https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65968
Bug ID: 65968
Summary: Failure to remove casts, cause poor code generation
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: law at redhat dot com
Target Milestone: ---
Failure to shorten the multiplies from int mode down to their more native short
modes causes poor code generation for this loop:
void f(short*a) {
a = __builtin_assume_aligned(a,128);
for (int i = 0; i < (1<<22); ++i) {
#ifdef EASY
a[i] *= a[i];
#else
int x = a[i];
x *= x;
a[i] = x;
#endif
}
}
With -DEASY, a nice little loop:
.L2:
movdqa (%rdi), %xmm0
addq $16, %rdi
pmullw %xmm0, %xmm0
movaps %xmm0, -16(%rdi)
cmpq %rdi, %rax
jne .L2
while without EASY, we get the uglier:
.L2:
movdqa (%rdi), %xmm0
addq $16, %rdi
movdqa %xmm0, %xmm2
movdqa %xmm0, %xmm1
pmullw %xmm0, %xmm2
pmulhw %xmm0, %xmm1
movdqa %xmm2, %xmm0
punpckhwd %xmm1, %xmm2
punpcklwd %xmm1, %xmm0
movdqa %xmm2, %xmm1
movdqa %xmm0, %xmm2
punpcklwd %xmm1, %xmm0
punpckhwd %xmm1, %xmm2
movdqa %xmm0, %xmm1
punpcklwd %xmm2, %xmm0
punpckhwd %xmm2, %xmm1
punpcklwd %xmm1, %xmm0
movaps %xmm0, -16(%rdi)
cmpq %rdi, %rax
jne .L2
The narrowing patterns currently in match.pd and proposed for match.pd at the
time of submitting this BZ handle plus/minus, but not multiply. When writing
the current patterns I saw regressions when mult handling was included.
Finding a way to avoid the regressions (should have filed BZs for them) while
still shortening for this case would be good.
Marc indicates that pattern along these lines:
(simplify
(vec_pack_trunc (widen_mult_lo @0 @1) (widen_mult_hi:c @0 @1))
(mult @0 @1))
Would help this specific case, but we may do better if we can do the type
narrowing before vectorization.
