[HLSL] Move length support out of the DirectX Backend (#121611)

## Changes
- Delete DirectX length intrinsic
- Delete HLSL length lang builtin
- Implement length algorithm entirely in the header.

## History
- In the past if an HLSL intrinsic lowered to either a spirv op code or
a DXIL opcode we represented it with intrinsics

## Why we are moving away?
- To make HLSL apis more portable the team decided that it makes sense
for some intrinsics to be defined only in the header.
- Since there tends to be more SPIRV opcodes than DXIL opcodes the plan
is to support SPIRV opcodes either with target specific builtins or via
pattern matching.

Author: farzonl

clang/include/clang/Basic/Builtins.td

@@ -4865,12 +4865,6 @@ def HLSLIsinf : LangBuiltin<"HLSL_LANG"> {
   let Prototype = "void(...)";
 }
 
-def HLSLLength : LangBuiltin<"HLSL_LANG"> {
-  let Spellings = ["__builtin_hlsl_length"];
-  let Attributes = [NoThrow, Const];
-  let Prototype = "void(...)";
-}
-
 def HLSLLerp : LangBuiltin<"HLSL_LANG"> {
   let Spellings = ["__builtin_hlsl_lerp"];
   let Attributes = [NoThrow, Const];

clang/lib/CodeGen/CGBuiltin.cpp

@@ -19334,20 +19334,6 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
         /*ReturnType=*/X->getType(), CGM.getHLSLRuntime().getLerpIntrinsic(),
         ArrayRef<Value *>{X, Y, S}, nullptr, "hlsl.lerp");
   }
-  case Builtin::BI__builtin_hlsl_length: {
-    Value *X = EmitScalarExpr(E->getArg(0));
-
-    assert(E->getArg(0)->getType()->hasFloatingRepresentation() &&
-           "length operand must have a float representation");
-    // if the operand is a scalar, we can use the fabs llvm intrinsic directly
-    if (!E->getArg(0)->getType()->isVectorType())
-      return EmitFAbs(*this, X);
-
-    return Builder.CreateIntrinsic(
-        /*ReturnType=*/X->getType()->getScalarType(),
-        CGM.getHLSLRuntime().getLengthIntrinsic(), ArrayRef<Value *>{X},
-        nullptr, "hlsl.length");
-  }
   case Builtin::BI__builtin_hlsl_normalize: {
     Value *X = EmitScalarExpr(E->getArg(0));
 

clang/lib/CodeGen/CGHLSLRuntime.h

@@ -77,7 +77,6 @@ class CGHLSLRuntime {
   GENERATE_HLSL_INTRINSIC_FUNCTION(Cross, cross)
   GENERATE_HLSL_INTRINSIC_FUNCTION(Degrees, degrees)
   GENERATE_HLSL_INTRINSIC_FUNCTION(Frac, frac)
-  GENERATE_HLSL_INTRINSIC_FUNCTION(Length, length)
   GENERATE_HLSL_INTRINSIC_FUNCTION(Lerp, lerp)
   GENERATE_HLSL_INTRINSIC_FUNCTION(Normalize, normalize)
   GENERATE_HLSL_INTRINSIC_FUNCTION(Rsqrt, rsqrt)

clang/lib/Headers/hlsl/hlsl_detail.h

@@ -13,6 +13,14 @@ namespace hlsl {
 
 namespace __detail {
 
+template <typename T, typename U> struct is_same {
+  static const bool value = false;
+};
+
+template <typename T> struct is_same<T, T> {
+  static const bool value = true;
+};
+
 template <bool B, typename T> struct enable_if {};
 
 template <typename T> struct enable_if<true, T> {
@@ -33,6 +41,21 @@ constexpr enable_if_t<sizeof(U) == sizeof(T), U> bit_cast(T F) {
   return __builtin_bit_cast(U, F);
 }
 
