[llvm] [GlobalISel] Optimized MatchData Handling (PR #92115)

[Amara Emerson via llvm-commits]

aemerson wrote:

Strangely, after 2 attempts at running everything with 10x repeats each time, I'm seeing performance *regressions*, if any change.

```
Program                                       compile_time
                                              lhs          rhs    diff
Bullet/bullet                                  10.98        11.00  0.2%
7zip/7zip-benchmark                            14.76        14.78  0.2%
lencod/lencod                                   4.65         4.66  0.2%
tramp3d-v4/tramp3d-v4                           5.68         5.69  0.1%
mafft/pairlocalalign                            2.68         2.68  0.0%
SPASS/SPASS                                     4.57         4.57  0.0%
kimwitu++/kc                                    6.65         6.65 -0.0%
consumer-typeset/consumer-typeset               3.59         3.59 -0.0%
sqlite3/sqlite3                                 2.49         2.49 -0.1%
                           Geomean difference                      0.1%
```

This is measured in an MacBook with release mode + assertions disabled, no LTO.

https://github.com/llvm/llvm-project/pull/92115