[CodeGen] Explicitly widen non-power-of-2 vector HFA base types.

[ab]

Summary

Currently, for the ppc64--gnu and aarch64 ABIs, we recognize:

typedef __attribute__((__ext_vector_type__(3))) float v3f32;
typedef __attribute__((__ext_vector_type__(16))) char v16i8;

struct HFA {
  v3f32 a;
  v16i8 b;
};

as an HFA. Since the base type is the first type encountered, we pass the HFA as:

[2 x <3 x float>]

Which leads to incorrect IR (relying on padding values) when the second field is used.

This patch teaches isHomogeneousAggregate to explicitly uses the widened vector (after size rounding).

I don't know if this approach is either sufficient or ideal; alternatives I considered:

• pass as [2 x <3 x float>], but bitcast the struct pointer itself to [2 x <4 x float>]*; this still relies on accessing padding.
• return an llvm::Type (instead of QualType) in isHomogeneousAggregate: there are a couple callers that rely on size/alignment info which should come from the clang::Type.
• reject non-power-of-2 vectors as HFA base types: this doesn't make sense, as we treat them everywhere else as the next-power-of-2 type.
• fixup the base type in the various isHomogeneousAggregate callers: repetitive changes all over the place.

[uweigand]

Makes sense to me.

[ab]

Thanks Ulrich. Tim?

[t.p.northover]

Sorry, I thought I'd already commented here. It looked good to me too.

Tim.

lib/CodeGen/TargetInfo.cpp

if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) {
     // Most ABIs only support float, double, and some vector type widths.
     if (!isHomogeneousAggregateBaseType(Ty))
       return false;
 
     // The base type must be the same for all members.  Types that
     // agree in both total size and mode (float vs. vector) are
     // treated as being equivalent here.
     const Type *TyPtr = Ty.getTypePtr();
-    if (!Base)
+    if (!Base) {
       Base = TyPtr;
+      // If it's a non-power-of-2 vector, its size is already a power-of-2,
+      // so make sure to widen it explicitly.
+      if (const VectorType *VT = Base->getAs<VectorType>()) {
+        QualType EltTy = VT->getElementType();
+        unsigned NumElements =
+            getContext().getTypeSize(VT) / getContext().getTypeSize(EltTy);
+        Base = getContext()
+                   .getVectorType(EltTy, NumElements, VT->getVectorKind())
+                   .getTypePtr();
+      }
+    }
 
     if (Base->isVectorType() != TyPtr->isVectorType() ||
         getContext().getTypeSize(Base) != getContext().getTypeSize(TyPtr))
       return false;
   }
   return Members > 0 && isHomogeneousAggregateSmallEnough(Base, Members);
 }
 

test/CodeGen/aarch64-arguments-hfa-v3.c

+// RUN: %clang_cc1 -triple arm64-apple-ios -target-feature +neon -target-abi darwinpcs -fallow-half-arguments-and-returns -emit-llvm -o - %s | FileCheck %s
+
+typedef __attribute__((__ext_vector_type__(16))) signed char int8x16_t;
+typedef __attribute__((__ext_vector_type__(3))) float float32x3_t;
+
+// CHECK: %struct.HFAv3 = type { [4 x <3 x float>] }
+typedef struct { float32x3_t arr[4]; } HFAv3;
+
+// CHECK: %struct.MixedHFAv3 = type { [3 x <3 x float>], <16 x i8> }
+typedef struct { float32x3_t arr[3]; int8x16_t b; } MixedHFAv3;
+
+// CHECK: define %struct.HFAv3 @test([4 x <4 x float>] %{{.*}}, [4 x <4 x float>] %{{.*}}, [4 x <4 x float>] %{{.*}})
+HFAv3 test(HFAv3 a0, HFAv3 a1, HFAv3 a2) {
+  return a2;
+}
+
+// CHECK: define %struct.MixedHFAv3 @test_mixed([4 x <4 x float>] %{{.*}}, [4 x <4 x float>] %{{.*}}, [4 x <4 x float>] %{{.*}})
+MixedHFAv3 test_mixed(MixedHFAv3 a0, MixedHFAv3 a1, MixedHFAv3 a2) {
+  return a2;
+}

