https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81803
Bug ID: 81803
Summary: Miscompilation at -O1 on mips64el
Product: gcc
Version: 7.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: aurelien at aurel32 dot net
Target Milestone: ---
Host: mips64el-unknown-linux-gnu
Target: mips64el-unknown-linux-gnu
Build: mips64el-unknown-linux-gnu
Created attachment 41965
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41965&action=edit
testcase
Since the switch to GCC 7 as the default compiler, Debian sees many
miscompilations on mips64el where variables which are stored on the stack and
smaller than 8 bytes long (e.g. uint16_t, bool) are wrongly reloaded with the
LD instruction.
I unfortunately haven't been able to get a self-contained reduced testcase, but
the attached testcase (from apparmor) should show the issue.
The aa_policy_cache_new function takes a uint16_t as the fourth argument. It
gets passed the max_caches, which also has a uint16_t type. When compiled with
g++ -Wfatal-errors -g -O2 -fstack-protector-strong -c -o testcase.o
testcase.ii the generated code contains:
164: dea50000 ld a1,0(s5)
164: R_MIPS_GOT_OFST .bss+0x38
164: R_MIPS_NONE *ABS*+0x38
164: R_MIPS_NONE *ABS*+0x38
168: dfa80000 ld a4,0(sp)
16c: 2406ff9c li a2,-100
170: 0320f809 jalr t9
170: R_MIPS_JALR aa_policy_cache_new
170: R_MIPS_NONE *ABS*
As you can see, a4 is loaded with the LD instruction, so the upper bytes are
just garbage and causes the aa_policy_cache_new to not function correctly.
