https://bugs.llvm.org/show_bug.cgi?id=48292
Bug ID: 48292
Summary: [aarch64] Inefficient code in __builtin_mul_overflow
with unsigned < 32 bits
Product: libraries
Version: 10.0
Hardware: Other
OS: Linux
Status: NEW
Severity: enhancement
Priority: P
Component: Backend: AArch64
Assignee: unassignedb...@nondot.org
Reporter: husseyde...@gmail.com
CC: arnaud.degrandmai...@arm.com,
llvm-bugs@lists.llvm.org, smithp...@googlemail.com,
ties.st...@arm.com
Given the following C code:
bool umul16_overflow(uint16_t a, uint16_t b, uint16_t *result)
{
return __builtin_mul_overflow(a,b,result);
}
aarch64 clang 10.0.1 -O2:
umul16_overflow:
// Cast args
and w9, w1, #0xffff
and w10, w0, #0xffff
// 16-bit to 32-bit long multiply
mul w9, w10, w9
// Store upper 16 bits in w10
lsr w10, w9, #16
// Store zero in x8??????????
mov x8, xzr
// Check if upper 16 bits (w10) are non-zero
cmp w10, #0
cset w10, ne
// Check if zero (x8) is non-zero??????????
// w8 = 0
cmp x8, #0
cset w8, ne
// Or the result against zero??????????
orr w0, w10, w8
// Store
strh w9, [x2]
ret
I have no idea what x8 is being used for here, it will always be zero. This is
the code without the redundant instructions:
umul16_overflow:
// Cast args
and w8, w1, #0xffff
and w9, w0, #0xffff
// 16-bit to 32-bit long multiply
mul w8, w9, w8
// Store upper 16 bits in w9
lsr w9, w8, #16
// Check if the upper 16 bits (w9) are non-zero
cmp w9, #0
cset w0, ne
// Store
strh w9, [x2]
ret
Note that we could simplify this even further by using cmp more creatively, but
this appears to be a different issue.
umul16_overflow:
// Cast args
and w8, w1, #0xffff
and w9, w0, #0xffff
// 16-bit to 32-bit long multiply
mul w8, w9, w8
// Check if the upper 16 bits are non-zero
cmp wzr, w8, lsr #16
cset w0, ne
// Store
strh w8, [x2]
ret
--
You are receiving this mail because:
You are on the CC list for the bug._______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
