[PATCH] D11393: [X86] Allow X86::COND_NE_OR_P and X86::COND_NP_OR_E to be reversed.
David Kreitzer via llvm-commits
llvm-commits at lists.llvm.org
Wed Jan 27 17:09:32 PST 2016
DavidKreitzer added inline comments.
Comment at: llvm/trunk/test/CodeGen/X86/fp-une-cmp.ll:51
@@ +50,3 @@
+; CHECK: %bb1
+; CHECK: jmp
> DavidKreitzer wrote:
> > If we invert the compound branch at the end of the entry block and place bb1 before bb2, we can eliminate the jmp at the end of bb1. Do you know why that isn't happening?
> The test case is explicitly added to test the ability for MachineBlockPlacement to break the topological order and reorder BB2 ahead of BB1 (BB1 is ahead of BB2 in source order) -- look at the profile annotation that BB1 is a really cold block -- this reordering is not possible without this patch.
> See also discussions in http://reviews.llvm.org/D11393?vs=on&id=45764&whitespace=ignore-most#toc
Thanks for the explanation! I missed the profile annotation - Please ignore my comment on this one.
More information about the llvm-commits