http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57837
Bug ID: 57837
Summary: ARM function pointer tailcall miscompilation
regression
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: me at williamgrant dot id.au
Created attachment 30469
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=30469&action=edit
arm.md fix
The fix for PR target/19599 in svn trunk r198928 causes predicated function
pointer tailcalls for some ARM targets to lose their predicate and execute
unconditionally, resulting in a crash or other misbehaviour. I was able to
reproduce the miscompilation, manifesting as a segfault, using -march=armv4t
-marm -O3, as below:
gcc -march=armv4t -marm -O3 -o bx-test bx-test.c
----
void baz() {};
void (*bar)() = baz;
void foo(int c) {
if (c == 1) {
foo(0);
}
bar();
}
int main(int argc, char **argv) {
foo(1);
}
----
Today's trunk gives the following asm:
----
foo:
cmp r0, #1
stmfd sp!, {r4, lr}
ldr r4, .L8
ldrne r3, [r4]
ldmnefd sp!, {r4, lr}
bx r3 @ indirect register sibling call
----
Note the bx to an r3 that is uninitialised when r0 == 1; it should actually be
a bxne. The bug in arm.md is fairly clear: the %? is missing from two bx
instructions, so the predicate is omitted. After identifying the bad code I
discovered that the issue was raised in review, but deemed irrelevant
(http://gcc.gnu.org/ml/gcc-patches/2013-05/msg01022.html).
The attached patch fixes the bug for me.
