https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116548
Bug ID: 116548
Summary: [avr] ivopts Introducing expensive loop condition
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: gjl at gcc dot gnu.org
Target Milestone: ---
uint8_t add1 (const uint8_t *bb, uint8_t nn)
{
uint8_t sum = 0;
do
{
sum += *bb++;
} while (--nn); // Why not just 2 instructions: decrement + branch
return sum;
}
$ avr-gcc -mmcu=avr4 -S -Os -dp
has an expensive and overly complicated loop condition:
add1:
/* prologue: function */
/* frame size = 0 */
/* stack size = 0 */
mov r20,r24 ; 47 [c=4 l=1] movqi_insn/0
mov r18,r24 ; 48 [c=4 l=1] movqi_insn/0
mov r19,r25 ; 49 [c=4 l=1] movqi_insn/0
ldi r24,0 ; 50 [c=4 l=1] movqi_insn/0
.L5:
movw r30,r18 ; 38 [c=4 l=1] *movhi/0
subi r18,-1 ; 39 [c=4 l=2] *addhi3_clobber/1
sbci r19,-1
ld r25,Z ; 40 [c=4 l=1] movqi_insn/3
add r24,r25 ; 41 [c=4 l=1] *addqi3/0
mov r25,r22 ; 42 [c=4 l=1] movqi_insn/0
sub r25,r18 ; 43 [c=4 l=1] *subqi3/0
add r25,r20 ; 55 [c=4 l=1] *op8.for.cczn.plus/1
brne .L5 ; 56 [c=4 l=1] branch_ZN
/* epilogue start */
ret ; 53 [c=0 l=1] return
In the loop we have R18 and R30 (Z) holding the current address.
The loop condition is:
Insn 42 = Move nn to R25.
Insn 43 = Subtract (low byte of) current address from R25.
Insn 55 = Add (low byte of) initial addressto R25
Insn 56 = branch if result != 0
This are 4 instructions, and the register pressure is: A reg that holds nn, a
reg that holds the current address, a reg that holds the start address, and a
reg to compute the condition.
Instead, the code could just DECrement nn in R20 and branch on != 0 which has
less code, less cycles and less register pressure, even in the case when the
start address and nn are needed after the loop.
With -fno-ivopts, the code is:
add1:
/* prologue: function */
/* frame size = 0 */
/* stack size = 0 */
movw r18,r24 ; 51 [c=4 l=1] *movhi/0
ldi r24,0 ; 44 [c=4 l=1] movqi_insn/0
.L5:
movw r30,r18 ; 35 [c=4 l=1] *movhi/0
ld r25,Z ; 36 [c=4 l=1] movqi_insn/3
subi r18,-1 ; 37 [c=4 l=2] *addhi3_clobber/1
sbci r19,-1
add r24,r25 ; 38 [c=4 l=1] *addqi3/0
subi r22,lo8(1) ; 49 [c=4 l=1] *op8.for.cczn.plus/0
brne .L5 ; 50 [c=4 l=1] branch_ZN
/* epilogue start */
ret ; 47 [c=0 l=1] return
Which uses nn in R22 with decrement (insn 49) and branch (insn 50).
Seems like ivopts cost model is off.
Target: avr
Configured with: ../../source/gcc-master/configure --target=avr --disable-nls
--with-dwarf2 --with-gnu-as --with-gnu-ld --disable-shared
--with-long-double=64 --enable-languages=c,c++
gcc version 15.0.0 20240829 (experimental) (GCC)
