[llvm] [MachineLICM] Allow hoisting loads from invariant address (PR #70796)

[Igor Kirillov via llvm-commits]

igogo-x86 wrote:

This patch affects several old tests, and I processed them differently:

1) Regression tests. I just updated the test lines:
* RISCV/rvv/fold-scalar-load-crash.ll
* X86/2009-04-25-CoalescerBug.ll
* X86/pr49393.ll
* X86/pr53842.ll

2) I need advice on what to do with these tests. They are testing some functionality but don't care about loads. I see three options - add a flag (as I did in my initial commit), update test lines as I did for regression tests or update tests by hoisting loads outside of loops on the IR level in a separate patch:

* X86/fma-commute-loop.ll
* AArch64/sinksplat.ll
* AArch64/ragreedy-local-interval-cost.ll
* AArch64/sinksplat.ll
* AArch64/zext-to-tbl.ll

3) The Test I don't understand, so I just updated the test line:
* X86/block-placement.ll

https://github.com/llvm/llvm-project/pull/70796