[PATCH] D34458: [TTI] Refine the cost of EXT in getUserCost()

[Haicheng Wu via Phabricator via llvm-commits]

haicheng added a comment.

In https://reviews.llvm.org/D34458#792223, @junbuml wrote:

> This looks good to me. It will be good if you can share any performance or code size impact  with this change?


Here is the performance and code size change of SPEC20xx on AArch64.

| Benchmark          | Performance (+ is better) | Code Size (+ is larger) |
| spec2000/gap       | +1.15%                    | 0%                      |
| spec2006/h264ref   | -2.55%                    | 0%                      |
| spec2017/leela     | +5.22%                    | 0%                      |
| spec2017/perlbench | -1.66%                    | 0%                      |
| spec2017/x264      | +1.70%                    | 0%                      |
| spec2017/xz        | -1.29%                    | 0%                      |

The regression in spec2006/h264ref is caused by a fully unrolled loop.  The loop has an indrect call inside it.  If transforming the indirect call to direct call, the performance can be improved by +10%, and then unrolling this loop can get additional +3% improvement.  We are working on it.


Repository:
  rL LLVM

https://reviews.llvm.org/D34458