[LVI] Fix LVI compile time regression around constantFoldUser()

[yamauchi]

Summary

Avoid checking each operand and calling getValueFromCondition() before calling
constantFoldUser() when the instruction type isn't supported by
constantFoldUser().

This fixes a large compile time regression in an internal build.

[yamauchi]

Cleaned up.

[sanjoy]

lib/Analysis/LazyValueInfo.cpp

1417

Please add an assert in constantFolderUser that this is true for its input. Also please s/Value *Val/User *U/

1495–1496

Don't you need to guard this too?

[yamauchi]

Addressed comments.

[sanjoy]

lib/Analysis/LazyValueInfo.cpp

1497

Shouldn't this be canConstantFoldUser(Usr) && usesOperand(Usr, Condition)? Otherwise we're not really short-circuiting anything.

[yamauchi]

Addressed comment.

[sanjoy]

lgtm

lib/Analysis/LazyValueInfo.cpp

1377

I'd call this isOperationFoldable since we're not checking if Usr is foldable or not but only looking at its operation to see if it could possibly be constant folded.

1446–1449

s/checking the operands/linearly iterating over the operands unnecessarily/

s/large phi/instructions with many operands/ (I don't think "phi" is relevant here -- this could also have happened with a large call, invoke or GEP instruction).

1492

s/ValUsesConditionAndFoldable/ValUsesConditionAndMayBeFoldable/

[yamauchi]

Done.