https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66933
Bug ID: 66933
Summary: [AVR] Shifted multiplication produces suboptimal asm
Product: gcc
Version: 5.2.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: a-gnu.org at 0au dot de
Target Milestone: ---
The function
uint8_t test(uint8_t a, uint8_t b) { return (a*b) >> 7; }
compiles into the following assembly, with e.g.
avr-gcc -mmcu=atmega328 -S -O3 test.c
(or any other optimization flag)
test:
mul r24,r22
movw r24,r0
clr __zero_reg__
lsl r24
mov r24,r25
rol r24
sbc r25,r25
ret
This has two obvious possible enhancements:
- It uses mul instead of fmul: fmul calculates (a*b)<<1,
so the high byte is already the correct return value of
the function
- After calculating the return value (wih "rol r24"),
there's an instruction "sbc r25, r25" that puts a
completely unneeded value in r25, if I'm not mistaken
it's 255 if (a*b) & 0x8000, else 0.
A better version that uses 8 instead of 12 cycles would be
test:
fmul r24, r22
mov r24, r1
clr __zero_reg__
ret
