[llvm-bugs] [Bug 50481] New: [ARM] Machine Outliner issue with LDRD/STRD in Thumb mode

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=50481

            Bug ID: 50481
           Summary: [ARM] Machine Outliner issue with LDRD/STRD in Thumb
                    mode
           Product: libraries
           Version: 12.0
          Hardware: PC
                OS: All
            Status: NEW
          Severity: release blocker
          Priority: P
         Component: Backend: ARM
          Assignee: unassignedbugs at nondot.org
          Reporter: yvan.roux at linaro.org
                CC: llvm-bugs at lists.llvm.org, smithp352 at googlemail.com,
                    Ties.Stuij at arm.com

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

https://github.com/rust-lang/rust/issues/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)

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20210526/871384fb/attachment.html>