[AMDGPU] Remove the AnnotateKernelFeatures pass (llvm#130198)

Previously the AnnotateKernelFeatures pass infers two attributes:
amdgpu-calls and amdgpu-stack-objects, which are used to help determine
if flat scratch init is allowed. PR llvm#118907 created the
amdgpu-no-flat-scratch-init attribute. Continuing with that work, this
patch makes use of this attribute to determine flat scratch init,
replacing amdgpu-calls and amdgpu-stack-objects. This also leads to the
removal of the AnnotateKernelFeatures pass.

Author: jwanggit86

llvm/lib/Target/AMDGPU/AMDGPU.h

@@ -95,11 +95,8 @@ void initializeAMDGPUDAGToDAGISelLegacyPass(PassRegistry &);
 
 void initializeAMDGPUAlwaysInlinePass(PassRegistry&);
 
-Pass *createAMDGPUAnnotateKernelFeaturesPass();
 Pass *createAMDGPUAttributorLegacyPass();
 void initializeAMDGPUAttributorLegacyPass(PassRegistry &);
-void initializeAMDGPUAnnotateKernelFeaturesPass(PassRegistry &);
-extern char &AMDGPUAnnotateKernelFeaturesID;
 
 // DPP/Iterative option enables the atomic optimizer with given strategy
 // whereas None disables the atomic optimizer.

llvm/lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

@@ -52,11 +52,6 @@ class AMDGPUAnnotateKernelFeatures : public CallGraphSCCPass {
 
 char AMDGPUAnnotateKernelFeatures::ID = 0;
 
-char &llvm::AMDGPUAnnotateKernelFeaturesID = AMDGPUAnnotateKernelFeatures::ID;
-
-INITIALIZE_PASS(AMDGPUAnnotateKernelFeatures, DEBUG_TYPE,
-                "Add AMDGPU function attributes", false, false)
-
 bool AMDGPUAnnotateKernelFeatures::addFeatureAttributes(Function &F) {
   bool HaveStackObjects = false;
   bool Changed = false;
@@ -131,7 +126,3 @@ bool AMDGPUAnnotateKernelFeatures::doInitialization(CallGraph &CG) {
   TM = &TPC->getTM<TargetMachine>();
   return false;
 }
-
-Pass *llvm::createAMDGPUAnnotateKernelFeaturesPass() {
-  return new AMDGPUAnnotateKernelFeatures();
-}

llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def

@@ -149,9 +149,3 @@ DUMMY_MACHINE_FUNCTION_PASS("amdgpu-regbanklegalize", AMDGPURegBankLegalizePass(
 DUMMY_MACHINE_FUNCTION_PASS("amdgpu-regbank-combiner", AMDGPURegBankCombinerPass())
 
 #undef DUMMY_MACHINE_FUNCTION_PASS
-
-
-#define DUMMY_CGSCC_PASS(NAME, CREATE_PASS)
-DUMMY_CGSCC_PASS("amdgpu-annotate-kernel-features", AMDGPUAnnotateKernelFeaturesPass())
-
-#undef DUMMY_CGSCC_PASS

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -515,7 +515,6 @@ extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
   initializeAMDGPUAlwaysInlinePass(*PR);
   initializeAMDGPUSwLowerLDSLegacyPass(*PR);
   initializeAMDGPUAttributorLegacyPass(*PR);
-  initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
   initializeAMDGPUAnnotateUniformValuesLegacyPass(*PR);
   initializeAMDGPUArgumentUsageInfoPass(*PR);
   initializeAMDGPUAtomicOptimizerPass(*PR);
@@ -1317,12 +1316,6 @@ void AMDGPUPassConfig::addIRPasses() {
 }
 
 void AMDGPUPassConfig::addCodeGenPrepare() {
-  if (TM->getTargetTriple().isAMDGCN()) {
-    // FIXME: This pass adds 2 hacky attributes that can be replaced with an
-    // analysis, and should be removed.
-    addPass(createAMDGPUAnnotateKernelFeaturesPass());
-  }
-
   if (TM->getTargetTriple().isAMDGCN() && EnableLowerKernelArguments)
     addPass(createAMDGPULowerKernelArgumentsPass());
 

llvm/lib/Target/AMDGPU/GCNSubtarget.cpp

@@ -601,12 +601,6 @@ GCNUserSGPRUsageInfo::GCNUserSGPRUsageInfo(const Function &F,
   const CallingConv::ID CC = F.getCallingConv();
   const bool IsKernel =
       CC == CallingConv::AMDGPU_KERNEL || CC == CallingConv::SPIR_KERNEL;
-  // FIXME: Should have analysis or something rather than attribute to detect
-  // calls.
-  const bool HasCalls = F.hasFnAttribute("amdgpu-calls");
-  // FIXME: This attribute is a hack, we just need an analysis on the function
-  // to look for allocas.
-  const bool HasStackObjects = F.hasFnAttribute("amdgpu-stack-objects");
 
   if (IsKernel && (!F.arg_empty() || ST.getImplicitArgNumBytes(F) != 0))
     KernargSegmentPtr = true;
@@ -629,12 +623,13 @@ GCNUserSGPRUsageInfo::GCNUserSGPRUsageInfo(const Function &F,
       DispatchID = true;
   }
 
-  // TODO: This could be refined a lot. The attribute is a poor way of
-  // detecting calls or stack objects that may require it before argument
-  // lowering.
   if (ST.hasFlatAddressSpace() && AMDGPU::isEntryFunctionCC(CC) &&
       (IsAmdHsaOrMesa || ST.enableFlatScratch()) &&
-      (HasCalls || HasStackObjects || ST.enableFlatScratch()) &&
+      // FlatScratchInit cannot be true for graphics CC if enableFlatScratch()
+      // is false.
+      (ST.enableFlatScratch() ||
+       (!AMDGPU::isGraphics(CC) &&
+        !F.hasFnAttribute("amdgpu-no-flat-scratch-init"))) &&
       !ST.flatScratchIsArchitected()) {
     FlatScratchInit = true;
   }