https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119279
Bug ID: 119279
Summary: Specifying frame pointer dependency in inline asm
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: jpoimboe at redhat dot com
CC: ak at gcc dot gnu.org, hjl.tools at gmail dot com, hpa at
zytor dot com,
jakub at gcc dot gnu.org, peterz at infradead dot org,
pinskia at gcc dot gnu.org, rguenth at gcc dot gnu.org,
torva...@linux-foundation.org, ubizjak at gmail dot com
Target Milestone: ---
Target: x86_64-*-*
Created attachment 60749
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=60749&action=edit
Reduced test case (gcc -O2 -fno-omit-frame-pointer -fno-optimize-sibling-calls
-c mmap.i -o mmap.o)
[ This was discussed previously in bug 117311, which was a documentation
request for existing (but unsupported) behavior. This bug here is not for
documenting existing behavior, but rather a request to define a supported
solution for specifying a frame pointer dependency in inline asm, whatever that
looks like. ]
In the Linux kernel on x86 it's common for inline asm to have a "call" to a
thunk which saves registers before calling out to another function. If the
inline asm is in a leaf function, or gets emitted before the function prologue,
the frame pointer doesn't get set up before the call.
The current unsupported workaround used throughout the kernel is to make the
stack pointer an in/out constraint:
register unsigned long current_stack_pointer asm("%rsp");
#define ASM_CALL_CONSTRAINT "+r" (current_stack_pointer)
We've also experimented with making __builtin_frame_address(0) an input
constraint, which also seems to work:
#define ASM_CALL_CONSTRAINT "r" (__builtin_frame_address(0))
Unfortunately these are unsupported hacks which just happen to work. It would
be much better to have a supported solution.
The attached test case results in the following:
0000000000000000 <pfn_modify_allowed>:
0: e8 00 00 00 00 call 5 <pfn_modify_allowed+0x5> 1:
R_X86_64_PLT32 __SCT__preempt_schedule-0x4
5: 48 83 3d 00 00 00 00 00 cmpq $0x0,0x0(%rip) # d
<pfn_modify_allowed+0xd> 8: R_X86_64_PC32
pfn_modify_allowed_pfn-0x5
d: 75 01 jne 10 <pfn_modify_allowed+0x10>
f: c3 ret
10: 55 push %rbp
11: 31 c0 xor %eax,%eax
13: 48 89 e5 mov %rsp,%rbp
16: e8 00 00 00 00 call 1b <pfn_modify_allowed+0x1b> 17:
R_X86_64_PLT32 capable-0x4
1b: 5d pop %rbp
1c: c3 ret