+template <typename T>
+constexpr enable_if_t<is_same<float, T>::value || is_same<half, T>::value, T>
+length_impl(T X) {
+  return __builtin_elementwise_abs(X);
+}
+
+template <typename T, int N>
+enable_if_t<is_same<float, T>::value || is_same<half, T>::value, T>
+length_vec_impl(vector<T, N> X) {
+  vector<T, N> XSquared = X * X;
+  T XSquaredSum = XSquared[0];
+  [unroll] for (int i = 1; i < N; ++i) XSquaredSum += XSquared[i];
+  return __builtin_elementwise_sqrt(XSquaredSum);
+}
+
 } // namespace __detail
 } // namespace hlsl
 #endif //_HLSL_HLSL_DETAILS_H_

clang/lib/Headers/hlsl/hlsl_intrinsics.h

@@ -1297,27 +1297,16 @@ float4 lerp(float4, float4, float4);
 ///
 /// Length is based on the following formula: sqrt(x[0]^2 + x[1]^2 + ...).
 
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
-_HLSL_BUILTIN_ALIAS(__builtin_hlsl_length)
-half length(half);
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
-_HLSL_BUILTIN_ALIAS(__builtin_hlsl_length)
-half length(half2);
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
-_HLSL_BUILTIN_ALIAS(__builtin_hlsl_length)
-half length(half3);
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
-_HLSL_BUILTIN_ALIAS(__builtin_hlsl_length)
-half length(half4);
+const inline half length(half X) { return __detail::length_impl(X); }
+const inline float length(float X) { return __detail::length_impl(X); }
 
-_HLSL_BUILTIN_ALIAS(__builtin_hlsl_length)
-float length(float);
-_HLSL_BUILTIN_ALIAS(__builtin_hlsl_length)
-float length(float2);
-_HLSL_BUILTIN_ALIAS(__builtin_hlsl_length)
-float length(float3);
-_HLSL_BUILTIN_ALIAS(__builtin_hlsl_length)
-float length(float4);
+template <int N> const inline half length(vector<half, N> X) {
+  return __detail::length_vec_impl(X);
+}
+
+template <int N> const inline float length(vector<float, N> X) {
+  return __detail::length_vec_impl(X);
+}
 
 //===----------------------------------------------------------------------===//
 // log builtins

clang/lib/Sema/SemaHLSL.cpp

@@ -2112,24 +2112,6 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
       return true;
     break;
   }
-  case Builtin::BI__builtin_hlsl_length: {
-    if (CheckFloatOrHalfRepresentations(&SemaRef, TheCall))
-      return true;
-    if (SemaRef.checkArgCount(TheCall, 1))
-      return true;
-
-    ExprResult A = TheCall->getArg(0);
-    QualType ArgTyA = A.get()->getType();
-    QualType RetTy;
-
-    if (auto *VTy = ArgTyA->getAs<VectorType>())
-      RetTy = VTy->getElementType();
-    else
-      RetTy = TheCall->getArg(0)->getType();
-
-    TheCall->setType(RetTy);
-    break;
-  }
   case Builtin::BI__builtin_hlsl_mad: {
     if (SemaRef.checkArgCount(TheCall, 3))
       return true;

