[ARM] Machine Outliner issue with LDRD/STRD in Thumb mode

[yroux]

Bugzilla Link	50481
Resolution	FIXED
Resolved on	Jun 14, 2021 22:51
Version	12.0
OS	All
Blocks	#48661
CC	@luqmana,@nikic,@simonwallis2,@smithp35,@tstellar
Fixed by commit(s)	6c78dbd a95bf58

Extended Description

This issue was reported and described by rust-lang community here:

rust-lang/rust#85351

The problem occurs when the machine outliner extracts a chunck of code which
contains a call and loads or stores a pair of registers from or into the
stack, such as:

bl foo
strd r0, r1, [sp, #​0]
strd r2, r3, [sp, #​8]

LR needs to be saved into the stack when jumping in such an outlined
function and stack's offsets should be changed accordingly, thus the
function should look like:

<OUTLINED_FUNCTION_0>:
str.w lr, [sp, #-8]!
bl foo
strd r0, r1, [sp, #​8]
strd r2, r3, [sp, #​16]
ldr.w lr, [sp], #​8
bx lr

But in the code generated by llvm-12 for Thumb2 targets, the stack offsets
are not patched. In fact, these offsets are patched in the MIR by the
Machine Outliner, but it assumes that immediates encoded in
AddrModeT2_i8s4 instructions are scaled like it is the case for the other
addressing modes.

I'll submit a small patch to fix the outliner which will change the scale
and bit number in that part, but maybe it'd be nice to do a bigger change
to make the encoding more consistent (there are already some fixme notes
about that in the code)

[yroux]

Proposed fix is here:

https://reviews.llvm.org/D103167

[yroux]

Fix commited into mainline as: https://reviews.llvm.org/rG6c78dbd4ca1f

[nikic]

Reopening this issue to track LLVM 12 backport. This should be safe to backport, right?

[yroux]

Ah yes sure, I think it safe to backport, sorry to have it marked as resolved too quickly ;)

[tstellar]

Merged: a95bf58