https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127237
Bug ID: 127237
Summary: [avr] IRA bloats code due to LRA
Product: gcc
Version: 17.0
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: ---
typedef __UINT32_TYPE__ T;
T and_not (T a, T b)
{
return a & ~b;
}
In this test case, b is passed in R18:SI and a is passed in R22:SI.
The expected code is that b is COMplemented in place, and the outcome is then
ANDed with a. That are 8 instructios, and it is what avr-gcc v14 and v15 are
doing:
$ avr-gcc-15 and-not.c -S -Os && cat and-not.s
com r18
com r19
com r20
com r21
and r22,r18
and r23,r19
and r24,r20
and r25,r21
With v16 and trunk however, IRA moves stuff around for no reason, adding 14
superfluous instructions and 2 stack slots:
$ avr-gcc-16 and-not.c -S -Os && cat and-not.s
push r28
push r29
/* stack size = 2 */
mov r30,r24
mov r31,r25
mov r24,r18
mov r25,r19
mov r26,r20
mov r27,r21
com r24
com r25
com r26
com r27
and r24,r22
and r25,r23
and r26,r30
and r27,r31
mov r22,r24
mov r23,r25
mov r24,r26
mov r25,r27
pop r29
pop r28
The obvious change that introduced this was the switch to LRA: avr-gcc v15
supports -mlra, and with that switch the code from v15 is the same bloat like
from v16 / trunk.
So for some reason, IRA lost capabilities just by enabling LRA instead of
reload.
