[Clang][AArch64] Enable _Float16 _Complex type

[paulwalker-arm]

clang/include/clang/AST/ASTContext.h

1117

Is this necessary? I don't see any other *ComplexTy types in this file (i.e. there is no DoubleComplexTy).

clang/lib/AST/ASTContext.cpp

1330

As above.

[peterwaller-arm]

Some comments.

clang/include/clang/AST/ASTContext.h

1117

Is this necessary?

clang/lib/AST/ASTContext.cpp

1330

Is this necessary?

6811–6812

This can be left unmodified.

clang/test/CodeGen/aarch64-complex-half-math.c

1

Can these test outputs be autogenerated with update_cc_test_checks.py?

2

I think this test needs a // REQUIRES: aarch64-registered-target to protect builds which don't have a LLVM_TARGETS_TO_BUILD which implies AArch64 is available.

5

My read is that this appears to be testing both -ffast-math and _Float16. Do we want to be testing both these things at once like this, can anyone comment?

My preference would be to have these test only the operators, and not the flags, and then maybe have a single test which does
https://clang.llvm.org/docs/LanguageExtensions.html#extensions-to-specify-floating-point-flags

#pragma clang fp contract(fast)

and verifies that the flags are present.

[fhahn]

clang/test/CodeGen/aarch64-complex-half-math.c

1

Do the tests need to depend on LLVM optimizations? In general this is undesirable as it makes them more fragile if things change on the LLVM side.

[aaron.ballman]

clang/include/clang/AST/ASTContext.h

1117

This is the wrong way to handle this -- complex types are stored in the ComplexTypes folding set and are created through ASTContext::getComplexType(). So I don't think this change is necessary.

clang/lib/AST/ASTContext.cpp

1330

Same here.

[MattDevereau]

Removed -O1 and -ffast-math flags

[peterwaller-arm]

LGTM.

clang/test/CodeGen/aarch64-complex-half-math.c

5

Discussed off-review: concluded for now that a no-optimizations test with no fast math flags seems reasonable, to address @fhahn's comment.

[aaron.ballman]

clang/test/CodeGen/aarch64-complex-half-math.c

4

There's nothing AArch64-specific about the change, so I'd expect some more general codegen tests.

Actually, I took a look to see what calls getFloatingTypeOfSizeWithinDomain() and I can't find any callers of that. Do you have a test case where you were hitting that particular unreachable?

[peterwaller-arm]

clang/test/CodeGen/aarch64-complex-half-math.c

4

Thanks for your observation @aaron.ballman.

It appears this has been working since D105331, which enabled it for X86.

Unless I'm missing something, the function in question appears to have been unused since 2010 (the last use was removed in d005ac937e4c08e0c607ddc42708f8bf2064c243 by @rjmccall), I presume getFloatingTypeOfSizeWithinDomain should be removed then.

[aaron.ballman]

clang/test/CodeGen/aarch64-complex-half-math.c

4

I presume getFloatingTypeOfSizeWithinDomain should be removed then.

Agreed; I've gone ahead and removed it in 841355c1e4e35fc02b5b171419979f5f9af0ebc8.

I think this patch can be abandoned, unless you think the extra test coverage is worth adding as an NFC commit?

[MattDevereau]

Abandoning this patch as it is redundant due to changes in D105331. This patch also drew attention to unused code that has since been removed in 841355c1e4e35fc02b5b171419979f5f9af0ebc8. Many thanks to @aaron.ballman

