[WebAssembly] Define local sp if llvm.stacksave is used

[kateinoigakukun]

Usually llvm.stacksave/stackrestore are used together with alloca but they can appear without it (e.g. alloca can be optimized away). WebAssembly's function local physical user sp register, which is referenced by llvm.stacksave is created while frame lowering and replaced with virtual register.
However the sp register was not created when llvm.stacksave is used without alloca, and it led MIR verification failure about use-before-def of sp virtual register.

Resolves #62235

[llvmbot]

@llvm/pr-subscribers-backend-webassembly

Changes

Usually llvm.stacksave/stackrestore are used together with alloca but they can appear without it (e.g. alloca can be optimized away). WebAssembly's function local physical user sp register, which is referenced by llvm.stacksave is created while frame lowering and replaced with virtual register.
However the sp register was not created when llvm.stacksave is used without alloca, and it led MIR verification failure about use-before-def of sp virtual register.

Resolves #62235

---

Full diff: https://github.com/llvm/llvm-project/pull/68133.diff

2 Files Affected:

• (modified) llvm/lib/Target/WebAssembly/WebAssemblyFrameLowering.cpp (+8-1)
• (modified) llvm/test/CodeGen/WebAssembly/userstack.ll (+12)

diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyFrameLowering.cpp b/llvm/lib/Target/WebAssembly/WebAssemblyFrameLowering.cpp
index e60f1397b993355..d26234e7e71d1d4 100644
--- a/llvm/lib/Target/WebAssembly/WebAssemblyFrameLowering.cpp
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyFrameLowering.cpp
@@ -147,7 +147,14 @@ bool WebAssemblyFrameLowering::needsPrologForEH(
 /// Unlike a machine stack pointer, the wasm user stack pointer is a global
 /// variable, so it is loaded into a register in the prolog.
 bool WebAssemblyFrameLowering::needsSP(const MachineFunction &MF) const {
-  return needsSPForLocalFrame(MF) || needsPrologForEH(MF);
+  auto &MRI = MF.getRegInfo();
+  // llvm.stacksave can explicitly read SP register and it can appear without
+  // dynamic alloca.
+  bool hasExplicitSPUse = any_of(MRI.use_operands(getSPReg(MF)), [](MachineOperand &MO) {
+    return !MO.isImplicit();
+  });
+
+  return needsSPForLocalFrame(MF) || needsPrologForEH(MF) || hasExplicitSPUse;
 }
 
 /// Returns true if the local user-space stack pointer needs to be written back
diff --git a/llvm/test/CodeGen/WebAssembly/userstack.ll b/llvm/test/CodeGen/WebAssembly/userstack.ll
index b92946d1a6af7f2..ea0fd49503380d6 100644
--- a/llvm/test/CodeGen/WebAssembly/userstack.ll
+++ b/llvm/test/CodeGen/WebAssembly/userstack.ll
@@ -550,6 +550,18 @@ define void @llvm_stack_builtins(i32 %alloc) noredzone {
  ret void
 }
 
+; Use of stacksave requires local SP definition even without dymamic alloca.
+; CHECK-LABEL: llvm_stacksave_noalloca:
+define void @llvm_stacksave_noalloca() noredzone {
+ ; CHECK: global.get $push[[L11:.+]]=, __stack_pointer{{$}}
+ %stack = call i8* @llvm.stacksave()
+
+ ; CHECK-NEXT: call use_i8_star, $pop[[L11:.+]]
+ call void @use_i8_star(i8* %stack)
+
+ ret void
+}
+
 ; Not actually using the alloca'd variables exposed an issue with register
 ; stackification, where copying the stack pointer into the frame pointer was
 ; moved after the stack pointer was updated for the dynamic alloca.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[dschuff]

Otherwise LGTM

[kateinoigakukun]

Thank you for reviewing quickly!