2, %xmm0
vmovups %xmm0, (%rdi)
ret
but if i add -DBUILTIN i get
foo:
vmovss (%rsi), %xmm0
vmovss (%rdx), %xmm1
vmovaps %xmm0, %xmm2
vfmadd213ss (%rcx), %xmm1, %xmm2
vfmadd132ss %xmm1, %xmm2, %xmm0
vmovss 4(%rdx), %xmm1
vmovss %xmm0, (%rdi)
vmovss 4(%rsi), %xmm0
vmovaps %xmm0, %xmm2
vfmadd213ss 4(%rcx), %xmm1, %xmm2
vfmadd132ss %xmm1, %xmm2, %xmm0
vmovss 8(%rdx), %xmm1
vmovss %xmm0, 4(%rdi)
vmovss 8(%rsi), %xmm0
vmovaps %xmm0, %xmm2
vfmadd213ss 8(%rcx), %xmm1, %xmm2
vfmadd132ss %xmm1, %xmm2, %xmm0
vmovss 12(%rdx), %xmm1
vmovss %xmm0, 8(%rdi)
vmovss 12(%rsi), %xmm0
vmovaps %xmm0, %xmm2
vfmadd213ss 12(%rcx), %xmm1, %xmm2
vfmadd132ss %xmm1, %xmm2, %xmm0
vmovss %xmm0, 12(%rdi)
ret
i expected identical results, the same happens on other targets.
