We can eliminate LDRcp, that's fine. However, I still don't like the separate hook and targetinstrinfo bit. Dan, can you have just a single isTriviallyReMaterializable hook that encompass all these?
Evan On Jun 18, 2007, at 9:42 AM, Chris Lattner wrote: > > On Jun 18, 2007, at 7:46 AM, Dan Gohman wrote: > >>> Is this necessary? ARM is already doing this by marking trivially >>> re- >>> materializable instructions with let isReMaterializable = 1. >> >> It's not necessary, but it seems cleaner :-). >> >> In the ARM target, other than instructions that are always trivially >> rematerialiable, there's only LDR, with a ReMaterializable clone >> named LDRcp. >> x86 has a variety of load instructions, so in order to work the way >> ARM does >> x86 would need a varity of clones, and then the clones wouldn't be >> able to >> use the automatic instruction selection, so there would still need >> to be some >> new custom logic for picking which clone to use for each constant. > > Ah, good point. Evan, do you think the LDRcp instruction can be > eliminated now? > > -Chris > _______________________________________________ > llvm-commits mailing list > llvm-commits@cs.uiuc.edu > http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits _______________________________________________ llvm-commits mailing list llvm-commits@cs.uiuc.edu http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
