https://github.com/spall updated
https://github.com/llvm/llvm-project/pull/129334
>From ebaef38c40cba83d4b446a5320d57c86038f8286 Mon Sep 17 00:00:00 2001
From: Sarah Spall <sarahsp...@microsoft.com>
Date: Fri, 28 Feb 2025 11:18:03 -0800
Subject: [PATCH 1/3] new max overloads + new min overloads
---
clang/lib/Headers/hlsl/hlsl_intrinsics.h | 360 +++++++++++++++++++++++
1 file changed, 360 insertions(+)
diff --git a/clang/lib/Headers/hlsl/hlsl_intrinsics.h
b/clang/lib/Headers/hlsl/hlsl_intrinsics.h
index 239d7a3f59b77..68acdf1720318 100644
--- a/clang/lib/Headers/hlsl/hlsl_intrinsics.h
+++ b/clang/lib/Headers/hlsl/hlsl_intrinsics.h
@@ -1678,6 +1678,30 @@ half3 max(half3, half3);
_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
half4 max(half4, half4);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half2 max(half2 p0, half p1) {
+ return __builtin_elementwise_max(p0, (half2)p1);
+}
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half2 max(half p0, half2 p1) {
+ return __builtin_elementwise_max((half2)p0, p1);
+}
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half3 max(half3 p0, half p1) {
+ return __builtin_elementwise_max(p0, (half3)p1);
+}
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half3 max(half p0, half3 p1) {
+ return __builtin_elementwise_max((half3)p0, p1);
+}
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half4 max(half4 p0, half p1) {
+ return __builtin_elementwise_max(p0, (half4)p1);
+}
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half4 max(half p0, half4 p1) {
+ return __builtin_elementwise_max((half4)p0, p1);
+}
#ifdef __HLSL_ENABLE_16_BIT
_HLSL_AVAILABILITY(shadermodel, 6.2)
@@ -1692,6 +1716,30 @@ int16_t3 max(int16_t3, int16_t3);
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int16_t4 max(int16_t4, int16_t4);
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t2 max(int16_t2 p0, int16_t p1) {
+ return __builtin_elementwise_max(p0, (int16_t2)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t2 max(int16_t p0, int16_t2 p1) {
+ return __builtin_elementwise_max((int16_t2)p0, p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t3 max(int16_t3 p0, int16_t p1) {
+ return __builtin_elementwise_max(p0, (int16_t3)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t3 max(int16_t p0, int16_t3 p1) {
+ return __builtin_elementwise_max((int16_t3)p0, p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t4 max(int16_t4 p0, int16_t p1) {
+ return __builtin_elementwise_max(p0, (int16_t4)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t4 max(int16_t p0, int16_t4 p1) {
+ return __builtin_elementwise_max((int16_t4)p0, p1);
+}
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
@@ -1705,6 +1753,30 @@ uint16_t3 max(uint16_t3, uint16_t3);
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint16_t4 max(uint16_t4, uint16_t4);
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t2 max(uint16_t2 p0, uint16_t p1) {
+ return __builtin_elementwise_max(p0, (uint16_t2)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t2 max(uint16_t p0, uint16_t2 p1) {
+ return __builtin_elementwise_max((uint16_t2)p0, p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t3 max(uint16_t3 p0, uint16_t p1) {
+ return __builtin_elementwise_max(p0, (uint16_t3)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t3 max(uint16_t p0, uint16_t3 p1) {
+ return __builtin_elementwise_max((uint16_t3)p0, p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t4 max(uint16_t4 p0, uint16_t p1) {
+ return __builtin_elementwise_max(p0, (uint16_t4)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t4 max(uint16_t p0, uint16_t4 p1) {