test/CodeGen/arm64-arguments.c

 
 // CHECK: [[HVAPTRPTR:%.*]] = bitcast i8* [[CURLIST]] to %struct.TooBigHVA**
 // CHECK: [[HVAPTR:%.*]] = load %struct.TooBigHVA*, %struct.TooBigHVA** [[HVAPTRPTR]]
   __builtin_va_list thelist;
   __builtin_va_start(thelist, n);
   struct TooBigHVA h = __builtin_va_arg(thelist, struct TooBigHVA);
   return h.d;
 }
+
+typedef __attribute__((__ext_vector_type__(3))) float float32x3_t;
+typedef struct { float32x3_t arr[4]; } HFAv3;
+
+float32x3_t test_hva_v3(int n, ...) {
+// CHECK-LABEL: define <3 x float> @test_hva_v3(i32 %n, ...)
+// CHECK: [[THELIST:%.*]] = alloca i8*
+// CHECK: [[CURLIST:%.*]] = load i8*, i8** [[THELIST]]
+
+  // HVA is not indirect, so occupies its full 16 bytes on the stack. but it
+  // must be properly aligned.
+// CHECK: [[ALIGN0:%.*]] = ptrtoint i8* [[CURLIST]] to i64
+// CHECK: [[ALIGN1:%.*]] = add i64 [[ALIGN0]], 15
+// CHECK: [[ALIGN2:%.*]] = and i64 [[ALIGN1]], -16
+// CHECK: [[ALIGNED_LIST:%.*]] = inttoptr i64 [[ALIGN2]] to i8*
+
+// CHECK: [[NEXTLIST:%.*]] = getelementptr inbounds i8, i8* [[ALIGNED_LIST]], i64 64
+// CHECK: store i8* [[NEXTLIST]], i8** [[THELIST]]
+
+// CHECK: bitcast i8* [[ALIGNED_LIST]] to %struct.HFAv3*
+  __builtin_va_list l;
+  __builtin_va_start(l, n);
+  HFAv3 r = __builtin_va_arg(l, HFAv3);
+  return r.arr[2];
+}
+
+float32x3_t test_hva_v3_call(HFAv3 *a) {
+// CHECK-LABEL: define <3 x float> @test_hva_v3_call(%struct.HFAv3* %a)
+// CHECK: call <3 x float> (i32, ...) @test_hva_v3(i32 1, [4 x <4 x float>] {{.*}})
+  return test_hva_v3(1, *a);
+}

test/CodeGen/ppc64le-aggregates.c

 struct v3f6 { float3 v[6]; };
 struct v3f7 { float3 v[7]; };
 struct v3f8 { float3 v[8]; };
 struct v3f9 { float3 v[9]; };
 
 struct v3fab { float3 a; float3 b; };
 struct v3fabc { float3 a; float3 b; float3 c; };
 
-// CHECK: define [1 x <3 x float>] @func_v3f1(<3 x float> inreg %x.coerce)
+// CHECK: define [1 x <4 x float>] @func_v3f1(<3 x float> inreg %x.coerce)
 struct v3f1 func_v3f1(struct v3f1 x) { return x; }
 
-// CHECK: define [2 x <3 x float>] @func_v3f2([2 x <3 x float>] %x.coerce)
+// CHECK: define [2 x <4 x float>] @func_v3f2([2 x <4 x float>] %x.coerce)
 struct v3f2 func_v3f2(struct v3f2 x) { return x; }
 
-// CHECK: define [3 x <3 x float>] @func_v3f3([3 x <3 x float>] %x.coerce)
+// CHECK: define [3 x <4 x float>] @func_v3f3([3 x <4 x float>] %x.coerce)
 struct v3f3 func_v3f3(struct v3f3 x) { return x; }
 
-// CHECK: define [4 x <3 x float>] @func_v3f4([4 x <3 x float>] %x.coerce)
+// CHECK: define [4 x <4 x float>] @func_v3f4([4 x <4 x float>] %x.coerce)
 struct v3f4 func_v3f4(struct v3f4 x) { return x; }
 
-// CHECK: define [5 x <3 x float>] @func_v3f5([5 x <3 x float>] %x.coerce)
+// CHECK: define [5 x <4 x float>] @func_v3f5([5 x <4 x float>] %x.coerce)
 struct v3f5 func_v3f5(struct v3f5 x) { return x; }
 
-// CHECK: define [6 x <3 x float>] @func_v3f6([6 x <3 x float>] %x.coerce)
+// CHECK: define [6 x <4 x float>] @func_v3f6([6 x <4 x float>] %x.coerce)
 struct v3f6 func_v3f6(struct v3f6 x) { return x; }
 
-// CHECK: define [7 x <3 x float>] @func_v3f7([7 x <3 x float>] %x.coerce)
+// CHECK: define [7 x <4 x float>] @func_v3f7([7 x <4 x float>] %x.coerce)
 struct v3f7 func_v3f7(struct v3f7 x) { return x; }
 
-// CHECK: define [8 x <3 x float>] @func_v3f8([8 x <3 x float>] %x.coerce)
+// CHECK: define [8 x <4 x float>] @func_v3f8([8 x <4 x float>] %x.coerce)
 struct v3f8 func_v3f8(struct v3f8 x) { return x; }
 
 // CHECK: define void @func_v3f9(%struct.v3f9* noalias sret %agg.result, %struct.v3f9* byval align 16 %x)
 struct v3f9 func_v3f9(struct v3f9 x) { return x; }
 
