[PATCH] R600/SI: Try to form (f64 [s|u]int_to_fp i1)

Matt Arsenault Matthew.Arsenault at amd.com
Wed Jan 14 09:53:40 PST 2015


This is un-optimized by the DAG combiner now to avoid
the from-i1 conversion. We get slightly better code
by doing this than materializing the weird constants
since there is no 64-bit select which end up getting split
up. The expanded pattern also shows up in fceil / ffloor
lowering.

http://reviews.llvm.org/D6970

Files:
  lib/Target/R600/SIISelLowering.cpp
  test/CodeGen/R600/fceil64.ll
  test/CodeGen/R600/ffloor.ll
  test/CodeGen/R600/sint_to_fp.f64.ll
  test/CodeGen/R600/uint_to_fp.f64.ll

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D6970.18163.patch
Type: text/x-patch
Size: 7731 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150114/0e379b9d/attachment.bin>


More information about the llvm-commits mailing list