[RISCV][GlobalISel] Add lowerCall for calling convention

[nitinjohnraj]

Summary

This patch implements minimal support for lowering function calls to callees with arguments and/or return values according to the RISC-V calling convention. Integer, pointer and aggregate types are supported.

Feedback is very much appreciated.

[lewis-revill]

Rebase, update test structure. Correct missing defs by adding CallReturnHandler.

[Joe]

If you're supporting a pointer value return type, should there be a test for this?

Other than that, this looks fine to me - although should wait until somebody more experienced looks at it.

Side note: Would it be worth, in the future, to outline the process of finding In/OutputArgs and ArgInfos to something like a computeOutgoingArgs/computeIncomingArgs that fills a list of ISD::In/OutputArg and ArgInfos? Or would this just obfuscate things?

[lewis-revill]

Add test for pointer return type

[lewis-revill]

Bug fix - add target flags to call instruction.

[lewis-revill]

Rebased, and updated where appropriate to continue to build correctly.

[arsenm]

llvm/lib/Target/RISCV/RISCVCallLowering.cpp

337

Same comment on the other half of the call lowering patches

[lewis-revill]

Use ValueAssigners to avoid rewriting generic infrastructure and use determineAndHandleAssignments instead.

[arsenm]

llvm/lib/Target/RISCV/RISCVCallLowering.cpp

263

Not sure why you're bothering to filter these out. splitToValueTypes should handle near everything (I'm aware of some assertions on weird non-byte element sized vectors)

[nitinjohnraj]

Rebased

[arsenm]

llvm/test/CodeGen/RISCV/GlobalISel/irtranslator/calls.ll

266

Use opaque pointers

268

Could use sret/byval tests. Also cases where the argument/return list is too long and needs to go to the stack

[eopXD]

llvm/lib/Target/RISCV/GISel/RISCVCallLowering.cpp

305 ↗

(On Diff #511197)

This seems to contradict with the description of this patch, "Simple non-aggregate types are supported."

[nitinjohnraj]

Added tests for aggregate types, sret and byval

[nitinjohnraj]

llvm/lib/Target/RISCV/GISel/RISCVCallLowering.cpp

305 ↗

(On Diff #511197)

You're right. I've updated the description and added a test accordingly.

llvm/test/CodeGen/RISCV/GlobalISel/irtranslator/calls.ll

266

Updated.

268

Could use sret/byval tests.

I've added these.

Also cases where the argument/return list is too long and needs to go to the stack

Are there other examples of a test like this?

[nitinjohnraj]

As stated in the parent revision, we will correctly handle byval, sret and cases where the number of arguments exceed the number of available registers in a future patch.

[arsenm]

llvm/lib/Target/RISCV/GISel/RISCVCallLowering.cpp

259 ↗

(On Diff #524377)

Should add some tail call cases. You should set IsTailCall to false if you aren’t handling them

[nitinjohnraj]

Set IsTailCall to false. We will handle tail calls in a future patch.