clang/include/clang/AST/ASTContext.h

public:
   CanQualType SatShortAccumTy, SatAccumTy, SatLongAccumTy;
   CanQualType SatUnsignedShortAccumTy, SatUnsignedAccumTy,
       SatUnsignedLongAccumTy;
   CanQualType SatShortFractTy, SatFractTy, SatLongFractTy;
   CanQualType SatUnsignedShortFractTy, SatUnsignedFractTy,
       SatUnsignedLongFractTy;
   CanQualType HalfTy; // [OpenCL 6.1.1.1], ARM NEON
   CanQualType BFloat16Ty;
-  CanQualType Float16Ty; // C11 extension ISO/IEC TS 18661-3
+  CanQualType Float16Ty, Float16ComplexTy; // C11 extension ISO/IEC TS 18661-3
   CanQualType VoidPtrTy, NullPtrTy;
   CanQualType DependentTy, OverloadTy, BoundMemberTy, UnknownAnyTy;
   CanQualType BuiltinFnTy;
   CanQualType PseudoObjectTy, ARCUnbridgedCastTy;
   CanQualType ObjCBuiltinIdTy, ObjCBuiltinClassTy, ObjCBuiltinSelTy;
   CanQualType ObjCBuiltinBoolTy;
 #define IMAGE_TYPE(ImgType, Id, SingletonId, Access, Suffix) \
   CanQualType SingletonId;

clang/lib/AST/ASTContext.cpp

void ASTContext::InitBuiltinTypes(const TargetInfo &Target,
   // GNU extension, __float128 for IEEE quadruple precision
   InitBuiltinType(Float128Ty,          BuiltinType::Float128);
 
   // __ibm128 for IBM extended precision
   InitBuiltinType(Ibm128Ty, BuiltinType::Ibm128);
 
   // C11 extension ISO/IEC TS 18661-3
   InitBuiltinType(Float16Ty,           BuiltinType::Float16);
+  InitBuiltinType(Float16ComplexTy,    BuiltinType::Float16);
 
   // ISO/IEC JTC1 SC22 WG14 N1169 Extension
   InitBuiltinType(ShortAccumTy,            BuiltinType::ShortAccum);
   InitBuiltinType(AccumTy,                 BuiltinType::Accum);
   InitBuiltinType(LongAccumTy,             BuiltinType::LongAccum);
   InitBuiltinType(UnsignedShortAccumTy,    BuiltinType::UShortAccum);
   InitBuiltinType(UnsignedAccumTy,         BuiltinType::UAccum);
   InitBuiltinType(UnsignedLongAccumTy,     BuiltinType::ULongAccum);

 /// point or a complex type (based on typeDomain/typeSize).
 /// 'typeDomain' is a real floating point or complex type.
 /// 'typeSize' is a real floating point or complex type.
 QualType ASTContext::getFloatingTypeOfSizeWithinDomain(QualType Size,
                                                        QualType Domain) const {
   FloatingRank EltRank = getFloatingRank(Size);
   if (Domain->isComplexType()) {
     switch (EltRank) {
-    case BFloat16Rank: llvm_unreachable("Complex bfloat16 is not supported");
-    case Float16Rank:
+    case BFloat16Rank:
+      llvm_unreachable("Complex bfloat16 is not supported");
     case HalfRank: llvm_unreachable("Complex half is not supported");
+    case Float16Rank:    return getComplexType(Float16Ty);
     case Ibm128Rank:     return getComplexType(Ibm128Ty);
     case FloatRank:      return getComplexType(FloatTy);
     case DoubleRank:     return getComplexType(DoubleTy);
     case LongDoubleRank: return getComplexType(LongDoubleTy);
     case Float128Rank:   return getComplexType(Float128Ty);
     }
   }
 