+ return __builtin_elementwise_max((uint16_t4)p0, p1);
+}
#endif
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
@@ -1715,6 +1787,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int3 max(int3, int3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int4 max(int4, int4);
+constexpr int2 max(int2 p0, int p1) {
+ return __builtin_elementwise_max(p0, (int2)p1);
+}
+constexpr int2 max(int p0, int2 p1) {
+ return __builtin_elementwise_max((int2)p0, p1);
+}
+constexpr int3 max(int3 p0, int p1) {
+ return __builtin_elementwise_max(p0, (int3)p1);
+}
+constexpr int3 max(int p0, int3 p1) {
+ return __builtin_elementwise_max((int3)p0, p1);
+}
+constexpr int4 max(int4 p0, int p1) {
+ return __builtin_elementwise_max(p0, (int4)p1);
+}
+constexpr int4 max(int p0, int4 p1) {
+ return __builtin_elementwise_max((int4)p0, p1);
+}
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint max(uint, uint);
@@ -1724,6 +1814,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint3 max(uint3, uint3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint4 max(uint4, uint4);
+constexpr uint2 max(uint2 p0, uint p1) {
+ return __builtin_elementwise_max(p0, (uint2)p1);
+}
+constexpr uint2 max(uint p0, uint2 p1) {
+ return __builtin_elementwise_max((uint2)p0, p1);
+}
+constexpr uint3 max(uint3 p0, uint p1) {
+ return __builtin_elementwise_max(p0, (uint3)p1);
+}
+constexpr uint3 max(uint p0, uint3 p1) {
+ return __builtin_elementwise_max((uint3)p0, p1);
+}
+constexpr uint4 max(uint4 p0, uint p1) {
+ return __builtin_elementwise_max(p0, (uint4)p1);
+}
+constexpr uint4 max(uint p0, uint4 p1) {
+ return __builtin_elementwise_max((uint4)p0, p1);
+}
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int64_t max(int64_t, int64_t);
@@ -1733,6 +1841,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int64_t3 max(int64_t3, int64_t3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int64_t4 max(int64_t4, int64_t4);
+constexpr int64_t2 max(int64_t2 p0, int64_t p1) {
+ return __builtin_elementwise_max(p0, (int64_t2)p1);
+}
+constexpr int64_t2 max(int64_t p0, int64_t2 p1) {
+ return __builtin_elementwise_max((int64_t2)p0, p1);
+}
+constexpr int64_t3 max(int64_t3 p0, int64_t p1) {
+ return __builtin_elementwise_max(p0, (int64_t3)p1);
+}
+constexpr int64_t3 max(int64_t p0, int64_t3 p1) {
+ return __builtin_elementwise_max((int64_t3)p0, p1);
+}
+constexpr int64_t4 max(int64_t4 p0, int64_t p1) {
+ return __builtin_elementwise_max(p0, (int64_t4)p1);
+}
+constexpr int64_t4 max(int64_t p0, int64_t4 p1) {
+ return __builtin_elementwise_max((int64_t4)p0, p1);
+}
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint64_t max(uint64_t, uint64_t);
@@ -1742,6 +1868,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint64_t3 max(uint64_t3, uint64_t3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint64_t4 max(uint64_t4, uint64_t4);
+constexpr uint64_t2 max(uint64_t2 p0, uint64_t p1) {
+ return __builtin_elementwise_max(p0, (uint64_t2)p1);
+}
+constexpr uint64_t2 max(uint64_t p0, uint64_t2 p1) {
+ return __builtin_elementwise_max((uint64_t2)p0, p1);
+}
+constexpr uint64_t3 max(uint64_t3 p0, uint64_t p1) {
+ return __builtin_elementwise_max(p0, (uint64_t3)p1);
+}
+constexpr uint64_t3 max(uint64_t p0, uint64_t3 p1) {
+ return __builtin_elementwise_max((uint64_t3)p0, p1);
+}
+constexpr uint64_t4 max(uint64_t4 p0, uint64_t p1) {
+ return __builtin_elementwise_max(p0, (uint64_t4)p1);
+}
+constexpr uint64_t4 max(uint64_t p0, uint64_t4 p1) {
+ return __builtin_elementwise_max((uint64_t4)p0, p1);
