https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109907
--- Comment #21 from Georg-Johann Lay <gjl at gcc dot gnu.org> ---
One more test:
unsigned char lfsr32_mpp_ge0 (unsigned long number)
{
unsigned char b = 0;
if (number >= 0) b--;
if (number & (1UL << 29)) b++;
if (number & (1UL << 13)) b++;
return b;
}
with -Os -mmcu=atmega128 -dp it generates:
lfsr32_ppp_ge0:
push r16 ; 85 [c=4 l=1] pushqi1/0
push r17 ; 86 [c=4 l=1] pushqi1/0
/* prologue: function */
/* frame size = 0 */
/* stack size = 2 */
movw r16,r22 ; 93 [c=4 l=1] *movhi/0
movw r18,r24 ; 94 [c=4 l=1] *movhi/0
movw r22,r18 ; 82 [c=4 l=2] *movsi/0
movw r20,r16
clr r20 ; 83 [c=16 l=4] *andsi3/1
clr r21
clr r22
andi r23,32
ldi r24,lo8(1) ; 68 [c=4 l=1] movqi_insn/1
sbrc r17,5 ; 84 [c=28 l=2] *sbrx_branchsi
rjmp .L18
or r20,r21 ; 75 [c=16 l=3] *cmpsi/0
or r20,r22
or r20,r23
breq .L19 ; 77 [c=12 l=1] *branch
ldi r24,0 ; 73 [c=4 l=1] movqi_insn/0
.L19:
neg r24 ; 72 [c=4 l=1] *negqi2
.L17:
/* epilogue start */
pop r17 ; 89 [c=4 l=1] popqi
pop r16 ; 90 [c=4 l=1] popqi
ret ; 91 [c=0 l=1] return_from_epilogue
.L18:
or r20,r21 ; 69 [c=16 l=3] *cmpsi/0
or r20,r22
or r20,r23
brne .L17 ; 71 [c=12 l=1] *branch
ldi r24,0 ; 67 [c=4 l=1] movqi_insn/0
rjmp .L17 ; 95 [c=4 l=1] jump
so it does arithmetic on 32-bit variables (one AND and two COMPAREs) in 28
instructions, use more stack and high register pressure. An optimal code would
require just 8 instructions and additional register pressure of just 1 byte for
the output:
lfsr32_mpp_ge0:
/* prologue: function */
/* frame size = 0 */
/* stack size = 0 */
clr R26 ;; b lives in R26, number lives in R25:R22.
sbrs R25, 7 ;; Skip next if number.31 = 1
dec R26
sbrc R25, 5 ;; Skip next if number.29 = 0
inc R26
sbrc R23, 5 ;; Skip next if number.13 = 0
inc R26
mov R24, r26
/* epilogue start */
ret
