[CodeGen][ARM] Fix FP16 vector coercion

[miyuki]

Summary

When a function argument or return type is a homogeneous aggregate
which contains an FP16 vector but the target does not support FP16
operations natively, the type must be converted into an array of
integer vectors by then front end (otherwise LLVM will handle FP16
vectors incorrectly by scalarizing them and promoting FP16 to float,
see https://reviews.llvm.org/D50507).

Currently the logic for checking whether or not a given homogeneous
aggregate contains FP16 vectors is incorrect: it only looks at the
type of the first vector.

This patch fixes the issue by adding a new method
ARMABIInfo::containsAnyFP16Vectors and using it. The traversal logic
of this method is largely the same as in
ABIInfo::isHomogeneousAggregate.

[ostannard]

This doesn't look like the right pace to fix this - the backend can handle vectors of i8 and i16, which are also not legal types, so why can't it correctly handle vectors of f16?

[miyuki]

In D63437#1546312, @ostannard wrote:

This doesn't look like the right pace to fix this - the backend can handle vectors of i8 and i16, which are also not legal types, so why can't it correctly handle vectors of f16?

IIRC, we decided to handle this in the frontend for consistency with scalar fp16 handling: https://reviews.llvm.org/D49987

[ostannard]

Ah, right. In that case, this LGTM.

cfe/trunk/lib/CodeGen/TargetInfo.cpp

private:
   ABIArgInfo classifyReturnType(QualType RetTy, bool isVariadic,
                                 unsigned functionCallConv) const;
   ABIArgInfo classifyArgumentType(QualType RetTy, bool isVariadic,
                                   unsigned functionCallConv) const;
   ABIArgInfo classifyHomogeneousAggregate(QualType Ty, const Type *Base,
                                           uint64_t Members) const;
   ABIArgInfo coerceIllegalVector(QualType Ty) const;
   bool isIllegalVectorType(QualType Ty) const;
+  bool containsAnyFP16Vectors(QualType Ty) const;
 
   bool isHomogeneousAggregateBaseType(QualType Ty) const override;
   bool isHomogeneousAggregateSmallEnough(const Type *Ty,
                                          uint64_t Members) const override;
 
   bool isEffectivelyAAPCS_VFP(unsigned callConvention, bool acceptHalf) const;
 
   void computeInfo(CGFunctionInfo &FI) const override;

 
 ABIArgInfo ARMABIInfo::classifyHomogeneousAggregate(QualType Ty,
                                                     const Type *Base,
                                                     uint64_t Members) const {
   assert(Base && "Base class should be set for homogeneous aggregate");
   // Base can be a floating-point or a vector.
   if (const VectorType *VT = Base->getAs<VectorType>()) {
     // FP16 vectors should be converted to integer vectors
-    if (!getTarget().hasLegalHalfType() &&
-        (VT->getElementType()->isFloat16Type() ||
-          VT->getElementType()->isHalfType())) {
+    if (!getTarget().hasLegalHalfType() && containsAnyFP16Vectors(Ty)) {
       uint64_t Size = getContext().getTypeSize(VT);
       llvm::Type *NewVecTy = llvm::VectorType::get(
           llvm::Type::getInt32Ty(getVMContext()), Size / 32);
       llvm::Type *Ty = llvm::ArrayType::get(NewVecTy, Members);
       return ABIArgInfo::getDirect(Ty, 0, nullptr, false);
     }
   }
   return ABIArgInfo::getDirect(nullptr, 0, nullptr, false);
if (isAndroid()) {
         return true;
       // Size should be greater than 32 bits.
       return Size <= 32;
     }
   }
   return false;
 }
 
+/// Return true if a type contains any 16-bit floating point vectors
+bool ARMABIInfo::containsAnyFP16Vectors(QualType Ty) const {
+  if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) {
+    uint64_t NElements = AT->getSize().getZExtValue();
+    if (NElements == 0)
+      return false;
+    return containsAnyFP16Vectors(AT->getElementType());
+  } else if (const RecordType *RT = Ty->getAs<RecordType>()) {
+    const RecordDecl *RD = RT->getDecl();
+
+    // If this is a C++ record, check the bases first.
+    if (const CXXRecordDecl *CXXRD = dyn_cast<CXXRecordDecl>(RD))
+      if (llvm::any_of(CXXRD->bases(), [this](const CXXBaseSpecifier &B) {
+            return containsAnyFP16Vectors(B.getType());
+          }))
+        return true;
+
+    if (llvm::any_of(RD->fields(), [this](FieldDecl *FD) {
+          return FD && containsAnyFP16Vectors(FD->getType());
+        }))
+      return true;
+
+    return false;
+  } else {
+    if (const VectorType *VT = Ty->getAs<VectorType>())
+      return (VT->getElementType()->isFloat16Type() ||
+              VT->getElementType()->isHalfType());
+    return false;
+  }
+}
+
 bool ARMABIInfo::isLegalVectorTypeForSwift(CharUnits vectorSize,
                                            llvm::Type *eltTy,
                                            unsigned numElts) const {
   if (!llvm::isPowerOf2_32(numElts))
     return false;
   unsigned size = getDataLayout().getTypeStoreSizeInBits(eltTy);
   if (size > 64)
     return false;