clang/test/CodeGenHLSL/builtins/length.hlsl

@@ -1,73 +1,130 @@
-// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
-// RUN:   dxil-pc-shadermodel6.3-library %s -fnative-half-type \
-// RUN:   -emit-llvm -disable-llvm-passes -o - | FileCheck %s \ 
-// RUN:   --check-prefixes=CHECK,NATIVE_HALF
-// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
-// RUN:   dxil-pc-shadermodel6.3-library %s -emit-llvm -disable-llvm-passes \
-// RUN:   -o - | FileCheck %s --check-prefixes=CHECK,NO_HALF
-
-// NATIVE_HALF: define noundef nofpclass(nan inf) half @
-// NATIVE_HALF: call reassoc nnan ninf nsz arcp afn half @llvm.fabs.f16(half
-// NO_HALF: call reassoc nnan ninf nsz arcp afn float @llvm.fabs.f32(float
-// NATIVE_HALF: ret half
-// NO_HALF: ret float
-half test_length_half(half p0)
-{
-  return length(p0);
-}
-// NATIVE_HALF: define noundef nofpclass(nan inf) half @
-// NATIVE_HALF: %hlsl.length = call reassoc nnan ninf nsz arcp afn half @llvm.dx.length.v2f16
-// NO_HALF: %hlsl.length = call reassoc nnan ninf nsz arcp afn float @llvm.dx.length.v2f32(
-// NATIVE_HALF: ret half %hlsl.length
-// NO_HALF: ret float %hlsl.length
-half test_length_half2(half2 p0)
-{
-  return length(p0);
-}
-// NATIVE_HALF: define noundef nofpclass(nan inf) half @
-// NATIVE_HALF: %hlsl.length = call reassoc nnan ninf nsz arcp afn half @llvm.dx.length.v3f16
-// NO_HALF: %hlsl.length = call reassoc nnan ninf nsz arcp afn float @llvm.dx.length.v3f32(
-// NATIVE_HALF: ret half %hlsl.length
-// NO_HALF: ret float %hlsl.length
-half test_length_half3(half3 p0)
-{
-  return length(p0);
-}
-// NATIVE_HALF: define noundef nofpclass(nan inf) half @
-// NATIVE_HALF: %hlsl.length = call reassoc nnan ninf nsz arcp afn half @llvm.dx.length.v4f16
-// NO_HALF: %hlsl.length = call reassoc nnan ninf nsz arcp afn float @llvm.dx.length.v4f32(
-// NATIVE_HALF: ret half %hlsl.length
-// NO_HALF: ret float %hlsl.length
-half test_length_half4(half4 p0)
-{
-  return length(p0);
-}
-
-// CHECK: define noundef nofpclass(nan inf) float @
-// CHECK: call reassoc nnan ninf nsz arcp afn float @llvm.fabs.f32(float
-// CHECK: ret float
-float test_length_float(float p0)
-{
-  return length(p0);
-}
-// CHECK: define noundef nofpclass(nan inf) float @
-// CHECK: %hlsl.length = call reassoc nnan ninf nsz arcp afn float @llvm.dx.length.v2f32(
-// CHECK: ret float %hlsl.length
-float test_length_float2(float2 p0)
-{
-  return length(p0);
-}
-// CHECK: define noundef nofpclass(nan inf) float @
-// CHECK: %hlsl.length = call reassoc nnan ninf nsz arcp afn float @llvm.dx.length.v3f32(
-// CHECK: ret float %hlsl.length
-float test_length_float3(float3 p0)
-{
-  return length(p0);
-}
-// CHECK: define noundef nofpclass(nan inf) float @
-// CHECK: %hlsl.length = call reassoc nnan ninf nsz arcp afn float @llvm.dx.length.v4f32(
-// CHECK: ret float %hlsl.length
-float test_length_float4(float4 p0)
-{
-  return length(p0);
-}
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: dxil-pc-shadermodel6.3-library %s -fnative-half-type \
+// RUN: -emit-llvm -O1 -o - | FileCheck %s
+
+// CHECK-LABEL: define noundef nofpclass(nan inf) half @_Z16test_length_halfDh(
+// CHECK-SAME: half noundef nofpclass(nan inf) [[P0:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[ELT_ABS_I:%.*]] = tail call reassoc nnan ninf nsz arcp afn noundef half @llvm.fabs.f16(half [[P0]])
