[CIR] Fix vector issues from latest rebase

Author: bcardosolopes

clang/lib/CIR/CodeGen/ABIInfo.h

@@ -10,6 +10,8 @@
 #define LLVM_CLANG_LIB_CIR_ABIINFO_H
 
 #include "clang/AST/Type.h"
+#include "clang/Basic/LangOptions.h"
+#include "clang/CIR/Dialect/IR/CIRDialect.h"
 
 namespace clang::CIRGen {
 
@@ -39,6 +41,13 @@ class ABIInfo {
   // Implement the Type::IsPromotableIntegerType for ABI specific needs. The
   // only difference is that this consideres bit-precise integer types as well.
   bool isPromotableIntegerTypeForABI(clang::QualType Ty) const;
+
+  /// Returns the optimal vector memory type based on the given vector type. For
+  /// example, on certain targets, a vector with 3 elements might be promoted to
+  /// one with 4 elements to improve performance.
+  virtual cir::VectorType
+  getOptimalVectorMemoryType(cir::VectorType T,
+                             const clang::LangOptions &Opt) const;
 };
 
 } // namespace clang::CIRGen

clang/lib/CIR/CodeGen/CIRGenExpr.cpp

@@ -173,9 +173,9 @@ static Address emitPointerWithAlignment(const Expr *expr,
             llvm_unreachable("NYI");
         }
 
-        auto ElemTy = cgf.convertTypeForMem(expr->getType()->getPointeeType());
+        auto eltTy = cgf.convertTypeForMem(expr->getType()->getPointeeType());
         addr = cgf.getBuilder().createElementBitCast(
-            cgf.getLoc(expr->getSourceRange()), addr, ElemTy);
+            cgf.getLoc(expr->getSourceRange()), addr, eltTy);
         if (CE->getCastKind() == CK_AddressSpaceConversion) {
           assert(!cir::MissingFeatures::addressSpace());
           llvm_unreachable("NYI");
@@ -616,6 +616,25 @@ void CIRGenFunction::emitStoreOfScalar(mlir::Value value, Address addr,
                                        LValueBaseInfo baseInfo,
                                        TBAAAccessInfo tbaaInfo, bool isInit,
                                        bool isNontemporal) {
+  assert(!cir::MissingFeatures::threadLocal() && "NYI");
+
+  auto eltTy = addr.getElementType();
+  if (const auto *clangVecTy = ty->getAs<clang::VectorType>()) {
+    // Boolean vectors use `iN` as storage type.
+    if (clangVecTy->isExtVectorBoolType()) {
+      llvm_unreachable("isExtVectorBoolType NYI");
+    }
+
+    // Handle vectors of size 3 like size 4 for better performance.
+    const auto vTy = cast<cir::VectorType>(eltTy);
+    auto newVecTy =
+        CGM.getABIInfo().getOptimalVectorMemoryType(vTy, getLangOpts());
+
+    if (vTy != newVecTy) {
+      llvm_unreachable("NYI");
+    }
+  }
+
   value = emitToMemory(value, ty);
 
   LValue atomicLValue =
@@ -626,26 +645,6 @@ void CIRGenFunction::emitStoreOfScalar(mlir::Value value, Address addr,
     return;
   }
 
-  mlir::Type SrcTy = value.getType();
-  if (const auto *ClangVecTy = ty->getAs<clang::VectorType>()) {
-    // TODO(CIR): this has fallen out of date with codegen
-    llvm_unreachable("NYI: Special treatment of 3-element vector store");
-    // auto VecTy = dyn_cast<cir::VectorType>(SrcTy);
-    // if (!CGM.getCodeGenOpts().PreserveVec3Type &&
-    //     ClangVecTy->getNumElements() == 3) {
-    //   // Handle vec3 special.
-    //   if (VecTy && VecTy.getSize() == 3) {
-    //     // Our source is a vec3, do a shuffle vector to make it a vec4.
-    //     value = builder.createVecShuffle(value.getLoc(), value,
-    //                                      ArrayRef<int64_t>{0, 1, 2, -1});
-    //     SrcTy = cir::VectorType::get(VecTy.getContext(), VecTy.getEltType(), 4);
-    //   }
-    //   if (addr.getElementType() != SrcTy) {
-    //     addr = addr.withElementType(SrcTy);
-    //   }
-    // }
-  }
-
   // Update the alloca with more info on initialization.
   assert(addr.getPointer() && "expected pointer to exist");
   auto SrcAlloca =
@@ -2917,40 +2916,36 @@ mlir::Value CIRGenFunction::emitLoadOfScalar(Address addr, bool isVolatile,
                                              LValueBaseInfo baseInfo,
                                              TBAAAccessInfo tbaaInfo,
                                              bool isNontemporal) {
-  // TODO(CIR): this has fallen out of sync with codegen
-  // Atomic operations have to be done on integral types
-  LValue atomicLValue =
-      LValue::makeAddr(addr, ty, getContext(), baseInfo, tbaaInfo);
-  if (ty->isAtomicType() || LValueIsSuitableForInlineAtomic(atomicLValue)) {
-    llvm_unreachable("NYI");
-  }
+  assert(!cir::MissingFeatures::threadLocal() && "NYI");
+  auto eltTy = addr.getElementType();
 
