This LGTM.

This small change LGTM,

Lined up all the pipes to keep them aligned.

Added a mainline target "-mtriple=thumbv8m.main -mcpu=cortex-m33" as suggested.

• Updated summary with description of the instructions not available in Thumb1.
• Merged the multiple --check-prefix= entries into a single --check-prefixes= entry.
• Changed CHECK-HWDIV-T2 to CHECK-HWDIV-ARM-OR-THUMB2.
• Changed CHECK-HW-DIV-T1 to CHECK-HWDIV-THUMB1.
• Tweaked comment about lack of 32-bit => 64-bit HW multiple instructions to say this prevents an optimisation.

Once again there are unrelated 8 failures showing on the pre-commit tests due to test timeouts.

Removed unnecessary lines from test case.

The failing pre-commit tests:

• ThreadSanitizer-x86_64::restore_stack.cpp
• libFuzzer::fuzzer-leak.test

are unrelated to this change (they do not involve any RWPI code)

It may be worth adding to the description some links to the Procedure Call Standard for the Arm Architecute which describe RWPI:

• Description of RWPI: https://github.com/ARM-software/abi-aa/blob/main/aapcs32/aapcs32.rst#read-write-position-independence-rwpi
• Use of R_ARM_SBREL32-type relocation: https://github.com/ARM-software/abi-aa/blob/main/aapcs32/aapcs32.rst#aapcs32-f5

Any feedback would be welcome on whether this was a suitable approach to fix this issue for Thumb2 execute-only code.

LGTM. Good to add missing tests for features that have been committed.

I feel that using a -g option make more sense.

This change fixed a failing build here, so it looks correct to me.

review request ping.

What are these disabled "R-UN" lines?

I have updated the patch with the suggested change to use a list of features to disable.

This change looks OK to me.

I currently don't know of any other optimisation phases where the debug information for the results of a phi node is incorrect; but I would not be surprised to find that once the loop rotate issue is fixed and the information is able to correctly propagate in this phase that other optimisation phases which copy phi nodes are also found to have a problem.

This change looks good to me, but I think it would be good if someone more experienced than me in the usage of debug metadata could give the final OK to this change.

Looking into what optimisations would be relevant to -Og (or rather what optimisation should be omitted because they are too disruptive to the debug experience) is something that I have on my road-map as something to look into and gather feedback on.

Looks good to me .... but I'll let Daniel give the final OK.

Correct spelling mistakes in comments.
Converted the test to a MIR testcase

I believe this change is good as a general back-stop check.

Updated PhiHasDebugValue() to avoid scanning the whole instructions stream by using a new function llvm::FindAllocaDbgValues(). This new function is similar to the exiting function llvm::FindAllocaDbgDeclare() but has the additional complication that the result of the find can find multiple entries when applied to a phi node.

• Moved the changes for ensuring the lowered load instruction in placed before the DEBUG_VALUE entry into a separate change: https://reviews.llvm.org/D23760
• Updated PhiHasDebugValue() to use llvm::any_of

Thanks for the review, I will update the patch taking into account of your comments.

In D16064#327293, @rafael wrote:

This fixes PR23020?

Fixed line length issue.
Renamed getFixupKindContainereSizeBytes() to getFixupKindContainereSizeInBytes()

This very old change never got committed (apologies to the reviewers) and the functionality has since been implement by friss in commit r230551

Changed SegmentSize to SegmentDescriptorSize.
Added comment to CIE test to explain that the DW_CFA_nop instructions are padding.

Thanks for review my patch. I agree that segment_descriptor-size does appear to be a more appropriate name.

Forgot to subscribe llvm-commits to the phabricator subscriber field, so repeating the initial summary as a comment in the hope this will make it to the mailing list.