[DAG] Peek through trunc when combining select into shifts.

[deadalnix]

Summary

This fixes a regression in D127115

Depends on D127115

[deadalnix]

This causes a regression in CodeGen/PowerPC/select_const.ll (at least in the no isel case).

[foad]

A trunc can change the value of the shift amount, so how is it correct to peek through it?

[RKSimon]

Yes, we're going to need a KnownBits check to ensure we were truncating away known zero bits

[deadalnix]

Check the truncate doesn't lose information.
Remove the zero extent case as it seems to only regress things.

[RKSimon]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

13184–13185

Please can you move this as an NFC pre-commit and then rebase the patch? Just to show that there are no changes to the logic in isTruncateOf.

[deadalnix]

Rebase on top of an NFC that moves isTruncateOf.

[deadalnix]

Rebase

[deadalnix]

rebase again

[nikic]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

2489

Outdated comment about zext.

2495

isTruncateOf() handles truncates and some kind of weird setcc pattern. Do we need/want that pattern here? Is there test coverage for it?

[deadalnix]

Rebase, add tests for setcc patterns in isTruncateOf.

[deadalnix]

Rebase, ping?

[nikic]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

2500

if (Known.countMaxActiveBits() < Sel.getScalarValueSizeInBits()) or so?

[deadalnix]

Use countMaxActiveBits

[nikic]

LGTM