Fixed string memory leak, replaced if-then-else with conditional expression. Added check in case there is no device code in an archive child, and we can skip processing it.

If Mehdi or someone else outside AMD could approve (or suggest more changes) that would be helpful. Thanks!

Made matching for ".cubin" exension explicit, and created a fallback to ".o" with a warning.

Removed local Err variable.

Fixed "//" in #include, fixed the parameters to loadArFile, fixed error message.

Removed some auto declarations to specific types.

Added tests and a comment to the UnbundleArchive function to describe what it does.

Added check that archive members are bitcode files, with a negative test. Added a negative test for a broken archive. Note: The error message "file too small to be an archive" is given regardless of the size of the input file. The error is reported if the archive library does not recognize the file.

In D81109#2072453, @jdoerfert wrote:

In D81109#2072395, @jsjodin wrote:

In D81109#2072178, @jdoerfert wrote:

Can you add a test for a broken archive.

Sure, I can use a non-archive file as an archive. I should also see what happens if .ll files are used inside an archive, I remember something bad happened when I tried it at some point.

A test and, if possible, an error message is probably preferable to "something bad" ;)

In D81109#2072178, @jdoerfert wrote:

Can you add a test for a broken archive.

Full context diff.

I don't know if this is needed anymore. We could probably abandon it if it does not seem useful.

Made changes to address comments by Hal Finkel.

Merged code with TOT and made full context patch.

Make TII mandatory (assert if given nullptr) in PSV constructor. Add TII as a member to the PSV manager so that it is not needed in the method calls.

Made the implementation of materializeImmediate more compact as Matt suggested.

Fixed issues raised by Matt.

Fixed remaining comments.

Fixed issues raised by Matt.

Also fixed succ_empty.

Removed the MachineRegion changes and submitted in a different patch. Fixed a few of the comments made by Matt.

Moved the code into the AMDGPU backend since there was little interest in a generic version. If someone wants to use this in the future, it is easy to lift it out into CodeGen again.

Added function to only run the structurizer on a region if there are non-uniform branch instructions present. The interface function should perhaps be called something more generic. Also synced with trunk.

Revert back to "Not Done", accidentally checked the box.

Cleaned up the code a bit and and clang-formatted everything.

Will create new diff and add correct mailing list.

Will create another patch.

Fix insertSelect to handle VCCZ and VCCNZ conditions.

Add missing file.

Moved the canFallThrough code to the header file.

Put implementation of canFallThrough to the header.

Fixed various bugs. The code now transforms DAGs, Loops and nested regions, and the generated code executes correctly, for some small examples.

Fix comment.

Fix whitespace

Add missing .h file.

In D23209#525350, @arsenm wrote:

What effect does this have on what instructions are emitted for the mask modifications? I noticed that sc emits fewer instructions to modify the exec mask. For example, we currently emit for if/then/else

if:
s_and_saveexec_b64 SaveExec, vcc
s_xor_b64 SaveExec, exec, SaveExec

else:
s_or_saveexec_b64 SaveExec, SaveExec
s_xor_b64 exec, exec, SaveExec

endif:
s_or_b64 exec, exec, SaveExec

While sc emits:
if:
s_and_saveexec_b64 SaveExec, vcc

else:
s_andn2_b64, exec, SaveExec, exec

endif:
s_mov_b64 exec, SaveExec

Switching this seems easy, but I tried with the current scheme, but I think it requires modification of how if.break and else.break are inserted

Remove virtual

Fixed various bugs related to maintaining live out information. Fixed merge issues.

Updated patch to use better debug locations when instructions are created.

Remove commented out code.

I see some commented out code that I missed to remove. I will fix that asap.

Added check in branch inversion to break if we encounter a block without a single predecessor.