[MLIR][NVVM] Reduce the scope of the LLVM_HAS_NVPTX_TARGET guard

[joker-eph]

Most of the code here does not depend on the NVPTX target. In particular the simple offload can just emit LLVM IR and we can use this without the NVVM backend being built, which can be useful for a frontend that just need to serialize the IR and leave it up to the runtime to JIT further.

[llvmbot]

@llvm/pr-subscribers-mlir
@llvm/pr-subscribers-mlir-gpu

@llvm/pr-subscribers-mlir-llvm

Author: Mehdi Amini (joker-eph)

Changes

Most of the code here does not depend on the NVPTX target. In particular the simple offload can just emit LLVM IR and we can use this without the NVVM backend being built, which can be useful for a frontend that just need to serialize the IR and leave it up to the runtime to JIT further.

---

Full diff: https://github.com/llvm/llvm-project/pull/97349.diff

2 Files Affected:

• (modified) mlir/lib/Dialect/GPU/Transforms/ModuleToBinary.cpp (-2)
• (modified) mlir/lib/Target/LLVM/NVVM/Target.cpp (+6-8)

diff --git a/mlir/lib/Dialect/GPU/Transforms/ModuleToBinary.cpp b/mlir/lib/Dialect/GPU/Transforms/ModuleToBinary.cpp
index 1e7596e8cc4af..62b438656f03f 100644
--- a/mlir/lib/Dialect/GPU/Transforms/ModuleToBinary.cpp
+++ b/mlir/lib/Dialect/GPU/Transforms/ModuleToBinary.cpp
@@ -48,9 +48,7 @@ void GpuModuleToBinaryPass::getDependentDialects(
   // Register all GPU related translations.
   registry.insert<gpu::GPUDialect>();
   registry.insert<LLVM::LLVMDialect>();
-#if LLVM_HAS_NVPTX_TARGET
   registry.insert<NVVM::NVVMDialect>();
-#endif
 #if MLIR_ENABLE_ROCM_CONVERSIONS
   registry.insert<ROCDL::ROCDLDialect>();
 #endif
diff --git a/mlir/lib/Target/LLVM/NVVM/Target.cpp b/mlir/lib/Target/LLVM/NVVM/Target.cpp
index acb903aa37caf..c32ccf90cbbbc 100644
--- a/mlir/lib/Target/LLVM/NVVM/Target.cpp
+++ b/mlir/lib/Target/LLVM/NVVM/Target.cpp
@@ -157,7 +157,6 @@ SerializeGPUModuleBase::loadBitcodeFiles(llvm::Module &module) {
   return std::move(bcFiles);
 }
 
-#if LLVM_HAS_NVPTX_TARGET
 namespace {
 class NVPTXSerializer : public SerializeGPUModuleBase {
 public:
@@ -532,6 +531,12 @@ NVPTXSerializer::moduleToObject(llvm::Module &llvmModule) {
   if (targetOptions.getCompilationTarget() == gpu::CompilationTarget::Offload)
     return SerializeGPUModuleBase::moduleToObject(llvmModule);
 
+#if ! LLVM_HAS_NVPTX_TARGET
+  getOperation()->emitError(
+      "The `NVPTX` target was not built. Please enable it when building LLVM.");
+  return std::nullopt;
+#endif // LLVM_HAS_NVPTX_TARGET
+
   // Emit PTX code.
   std::optional<llvm::TargetMachine *> targetMachine =
       getOrCreateTargetMachine();
@@ -569,7 +574,6 @@ NVPTXSerializer::moduleToObject(llvm::Module &llvmModule) {
   return compileToBinary(*serializedISA);
 #endif // MLIR_ENABLE_NVPTXCOMPILER
 }
-#endif // LLVM_HAS_NVPTX_TARGET
 
 std::optional<SmallVector<char, 0>>
 NVVMTargetAttrImpl::serializeToObject(Attribute attribute, Operation *module,
@@ -581,15 +585,9 @@ NVVMTargetAttrImpl::serializeToObject(Attribute attribute, Operation *module,
     module->emitError("Module must be a GPU module.");
     return std::nullopt;
   }
-#if LLVM_HAS_NVPTX_TARGET
   NVPTXSerializer serializer(*module, cast<NVVMTargetAttr>(attribute), options);
   serializer.init();
   return serializer.run();
-#else
-  module->emitError(
-      "The `NVPTX` target was not built. Please enable it when building LLVM.");
-  return std::nullopt;
-#endif // LLVM_HAS_NVPTX_TARGET
 }
 
 Attribute

[github-actions]

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

[fabianmcg]

LGTM, see minor comment on ModuleToBinary