[GlobalOpt] Extend logic in SRA heuristic to skip stores of initializer.

[fhahn]

Summary

If all stores only store the initializer value of a global, consider it
as not stored in the heuristic. GlobalOpt will remove such stores later
on.

Depends on D129857.

[fhahn]

Rebase and ping :)

[nikic]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

400

Rebase over https://github.com/llvm/llvm-project/commit/0ecef88cb167dbd19023a932e7abac09059c3e4e?

408

Shouldn't we be comparing with the initializer at the specific offset?

[fhahn]

Update to use ConstantFoldLoadFromConst to check the initializer value in the specifi field.

[fhahn]

Rebase on top of 0ecef88.

[fhahn]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

400

Done!

408

Yeah, I updated the code to use ConstantFoldLoadFromConst

[nikic]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

408

Why can't this reuse the offset calculation from above?

[fhahn]

Use existing offset directly.

[fhahn]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

408

Missed that, should be updated now, thanks!

[nikic]

LGTM, though I think it might be cleaner if we collected the initializer part during the construction of GlobalPart, rather than calculating it twice, here and during the SRA split.

[fhahn]

In D144476#4171586, @nikic wrote:

LGTM, though I think it might be cleaner if we collected the initializer part during the construction of GlobalPart, rather than calculating it twice, here and during the SRA split.

Posted D145490 as follow-up