[PATCH] D41300: [ARM] Fix PR35379 - incorrect unwind information when compiling with -Oz

[Ryan Prichard via Phabricator via llvm-commits]

rprichard added a comment.

If I understand correctly, with this fix, we'll output something like this for the original test case (https://godbolt.org/g/rJRr75):

  work:
    .save {r11, lr}
    push {r9, r10, r11, lr}
    .setfp r11, sp, #8
    add r11, sp, #8
    ...

The .save directive apparently only accounts for 8 bytes of SP adjustment, but SP was decremented by 16. If I assemble it, then the `readelf -u` output looks broken:

  $ arm-linux-androideabi-gcc -c test.s && readelf -u test.o
  
  Unwind table index '.ARM.exidx' at offset 0x64 contains 1 entries:
  
  0x0 <work>: @0x0
    Compact model index: 1
    0x9b      vsp = r11
    0x41      vsp = vsp - 8
    0x84 0x80 pop {r11, r14}
    0xb0      finish
    0xb0      finish

(r11 points at the saved {r11, r14} registers. The `vsp = vsp - 8` adjustment is wrong.)

If the tryFoldSPUpdateIntoPushPop optimization is disabled (e.g. by compiling with -Os instead), there is a separate 8-byte stack adjustment, which is marked with a .pad directive:

  .pad #8
  sub sp, sp, #8

It looks like EmitUnwindingInstruction also emits the .pad directive. If I add a `.pad #8` directive to my assembly file, immediately after the `.save` directive, then the `readelf -u` output is fixed. (i.e. The `vsp = vsp - 8` line disappears.)


Repository:
  rL LLVM

https://reviews.llvm.org/D41300