Implement some of the review suggestions.

Author: AlexVlx

llvm/lib/Target/AMDGPU/AMDGPUExpandFeaturePredicates.cpp

@@ -13,12 +13,10 @@
 // (AMDGCNSPIRV). These placeholder globals are used to guide target specific
 // lowering, once the concrete target is known, by way of constant folding their
 // value all the way into a terminator (i.e. a controlled block) or into a no
-// live use scenario. The pass makes a best effort attempt to look through
-// calls, i.e. a constant evaluatable passthrough of a predicate value will
-// generally work, however we hard fail if the folding fails, to avoid obtuse
-// BE errors or opaque run time errors. This pass should run as early as
-// possible / immediately after Clang CodeGen, so that the optimisation pipeline
-// and the BE operate with concrete target data.
+// live use scenario. We hard fail if the folding fails, to avoid obtuse BE
+// errors or opaque run time errors. This pass should run as early as possible /
+// immediately after Clang CodeGen, so that the optimisation pipeline and the BE
+// operate with concrete target data.
 //===----------------------------------------------------------------------===//
 
 #include "AMDGPU.h"
@@ -50,13 +48,13 @@ template <typename C> void collectUsers(Value *V, C &Container) {
 }
 
 inline void setPredicate(const GCNSubtarget &ST, GlobalVariable *P) {
-  const auto IsFeature = P->getName().starts_with("llvm.amdgcn.has");
-  const auto Offset =
+  const bool IsFeature = P->getName().starts_with("llvm.amdgcn.has");
+  const size_t Offset =
       IsFeature ? sizeof("llvm.amdgcn.has") : sizeof("llvm.amdgcn.is");
 
-  auto PV = P->getName().substr(Offset).str();
+  std::string PV = P->getName().substr(Offset).str();
   if (IsFeature) {
-    auto Dx = PV.find(',');
+    size_t Dx = PV.find(',');
     while (Dx != std::string::npos) {
       PV.insert(++Dx, {'+'});
 
@@ -65,7 +63,7 @@ inline void setPredicate(const GCNSubtarget &ST, GlobalVariable *P) {
     PV.insert(PV.cbegin(), '+');
   }
 
-  auto *PTy = P->getValueType();
+  Type *PTy = P->getValueType();
   P->setLinkage(GlobalValue::PrivateLinkage);
   P->setExternallyInitialized(false);
 
@@ -103,15 +101,15 @@ std::pair<PreservedAnalyses, bool> handlePredicate(const GCNSubtarget &ST,
     auto *I = *ToFold.begin();
     ToFold.erase(I);
 
+    I->dropDroppableUses();
+
     if (auto *C = ConstantFoldInstruction(I, P->getDataLayout())) {
       collectUsers(I, ToFold);
       I->replaceAllUsesWith(C);
       I->eraseFromParent();
       continue;
     } else if (I->isTerminator() && ConstantFoldTerminator(I->getParent())) {
       continue;
-    } else if (I->users().empty()) {
-      continue;
     }
 
     return unfoldableFound(I->getParent()->getParent(), P, I);

llvm/test/CodeGen/AMDGPU/amdgpu-expand-feature-predicates.ll

@@ -121,7 +121,7 @@ define amdgpu_kernel void @kernel(ptr addrspace(1) %p.coerce, i32 %x) {
 ; GFX1010-NEXT:    br label %[[IF_END6]]
 ; GFX1010:       [[IF_END6]]:
 ; GFX1010-NEXT:    call void @llvm.assume(i1 true)
-; GFX1010-NEXT:    call void @llvm.assume(i1 false)
+; GFX1010-NEXT:    call void @llvm.assume(i1 true)
 ; GFX1010-NEXT:    br label %[[FOR_COND]]
 ; GFX1010:       [[FOR_COND]]:
 ; GFX1010-NEXT:    [[DOTPROMOTED:%.*]] = load i32, ptr [[TMP1]], align 4
@@ -167,7 +167,7 @@ define amdgpu_kernel void @kernel(ptr addrspace(1) %p.coerce, i32 %x) {
 ; GFX1101-NEXT:    call void @llvm.amdgcn.s.ttracedata.imm(i16 1)
 ; GFX1101-NEXT:    br label %[[IF_END6]]
 ; GFX1101:       [[IF_END6]]:
-; GFX1101-NEXT:    call void @llvm.assume(i1 false)
+; GFX1101-NEXT:    call void @llvm.assume(i1 true)
 ; GFX1101-NEXT:    call void @llvm.assume(i1 true)
 ; GFX1101-NEXT:    br label %[[FOR_COND:.*]]
 ; GFX1101:       [[FOR_COND]]:
@@ -278,7 +278,7 @@ define amdgpu_kernel void @kernel(ptr addrspace(1) %p.coerce, i32 %x) {
 ; GFX1201-W64-NEXT:    call void @llvm.amdgcn.s.ttracedata.imm(i16 1)
 ; GFX1201-W64-NEXT:    br label %[[IF_END11]]
 ; GFX1201-W64:       [[IF_END11]]:
-; GFX1201-W64-NEXT:    call void @llvm.assume(i1 false)
+; GFX1201-W64-NEXT:    call void @llvm.assume(i1 true)
 ; GFX1201-W64-NEXT:    [[DOTPROMOTED9:%.*]] = load i32, ptr [[TMP1]], align 4
 ; GFX1201-W64-NEXT:    [[SUB13_PEEL:%.*]] = sub nsw i32 [[DOTPROMOTED9]], [[X]]
 ; GFX1201-W64-NEXT:    store i32 [[SUB13_PEEL]], ptr [[TMP1]], align 4