This combine does create the strict_fsub in terms of strict_fadd while preserving the chain-protected ordering. Would it still be appropriate to have this transform and others like it in the DAGCombiner?

It seems one of the issues that my tests revealed was already remedied in very recent changes, causing my test to fail. I changed the test to reflect that while also taking steven's recommendations.

Took steven.zhang's suggestion, added REQUIRES line to diff. Hopefully addressed harbormaster concerns.

Rebased for harbormaster.

LGTM, just give the test a good name.

Changed directory.

Mooted by D70867.

Sorry for the delay- I had to investigate some of the fragility so that my test case reduced to this pass would still trigger the issue. However, I did manage to get a test case that is only a few lines long that would trigger the issue without the update.

ping

ping

ping.

ping.

Ensured that all changes to LLVM source are clang-formatted.

Added libcall lowering for fptosi and fptoui constrained intrinsics. They take a similar lowering route to STRICT_FP_ROUND.

Updated tests to have nounwind attribute so that we don't have to worry about https://reviews.llvm.org/D67016 non-relevant .cfi.

Still have some test changes that I have to make- slight diff change incoming soon

Thanks to jsji for making me realize that a lot of the functionality I implemented in this patch already existed. I converted the ExpandFloatRes/Op_STRICT_* functions to use ExpandChainLibCall and eliminated the LibCallifyStrictFP and MakeLibCallChained functions from the previous diffs.

Variety of adjustments, see comment replies (still incoming if I just submitted this)

Ping.

Ping.

Diff did not include my test case- correcting that.

ping

Ping.

Full context.

Sorry Hal, I must have accidentally released an incomplete draft yesterday instead of simply deleting it like I meant to.

That is a good observation, Craig. Changed.

Added space (i= 0 -> i = 0)

Modified the SystemZ test for this fix so that it also works for SystemZ Z13, which is afflicted by a different errror involving function return types at an earlier stage in compilation. To circumvent this, all v3f64 tests were changed so that they took a pointer parameter, modified the value at that pointer, and returned void. These still test StrictFP code generation while avoiding an unrelated platform specific error.