+}
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
float max(float, float);
@@ -1751,6 +1895,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
float3 max(float3, float3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
float4 max(float4, float4);
+constexpr float2 max(float2 p0, float p1) {
+ return __builtin_elementwise_max(p0, (float2)p1);
+}
+constexpr float2 max(float p0, float2 p1) {
+ return __builtin_elementwise_max((float2)p0, p1);
+}
+constexpr float3 max(float3 p0, float p1) {
+ return __builtin_elementwise_max(p0, (float3)p1);
+}
+constexpr float3 max(float p0, float3 p1) {
+ return __builtin_elementwise_max((float3)p0, p1);
+}
+constexpr float4 max(float4 p0, float p1) {
+ return __builtin_elementwise_max(p0, (float4)p1);
+}
+constexpr float4 max(float p0, float4 p1) {
+ return __builtin_elementwise_max((float4)p0, p1);
+}
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
double max(double, double);
@@ -1760,6 +1922,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
double3 max(double3, double3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
double4 max(double4, double4);
+constexpr double2 max(double2 p0, double p1) {
+ return __builtin_elementwise_max(p0, (double2)p1);
+}
+constexpr double2 max(double p0, double2 p1) {
+ return __builtin_elementwise_max((double2)p0, p1);
+}
+constexpr double3 max(double3 p0, double p1) {
+ return __builtin_elementwise_max(p0, (double3)p1);
+}
+constexpr double3 max(double p0, double3 p1) {
+ return __builtin_elementwise_max((double3)p0, p1);
+}
+constexpr double4 max(double4 p0, double p1) {
+ return __builtin_elementwise_max(p0, (double4)p1);
+}
+constexpr double4 max(double p0, double4 p1) {
+ return __builtin_elementwise_max((double4)p0, p1);
+}
//===----------------------------------------------------------------------===//
// min builtins
@@ -1782,6 +1962,30 @@ half3 min(half3, half3);
_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
half4 min(half4, half4);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half2 min(half2 p0, half p1) {
+ return __builtin_elementwise_min(p0, (half2)p1);
+}
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half2 min(half p0, half2 p1) {
+ return __builtin_elementwise_min((half2)p0, p1);
+}
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half3 min(half3 p0, half p1) {
+ return __builtin_elementwise_min(p0, (half3)p1);
+}
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half3 min(half p0, half3 p1) {
+ return __builtin_elementwise_min((half3)p0, p1);
+}
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half4 min(half4 p0, half p1) {
+ return __builtin_elementwise_min(p0, (half4)p1);
+}
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+constexpr half4 min(half p0, half4 p1) {
+ return __builtin_elementwise_min((half4)p0, p1);
+}
#ifdef __HLSL_ENABLE_16_BIT
_HLSL_AVAILABILITY(shadermodel, 6.2)
@@ -1796,6 +2000,30 @@ int16_t3 min(int16_t3, int16_t3);
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int16_t4 min(int16_t4, int16_t4);
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t2 min(int16_t2 p0, int16_t p1) {
+ return __builtin_elementwise_min(p0, (int16_t2)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t2 min(int16_t p0, int16_t2 p1) {
+ return __builtin_elementwise_min((int16_t2)p0, p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t3 min(int16_t3 p0, int16_t p1) {
+ return __builtin_elementwise_min(p0, (int16_t3)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t3 min(int16_t p0, int16_t3 p1) {
