https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85628
Bug ID: 85628
Summary: Make better use of BFI (BFXIL)
Product: gcc
Version: unknown
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
Target: aarch64
The testcase:
unsigned long long combine(unsigned long long a, unsigned long long b) {
return (a & 0xffffffff00000000ll) | (b & 0x00000000ffffffffll);
}
void read2(unsigned long long a, unsigned long long b, unsigned long long *c,
unsigned long long *d) {
*c = combine(a, b); *d = combine(b, a);
}
on aarch64 with -O2 currently generates:
combine:
bfi x0, x1, 0, 32
ret
read2:
and x5, x1, 4294967295
and x4, x0, -4294967296
orr x4, x4, x5
and x1, x1, -4294967296
and x0, x0, 4294967295
str x4, [x2]
orr x0, x0, x1
str x0, [x3]
ret
With LLVM it does a better job:
combine: // @combine
bfxil x0, x1, #0, #32
ret
read2: // @read2
mov x8, x0
bfxil x8, x1, #0, #32
bfxil x1, x0, #0, #32
str x8, [x2]
str x1, [x3]
ret
This should just be a matter of adding the necessary patterns in aarch64.md.
Combine already tries to match:
(set (reg:DI 105)
(ior:DI (and:DI (reg/v:DI 97 [ b ])
(const_int -4294967296 [0xffffffff00000000]))
(and:DI (reg/v:DI 96 [ a ])
(const_int 4294967295 [0xffffffff]))))
