[RISC-V] Only emit multiples of 16 as immediate for cm.push

[nemanjai]

An immediate that isn't a multiple of 16 is meaningless and we should always emit the next higher multiple. Otherwise the value is simply truncated of course which leads to clobbering CSR's on the stack.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Nemanja Ivanovic (nemanjai)

Changes

An immediate that isn't a multiple of 16 is meaningless and we should always emit the next higher multiple. Otherwise the value is simply truncated of course which leads to clobbering CSR's on the stack.

---

Full diff: https://github.com/llvm/llvm-project/pull/84935.diff

2 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVFrameLowering.cpp (+1-1)
• (modified) llvm/test/CodeGen/RISCV/zcmp-additional-stack.ll (+2-1)

diff --git a/llvm/lib/Target/RISCV/RISCVFrameLowering.cpp b/llvm/lib/Target/RISCV/RISCVFrameLowering.cpp
index 8bac41372b5a83..4f3d915818d1ae 100644
--- a/llvm/lib/Target/RISCV/RISCVFrameLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVFrameLowering.cpp
@@ -555,7 +555,7 @@ void RISCVFrameLowering::emitPrologue(MachineFunction &MF,
       FirstFrameSetup->getOpcode() == RISCV::CM_PUSH) {
     // Use available stack adjustment in push instruction to allocate additional
     // stack space.
-    uint64_t Spimm = std::min(StackSize, (uint64_t)48);
+    uint64_t Spimm = alignTo(std::min(StackSize, (uint64_t)48), 16);
     FirstFrameSetup->getOperand(1).setImm(Spimm);
     StackSize -= Spimm;
   }
diff --git a/llvm/test/CodeGen/RISCV/zcmp-additional-stack.ll b/llvm/test/CodeGen/RISCV/zcmp-additional-stack.ll
index e5c2e0180ee0a6..738d1b07e2210c 100644
--- a/llvm/test/CodeGen/RISCV/zcmp-additional-stack.ll
+++ b/llvm/test/CodeGen/RISCV/zcmp-additional-stack.ll
@@ -3,7 +3,8 @@
 define ptr @func(ptr %s, i32 %_c, ptr %incdec.ptr, i1 %0, i8 %conv14) #0 {
 ; RV32-LABEL: func:
 ; RV32:       # %bb.0: # %entry
-; RV32-NEXT:    cm.push {ra, s0-s1}, -24
+; RV32-NEXT:    cm.push {ra, s0-s1}, -32
+; RV32-NEXT:    addi sp, sp, 8
 ; RV32-NEXT:    .cfi_def_cfa_offset 24
 ; RV32-NEXT:    .cfi_offset ra, -12
 ; RV32-NEXT:    .cfi_offset s0, -8

[francisvm]

CM_PUSH LGTM. Do we need to do the same for CM_POP in emitEpilogue?

[topperc]

Does this mean we can move the sp by more than StackSize? Doesn't that make the position of sp out of sync with the rest of the code?

[nemanjai]

Well, if the amount of space in the offset for the cm.push is more than the additional stack space required, then yes we could. I don't follow why this is a problem though. Or maybe I am misunderstanding what the StackSize variable actually represents - the way I have interpreted it is "The amount of stack space required for parameters and local objects (not including the space to spill CSR's)". Of course, if my interpretation is correct, I'd certainly make the argument that the variable is poorly named, but hey.

[jupiter-zzp]

We can move the sp by more than StackSize here, and if it does, the stack pointer will be readjusted in the following if statement where StackSize != 0.

[nemanjai]

CM_PUSH LGTM. Do we need to do the same for CM_POP in emitEpilogue?

Argh! Yes indeed! I don't know why I missed that (other than I am porting this from a downstream fix I did which is on LLVM 17 where the fix is in a separate function that computes the immediate and is called in both places).
In any case, I'll fix it.

[topperc]

Aren't these cfi directives wrong, since the stack pointer was moved by 32 not 24?

[nemanjai]

They most certainly appear to be wrong.