[HLSL] add pow library function

[bob80905]

Summary

This change exposes the pow library function for HLSL, only available for floating point types.
The pow function is supported for all scalar, vector, and matrix types that contain floating point types.

The full documentation of the HLSL pow function is available here:
https://learn.microsoft.com/en-us/windows/win32/direct3dhlsl/dx-graphics-hlsl-pow

[bob80905]

• remove non-floating-point func decls from header file

[bob80905]

• add eof newline

[beanz]

clang/test/CodeGenHLSL/builtins/pow.hlsl

7

Does this need to set -D__HLSL_ENABLE_16_BIT? If 16-bit types are disabled, this should result in half being 32-bit right?

13

nit: can you clang-format this source?

[bob80905]

clang/test/CodeGenHLSL/builtins/pow.hlsl

7

The define is necessary (the test fails otherwise) because without the definition, the definition of "half4" wouldn't exist (half4 is defined in hlsl_intrinsics.h).
To your second question, yes, I think that's correct.

13

Will do

[bogner]

Looking at https://learn.microsoft.com/en-us/windows/win32/direct3dhlsl/dx-graphics-hlsl-pow, it seems a bit underspecified, but do we need to worry about differences between that spec and IEEE behaviour here? For example, it says there that x<0, y=any -> NaN, but we'll only generate NaN for non-integer y in these cases as per https://en.cppreference.com/w/c/numeric/math/pow

clang/test/CodeGenHLSL/builtins/pow.hlsl

11

Why check the function name in the no-half case but not the other check?

[bob80905]

• add func name to checks, clang format

[bob80905]

clang/test/CodeGenHLSL/builtins/pow.hlsl

11

This is the pattern I've been using since starting on the upstreaming effort, so I've been copying this pattern. I don't see a reason why the function name shouldn't be included, so I'll add it here.

[bogner]

In D156178#4530161, @bogner wrote:

Looking at https://learn.microsoft.com/en-us/windows/win32/direct3dhlsl/dx-graphics-hlsl-pow, it seems a bit underspecified, but do we need to worry about differences between that spec and IEEE behaviour here? For example, it says there that x<0, y=any -> NaN, but we'll only generate NaN for non-integer y in these cases as per https://en.cppreference.com/w/c/numeric/math/pow

FWIW it looks like the behaviour here matches what DXC does in practice today, so discussing the discrepancy here probably doesn't need to block getting this in

[bob80905]

• clang format pow.hlsl

clang/lib/Headers/hlsl/hlsl_intrinsics.h

 __attribute__((clang_builtin_alias(__builtin_elementwise_bitreverse)))
 uint64_t reversebits(uint64_t);
 __attribute__((clang_builtin_alias(__builtin_elementwise_bitreverse)))
 uint64_t2 reversebits(uint64_t2);
 __attribute__((clang_builtin_alias(__builtin_elementwise_bitreverse)))
 uint64_t3 reversebits(uint64_t3);
 __attribute__((clang_builtin_alias(__builtin_elementwise_bitreverse)))
 uint64_t4 reversebits(uint64_t4);
+
+// pow builtins
+#ifdef __HLSL_ENABLE_16_BIT
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow)))
+half pow(half, half);
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow)))
+half2 pow(half2, half2);
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow)))
+half3 pow(half3, half3);
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow)))
+half4 pow(half4, half4);
+#endif
+
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow))) float
+pow(float, float);
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow)))
+float2 pow(float2, float2);
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow)))
+float3 pow(float3, float3);
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow)))
+float4 pow(float4, float4);
+
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow))) double
+pow(double, double);
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow)))
+double2 pow(double2, double2);
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow)))
+double3 pow(double3, double3);
+__attribute__((clang_builtin_alias(__builtin_elementwise_pow)))
+double4 pow(double4, double4);
+
 } // namespace hlsl
 #endif //_HLSL_HLSL_INTRINSICS_H_