clang/test/CodeGen/aarch64-complex-half-math.c

+// RUN: %clang_cc1 %s -O1 -emit-llvm -triple aarch64-unknown-unknown -ffast-math -o - | FileCheck %s --check-prefix=AARCH64
+
+_Float16 _Complex add_float_rr(_Float16 a, _Float16 b) {
+  // AARCH64-LABEL: @add_float_rr(
+  // AARCH64: fadd reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fadd
+  // AARCH64: ret { half, half }
+  return a + b;
+}
+_Float16 _Complex add_float_cr(_Float16 _Complex a, _Float16 b) {
+  // AARCH64-LABEL: @add_float_cr(
+  // AARCH64: fadd reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fadd
+  // AARCH64: ret { half, half }
+  return a + b;
+}
+_Float16 _Complex add_float_rc(_Float16 a, _Float16 _Complex b) {
+  // AARCH64-LABEL: @add_float_rc(
+  // AARCH64: fadd reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fadd
+  // AARCH64: ret { half, half }
+  return a + b;
+}
+_Float16 _Complex add_float_cc(_Float16 _Complex a, _Float16 _Complex b) {
+  // AARCH64-LABEL: @add_float_cc(
+  // AARCH64: fadd reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fadd reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fadd
+  // AARCH64: ret { half, half }
+  return a + b;
+}
+
+_Float16 _Complex sub_float_rr(_Float16 a, _Float16 b) {
+  // AARCH64-LABEL: @sub_float_rr(
+  // AARCH64: fsub reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fsub
+  // AARCH64: ret { half, half }
+  return a - b;
+}
+_Float16 _Complex sub_float_cr(_Float16 _Complex a, _Float16 b) {
+  // AARCH64-LABEL: @sub_float_cr(
+  // AARCH64: fsub reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fsub
+  // AARCH64: ret { half, half }
+  return a - b;
+}
+_Float16 _Complex sub_float_rc(_Float16 a, _Float16 _Complex b) {
+  // AARCH64-LABEL: @sub_float_rc(
+  // AARCH64: fsub reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fneg reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fsub
+  // AARCH64: ret { half, half }
+  return a - b;
+}
+_Float16 _Complex sub_float_cc(_Float16 _Complex a, _Float16 _Complex b) {
+  // AARCH64-LABEL: @sub_float_cc(
+  // AARCH64: fsub reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fsub reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fsub
+  // AARCH64: ret { half, half }
+  return a - b;
+}
+
+_Float16 _Complex mul_float_rr(_Float16 a, _Float16 b) {
+  // AARCH64-LABEL: @mul_float_rr(
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fmul
+  // AARCH64: ret { half, half }
+  return a * b;
+}
+_Float16 _Complex mul_float_cr(_Float16 _Complex a, _Float16 b) {
+  // AARCH64-LABEL: @mul_float_cr(
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fmul
+  // AARCH64: ret { half, half }
+  return a * b;
+}
+_Float16 _Complex mul_float_rc(_Float16 a, _Float16 _Complex b) {
+  // AARCH64-LABEL: @mul_float_rc(
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fmul
+  // AARCH64: ret { half, half }
+  return a * b;
+}
+_Float16 _Complex mul_float_cc(_Float16 _Complex a, _Float16 _Complex b) {
+  // AARCH64-LABEL: @mul_float_cc(
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fadd reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fsub reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fmul
+  // AARCH64: ret { half, half }
+  return a * b;
+}
+
+_Float16 _Complex div_float_rr(_Float16 a, _Float16 b) {
+  // AARCH64-LABEL: @div_float_rr(
+  // AARCH64: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64: ret { half, half }
+  return a / b;
+}
+_Float16 _Complex div_float_cr(_Float16 _Complex a, _Float16 b) {
+  // AARCH64-LABEL: @div_float_cr(
+  // AARCH64: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64: ret { half, half }
+  return a / b;
+}
+_Float16 _Complex div_float_rc(_Float16 a, _Float16 _Complex b) {
+  // AARCH64-LABEL: @div_float_rc(
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fadd reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fneg reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64: ret { half, half }
+  return a / b;
+}
+_Float16 _Complex div_float_cc(_Float16 _Complex a, _Float16 _Complex b) {
+  // AARCH64-LABEL: @div_float_cc(
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fadd reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fadd reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fmul reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fsub reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64-NOT: fdiv reassoc nnan ninf nsz arcp afn half
+  // AARCH64: ret { half, half }
+  return a / b;
+}