+ return __builtin_elementwise_min((int16_t3)p0, p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t4 min(int16_t4 p0, int16_t p1) {
+ return __builtin_elementwise_min(p0, (int16_t4)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr int16_t4 min(int16_t p0, int16_t4 p1) {
+ return __builtin_elementwise_min((int16_t4)p0, p1);
+}
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
@@ -1809,6 +2037,30 @@ uint16_t3 min(uint16_t3, uint16_t3);
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint16_t4 min(uint16_t4, uint16_t4);
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t2 min(uint16_t2 p0, uint16_t p1) {
+ return __builtin_elementwise_min(p0, (uint16_t2)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t2 min(uint16_t p0, uint16_t2 p1) {
+ return __builtin_elementwise_min((uint16_t2)p0, p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t3 min(uint16_t3 p0, uint16_t p1) {
+ return __builtin_elementwise_min(p0, (uint16_t3)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t3 min(uint16_t p0, uint16_t3 p1) {
+ return __builtin_elementwise_min((uint16_t3)p0, p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t4 min(uint16_t4 p0, uint16_t p1) {
+ return __builtin_elementwise_min(p0, (uint16_t4)p1);
+}
+_HLSL_AVAILABILITY(shadermodel, 6.2)
+constexpr uint16_t4 min(uint16_t p0, uint16_t4 p1) {
+ return __builtin_elementwise_min((uint16_t4)p0, p1);
+}
#endif
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
@@ -1819,6 +2071,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int3 min(int3, int3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int4 min(int4, int4);
+constexpr int2 min(int2 p0, int p1) {
+ return __builtin_elementwise_min(p0, (int2)p1);
+}
+constexpr int2 min(int p0, int2 p1) {
+ return __builtin_elementwise_min((int2)p0, p1);
+}
+constexpr int3 min(int3 p0, int p1) {
+ return __builtin_elementwise_min(p0, (int3)p1);
+}
+constexpr int3 min(int p0, int3 p1) {
+ return __builtin_elementwise_min((int3)p0, p1);
+}
+constexpr int4 min(int4 p0, int p1) {
+ return __builtin_elementwise_min(p0, (int4)p1);
+}
+constexpr int4 min(int p0, int4 p1) {
+ return __builtin_elementwise_min((int4)p0, p1);
+}
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint min(uint, uint);
@@ -1828,6 +2098,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint3 min(uint3, uint3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint4 min(uint4, uint4);
+constexpr uint2 min(uint2 p0, uint p1) {
+ return __builtin_elementwise_min(p0, (uint2)p1);
+}
+constexpr uint2 min(uint p0, uint2 p1) {
+ return __builtin_elementwise_min((uint2)p0, p1);
+}
+constexpr uint3 min(uint3 p0, uint p1) {
+ return __builtin_elementwise_min(p0, (uint3)p1);
+}
+constexpr uint3 min(uint p0, uint3 p1) {
+ return __builtin_elementwise_min((uint3)p0, p1);
+}
+constexpr uint4 min(uint4 p0, uint p1) {
+ return __builtin_elementwise_min(p0, (uint4)p1);
+}
+constexpr uint4 min(uint p0, uint4 p1) {
+ return __builtin_elementwise_min((uint4)p0, p1);
+}
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
float min(float, float);
@@ -1837,6 +2125,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
float3 min(float3, float3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
float4 min(float4, float4);
+constexpr float2 min(float2 p0, float p1) {
+ return __builtin_elementwise_min(p0, (float2)p1);
+}
+constexpr float2 min(float p0, float2 p1) {
+ return __builtin_elementwise_min((float2)p0, p1);
+}
+constexpr float3 min(float3 p0, float p1) {
+ return __builtin_elementwise_min(p0, (float3)p1);
+}
+constexpr float3 min(float p0, float3 p1) {
