https://bugs.llvm.org/show_bug.cgi?id=41575
Bug ID: 41575
Summary: r7 not saved/restored around inline asm blocks that
kill it
Product: libraries
Version: trunk
Hardware: PC
OS: Linux
Status: NEW
Severity: enhancement
Priority: P
Component: Backend: ARM
Assignee: [email protected]
Reporter: [email protected]
CC: [email protected], [email protected],
[email protected]
With clang trunk:
$ cat /tmp/repro.c
void escape(void *);
void Bar(int p0, int p1, int p2, int p3, int p4, int p5) {
register unsigned magic asm("r7") = 0x000f0002;
asm volatile("svc 0\n" : : "r"(magic) : "memory");
int local = p5;
escape(&local);
}
$ clang --target=arm-linux-gnueabihf -march=armv7-a -mthumb /tmp/repro.c -S -o
- -O2 -emit-llvm
<< snip >>
; Function Attrs: nounwind
define dso_local void @Bar(i32 %p0, i32 %p1, i32 %p2, i32 %p3, i32 %p4, i32
%p5) local_unnamed_addr #0 {
entry:
%local = alloca i32, align 4
tail call void asm sideeffect "svc 0\0A", "{r7},~{memory}"(i32 983042) #3,
!srcloc !3
%0 = bitcast i32* %local to i8*
call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %0) #3
store i32 %p5, i32* %local, align 4, !tbaa !4
call void @escape(i8* nonnull %0) #3
call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %0) #3
ret void
}
$ clang --target=arm-linux-gnueabihf -march=armv7-a -mthumb /tmp/repro.c -S -o
- -O2
<< snip >>
Bar:
.fnstart
@ %bb.0: @ %entry
.save {r6, r7, lr}
push {r6, r7, lr}
.setfp r7, sp, #4
add r7, sp, #4
.pad #4
sub sp, #4
movs r7, #2
movt r7, #15
@APP
svc #0
@NO_APP
ldr r0, [r7, #12]
str r0, [sp]
mov r0, sp
bl escape
add sp, #4
pop {r6, r7, pc}
So, we set up r7 as our fp, it gets clobbered by `magic` for inline asm to do
stuff with, and we proceed to use the `magic` value as though it was still our
pc in a later ldr.
As one who's not super familiar with our rules around inline asm, I'd assume
that `tail call void asm sideeffect "svc 0\0A", "{r7},~{memory}"(i32 983042)
#3, !srcloc !3` should emit code to save/restore r7, instead of leaving it in a
clobbered state.
FWIW, we have code that looks very similar to the non-reduced repro in
compiler-rt's clear_cache.c. That appears to work because r7 is a callee-saved
register, so I presume we save/restore everything properly there (and don't use
r7 as a fp).
--
You are receiving this mail because:
You are on the CC list for the bug._______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
