[llvm] [DFAJumpThreading] Constraint the number of cloned instructions (PR #161632)

[Hongyu Chen via llvm-commits]

XChy wrote:

> How many transformations are restricted by this (would it be possible to check stats on the test suite)? 

I have checked the statistics on the test suite. Only reduce one transformation in the outlier `libclamav_nsis_LZMADecode.c`.

> I fear that restricting possibly beneficial transformations to avoid second-order effects might not be the best way forward.

Passes like inlining and loop unrolling would prevent massive expansion; otherwise, the compile time will blow up in later optimizations, although there are actually benefits after expansion. That's an acceptable trade-off for me.

The outlier libclamav_nsis_LZMADecode.c grows from about 1K lines to over 4K lines after DFAJumpThreading, that's unacceptable to me.


https://github.com/llvm/llvm-project/pull/161632