-// CHECK: define [2 x <3 x float>] @func_v3fab([2 x <3 x float>] %x.coerce)
+// CHECK: define [2 x <4 x float>] @func_v3fab([2 x <4 x float>] %x.coerce)
 struct v3fab func_v3fab(struct v3fab x) { return x; }
 
-// CHECK: define [3 x <3 x float>] @func_v3fabc([3 x <3 x float>] %x.coerce)
+// CHECK: define [3 x <4 x float>] @func_v3fabc([3 x <4 x float>] %x.coerce)
 struct v3fabc func_v3fabc(struct v3fabc x) { return x; }
 
 // CHECK-LABEL: @call_v3f1
 // CHECK: %[[TMP:[^ ]+]] = load <3 x float>, <3 x float>* getelementptr inbounds (%struct.v3f1, %struct.v3f1* @global_v3f1, i32 0, i32 0, i32 0), align 1
-// CHECK: call [1 x <3 x float>] @func_v3f1(<3 x float> inreg %[[TMP]])
+// CHECK: call [1 x <4 x float>] @func_v3f1(<3 x float> inreg %[[TMP]])
 struct v3f1 global_v3f1;
 void call_v3f1(void) { global_v3f1 = func_v3f1(global_v3f1); }
 
 // CHECK-LABEL: @call_v3f2
-// CHECK: %[[TMP:[^ ]+]] = load [2 x <3 x float>], [2 x <3 x float>]* getelementptr inbounds (%struct.v3f2, %struct.v3f2* @global_v3f2, i32 0, i32 0), align 1
-// CHECK: call [2 x <3 x float>] @func_v3f2([2 x <3 x float>] %[[TMP]])
+// CHECK: %[[TMP:[^ ]+]] = load [2 x <4 x float>], [2 x <4 x float>]* bitcast (%struct.v3f2* @global_v3f2 to [2 x <4 x float>]*), align 16
+// CHECK: call [2 x <4 x float>] @func_v3f2([2 x <4 x float>] %[[TMP]])
 struct v3f2 global_v3f2;
 void call_v3f2(void) { global_v3f2 = func_v3f2(global_v3f2); }
 
 // CHECK-LABEL: @call_v3f3
-// CHECK: %[[TMP:[^ ]+]] = load [3 x <3 x float>], [3 x <3 x float>]* getelementptr inbounds (%struct.v3f3, %struct.v3f3* @global_v3f3, i32 0, i32 0), align 1
-// CHECK: call [3 x <3 x float>] @func_v3f3([3 x <3 x float>] %[[TMP]])
+// CHECK: %[[TMP:[^ ]+]] = load [3 x <4 x float>], [3 x <4 x float>]* bitcast (%struct.v3f3* @global_v3f3 to [3 x <4 x float>]*), align 16
+// CHECK: call [3 x <4 x float>] @func_v3f3([3 x <4 x float>] %[[TMP]])
 struct v3f3 global_v3f3;
 void call_v3f3(void) { global_v3f3 = func_v3f3(global_v3f3); }
 
 // CHECK-LABEL: @call_v3f4
-// CHECK: %[[TMP:[^ ]+]] = load [4 x <3 x float>], [4 x <3 x float>]* getelementptr inbounds (%struct.v3f4, %struct.v3f4* @global_v3f4, i32 0, i32 0), align 1
-// CHECK: call [4 x <3 x float>] @func_v3f4([4 x <3 x float>] %[[TMP]])
+// CHECK: %[[TMP:[^ ]+]] = load [4 x <4 x float>], [4 x <4 x float>]* bitcast (%struct.v3f4* @global_v3f4 to [4 x <4 x float>]*), align 16
+// CHECK: call [4 x <4 x float>] @func_v3f4([4 x <4 x float>] %[[TMP]])
 struct v3f4 global_v3f4;
 void call_v3f4(void) { global_v3f4 = func_v3f4(global_v3f4); }
 
 // CHECK-LABEL: @call_v3f5
-// CHECK: %[[TMP:[^ ]+]] = load [5 x <3 x float>], [5 x <3 x float>]* getelementptr inbounds (%struct.v3f5, %struct.v3f5* @global_v3f5, i32 0, i32 0), align 1
-// CHECK: call [5 x <3 x float>] @func_v3f5([5 x <3 x float>] %[[TMP]])
+// CHECK: %[[TMP:[^ ]+]] = load [5 x <4 x float>], [5 x <4 x float>]* bitcast (%struct.v3f5* @global_v3f5 to [5 x <4 x float>]*), align 16
+// CHECK: call [5 x <4 x float>] @func_v3f5([5 x <4 x float>] %[[TMP]])
 struct v3f5 global_v3f5;
 void call_v3f5(void) { global_v3f5 = func_v3f5(global_v3f5); }
 
