https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121623
Bug ID: 121623
Summary: [ia64] bootstrap comparison failure (differently
ordered st8s in MMI for gcc/trans-mem.o), stage 2 -> 3
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: bootstrap
Assignee: unassigned at gcc dot gnu.org
Reporter: frank.scheiner at web dot de
Target Milestone: ---
Host: ia64-t2-linux-gnu
Target: ia64-t2-linux-gnu
Build: ia64-t2-linux-gnu
Created attachment 62166
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=62166&action=edit
Patch to ignore bogus error during comparison
I pushed two bootstraps and testsuite runs for c and c++ through on my rx2800
i2 recently:
* GCC 15.1.0:
https://gcc.gnu.org/pipermail/gcc-testresults/2025-August/855566.html
* GCC 15.2.0
https://gcc.gnu.org/pipermail/gcc-testresults/2025-August/855308.html
For comparison (but for c, c++, fortran and objc):
* gcc-15-20240528 w/LRA enabled:
https://gcc.gnu.org/pipermail/gcc-testresults/2024-June/817268.html
For the bootstrap to succeed a patch is needed to overrule the following
comparison failure between stage 2 and stage 3 compile results for
`gcc/trans-mem.o` (patch's attached):
```
bash-5.2# date; time make -j9 bootstrap 2>&1 | tee
/srv/ssd/logs/15.1.0-RC-20250418/make.log; echo $?; date
[...]
make[2]: Entering directory '/dev/shm/gcc-15.1.0-RC-20250418/gcc.build.lnx'
make[3]: Entering directory '/dev/shm/gcc-15.1.0-RC-20250418/gcc.build.lnx'
rm -f stage_current
make[3]: Leaving directory '/dev/shm/gcc-15.1.0-RC-20250418/gcc.build.lnx'
Comparing stages 2 and 3
Bootstrap comparison failure!
gcc/trans-mem.o differs
make[2]: *** [Makefile:23407: compare] Error 1
make[2]: Leaving directory '/dev/shm/gcc-15.1.0-RC-20250418/gcc.build.lnx'
make[1]: *** [Makefile:23387: stage3-bubble] Error 2
make[1]: Leaving directory '/dev/shm/gcc-15.1.0-RC-20250418/gcc.build.lnx'
make: *** [Makefile:23450: bootstrap] Error 2
real 331m2.512s
user 2360m28.519s
sys 27m9.219s
0
Sun Apr 20 17:28:52 UTC 2025
bash-5.2#x
```
It took some time to determine the actual reason for this miscompare - which
also played into not having testsuite results for the GCC 15 RCs - but the
actual difference is the order of two stores:
```
root@darkstar:/srv/ssd/gcc-15.1.0-RC-20250423/gcc.build.lnx# objdump -l -d
stage2-gcc/trans-mem.o.stripped > stage2-gcc/trans-mem.o.stripped.s
root@darkstar:/srv/ssd/gcc-15.1.0-RC-20250423/gcc.build.lnx# objdump -l -d
stage3-gcc/trans-mem.o.stripped > stage3-gcc/trans-mem.o.stripped.s
root@darkstar:/srv/ssd/gcc-15.1.0-RC-20250423/gcc.build.lnx# diff -Nur
stage2-gcc/trans-mem.o.stripped.s stage3-gcc/trans-mem.o.stripped.s
--- stage2-gcc/trans-mem.o.stripped.s 2025-04-25 17:42:30.312500000 +0000
+++ stage3-gcc/trans-mem.o.stripped.s 2025-04-25 17:43:11.687500000 +0000
@@ -1,5 +1,5 @@
-stage2-gcc/trans-mem.o.stripped: file format elf64-ia64-little
+stage3-gcc/trans-mem.o.stripped: file format elf64-ia64-little
Disassembly of section .text:
@@ -581,8 +581,8 @@
bf0: 09 18 21 49 00 21 [MMI] adds r35=72,r36
bf6: 10 83 30 00 42 00 adds r49=16,r12
bfc: 06 00 00 84 mov r48=r0;;
- c00: 08 00 00 1e 98 11 [MMI] st8 [r15]=r0
- c06: 80 00 38 30 2b 00 st8 [r14]=r0,8
+ c00: 08 40 00 1c 98 15 [MMI] st8 [r14]=r0,8
+ c06: 00 00 3c 30 23 00 st8 [r15]=r0
c0c: 00 00 04 00 nop.i 0x0
c10: 09 78 80 42 00 21 [MMI] adds r15=32,r33
c16: f0 02 bc 30 20 c0 ld8 r47=[r47]
```
...as found out with Ski performing the above commands.
These stores look independent, so the order inside a bundle should be
negligible, also considering that instructions in a bundle are executed in
parallel. Hence I figured both compile results are actually valid. This is also
the only file and place where it happens. Why it happens is yet unclear.
This wasn't present in gcc-15-20240528 yet AFAIR, but GCC 15.1.0 RC1
(15.1.0-RC-20250418) had it already and it's there ever since (also observed
for GCC 15.1.0 and 15.2.0).
Unfortunately I have no idea where to actually start looking in the compiler
code to figure that one out.
