[llvm] [MachinePipeliner] Improve loop carried dependence analysis (PR #94185)

[Yuta Mukai via llvm-commits]

ytmukai wrote:

@iajbar The number of pipelined loops has been recounted to exclude duplicates due to inlining, etc. The optimization flag is -O3.

| Both pipelined | Failed without this patch | Failed with this patch |
| --- | --- | --- |
| 1227 | 39 | 27 |

The loops that cannot be pipelined with this patch are as follows:

```
llvm-test-suite/MicroBenchmarks/ImageProcessing/Dither/orderedDitherKernel.c:33:7
llvm-test-suite/MicroBenchmarks/ImageProcessing/Dither/orderedDitherKernel.c:52:7
llvm-test-suite/MicroBenchmarks/ImageProcessing/Dither/orderedDitherKernel.c:65:7
llvm-test-suite/MultiSource/Applications/JM/lencod/transform8x8.c:1563:5
llvm-test-suite/MultiSource/Applications/oggenc/oggenc.c:54181:5
llvm-test-suite/MultiSource/Applications/oggenc/oggenc.c:54319:5
llvm-test-suite/MultiSource/Applications/oggenc/oggenc.c:54321:5
llvm-test-suite/MultiSource/Applications/oggenc/oggenc.c:54322:5
llvm-test-suite/MultiSource/Applications/sqlite3/sqlite3.c:15312:5
llvm-test-suite/MultiSource/Benchmarks/7zip/C/LzFind.c:293:3
llvm-test-suite/MultiSource/Benchmarks/7zip/CPP/7zip/Crypto/HmacSha1.cpp:70:5
llvm-test-suite/MultiSource/Benchmarks/Bullet/btSoftBody.cpp:1798:2
llvm-test-suite/MultiSource/Benchmarks/Bullet/btSoftBody.cpp:1813:2
llvm-test-suite/MultiSource/Benchmarks/Bullet/btSoftBody.cpp:1819:3
llvm-test-suite/MultiSource/Benchmarks/Bullet/btSoftBody.cpp:1857:2
llvm-test-suite/MultiSource/Benchmarks/Bullet/btSoftBody.cpp:1863:3
llvm-test-suite/MultiSource/Benchmarks/MiBench/consumer-lame/lame.c:1121:5
llvm-test-suite/MultiSource/Benchmarks/MiBench/consumer-lame/takehiro.c:739:6
llvm-test-suite/MultiSource/Benchmarks/MiBench/office-ispell/correct.c:1362:7
llvm-test-suite/MultiSource/Benchmarks/MiBench/office-ispell/correct.c:1367:7
llvm-test-suite/MultiSource/Benchmarks/MiBench/office-ispell/correct.c:1375:7
llvm-test-suite/MultiSource/Benchmarks/MiBench/office-ispell/correct.c:1380:7
llvm-test-suite/MultiSource/Benchmarks/Prolangs-C/TimberWolfMC/upin.c:175:5
llvm-test-suite/MultiSource/Benchmarks/Prolangs-C/TimberWolfMC/upin.c:178:5
llvm-test-suite/SingleSource/Benchmarks/Misc/ReedSolomon.c:210:13
llvm-test-suite/SingleSource/Benchmarks/Misc/ReedSolomon.c:239:13
llvm-test-suite/SingleSource/Benchmarks/Misc/ReedSolomon.c:264:10
```

For these, I confirmed that there were no problems in detecting the dependences as in the previous post.

The following loops are incorrectly pipelined without this patch.

https://github.com/llvm/llvm-test-suite/blob/1f917f95918479727e727c710348d9bf674588fc/MultiSource/Applications/oggenc/oggenc.c#L54321-L54322

The schedule result is as follows:

```
// The values in the first brackets are the scheduled stages
cycle 0 (0) (0) %74:intregs = PHI %72:intregs, %bb.25, %75:intregs, %bb.26
cycle 0 (1) (4) %397:intregs = F2_sfsub %396:intregs, %395:intregs, implicit $usr, debug-location !18418; llvm-test-suite/MultiSource/Applications/oggenc/oggenc.c:54321:46
cycle 0 (0) (1) %395:intregs = L2_loadri_io %74:intregs, 0, debug-location !18417 :: (load (s32) from %ir.lsr.iv473); llvm-test-suite/MultiSource/Applications/oggenc/oggenc.c:54321:49
cycle 0 (0) (3) %396:intregs = L2_loadri_io %74:intregs, -4, debug-location !18418 :: (load (s32) from %ir.cgep691); llvm-test-suite/MultiSource/Applications/oggenc/oggenc.c:54321:46
cycle 0 (0) (2) %75:intregs = A2_addi %74:intregs, -4
cycle 1 (1) (5) S2_storeri_io %74:intregs, -4, %397:intregs, debug-location !18418 :: (store (s32) into %ir.cgep691); llvm-test-suite/MultiSource/Applications/oggenc/oggenc.c:54321:46
```

The store for `g1[g1_order-i]` must precede the load for `g1[g1_order-i+1]` of the next iteration. However, the store is scheduled after the previous stage (iteration) of the load.
These loops will no longer be pipelined as the dependences are correctly recognised by this patch.


https://github.com/llvm/llvm-project/pull/94185