cfe/trunk/test/CodeGen/arm-vfp16-arguments2.cpp

+// RUN: %clang_cc1 -triple armv7a--none-eabi -target-abi aapcs \
+// RUN:   -mfloat-abi soft -target-feature +neon -emit-llvm -o - -O1 %s \
+// RUN:   | FileCheck %s --check-prefix=CHECK-SOFT
+// RUN: %clang_cc1 -triple armv7a--none-eabi -target-abi aapcs \
+// RUN:   -mfloat-abi hard -target-feature +neon -emit-llvm -o - -O1 %s \
+// RUN:   | FileCheck %s --check-prefix=CHECK-HARD
+// RUN: %clang_cc1 -triple armv7a--none-eabi -target-abi aapcs \
+// RUN:   -mfloat-abi hard -target-feature +neon -target-feature +fullfp16 \
+// RUN:   -emit-llvm -o - -O1 %s \
+// RUN:   | FileCheck %s --check-prefix=CHECK-FULL
+
+typedef float float32_t;
+typedef __fp16 float16_t;
+typedef __attribute__((neon_vector_type(2))) float32_t float32x2_t;
+typedef __attribute__((neon_vector_type(4))) float16_t float16x4_t;
+
+struct S1 {
+  float32x2_t M1;
+  float16x4_t M2;
+};
+
+struct B1 { float32x2_t M; };
+struct B2 { float16x4_t M; };
+
+struct S2 : B1, B2 {};
+
+struct S3 : B1 {
+  float16x4_t M;
+};
+
+struct S4 : B1 {
+  B2 M[1];
+};
+
+// S5 does not contain any FP16 vectors
+struct S5 : B1 {
+  B1 M[1];
+};
+
+// CHECK-SOFT: define void @_Z2f12S1(%struct.S1* noalias nocapture sret %agg.result, [2 x i64] %s1.coerce)
+// CHECK-HARD: define arm_aapcs_vfpcc [2 x <2 x i32>] @_Z2f12S1([2 x <2 x i32>] returned %s1.coerce)
+// CHECK-FULL: define arm_aapcs_vfpcc %struct.S1 @_Z2f12S1(%struct.S1 returned %s1.coerce)
+struct S1 f1(struct S1 s1) { return s1; }
+
+// CHECK-SOFT: define void @_Z2f22S2(%struct.S2* noalias nocapture sret %agg.result, [4 x i32] %s2.coerce)
+// CHECK-HARD: define arm_aapcs_vfpcc [2 x <2 x i32>] @_Z2f22S2([2 x <2 x i32>] returned %s2.coerce)
+// CHECK-FULL: define arm_aapcs_vfpcc %struct.S2 @_Z2f22S2(%struct.S2 returned %s2.coerce)
+struct S2 f2(struct S2 s2) { return s2; }
+
+// CHECK-SOFT: define void @_Z2f32S3(%struct.S3* noalias nocapture sret %agg.result, [2 x i64] %s3.coerce)
+// CHECK-HARD: define arm_aapcs_vfpcc [2 x <2 x i32>] @_Z2f32S3([2 x <2 x i32>] returned %s3.coerce)
+// CHECK-FULL: define arm_aapcs_vfpcc %struct.S3 @_Z2f32S3(%struct.S3 returned %s3.coerce)
+struct S3 f3(struct S3 s3) { return s3; }
+
+// CHECK-SOFT: define void @_Z2f42S4(%struct.S4* noalias nocapture sret %agg.result, [2 x i64] %s4.coerce)
+// CHECK-HARD: define arm_aapcs_vfpcc [2 x <2 x i32>] @_Z2f42S4([2 x <2 x i32>] returned %s4.coerce)
+// CHECK-FULL: define arm_aapcs_vfpcc %struct.S4 @_Z2f42S4(%struct.S4 returned %s4.coerce)
+struct S4 f4(struct S4 s4) { return s4; }
+
+// CHECK-SOFT: define void @_Z2f52S5(%struct.S5* noalias nocapture sret %agg.result, [2 x i64] %s5.coerce)
+// CHECK-HARD: define arm_aapcs_vfpcc %struct.S5 @_Z2f52S5(%struct.S5 returned %s5.coerce)
+// CHECK-FULL: define arm_aapcs_vfpcc %struct.S5 @_Z2f52S5(%struct.S5 returned %s5.coerce)
+struct S5 f5(struct S5 s5) { return s5; }