[llvm] [RISCV] Fix musttail with indirect arguments by forwarding incoming pointers (PR #185094)

[Folkert de Vries via llvm-commits]

================
@@ -0,0 +1,143 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=riscv32 %s -o - | FileCheck %s --check-prefix=RV32
+; RUN: llc -mtriple=riscv64 %s -o - | FileCheck %s --check-prefix=RV64
+
+; Test that musttail with indirect args (fp128 on RV32) forwards the incoming
+; pointer instead of creating a new stack temporary. Without this fix, the
+; pointer would dangle after the tail call deallocates the caller's frame.
+
+declare i32 @callee_musttail_indirect(fp128 %a)
+
+; fp128 is indirect on RV32 (too large for registers), direct on RV64.
+; On RV32, musttail must forward the incoming indirect pointer (a0) directly.
+define i32 @caller_musttail_indirect(fp128 %a) nounwind {
+; RV32-LABEL: caller_musttail_indirect:
+; RV32:       # %bb.0:
+; RV32-NEXT:    tail callee_musttail_indirect
+;
+; RV64-LABEL: caller_musttail_indirect:
+; RV64:       # %bb.0:
+; RV64-NEXT:    tail callee_musttail_indirect
+  %call = musttail call i32 @callee_musttail_indirect(fp128 %a)
+  ret i32 %call
+}
+
+; Verify that non-musttail tail call with indirect args does NOT tail call
+; (this is the PR #184972 fix - indirect args are unsafe for regular tail calls).
+define void @caller_no_musttail_indirect() nounwind {
+; RV32-LABEL: caller_no_musttail_indirect:
+; RV32:       # %bb.0:
+; RV32-NEXT:    addi sp, sp, -32
+; RV32-NEXT:    sw ra, 28(sp) # 4-byte Folded Spill
+; RV32-NEXT:    lui a1, 262128
+; RV32-NEXT:    mv a0, sp
+; RV32-NEXT:    sw zero, 0(sp)
+; RV32-NEXT:    sw zero, 4(sp)
+; RV32-NEXT:    sw zero, 8(sp)
+; RV32-NEXT:    sw a1, 12(sp)
+; RV32-NEXT:    call callee_musttail_indirect
+; RV32-NEXT:    lw ra, 28(sp) # 4-byte Folded Reload
+; RV32-NEXT:    addi sp, sp, 32
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: caller_no_musttail_indirect:
+; RV64:       # %bb.0:
+; RV64-NEXT:    lui a1, 16383
+; RV64-NEXT:    slli a1, a1, 36
+; RV64-NEXT:    li a0, 0
+; RV64-NEXT:    tail callee_musttail_indirect
+  %call = tail call i32 @callee_musttail_indirect(fp128 0xL00000000000000003FFF000000000000)
+  ret void
+}
+
+; Verify that non-musttail tail call forwarding an indirect arg from the
+; caller's own parameters also does NOT tail call (the arg lives on the
+; caller's frame, which would be deallocated).
+define i32 @caller_no_musttail_forward_indirect(fp128 %a) nounwind {
+; RV32-LABEL: caller_no_musttail_forward_indirect:
+; RV32:       # %bb.0:
+; RV32-NEXT:    addi sp, sp, -32
+; RV32-NEXT:    sw ra, 28(sp) # 4-byte Folded Spill
+; RV32-NEXT:    lw a1, 0(a0)
+; RV32-NEXT:    lw a2, 4(a0)
+; RV32-NEXT:    lw a3, 8(a0)
+; RV32-NEXT:    lw a4, 12(a0)
+; RV32-NEXT:    mv a0, sp
+; RV32-NEXT:    sw a1, 0(sp)
+; RV32-NEXT:    sw a2, 4(sp)
+; RV32-NEXT:    sw a3, 8(sp)
+; RV32-NEXT:    sw a4, 12(sp)
+; RV32-NEXT:    call callee_musttail_indirect
+; RV32-NEXT:    lw ra, 28(sp) # 4-byte Folded Reload
+; RV32-NEXT:    addi sp, sp, 32
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: caller_no_musttail_forward_indirect:
+; RV64:       # %bb.0:
+; RV64-NEXT:    tail callee_musttail_indirect
+  %call = tail call i32 @callee_musttail_indirect(fp128 %a)
+  ret i32 %call
+}
+
+; Test musttail with two indirect fp128 args on RV32. Both pointers must be
+; forwarded. Exercises the DenseMap with two distinct OrigArgIndex values.
+declare i32 @callee_musttail_two_indirect(fp128 %a, fp128 %b)
+
+define i32 @caller_musttail_two_indirect(fp128 %a, fp128 %b) nounwind {
+; RV32-LABEL: caller_musttail_two_indirect:
+; RV32:       # %bb.0:
+; RV32-NEXT:    tail callee_musttail_two_indirect
+;
+; RV64-LABEL: caller_musttail_two_indirect:
+; RV64:       # %bb.0:
+; RV64-NEXT:    tail callee_musttail_two_indirect
+  %call = musttail call i32 @callee_musttail_two_indirect(fp128 %a, fp128 %b)
----------------
folkertdev wrote:

I think you should also test flipping the arguments

```
musttail call i32 @callee_musttail_two_indirect(fp128 %b, fp128 %a)
```



https://github.com/llvm/llvm-project/pull/185094