[PATCH] D90776: [FastISel] Sink more materializations to first use

[Paul Robinson via Phabricator via llvm-commits]

probinson added a comment.

I prototyped a patch that would

- call flushLocalValueMap() at the top of selectInstruction()
- turn off the instruction-sinking stuff
- not erase the DbgLoc applied to local value instructions

I also added statistics for the number of local values added to the map, and the number of lookups that got a hit (assumed to mean a reuse).

| stat                  | no patch    | with patch  | delta           |
| local values recorded | 4686197           | 4711277           | +25080 (+0.5%)  |
| local values reused   | 1524558           | 1283825           | -240733 (-16%)  |
| reuse rate            | 32.5%       | 27.3%       | -5.2%           |
| .text size            | 127,537,392 | 127,507,888 | -29505 (-0.02%) |
| Line-0 bytes          | 9.071,744   | 8,935,299   | -136445 (-1.5%) |
|

This tells me that comparatively few values were reuses across instruction boundaries (~5%).  Despite those reuses going away, overall .text size still went down, which we could assume is due to fewer spills/reloads; I don't have stats there, we'd have to see what regalloc is willing to report.
There's also a pleasant decrease in the number of bytes with line-0 attributions.  I noticed in small examples that reloads at the top of a block still get line 0, which seems not unfair.

I don't have build-time stats yet, but as the patch removes the old sinking code, I'd anticipate that build times would go down.

Once I have that data, I'll clean up the prototype patch and post that.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D90776/new/

https://reviews.llvm.org/D90776