[llvm-commits] [llvm] r163302 - in /llvm/trunk: lib/Transforms/Utils/SimplifyCFG.cpp test/Transforms/SimplifyCFG/switch_create.ll test/Transforms/SimplifyCFG/switch_to_lookup_table.ll

[Hans Wennborg]

On Mon, Sep 10, 2012 at 7:19 PM, Evan Cheng <evan.cheng at apple.com> wrote:
> The *right* longer term solution is to expose more target info to LLVM IR
> passes. It's going to be more and more important especially with recent
> interest in vectorization. I am fairly certain we'll start exposing some
> target "instruction cost" (to start) in 2013.
>
> That said, I think we will need a short term compromise. Hans, can you
> experiment with moving this to codegenprep? I'd like to see what are the
> interaction with other IR optimizations.

I've been looking into moving this to codegenprep today. One of the
issues that came up is that if we move out of SimplifyCFG, we don't
get the benefit from some of the clean-ups it does, e.g. removing
unreachable blocks, merging blocks, etc. It seems that the codegen
passes manage to clean this up pretty well, but the output on the IR
level is not as good.

Another thing is that I couldn't find a way to run codegenprep via opt
so that TargetLowerInfo is available. I guess this means I wouldn't be
able to test my transformation via opt, but should use llc instead? Or
I guess I could turn the logic around and do my transformation
*unless* there is TLI available which says we shouldn't do it.

Talking to Duncan on IRC, he thinks we should not move this to
CodeGenPrepare, but instead look into converting the table look-ups
back into switches for those targets that need it, and that this would
benefit other loads from constants as well. We could do this quite
easily in CodeGenPrepare, I think. It would look something like the
attached (work in progress) patch.

Please let me know what you think.

Thanks,
Hans
-------------- next part --------------
A non-text attachment was scrubbed...
Name: load_to_switch.patch
Type: application/octet-stream
Size: 4157 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20120911/29be71fa/attachment.obj>