| Issue |
244792
|
| Summary |
[RISCV] RVV callee-saved registers restored from the wrong address in realigned frames
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
petar-jovanovic
|
When a function has callee-saved vector registers (a `riscv_vector_cc` function, or an interrupt handler with V) and its stack is realigned, the prologue saves those registers **before** it realigns SP and sets up the base pointer. The epilogue then reloads them relative to the realigned SP or BP.
- **Frame addressed from SP:** whenever the realignment moves SP, the reload reads from a lower address than the save. The caller's vector registers come back corrupted (confirmed under qemu-riscv64), and whether a call fails depends on the SP at function entry.
- **Frame addressed from BP** (realignment plus a variable-sized object): the prologue stores through `s1` before `mv s1, sp`, i.e. through the caller's `s1`.
This is a silent miscompile: no assertion fires and `-verify-machineinstrs` is clean. The CFI for these registers is also wrong.
### Reproducer
```llvm
define riscv_vector_cc void @realign_sp() nounwind {
%local = alloca i64, align 64
store volatile i64 0, ptr %local
call void asm sideeffect "", "~{v24}"()
ret void
}
define riscv_vector_cc void @realign_bp(i64 %n) nounwind {
%local = alloca i64, align 64
%vla = alloca i8, i64 %n
store volatile i64 0, ptr %local
store volatile i8 0, ptr %vla
call void asm sideeffect "", "~{v24}"()
ret void
}
```
```
llc -mtriple=riscv64 -mattr=+v repro.ll -o -
```
Output (excerpt, annotated):
```
realign_sp:
...
addi a0, sp, 112
vs1r.v v24, (a0) # save, before the realignment
andi sp, sp, -64 # realign: SP moves down by 0, 16, 32 or 48
...
addi a0, sp, 112
vl1r.v v24, (a0) # reload, relative to the realigned SP
realign_bp:
...
addi a1, s1, 112
vs1r.v v24, (a1) # s1 still holds the caller's value
andi sp, sp, -64
mv s1, sp # BP is only set up here
...
addi a0, s1, 112
vl1r.v v24, (a0)
```
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs