[clang][CodeGen][X86_64] Honor per-function AVX ABI in C/C++ call paths, maintain old psABI for PlayStation. (#193298)

Fixes #64706

Wire per function x86 AVX ABI level into CodeGen arrangement methods so
__attribute(__target("avx"))) / __attribute(__target("avx512f"))) on
methods, ctors, and free-functions affects ABI lowering consistently.
    
Specifically:
- Added X86AVXABILevel member to CGFunctionInfo.
- Populated X86AVXABILevel member in CGFunctionInfo objects via
arrangement methods declared in CodeGenTypes.h.
- Respect CGFunctionInfo AVX Level in X86_64ABIInfo::computeInfo.
- Add/extend regression tests for:
- free-function target-attribute AVX ABI lowering
- C++ method/ctor target-attribute AVX ABI lowering
- PS4/PS5 legacy ABI behavior (no per-function AVX ABI change)

---------

Co-authored-by: Aaron Ballman <aaron@aaronballman.com>

Author: freaknbigpanda

clang/docs/ReleaseNotes.md

@@ -50,6 +50,13 @@ latest release, please see the [Clang Web Site](https://clang.llvm.org) or the
 
 ### ABI Changes in This Version
 
+- Except on PlayStation, Clang now derives the x86-64 System V AVX ABI level
+for 256- and 512-bit vector arguments and returns from effective per-function
+target features. Features and `arch=` CPUs that imply AVX or AVX512F are
+honored, and calls use the caller's features, matching GCC. Per-function
+features cannot lower the translation-unit ABI level;
+`-fclang-abi-compat=23` restores the previous behavior. (#GH193298)
+
 ### AST Dumping Potentially Breaking Changes
 
 ### Clang Frontend Potentially Breaking Changes

clang/include/clang/Basic/ABIVersions.def

@@ -140,6 +140,12 @@ ABI_VER_MAJOR(21)
 /// local classes as nested-name entities instead of local-name entities.
 ABI_VER_MAJOR(22)
 
+/// Attempt to be ABI-compatible with code generated by Clang 23.0.x.
+/// This causes clang to:
+///   - Ignore per-function target attributes when determining the x86 AVX ABI
+///     level.
+ABI_VER_MAJOR(23)
+
 /// Conform to the underlying platform's C and C++ ABIs as closely as we can.
 ABI_VER_LATEST(Latest)
 

clang/include/clang/Basic/LangOptions.def

@@ -426,7 +426,7 @@ LANGOPT(ForceEmitVTables, 1, 0, NotCompatible, "whether to emit all vtables")
 LANGOPT(AllowEditorPlaceholders, 1, 0, Benign,
         "allow editor placeholders in source")
 
-ENUM_LANGOPT(ClangABICompat, ClangABI, 4, ClangABI::Latest, NotCompatible,
+ENUM_LANGOPT(ClangABICompat, ClangABI, 5, ClangABI::Latest, NotCompatible,
              "version of Clang that we should attempt to be ABI-compatible "
              "with")
 

clang/include/clang/CodeGen/CGFunctionInfo.h

@@ -17,10 +17,9 @@
 
 #include "clang/AST/CanonicalType.h"
 #include "clang/AST/CharUnits.h"
-#include "clang/AST/Decl.h"
 #include "clang/AST/Type.h"
-#include "llvm/IR/DerivedTypes.h"
 #include "llvm/ADT/FoldingSet.h"
+#include "llvm/IR/DerivedTypes.h"
 #include "llvm/Support/TrailingObjects.h"
 #include <cassert>
 
@@ -653,6 +652,10 @@ class CGFunctionInfo final
   /// Log 2 of the maximum vector width.
   unsigned MaxVectorWidth : 4;
 
+  /// X86 AVX Level, can be different from global / module level AVX level
+  /// because of target attributes.
+  unsigned X86ABIAVXLevel = 0;
+
   RequiredArgs Required;
 
   /// The struct representing all arguments passed in memory.  Only used when
@@ -683,7 +686,8 @@ class CGFunctionInfo final
 public:
   static CGFunctionInfo *
   create(unsigned llvmCC, bool instanceMethod, bool chainCall,
-         bool delegateCall, const FunctionType::ExtInfo &extInfo,
+         bool delegateCall, unsigned X86ABIAVXLevel,
+         const FunctionType::ExtInfo &extInfo,
          ArrayRef<ExtParameterInfo> paramInfos, CanQualType resultType,
          ArrayRef<CanQualType> argTypes, RequiredArgs required);
   void operator delete(void *p) { ::operator delete(p); }
@@ -809,6 +813,8 @@ class CGFunctionInfo final
     MaxVectorWidth = llvm::countr_zero(Width) + 1;
   }
 
