[PATCH] D44644: [X86] Use silvermont cost model overrides for goldmont as well.

Sat Mar 24 17:43:16 PDT 2018

craig.topper updated this revision to Diff 139731.
craig.topper added a comment.

Introduce a GLM specific table the overrides FDIV. Override FSQRT for both GLM and SLM. It appears for packed operations only half of the 128-bit vector is produced at a time for both SLM and GLM as the throughput is twice the scalar throughput. The default SSE42 throughputs we were getting otherwise don't match that behavior.

Throughput data for GLM was taken from table 16-17 in the latest Intel Optimization Manual.

Repository:
  rL LLVM

https://reviews.llvm.org/D44644

Files:
  lib/Target/X86/X86Subtarget.h
  lib/Target/X86/X86TargetTransformInfo.cpp
  test/Analysis/CostModel/X86/arith-fp.ll
  test/Analysis/CostModel/X86/arith.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D44644.139731.patch
Type: text/x-patch
Size: 62956 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180325/46c824ce/attachment-0001.bin>