clang/test/CodeGenHLSL/builtins/pow.hlsl

+// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -x hlsl -triple \
+// RUN:   dxil-pc-shadermodel6.3-library %s -fnative-half-type \
+// RUN:   -emit-llvm -disable-llvm-passes -O3 -o - | FileCheck %s
+// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -x hlsl -triple \
+// RUN:   dxil-pc-shadermodel6.3-library %s -emit-llvm -disable-llvm-passes \
+// RUN:   -D__HLSL_ENABLE_16_BIT -o - | FileCheck %s --check-prefix=NO_HALF
+
+// CHECK: define noundef half @
+// CHECK: call half @llvm.pow.f16(
+// NO_HALF: define noundef float @"?test_pow_half@@YA$halff@$halff@0@Z"(
+// NO_HALF: call float @llvm.pow.f32(
+half test_pow_half(half p0, half p1)
+{
+    return pow(p0, p1);
+}
+// CHECK: define noundef <2 x half> @"?test_pow_half2@@YAT?$__vector@$f16@$01@__clang@@T12@0@Z"(
+// CHECK: call <2 x half> @llvm.pow.v2f16
+// NO_HALF: define noundef <2 x float> @"?test_pow_float2@@YAT?$__vector@M$01@__clang@@T12@0@Z"(
+// NO_HALF: call <2 x float> @llvm.pow.v2f32(
+half2 test_pow_half2(half2 p0, half2 p1)
+{
+    return pow(p0, p1);
+}
+// CHECK: define noundef <3 x half> @"?test_pow_half3@@YAT?$__vector@$f16@$02@__clang@@T12@0@Z"(
+// CHECK: call <3 x half> @llvm.pow.v3f16
+// NO_HALF: define noundef <3 x float> @"?test_pow_float3@@YAT?$__vector@M$02@__clang@@T12@0@Z"(
+// NO_HALF: call <3 x float> @llvm.pow.v3f32(
+half3 test_pow_half3(half3 p0, half3 p1)
+{
+    return pow(p0, p1);
+}
+// CHECK: define noundef <4 x half> @"?test_pow_half4@@YAT?$__vector@$f16@$03@__clang@@T12@0@Z"(
+// CHECK: call <4 x half> @llvm.pow.v4f16
+// NO_HALF: define noundef <4 x float> @"?test_pow_float4@@YAT?$__vector@M$03@__clang@@T12@0@Z"(
+// NO_HALF: call <4 x float> @llvm.pow.v4f32(
+half4 test_pow_half4(half4 p0, half4 p1)
+{
+    return pow(p0, p1);
+}
+
+// CHECK: define noundef float @"?test_pow_float@@YAMMM@Z"(
+// CHECK: call float @llvm.pow.f32(
+float test_pow_float(float p0, float p1)
+{
+    return pow(p0, p1);
+}
+// CHECK: define noundef <2 x float> @"?test_pow_float2@@YAT?$__vector@M$01@__clang@@T12@0@Z"(
+// CHECK: call <2 x float> @llvm.pow.v2f32
+float2 test_pow_float2(float2 p0, float2 p1)
+{
+    return pow(p0, p1);
+}
+// CHECK: define noundef <3 x float> @"?test_pow_float3@@YAT?$__vector@M$02@__clang@@T12@0@Z"(
+// CHECK: call <3 x float> @llvm.pow.v3f32
+float3 test_pow_float3(float3 p0, float3 p1)
+{
+    return pow(p0, p1);
+}
+// CHECK: define noundef <4 x float> @"?test_pow_float4@@YAT?$__vector@M$03@__clang@@T12@0@Z"(
+// CHECK: call <4 x float> @llvm.pow.v4f32
+float4 test_pow_float4(float4 p0, float4 p1)
+{
+    return pow(p0, p1);
+}
+
+// CHECK: define noundef double @"?test_pow_double@@YANNN@Z"(
+// CHECK: call double @llvm.pow.f64(
+double test_pow_double(double p0, double p1)
+{
+    return pow(p0, p1);
+}
+// CHECK: define noundef <2 x double> @"?test_pow_double2@@YAT?$__vector@N$01@__clang@@T12@0@Z"(
+// CHECK: call <2 x double> @llvm.pow.v2f64
+double2 test_pow_double2(double2 p0, double2 p1)
+{
+    return pow(p0, p1);
+}
+// CHECK: define noundef <3 x double> @"?test_pow_double3@@YAT?$__vector@N$02@__clang@@T12@0@Z"(
+// CHECK: call <3 x double> @llvm.pow.v3f64
+double3 test_pow_double3(double3 p0, double3 p1)
+{
+    return pow(p0, p1);
+}
+// CHECK: define noundef <4 x double> @"?test_pow_double4@@YAT?$__vector@N$03@__clang@@T12@0@Z"(
+// CHECK: call <4 x double> @llvm.pow.v4f64
+double4 test_pow_double4(double4 p0, double4 p1)
+{
+    return pow(p0, p1);
+}