+  unsigned getX86ABIAVXLevel() const { return X86ABIAVXLevel; }
+
   void Profile(llvm::FoldingSetNodeID &ID) {
     ID.AddInteger(getASTCallingConvention());
     ID.AddBoolean(InstanceMethod);
@@ -821,6 +827,7 @@ class CGFunctionInfo final
     ID.AddInteger(RegParm);
     ID.AddBoolean(NoCfCheck);
     ID.AddBoolean(CmseNSCall);
+    ID.AddInteger(X86ABIAVXLevel);
     ID.AddInteger(Required.getOpaqueData());
     ID.AddBoolean(HasExtParameterInfos);
     if (HasExtParameterInfos) {
@@ -833,6 +840,7 @@ class CGFunctionInfo final
   }
   static void Profile(llvm::FoldingSetNodeID &ID, bool InstanceMethod,
                       bool ChainCall, bool IsDelegateCall,
+                      unsigned X86ABIAVXLevel,
                       const FunctionType::ExtInfo &info,
                       ArrayRef<ExtParameterInfo> paramInfos,
                       RequiredArgs required, CanQualType resultType,
@@ -848,6 +856,7 @@ class CGFunctionInfo final
     ID.AddInteger(info.getRegParm());
     ID.AddBoolean(info.getNoCfCheck());
     ID.AddBoolean(info.getCmseNSCall());
+    ID.AddInteger(X86ABIAVXLevel);
     ID.AddInteger(required.getOpaqueData());
     ID.AddBoolean(!paramInfos.empty());
     if (!paramInfos.empty()) {

clang/include/clang/CodeGen/CodeGenABITypes.h

@@ -81,20 +81,22 @@ const CGFunctionInfo &
 arrangeCXXMethodCall(CodeGenModule &CGM, CanQualType returnType,
                      ArrayRef<CanQualType> argTypes, FunctionType::ExtInfo info,
                      ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos,
-                     RequiredArgs args);
+                     RequiredArgs args, const FunctionDecl *CallerFD);
 
 const CGFunctionInfo &arrangeFreeFunctionCall(
     CodeGenModule &CGM, CanQualType returnType, ArrayRef<CanQualType> argTypes,
     FunctionType::ExtInfo info,
-    ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos,
-    RequiredArgs args);
+    ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos, RequiredArgs args,
+    const FunctionDecl *CallerFD);
 
 // An overload with an empty `paramInfos`
 inline const CGFunctionInfo &
 arrangeFreeFunctionCall(CodeGenModule &CGM, CanQualType returnType,
                         ArrayRef<CanQualType> argTypes,
-                        FunctionType::ExtInfo info, RequiredArgs args) {
-  return arrangeFreeFunctionCall(CGM, returnType, argTypes, info, {}, args);
+                        FunctionType::ExtInfo info, RequiredArgs args,
+                        const FunctionDecl *CallerFD) {
+  return arrangeFreeFunctionCall(CGM, returnType, argTypes, info, {}, args,
+                                 CallerFD);
 }
 
 /// Returns the implicit arguments to add to a complete, non-delegating C++

clang/lib/CodeGen/ABIInfo.h

@@ -26,6 +26,7 @@ class FixedVectorType;
 namespace clang {
 class ASTContext;
 class CodeGenOptions;
+class FunctionDecl;
 class TargetInfo;
 
 namespace CodeGen {
@@ -69,6 +70,14 @@ class ABIInfo {
   /// functions.
   llvm::CallingConv::ID getRuntimeCC() const { return RuntimeCC; }
 
+  // Get X86ABIAVXLevel for the given FunctionDecl and ExtInfo.
+  // This can be different than the global / module level X86ABIAVXLevel
+  // due to function attributes.
+  virtual unsigned getX86ABIAVXLevel(const FunctionDecl *,
+                                     const FunctionType::ExtInfo &) const {
+    return 0;
+  }
+
   virtual void computeInfo(CodeGen::CGFunctionInfo &FI) const = 0;
 
   /// EmitVAArg - Emit the target dependent code to load a value of

clang/lib/CodeGen/CGCUDARuntime.cpp

@@ -53,7 +53,8 @@ static llvm::Value *emitGetParamBuf(CodeGenFunction &CGF,
   Args.add(RValue::get(CGF.CGM.getSize(Offset)),
            CGF.getContext().getSizeType());
   const CGFunctionInfo &CallInfo = CGF.CGM.getTypes().arrangeFreeFunctionCall(
-      Args, GetParamBufProto, /*ChainCall=*/false);
+      Args, GetParamBufProto, /*ChainCall=*/false,
+      CGF.getCurrentFunctionDecl());
   auto Ret = CGF.EmitCall(CallInfo, Callee, /*ReturnValue=*/{}, Args);
 
   return Ret.getScalarVal();
@@ -116,7 +117,7 @@ RValue CGCUDARuntime::EmitCUDADeviceKernelCallExpr(
   LaunchCallArgs[1] = CallArg(RValue::get(Config), CGM.getContext().VoidPtrTy);
   const CGFunctionInfo &LaunchCallInfo = CGM.getTypes().arrangeFreeFunctionCall(
       LaunchCallArgs, LaunchCalleeFuncTy->getAs<FunctionProtoType>(),
-      /*ChainCall=*/false);
+      /*ChainCall=*/false, CGF.getCurrentFunctionDecl());
   CGF.EmitCall(LaunchCallInfo, LaunchCallee, ReturnValue, LaunchCallArgs,
                CallOrInvoke,
                /*IsMustTail=*/false, E->getExprLoc());