[PATCH] D10579: AArch64: Use CMP; CCMP sequences for and/or/setcc trees.
James Molloy
james.molloy at arm.com
Tue Jul 14 03:38:28 PDT 2015
jmolloy added a comment.
Apart from the goto comment which Renato already mentioned, this looks OK to me.
I've run it through all our tests and this time there were no conformance regressions (woohoo!). LNT diff looks like:
| **Performance Regressions - Execution Time** | **Delta** |
| MultiSource/Applications/sqlite3/sqlite3 | 4.53% |
| MultiSource/Benchmarks/Olden/bh/bh | 3.85% |
| SingleSource/Benchmarks/Shootout-C++/lists | 3.20% |
| MultiSource/Applications/lua/lua | 2.70% |
| **Performance Improvements - Execution Time** | **Delta ** |
| MultiSource/Benchmarks/ASC_Sequoia/AMGmk/AMGmk | -10.44% |
| MultiSource/Benchmarks/VersaBench/bmm/bmm | -1.15% |
That's on a Cortex-A57.
================
Comment at: lib/Target/AArch64/AArch64ISelLowering.cpp:1480
@@ +1479,3 @@
+ AArch64CC = changeIntCCToAArch64CC(CC);
+ goto CreateCCNode;
+ }
----------------
rengolin wrote:
> I don't think this is a really good use case for a goto.
Yeah I really don't like the goto here, sorry.
Repository:
rL LLVM
http://reviews.llvm.org/D10579
More information about the llvm-commits
mailing list