[llvm] [SLP] Bail out on store-to-load forwarding hazards (PR #199606)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> > What is the compile time effect?
> 
> Compile-time impact is negligible. I measured retired instructions (perf stat -e instructions:u) for compiling all 635 CTMark TUs with a base vs. patched release/no-asserts clang: total change is +0.01% (+0.006% and +0.021% across two runs, i.e. within noise). The largest per-file deltas are ~+1% on a few small 7zip TUs; the aggregate is flat. The store→load-forwarding scan now walks only the tree's load/gather nodes rather than doing a whole-tree operand BFS, which keeps the added work bounded.

Give a link

https://github.com/llvm/llvm-project/pull/199606