+ return __builtin_elementwise_min((float3)p0, p1);
+}
+constexpr float4 min(float4 p0, float p1) {
+ return __builtin_elementwise_min(p0, (float4)p1);
+}
+constexpr float4 min(float p0, float4 p1) {
+ return __builtin_elementwise_min((float4)p0, p1);
+}
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int64_t min(int64_t, int64_t);
@@ -1846,6 +2152,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int64_t3 min(int64_t3, int64_t3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int64_t4 min(int64_t4, int64_t4);
+constexpr int64_t2 min(int64_t2 p0, int64_t p1) {
+ return __builtin_elementwise_min(p0, (int64_t2)p1);
+}
+constexpr int64_t2 min(int64_t p0, int64_t2 p1) {
+ return __builtin_elementwise_min((int64_t2)p0, p1);
+}
+constexpr int64_t3 min(int64_t3 p0, int64_t p1) {
+ return __builtin_elementwise_min(p0, (int64_t3)p1);
+}
+constexpr int64_t3 min(int64_t p0, int64_t3 p1) {
+ return __builtin_elementwise_min((int64_t3)p0, p1);
+}
+constexpr int64_t4 min(int64_t4 p0, int64_t p1) {
+ return __builtin_elementwise_min(p0, (int64_t4)p1);
+}
+constexpr int64_t4 min(int64_t p0, int64_t4 p1) {
+ return __builtin_elementwise_min((int64_t4)p0, p1);
+}
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint64_t min(uint64_t, uint64_t);
@@ -1855,6 +2179,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint64_t3 min(uint64_t3, uint64_t3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint64_t4 min(uint64_t4, uint64_t4);
+constexpr uint64_t2 min(uint64_t2 p0, uint64_t p1) {
+ return __builtin_elementwise_min(p0, (uint64_t2)p1);
+}
+constexpr uint64_t2 min(uint64_t p0, uint64_t2 p1) {
+ return __builtin_elementwise_min((uint64_t2)p0, p1);
+}
+constexpr uint64_t3 min(uint64_t3 p0, uint64_t p1) {
+ return __builtin_elementwise_min(p0, (uint64_t3)p1);
+}
+constexpr uint64_t3 min(uint64_t p0, uint64_t3 p1) {
+ return __builtin_elementwise_min((uint64_t3)p0, p1);
+}
+constexpr uint64_t4 min(uint64_t4 p0, uint64_t p1) {
+ return __builtin_elementwise_min(p0, (uint64_t4)p1);
+}
+constexpr uint64_t4 min(uint64_t p0, uint64_t4 p1) {
+ return __builtin_elementwise_min((uint64_t4)p0, p1);
+}
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
double min(double, double);
@@ -1864,6 +2206,24 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
double3 min(double3, double3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
double4 min(double4, double4);
+constexpr double2 min(double2 p0, double p1) {
+ return __builtin_elementwise_min(p0, (double2)p1);
+}
+constexpr double2 min(double p0, double2 p1) {
+ return __builtin_elementwise_min((double2)p0, p1);
+}
+constexpr double3 min(double3 p0, double p1) {
+ return __builtin_elementwise_min(p0, (double3)p1);
+}
+constexpr double3 min(double p0, double3 p1) {
+ return __builtin_elementwise_min((double3)p0, p1);
+}
+constexpr double4 min(double4 p0, double p1) {
+ return __builtin_elementwise_min(p0, (double4)p1);
+}
+constexpr double4 min(double p0, double4 p1) {
+ return __builtin_elementwise_min((double4)p0, p1);
+}
//===----------------------------------------------------------------------===//
// normalize builtins
>From 4f121bfbce25eb42a03ab4105989037be10bacca Mon Sep 17 00:00:00 2001
From: Sarah Spall <sarahsp...@microsoft.com>
Date: Fri, 28 Feb 2025 14:46:21 -0800
Subject: [PATCH 2/3] tests
---
clang/test/CodeGenHLSL/builtins/max.hlsl | 32 ++++++++++++++++++++++++
clang/test/CodeGenHLSL/builtins/min.hlsl | 32 ++++++++++++++++++++++++
2 files changed, 64 insertions(+)
diff --git a/clang/test/CodeGenHLSL/builtins/max.hlsl