 // CHECK-LABEL: @call_v3f6
-// CHECK: %[[TMP:[^ ]+]] = load [6 x <3 x float>], [6 x <3 x float>]* getelementptr inbounds (%struct.v3f6, %struct.v3f6* @global_v3f6, i32 0, i32 0), align 1
-// CHECK: call [6 x <3 x float>] @func_v3f6([6 x <3 x float>] %[[TMP]])
+// CHECK: %[[TMP:[^ ]+]] = load [6 x <4 x float>], [6 x <4 x float>]* bitcast (%struct.v3f6* @global_v3f6 to [6 x <4 x float>]*), align 16
+// CHECK: call [6 x <4 x float>] @func_v3f6([6 x <4 x float>] %[[TMP]])
 struct v3f6 global_v3f6;
 void call_v3f6(void) { global_v3f6 = func_v3f6(global_v3f6); }
 
 // CHECK-LABEL: @call_v3f7
-// CHECK: %[[TMP:[^ ]+]] = load [7 x <3 x float>], [7 x <3 x float>]* getelementptr inbounds (%struct.v3f7, %struct.v3f7* @global_v3f7, i32 0, i32 0), align 1
-// CHECK: call [7 x <3 x float>] @func_v3f7([7 x <3 x float>] %[[TMP]])
+// CHECK: %[[TMP:[^ ]+]] = load [7 x <4 x float>], [7 x <4 x float>]* bitcast (%struct.v3f7* @global_v3f7 to [7 x <4 x float>]*), align 16
+// CHECK: call [7 x <4 x float>] @func_v3f7([7 x <4 x float>] %[[TMP]])
 struct v3f7 global_v3f7;
 void call_v3f7(void) { global_v3f7 = func_v3f7(global_v3f7); }
 
 // CHECK-LABEL: @call_v3f8
-// CHECK: %[[TMP:[^ ]+]] = load [8 x <3 x float>], [8 x <3 x float>]* getelementptr inbounds (%struct.v3f8, %struct.v3f8* @global_v3f8, i32 0, i32 0), align 1
-// CHECK: call [8 x <3 x float>] @func_v3f8([8 x <3 x float>] %[[TMP]])
+// CHECK: %[[TMP:[^ ]+]] = load [8 x <4 x float>], [8 x <4 x float>]* bitcast (%struct.v3f8* @global_v3f8 to [8 x <4 x float>]*), align 16
+// CHECK: call [8 x <4 x float>] @func_v3f8([8 x <4 x float>] %[[TMP]])
 struct v3f8 global_v3f8;
 void call_v3f8(void) { global_v3f8 = func_v3f8(global_v3f8); }
 
 // CHECK-LABEL: @call_v3f9
 // CHECK: call void @func_v3f9(%struct.v3f9* sret %{{[^ ]+}}, %struct.v3f9* byval align 16 @global_v3f9)
 struct v3f9 global_v3f9;
 void call_v3f9(void) { global_v3f9 = func_v3f9(global_v3f9); }
 
 // CHECK-LABEL: @call_v3fab
-// CHECK: %[[TMP:[^ ]+]] = load [2 x <3 x float>], [2 x <3 x float>]* bitcast (%struct.v3fab* @global_v3fab to [2 x <3 x float>]*)
-// CHECK: call [2 x <3 x float>] @func_v3fab([2 x <3 x float>] %[[TMP]])
+// CHECK: %[[TMP:[^ ]+]] = load [2 x <4 x float>], [2 x <4 x float>]* bitcast (%struct.v3fab* @global_v3fab to [2 x <4 x float>]*), align 16
+// CHECK: call [2 x <4 x float>] @func_v3fab([2 x <4 x float>] %[[TMP]])
 struct v3fab global_v3fab;
 void call_v3fab(void) { global_v3fab = func_v3fab(global_v3fab); }
 
 // CHECK-LABEL: @call_v3fabc
-// CHECK: %[[TMP:[^ ]+]] = load [3 x <3 x float>], [3 x <3 x float>]* bitcast (%struct.v3fabc* @global_v3fabc to [3 x <3 x float>]*)
-// CHECK: call [3 x <3 x float>] @func_v3fabc([3 x <3 x float>] %[[TMP]])
+// CHECK: %[[TMP:[^ ]+]] = load [3 x <4 x float>], [3 x <4 x float>]* bitcast (%struct.v3fabc* @global_v3fabc to [3 x <4 x float>]*), align 16
+// CHECK: call [3 x <4 x float>] @func_v3fabc([3 x <4 x float>] %[[TMP]])
 struct v3fabc global_v3fabc;
 void call_v3fabc(void) { global_v3fabc = func_v3fabc(global_v3fabc); }
 
 
 // Test returning small aggregates.
 
 struct s1 { char c[1]; };
 struct s2 { char c[2]; };