[PATCH] D70401: [RISCV] Implement ilp32e ABI

Tue Nov 19 22:24:14 PST 2019

Jim marked an inline comment as done.
Jim added inline comments.

================
Comment at: llvm/test/CodeGen/RISCV/callee-saved-fpr64s.ll:11
+; This test currently fails because the machine code does not validate:
+; RUN: not llc -mtriple=riscv32 -mattr=+d -target-abi ilp32e -verify-machineinstrs < %s
+; It will need FileCheck %s -check-prefix=ILP32-LP64-NO-D
----------------
shiva0217 wrote:
> lenary wrote:
> > @shiva0217 I think this test is failing because of the base pointer patch, but I'm not sure. Can you look at the issue? It thinks that x8 gets killed by a store (which I don't think should be using x8), and therefore x8 is not live when we come to the epilog. It's a super confusing issue.
> Hi @lenary, it seems that hasBP() return false in this case, the issue trigger by register allocation allocating x8 which should be preserved. I'm not sure why it will happen, I try to write a simple C code to reproduce the case but fail to do that. Could you obtain the C code for the test case?
It seems that RISCVRegisterInfo::getReservedRegs doesn't add x8(fp) into reserved registers (TFI->hasFP(MF) return false), then x8 is a candidate register for register allocation. After register allocation, some of fpr64 splitted into stack that makes stack need to be realign (MaxAlignment(8) > StackAlignment(4)), therefore x8 should be used as frame pointer (TFI->hasFP(MF) return true). In emitting epilogue, instructions for fp adjustment is inserted.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D70401/new/

https://reviews.llvm.org/D70401