-  auto ElemTy = addr.getElementType();
+  if (const auto *clangVecTy = ty->getAs<clang::VectorType>()) {
+    // Boolean vectors use `iN` as storage type.
+    if (clangVecTy->isExtVectorBoolType()) {
+      llvm_unreachable("NYI");
+    }
 
-  if (const auto *ClangVecTy = ty->getAs<clang::VectorType>()) {
     // Handle vectors of size 3 like size 4 for better performance.
-    const auto VTy = cast<cir::VectorType>(ElemTy);
+    const auto vTy = cast<cir::VectorType>(eltTy);
+    auto newVecTy =
+        CGM.getABIInfo().getOptimalVectorMemoryType(vTy, getLangOpts());
 
-    // TODO(CIR): this has fallen out of sync with codegen
-    llvm_unreachable("NYI: Special treatment of 3-element vector store");
-    // if (!CGM.getCodeGenOpts().PreserveVec3Type &&
-    //     ClangVecTy->getNumElements() == 3) {
-    //   auto loc = addr.getPointer().getLoc();
-    //   auto vec4Ty = cir::VectorType::get(VTy.getContext(), VTy.getEltType(), 4);
-    //   Address Cast = addr.withElementType(vec4Ty);
-    //   // Now load value.
-    //   mlir::Value V = builder.createLoad(loc, Cast);
+    if (vTy != newVecTy) {
+      llvm_unreachable("NYI");
+    }
+  }
 
-      // // Shuffle vector to get vec3.
-      // V = builder.createVecShuffle(loc, V, ArrayRef<int64_t>{0, 1, 2});
-      // return emitFromMemory(V, ty);
-    // }
+  LValue atomicLValue =
+      LValue::makeAddr(addr, ty, getContext(), baseInfo, tbaaInfo);
+  if (ty->isAtomicType() || LValueIsSuitableForInlineAtomic(atomicLValue)) {
+    llvm_unreachable("NYI");
   }
 
+  // TODO(cir): modernize this with addr.withElementType(convertTypeForLoadStore
   auto Ptr = addr.getPointer();
-  if (mlir::isa<cir::VoidType>(ElemTy)) {
-    ElemTy = cir::IntType::get(&getMLIRContext(), 8, true);
-    auto ElemPtrTy = cir::PointerType::get(&getMLIRContext(), ElemTy);
+  if (mlir::isa<cir::VoidType>(eltTy)) {
+    eltTy = cir::IntType::get(&getMLIRContext(), 8, true);
+    auto ElemPtrTy = cir::PointerType::get(&getMLIRContext(), eltTy);
     Ptr = builder.create<cir::CastOp>(loc, ElemPtrTy, cir::CastKind::bitcast,
                                       Ptr);
   }
@@ -2962,7 +2957,6 @@ mlir::Value CIRGenFunction::emitLoadOfScalar(Address addr, bool isVolatile,
   CGM.decorateOperationWithTBAA(loadOp, tbaaInfo);
 
   assert(!cir::MissingFeatures::emitScalarRangeCheck() && "NYI");
-
   return emitFromMemory(loadOp, ty);
 }
 

clang/lib/CIR/CodeGen/CIRGenModule.cpp

@@ -494,6 +494,10 @@ void CIRGenModule::setDSOLocal(CIRGlobalValueInterface GV) const {
   GV.setDSOLocal(shouldAssumeDSOLocal(*this, GV));
 }
 
+const ABIInfo &CIRGenModule::getABIInfo() {
+  return getTargetCIRGenInfo().getABIInfo();
+}
+
 void CIRGenModule::emitGlobal(GlobalDecl GD) {
   llvm::TimeTraceScope scope("build CIR Global", [&]() -> std::string {
     auto *ND = dyn_cast<NamedDecl>(GD.getDecl());

clang/lib/CIR/CodeGen/CIRGenModule.h

@@ -471,8 +471,8 @@ class CIRGenModule : public CIRGenTypeCache {
   /// NOTE: This should only be called for definitions.
   void setCommonAttributes(GlobalDecl GD, mlir::Operation *GV);
 
