probe-stack=inline-asm will produce invalid uwtables

[nagisa]

Bugzilla Link	49600
Version	trunk
OS	All
CC	@topperc,@RKSimon,@phoebewang,@rotateright

Extended Description

Given a function as such

; RUN: llc < %s
define void @​big_stack() "probe-stack"="inline-asm" uwtable {
start:
%_two_page_stack = alloca [8192 x i8], align 1
ret void
}

the following assembly will be generated:

big_stack:
.cfi_startproc
subq $4096, %rsp
movq $0, (%rsp)
subq $3968, %rsp
.cfi_def_cfa_offset 8072
addq $8064, %rsp
.cfi_def_cfa_offset 8
retq

Here the unwind tables are not accurate while stack probing is ongoing – the rsp is adjusted, but not the cfa_offsets. And so attempts to obtain a stack trace will fail if the current instruction is somewhere in between the instructions implementing the stack probing.

This also occurs with the non-unrolled implementation of the stack probing:

; RUN: llc < %s
define void @​big_stack() "probe-stack"="inline-asm" uwtable {
start:
%_two_page_stack = alloca [64000 x i8], align 1
ret void
}

--->

big_stack:
.cfi_startproc
movq %rsp, %r11
subq $61440, %r11
.LBB0_1:
subq $4096, %rsp
movq $0, (%rsp)
cmpq %r11, %rsp
jne .LBB0_1
subq $2432, %rsp
.cfi_def_cfa_offset 63880
addq $63872, %rsp
.cfi_def_cfa_offset 8
retq

however in the loop case the solution needs to involve allocation of a separate register as insertion of .cfi directives in a loop won't help in any way.

[nagisa]

The correct assembly for the unrolled case would probably look a lot like this:

big_stack:
.cfi_startproc
subq $4096, %rsp
.cfi_def_cfa_offset 4096
movq $0, (%rsp)
subq $3968, %rsp
.cfi_def_cfa_offset 8072
addq $8064, %rsp
.cfi_def_cfa_offset 8
retq

or an equivalent using .cfi_adjust_cfa_offset directives.

[erikdesjardins]

Update: the unrolled/inline case was fixed in https://reviews.llvm.org/D99579 / 1c268a8.

The loop case is still unfixed.

[erikdesjardins]

Loop case fixed by https://reviews.llvm.org/D116628 / a8ac117.

I believe this can now be closed.