Files
clang-p2996/llvm/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp
Florian Hahn 0f82469314 [Passes] Run SimpleLoopUnswitch after introducing invariant branches. (#81271)
IndVars may be able to replace a loop dependent condition with a loop
invariant one, but loop-unswitch runs before IndVars, so the invariant
check remains in the loop.

For an example, consider a read-only loop with a bounds check:
https://godbolt.org/z/8cdj4qhbG

This patch uses a approach similar to the way extra cleanup passes are
run on demand after vectorization (added in acea6e9cfa).

It introduces a new ShouldRunExtraSimpleLoopUnswitch analysis marker,
which IndVars can use to indicate that extra unswitching is beneficial.

ExtraSimpleLoopUnswitchPassManager uses this analysis to determine
whether to run its passes on a loop.

Compile-time impact (geomean) ranges from +0.0% to 0.02%
https://llvm-compile-time-tracker.com/compare.php?from=138c0beb109ffe47f75a0fe8c4dc2cdabe8a6532&to=19e6e99eeb280d426907ea73a21b139ba7225627&stat=instructions%3Au

Compile-time impact (geomean) of unconditionally running
SimpleLoopUnswitch ranges from +0.05% - +0.16%

https://llvm-compile-time-tracker.com/compare.php?from=138c0beb109ffe47f75a0fe8c4dc2cdabe8a6532&to=2930dfd5accdce2e6f8d5146ae4d626add2065a2&stat=instructions:u

Unconditionally running SimpleLoopUnswitch seems to indicate that there
are multiple other scenarios where we fail to run unswitching when
opportunities remain.


Fixes https://github.com/llvm/llvm-project/issues/85551.

PR: https://github.com/llvm/llvm-project/pull/81271
2024-04-12 22:07:29 +01:00

152 KiB