-  // TODO: this obviously overlaps with
   const TargetCIRGenInfo &getTargetCIRGenInfo();
+  const ABIInfo &getABIInfo();
 
   /// Helpers to convert Clang's SourceLocation to a MLIR Location.
   mlir::Location getLoc(clang::SourceLocation SLoc);

clang/lib/CIR/CodeGen/TargetInfo.cpp

@@ -318,6 +318,15 @@ static bool classifyReturnType(const CIRGenCXXABI &CXXABI,
 
 CIRGenCXXABI &ABIInfo::getCXXABI() const { return CGT.getCXXABI(); }
 
+cir::VectorType
+ABIInfo::getOptimalVectorMemoryType(cir::VectorType T,
+                                    const clang::LangOptions &Opt) const {
+  if (T.getSize() == 3 && !Opt.PreserveVec3Type) {
+    llvm_unreachable("NYI");
+  }
+  return T;
+}
+
 clang::ASTContext &ABIInfo::getContext() const { return CGT.getContext(); }
 
 cir::ABIArgInfo X86_64ABIInfo::getIndirectResult(QualType Ty,

clang/lib/CIR/Dialect/Transforms/TargetLowering/ABIInfo.h

@@ -27,6 +27,7 @@ class LowerTypes;
 
 /// Target specific hooks for defining how a type should be passed or returned
 /// from functions.
+/// FIXME(cir): this needs to be merged with clang/lib/CIR/CodeGen/ABIInfo.h
 class ABIInfo {
 protected:
   LowerTypes <

clang/test/CIR/CodeGen/OpenCL/printf.cl

@@ -14,7 +14,6 @@
 // RUN: FileCheck -input-file=%t.30fp64.ll -check-prefixes=LLVM-FP64,LLVM-ALL %s
 // RUN: %clang_cc1 -fclangir -no-enable-noundef-analysis -cl-std=CL3.0 -cl-ext=-__opencl_c_fp64,-cl_khr_fp64 -triple spirv64-unknown-unknown -disable-llvm-passes -emit-llvm -fno-clangir-call-conv-lowering -o %t.30nofp64.ll %s
 // RUN: FileCheck -input-file=%t.30nofp64.ll -check-prefixes=LLVM-NOFP64,LLVM-ALL %s
-// XFAIL: *
 
 typedef __attribute__((ext_vector_type(2))) float float2;
 typedef __attribute__((ext_vector_type(2))) half half2;

clang/test/CIR/CodeGen/builtins-elementwise.c

@@ -3,7 +3,6 @@
 // RUN: %clang_cc1 -triple aarch64-none-linux-android24  -fclangir \
 // RUN:  -emit-llvm  %s -o %t.ll
 // RUN: FileCheck  --check-prefix=LLVM --input-file=%t.ll %s
-// XFAIL: *
 
 typedef int vint4 __attribute__((ext_vector_type(4)));
 typedef float vfloat4 __attribute__((ext_vector_type(4)));

clang/test/CIR/CodeGen/vectype.cpp

@@ -1,5 +1,4 @@
 // RUN: %clang_cc1 -std=c++17 -triple x86_64-unknown-linux-gnu -fclangir -emit-cir %s -o - | FileCheck %s
-// XFAIL: *
 
 typedef int vi4 __attribute__((vector_size(16)));
 typedef double vd2 __attribute__((vector_size(16)));

clang/test/CIR/Lowering/ThroughMLIR/vectype.cpp

@@ -1,6 +1,5 @@
 // RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -fno-clangir-direct-lowering -emit-mlir %s -o %t.mlir
 // RUN: FileCheck --input-file=%t.mlir %s
-// XFAIL: *
 
 typedef int vi4 __attribute__((vector_size(16)));
 

clang/test/CIR/Lowering/vectype.cpp

@@ -2,7 +2,6 @@
 // RUN: cir-opt %t.cir -cir-to-llvm -o %t.mlir
 // RUN: %clang_cc1 -std=c++17 -triple x86_64-unknown-linux-gnu -fclangir -emit-llvm %s -o %t.ii
 // RUN: FileCheck --input-file=%t.mlir %s
-// XFAIL: *
 
 typedef int vi4 __attribute__((vector_size(16)));
 typedef double vd2 __attribute__((vector_size(16)));