https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110093
Bug ID: 110093
Summary: [12/13/14 Regression][avr] Move frenzy leading to code
bloat
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: gjl at gcc dot gnu.org
Target Milestone: ---
So here is a C test case:
long add (long aa, long bb, long cc)
{
if (cc < 0)
return aa - cc;
return aa + bb;
}
$ avr-gcc-8 -S -Os -dp
compiles this to following assembly:
add:
push r14 ; 30 [c=4 l=1] pushqi1/0
push r15 ; 31 [c=4 l=1] pushqi1/0
push r16 ; 32 [c=4 l=1] pushqi1/0
push r17 ; 33 [c=4 l=1] pushqi1/0
/* prologue: function */
/* frame size = 0 */
/* stack size = 4 */
sbrs r17,7 ; 42 [c=28 l=2] *sbrx_and_branchsi
rjmp .L2
sub r22,r14 ; 11 [c=16 l=4] subsi3/0
sbc r23,r15
sbc r24,r16
sbc r25,r17
.L1:
/* epilogue start */
pop r17 ; 36 [c=4 l=1] popqi
pop r16 ; 37 [c=4 l=1] popqi
pop r15 ; 38 [c=4 l=1] popqi
pop r14 ; 39 [c=4 l=1] popqi
ret ; 40 [c=0 l=1] return_from_epilogue
.L2:
add r22,r18 ; 16 [c=16 l=4] addsi3/0
adc r23,r19
adc r24,r20
adc r25,r21
rjmp .L1 ; 43 [c=4 l=1] jump
Notice that the operations on aa in SI:22 can be done in place, no moves
needed. The superfluous PUSHes and POPs is PR109910, which is yet another
issue...
The code from above then deteriorates with v12, v13, v14 20230501 to a move
bonanza that starts moving stuff for no reason, leading to high register
pressure, required stack increases from 4 bytes to 14 bytes, code size
increase from 20 instructions to 56:
add:
push r4 ; 85 [c=4 l=1] pushqi1/0
push r5 ; 86 [c=4 l=1] pushqi1/0
push r6 ; 87 [c=4 l=1] pushqi1/0
push r7 ; 88 [c=4 l=1] pushqi1/0
push r8 ; 89 [c=4 l=1] pushqi1/0
push r9 ; 90 [c=4 l=1] pushqi1/0
push r10 ; 91 [c=4 l=1] pushqi1/0
push r11 ; 92 [c=4 l=1] pushqi1/0
push r12 ; 93 [c=4 l=1] pushqi1/0
push r13 ; 94 [c=4 l=1] pushqi1/0
push r14 ; 95 [c=4 l=1] pushqi1/0
push r15 ; 96 [c=4 l=1] pushqi1/0
push r16 ; 97 [c=4 l=1] pushqi1/0
push r17 ; 98 [c=4 l=1] pushqi1/0
/* prologue: function */
/* frame size = 0 */
/* stack size = 14 */
.L__stack_usage = 14
mov r4,r22 ; 68 [c=4 l=1] movqi_insn/0
mov r5,r23 ; 69 [c=4 l=1] movqi_insn/0
mov r6,r24 ; 70 [c=4 l=1] movqi_insn/0
mov r7,r25 ; 71 [c=4 l=1] movqi_insn/0
mov r8,r18 ; 72 [c=4 l=1] movqi_insn/0
mov r9,r19 ; 73 [c=4 l=1] movqi_insn/0
mov r10,r20 ; 74 [c=4 l=1] movqi_insn/0
mov r11,r21 ; 75 [c=4 l=1] movqi_insn/0
mov r12,r14 ; 78 [c=4 l=1] movqi_insn/0
mov r13,r15 ; 79 [c=4 l=1] movqi_insn/0
mov r14,r16 ; 80 [c=4 l=1] movqi_insn/0
mov r15,r17 ; 81 [c=4 l=1] movqi_insn/0
mov r25,r7 ; 66 [c=4 l=4] *movsi/0
mov r24,r6
mov r23,r5
mov r22,r4
sbrs r15,7 ; 117 [c=28 l=2] *sbrx_and_branchsi
rjmp .L2
sub r22,r12 ; 67 [c=16 l=4] *subsi3/0
sbc r23,r13
sbc r24,r14
sbc r25,r15
.L1:
/* epilogue start */
pop r17 ; 101 [c=4 l=1] popqi
pop r16 ; 102 [c=4 l=1] popqi
pop r15 ; 103 [c=4 l=1] popqi
pop r14 ; 104 [c=4 l=1] popqi
pop r13 ; 105 [c=4 l=1] popqi
pop r12 ; 106 [c=4 l=1] popqi
pop r11 ; 107 [c=4 l=1] popqi
pop r10 ; 108 [c=4 l=1] popqi
pop r9 ; 109 [c=4 l=1] popqi
