Updated differential. This moves the actions to near where D69949 had them, guards them with TM.Options.UnsafeFPMath (I can't believe I didn't do that in the first patch), and changes tests accordingly.

Oops, sorry about the typo in the URL. When ready, could someone please land for me? Thanks!

This addresses @jsji's note about the potential for stack slots being let through by using isPhysicalRegister. It changes this pattern and the one from D69497 to use !isVirtualRegister.

In D71170#1774538, @steven.zhang wrote:

LGTM, as early return makes the code more clear. Thank you for the clean up.

In D71170#1774445, @steven.zhang wrote:

Please attach the test if it is NOT NFC patch, otherwise, adding a tag [NFC] in the title, thank you.

@lkail yes, new contributor here so if you can land it or tell me how to go about that? Thanks

See attached for an updated diff. This also changes the name of FeedImmed to DefImmed to be congruent with DefReg[12].

Per @jsji's suggestion, this checks if the operands to DefMI are both the same virtual register, which enables optimization in more cases.

Updated revision. Note, I also had to update test/CodeGen/PowerPC/load-shuffle-and-shuffle-store.ll since this will omit the optimization in some of those cases.

Thanks. Should have checked :) This issue is causing buffer overflows in some applications compiled with Clang 9, so hopefully your fix can be ok'd soon.