[Verifier][AMDGPU] Restrict targets for chain calling convention.

[jofrn]

The chain calling convention is invalid in particular situations like being on an architecture it cannot compile to.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

@llvm/pr-subscribers-llvm-ir

Author: None (jofrn)

Changes

The chain calling convention is invalid in particular situations like being on an architecture it cannot compile to.

---

Full diff: https://github.com/llvm/llvm-project/pull/157469.diff

2 Files Affected:

• (modified) llvm/lib/IR/Verifier.cpp (+8-2)
• (added) llvm/test/CodeGen/AMDGPU/amdgpu-cs-chain-invalid-arch.ll (+6)

diff --git a/llvm/lib/IR/Verifier.cpp b/llvm/lib/IR/Verifier.cpp
index 1d3c379f461fa..63ae083fec8e8 100644
--- a/llvm/lib/IR/Verifier.cpp
+++ b/llvm/lib/IR/Verifier.cpp
@@ -2938,10 +2938,16 @@ void Verifier::visitFunction(const Function &F) {
           "Calling convention parameter requires byval", &F);
     break;
   }
-  case CallingConv::AMDGPU_KERNEL:
-  case CallingConv::SPIR_KERNEL:
   case CallingConv::AMDGPU_CS_Chain:
   case CallingConv::AMDGPU_CS_ChainPreserve:
+  {
+    auto TT = M.getTargetTriple().str();
+    if (TT.find("gfx1200") || TT.find("gfx942"))
+      Check(false, "Chain calling convention is invalid on this target", &F);
+  }
+    [[fallthrough]];
+  case CallingConv::AMDGPU_KERNEL:
+  case CallingConv::SPIR_KERNEL:
     Check(F.getReturnType()->isVoidTy(),
           "Calling convention requires void return type", &F);
     [[fallthrough]];
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-cs-chain-invalid-arch.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-cs-chain-invalid-arch.ll
new file mode 100644
index 0000000000000..c6f8a09eaf5b6
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-cs-chain-invalid-arch.ll
@@ -0,0 +1,6 @@
+; RUN: not llc -mtriple=amdgcn -mcpu=gfx1200 -o - < %s 2>&1 | FileCheck %s
+
+define amdgpu_cs_chain void @test_alloca() {
+; CHECK: Chain calling convention is invalid on this target
+  ret void
+}

[github-actions]

⚠️ C/C++ code formatter, clang-format found issues in your code. ⚠️

You can test this locally with the following command:

git-clang-format --diff origin/main HEAD --extensions cpp -- llvm/lib/IR/Verifier.cpp

⚠️
The reproduction instructions above might return results for more than one PR
in a stack if you are using a stacked PR workflow. You can limit the results by
changing origin/main to the base branch/commit you want to compare against.
⚠️

View the diff from clang-format here.

diff --git a/llvm/lib/IR/Verifier.cpp b/llvm/lib/IR/Verifier.cpp
index 4f2aa4005..bf05113c8 100644
--- a/llvm/lib/IR/Verifier.cpp
+++ b/llvm/lib/IR/Verifier.cpp
@@ -2933,8 +2933,7 @@ void Verifier::visitFunction(const Function &F) {
     break;
   }
   case CallingConv::AMDGPU_CS_Chain:
-  case CallingConv::AMDGPU_CS_ChainPreserve:
-  {
+  case CallingConv::AMDGPU_CS_ChainPreserve: {
     auto TT = M.getTargetTriple().str();
     if (TT.find("gfx1200") || TT.find("gfx942"))
       Check(false, "Chain calling convention is invalid on this target", &F);

[arsenm]

You cannot rely on this subtarget check

[arsenm]

I would just fix this restriction, this seems totally artificial

[jofrn]

#132711 fixes the restriction, but they mentioned that the IR is invalid for this architecture, so the PR was never approved.

[arsenm]

I reject the notion that IR can be invalid for an architecture

[shiltian]

This is unacceptable. We should never do this.

[jofrn]

What other way does one check for subtarget?

[arsenm]

The IR verifier cannot and should not check for the subtarget. We cannot have subtarget dependent IR valid rules