pop r8 ; 110 [c=4 l=1] popqi
pop r7 ; 111 [c=4 l=1] popqi
pop r6 ; 112 [c=4 l=1] popqi
pop r5 ; 113 [c=4 l=1] popqi
pop r4 ; 114 [c=4 l=1] popqi
ret ; 115 [c=0 l=1] return_from_epilogue
.L2:
add r22,r8 ; 65 [c=16 l=4] *addsi3/0
adc r23,r9
adc r24,r10
adc r25,r11
rjmp .L1 ; 118 [c=4 l=1] jump
Then finally, with v14 20230602, crazyness increases even more to even requires
a stack frame and a frame pointer. Register allocator starts to move stuff to
a stack slot and back again. Code size increases again from 56 instructions to
68, more stack usage:
add:
push r4 ; 84 [c=4 l=1] pushqi1/0
push r5 ; 85 [c=4 l=1] pushqi1/0
push r6 ; 86 [c=4 l=1] pushqi1/0
push r7 ; 87 [c=4 l=1] pushqi1/0
push r8 ; 88 [c=4 l=1] pushqi1/0
push r9 ; 89 [c=4 l=1] pushqi1/0
push r10 ; 90 [c=4 l=1] pushqi1/0
push r11 ; 91 [c=4 l=1] pushqi1/0
push r14 ; 92 [c=4 l=1] pushqi1/0
push r15 ; 93 [c=4 l=1] pushqi1/0
push r16 ; 94 [c=4 l=1] pushqi1/0
push r17 ; 95 [c=4 l=1] pushqi1/0
push r28 ; 96 [c=4 l=1] pushqi1/0
push r29 ; 97 [c=4 l=1] pushqi1/0
; SP -= 4 ; 101 [c=4 l=2] *addhi3_sp
rcall .
rcall .
in r28,__SP_L__ ; 127 [c=4 l=2] *movhi/7
in r29,__SP_H__
/* prologue: function */
/* frame size = 4 */
/* stack size = 18 */
.L__stack_usage = 18
mov r8,r22 ; 69 [c=4 l=1] movqi_insn/0
mov r9,r23 ; 70 [c=4 l=1] movqi_insn/0
mov r10,r24 ; 71 [c=4 l=1] movqi_insn/0
mov r11,r25 ; 72 [c=4 l=1] movqi_insn/0
std Y+1,r18 ; 73 [c=4 l=1] movqi_insn/2
std Y+2,r19 ; 74 [c=4 l=1] movqi_insn/2
std Y+3,r20 ; 75 [c=4 l=1] movqi_insn/2
std Y+4,r21 ; 76 [c=4 l=1] movqi_insn/2
mov r4,r14 ; 77 [c=4 l=1] movqi_insn/0
mov r5,r15 ; 78 [c=4 l=1] movqi_insn/0
mov r6,r16 ; 79 [c=4 l=1] movqi_insn/0
mov r7,r17 ; 80 [c=4 l=1] movqi_insn/0
sbrs r7,7 ; 124 [c=28 l=2] *sbrx_and_branchsi
rjmp .L2
mov r25,r11 ; 67 [c=4 l=4] *movsi/0
mov r24,r10
mov r23,r9
mov r22,r8
sub r22,r4 ; 68 [c=16 l=4] *subsi3/0
sbc r23,r5
sbc r24,r6
sbc r25,r7
.L1:
/* epilogue start */
; SP += 4 ; 107 [c=4 l=4] *addhi3_sp
pop __tmp_reg__
pop __tmp_reg__
pop __tmp_reg__
pop __tmp_reg__
pop r29 ; 108 [c=4 l=1] popqi
pop r28 ; 109 [c=4 l=1] popqi
pop r17 ; 110 [c=4 l=1] popqi
pop r16 ; 111 [c=4 l=1] popqi
pop r15 ; 112 [c=4 l=1] popqi
pop r14 ; 113 [c=4 l=1] popqi
pop r11 ; 114 [c=4 l=1] popqi
pop r10 ; 115 [c=4 l=1] popqi
pop r9 ; 116 [c=4 l=1] popqi
pop r8 ; 117 [c=4 l=1] popqi
pop r7 ; 118 [c=4 l=1] popqi
pop r6 ; 119 [c=4 l=1] popqi
pop r5 ; 120 [c=4 l=1] popqi
pop r4 ; 121 [c=4 l=1] popqi
ret ; 122 [c=0 l=1] return_from_epilogue
.L2:
ldd r22,Y+1 ; 65 [c=16 l=4] *movsi/2
ldd r23,Y+2
ldd r24,Y+3
ldd r25,Y+4
add r22,r8 ; 66 [c=16 l=4] *addsi3/0
adc r23,r9
adc r24,r10
adc r25,r11
rjmp .L1 ; 125 [c=4 l=1] jump
So we have the following results:
Optimal code: Size 12 Instr, no Stack
avr-gcc v8: Size 20 Instr, 4 Stack
avr-gcc v12, v13, v14 (May) Size 56 Instr, 14 Stack
avr-gcc v14 (June) Size 68 Instr, 18 Stack + Frame Pointer
Target: avr
Configured with: --target=avr --disable-nls --with-gnu-as --with-gnu-ld
--disable-shared --enable-languages=c,c++