+// CHECK-NEXT:    ret half [[ELT_ABS_I]]
+//
+half test_length_half(half p0)
+{
+  return length(p0);
+}
+
+// CHECK-LABEL: define noundef nofpclass(nan inf) half @_Z17test_length_half2Dv2_Dh(
+// CHECK-SAME: <2 x half> noundef nofpclass(nan inf) [[P0:%.*]]) local_unnamed_addr #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[MUL_I:%.*]] = fmul reassoc nnan ninf nsz arcp afn <2 x half> [[P0]], [[P0]]
+// CHECK-NEXT:    [[VECEXT_I:%.*]] = extractelement <2 x half> [[MUL_I]], i64 0
+// CHECK-NEXT:    [[VECEXT1_I:%.*]] = extractelement <2 x half> [[MUL_I]], i64 1
+// CHECK-NEXT:    [[ADD_I:%.*]] = fadd reassoc nnan ninf nsz arcp afn half [[VECEXT_I]], [[VECEXT1_I]]
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call reassoc nnan ninf nsz arcp afn noundef half @llvm.sqrt.f16(half [[ADD_I]])
+// CHECK-NEXT:    ret half [[TMP0]]
+//
+half test_length_half2(half2 p0)
+{
+  return length(p0);
+}
+
+// CHECK-LABEL: define noundef nofpclass(nan inf) half @_Z17test_length_half3Dv3_Dh(
+// CHECK-SAME: <3 x half> noundef nofpclass(nan inf) [[P0:%.*]]) local_unnamed_addr #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[MUL_I:%.*]] = fmul reassoc nnan ninf nsz arcp afn <3 x half> [[P0]], [[P0]]
+// CHECK-NEXT:    [[VECEXT_I:%.*]] = extractelement <3 x half> [[MUL_I]], i64 0
+// CHECK-NEXT:    [[VECEXT1_I:%.*]] = extractelement <3 x half> [[MUL_I]], i64 1
+// CHECK-NEXT:    [[ADD_I:%.*]] = fadd reassoc nnan ninf nsz arcp afn half [[VECEXT1_I]], [[VECEXT_I]]
+// CHECK-NEXT:    [[VECEXT1_I_1:%.*]] = extractelement <3 x half> [[MUL_I]], i64 2
+// CHECK-NEXT:    [[ADD_I_1:%.*]] = fadd reassoc nnan ninf nsz arcp afn half [[VECEXT1_I_1]], [[ADD_I]]
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call reassoc nnan ninf nsz arcp afn noundef half @llvm.sqrt.f16(half [[ADD_I_1]])
+// CHECK-NEXT:    ret half [[TMP0]]
+//
+half test_length_half3(half3 p0)
+{
+  return length(p0);
+}
+
+// CHECK-LABEL: define noundef nofpclass(nan inf) half @_Z17test_length_half4Dv4_Dh(
+// CHECK-SAME: <4 x half> noundef nofpclass(nan inf) [[P0:%.*]]) local_unnamed_addr #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[MUL_I:%.*]] = fmul reassoc nnan ninf nsz arcp afn <4 x half> [[P0]], [[P0]]
+// CHECK-NEXT:    [[VECEXT_I:%.*]] = extractelement <4 x half> [[MUL_I]], i64 0
+// CHECK-NEXT:    [[VECEXT1_I:%.*]] = extractelement <4 x half> [[MUL_I]], i64 1
+// CHECK-NEXT:    [[ADD_I:%.*]] = fadd reassoc nnan ninf nsz arcp afn half [[VECEXT1_I]], [[VECEXT_I]]
+// CHECK-NEXT:    [[VECEXT1_I_1:%.*]] = extractelement <4 x half> [[MUL_I]], i64 2
+// CHECK-NEXT:    [[ADD_I_1:%.*]] = fadd reassoc nnan ninf nsz arcp afn half [[VECEXT1_I_1]], [[ADD_I]]
+// CHECK-NEXT:    [[VECEXT1_I_2:%.*]] = extractelement <4 x half> [[MUL_I]], i64 3
+// CHECK-NEXT:    [[ADD_I_2:%.*]] = fadd reassoc nnan ninf nsz arcp afn half [[VECEXT1_I_2]], [[ADD_I_1]]
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call reassoc nnan ninf nsz arcp afn noundef half @llvm.sqrt.f16(half [[ADD_I_2]])
+// CHECK-NEXT:    ret half [[TMP0]]
+//
+half test_length_half4(half4 p0)
+{
+  return length(p0);
+}
+
+
+// CHECK-LABEL: define noundef nofpclass(nan inf) float @_Z17test_length_floatf(
+// CHECK-SAME: float noundef nofpclass(nan inf) [[P0:%.*]]) local_unnamed_addr #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[ELT_ABS_I:%.*]] = tail call reassoc nnan ninf nsz arcp afn noundef float @llvm.fabs.f32(float [[P0]])
+// CHECK-NEXT:    ret float [[ELT_ABS_I]]
+//
+float test_length_float(float p0)
+{
+  return length(p0);
+}
+
+// CHECK-LABEL: define noundef nofpclass(nan inf) float @_Z18test_length_float2Dv2_f(
+// CHECK-SAME: <2 x float> noundef nofpclass(nan inf) [[P0:%.*]]) local_unnamed_addr #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[MUL_I:%.*]] = fmul reassoc nnan ninf nsz arcp afn <2 x float> [[P0]], [[P0]]
+// CHECK-NEXT:    [[VECEXT_I:%.*]] = extractelement <2 x float> [[MUL_I]], i64 0
+// CHECK-NEXT:    [[VECEXT1_I:%.*]] = extractelement <2 x float> [[MUL_I]], i64 1
+// CHECK-NEXT:    [[ADD_I:%.*]] = fadd reassoc nnan ninf nsz arcp afn float [[VECEXT_I]], [[VECEXT1_I]]
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call reassoc nnan ninf nsz arcp afn noundef float @llvm.sqrt.f32(float [[ADD_I]])
+// CHECK-NEXT:    ret float [[TMP0]]
+//
+float test_length_float2(float2 p0)
+{
+  return length(p0);
+}
+
+// CHECK-LABEL: define noundef nofpclass(nan inf) float @_Z18test_length_float3Dv3_f(
+// CHECK-SAME: <3 x float> noundef nofpclass(nan inf) [[P0:%.*]]) local_unnamed_addr #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[MUL_I:%.*]] = fmul reassoc nnan ninf nsz arcp afn <3 x float> [[P0]], [[P0]]
+// CHECK-NEXT:    [[VECEXT_I:%.*]] = extractelement <3 x float> [[MUL_I]], i64 0
+// CHECK-NEXT:    [[VECEXT1_I:%.*]] = extractelement <3 x float> [[MUL_I]], i64 1
+// CHECK-NEXT:    [[ADD_I:%.*]] = fadd reassoc nnan ninf nsz arcp afn float [[VECEXT1_I]], [[VECEXT_I]]
+// CHECK-NEXT:    [[VECEXT1_I_1:%.*]] = extractelement <3 x float> [[MUL_I]], i64 2
+// CHECK-NEXT:    [[ADD_I_1:%.*]] = fadd reassoc nnan ninf nsz arcp afn float [[VECEXT1_I_1]], [[ADD_I]]
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call reassoc nnan ninf nsz arcp afn noundef float @llvm.sqrt.f32(float [[ADD_I_1]])
+// CHECK-NEXT:    ret float [[TMP0]]
+//
+float test_length_float3(float3 p0)
+{
+  return length(p0);
+}
+
+
+// CHECK-LABEL: define noundef nofpclass(nan inf) float @_Z18test_length_float4Dv4_f(
+// CHECK-SAME: <4 x float> noundef nofpclass(nan inf) [[P0:%.*]]) local_unnamed_addr #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[MUL_I:%.*]] = fmul reassoc nnan ninf nsz arcp afn <4 x float> [[P0]], [[P0]]
+// CHECK-NEXT:    [[VECEXT_I:%.*]] = extractelement <4 x float> [[MUL_I]], i64 0
+// CHECK-NEXT:    [[VECEXT1_I:%.*]] = extractelement <4 x float> [[MUL_I]], i64 1
+// CHECK-NEXT:    [[ADD_I:%.*]] = fadd reassoc nnan ninf nsz arcp afn float [[VECEXT1_I]], [[VECEXT_I]]
+// CHECK-NEXT:    [[VECEXT1_I_1:%.*]] = extractelement <4 x float> [[MUL_I]], i64 2
+// CHECK-NEXT:    [[ADD_I_1:%.*]] = fadd reassoc nnan ninf nsz arcp afn float [[VECEXT1_I_1]], [[ADD_I]]
+// CHECK-NEXT:    [[VECEXT1_I_2:%.*]] = extractelement <4 x float> [[MUL_I]], i64 3
+// CHECK-NEXT:    [[ADD_I_2:%.*]] = fadd reassoc nnan ninf nsz arcp afn float [[VECEXT1_I_2]], [[ADD_I_1]]
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call reassoc nnan ninf nsz arcp afn noundef float @llvm.sqrt.f32(float [[ADD_I_2]])
+// CHECK-NEXT:    ret float [[TMP0]]
+//
+float test_length_float4(float4 p0)
+{
+  return length(p0);
+}