b/clang/test/CodeGenHLSL/builtins/max.hlsl
index 0b767335556ee..6b5fb6ae59534 100644
--- a/clang/test/CodeGenHLSL/builtins/max.hlsl
+++ b/clang/test/CodeGenHLSL/builtins/max.hlsl
@@ -18,6 +18,9 @@ int16_t3 test_max_short3(int16_t3 p0, int16_t3 p1) { return
max(p0, p1); }
// NATIVE_HALF-LABEL: define noundef <4 x i16> @_Z15test_max_short4
// NATIVE_HALF: call <4 x i16> @llvm.smax.v4i16
int16_t4 test_max_short4(int16_t4 p0, int16_t4 p1) { return max(p0, p1); }
+// NATIVE_HALF-LABEL: define noundef <4 x i16> {{.*}}test_max_short4_mismatch
+// NATIVE_HALF: call <4 x i16> @llvm.smax.v4i16
+int16_t4 test_max_short4_mismatch(int16_t4 p0, int16_t p1) { return max(p0,
p1); }
// NATIVE_HALF-LABEL: define noundef i16 @_Z15test_max_ushort
// NATIVE_HALF: call i16 @llvm.umax.i16(
@@ -31,6 +34,9 @@ uint16_t3 test_max_ushort3(uint16_t3 p0, uint16_t3 p1) {
return max(p0, p1); }
// NATIVE_HALF-LABEL: define noundef <4 x i16> @_Z16test_max_ushort4
// NATIVE_HALF: call <4 x i16> @llvm.umax.v4i16
uint16_t4 test_max_ushort4(uint16_t4 p0, uint16_t4 p1) { return max(p0, p1); }
+// NATIVE_HALF-LABEL: define noundef <4 x i16> {{.*}}test_max_ushort4_mismatch
+// NATIVE_HALF: call <4 x i16> @llvm.umax.v4i16
+uint16_t4 test_max_ushort4_mismatch(uint16_t4 p0, uint16_t p1) { return
max(p0, p1); }
#endif
// CHECK-LABEL: define noundef i32 @_Z12test_max_int
@@ -45,6 +51,9 @@ int3 test_max_int3(int3 p0, int3 p1) { return max(p0, p1); }
// CHECK-LABEL: define noundef <4 x i32> @_Z13test_max_int4
// CHECK: call <4 x i32> @llvm.smax.v4i32
int4 test_max_int4(int4 p0, int4 p1) { return max(p0, p1); }
+// CHECK-LABEL: define noundef <4 x i32> {{.*}}test_max_int4_mismatch
+// CHECK: call <4 x i32> @llvm.smax.v4i32
+int4 test_max_int4_mismatch(int4 p0, int p1) { return max(p0, p1); }
// CHECK-LABEL: define noundef i32 @_Z13test_max_uint
// CHECK: call i32 @llvm.umax.i32(
@@ -58,6 +67,9 @@ uint3 test_max_uint3(uint3 p0, uint3 p1) { return max(p0,
p1); }
// CHECK-LABEL: define noundef <4 x i32> @_Z14test_max_uint4
// CHECK: call <4 x i32> @llvm.umax.v4i32
uint4 test_max_uint4(uint4 p0, uint4 p1) { return max(p0, p1); }
+// CHECK-LABEL: define noundef <4 x i32> {{.*}}test_max_uint4_mismatch
+// CHECK: call <4 x i32> @llvm.umax.v4i32
+uint4 test_max_uint4_mismatch(uint4 p0, uint p1) { return max(p0, p1); }
// CHECK-LABEL: define noundef i64 @_Z13test_max_long
// CHECK: call i64 @llvm.smax.i64(
@@ -71,6 +83,9 @@ int64_t3 test_max_long3(int64_t3 p0, int64_t3 p1) { return
max(p0, p1); }
// CHECK-LABEL: define noundef <4 x i64> @_Z14test_max_long4
// CHECK: call <4 x i64> @llvm.smax.v4i64
int64_t4 test_max_long4(int64_t4 p0, int64_t4 p1) { return max(p0, p1); }
+// CHECK-LABEL: define noundef <4 x i64> {{.*}}test_max_long4_mismatch
+// CHECK: call <4 x i64> @llvm.smax.v4i64
+int64_t4 test_max_long4_mismatch(int64_t4 p0, int64_t p1) { return max(p0,
p1); }
// CHECK-LABEL: define noundef i64 @_Z14test_max_ulong
// CHECK: call i64 @llvm.umax.i64(
@@ -84,6 +99,9 @@ uint64_t3 test_max_ulong3(uint64_t3 p0, uint64_t3 p1) {
return max(p0, p1); }
// CHECK-LABEL: define noundef <4 x i64> @_Z15test_max_ulong4
// CHECK: call <4 x i64> @llvm.umax.v4i64
uint64_t4 test_max_ulong4(uint64_t4 p0, uint64_t4 p1) { return max(p0, p1); }
+// CHECK-LABEL: define noundef <4 x i64> {{.*}}test_max_ulong4_mismatch
+// CHECK: call <4 x i64> @llvm.umax.v4i64
+uint64_t4 test_max_ulong4_mismatch(uint64_t4 p0, uint64_t p1) { return max(p0,
p1); }
// NATIVE_HALF-LABEL: define noundef nofpclass(nan inf) half @_Z13test_max_half
// NATIVE_HALF: call reassoc nnan ninf nsz arcp afn half @llvm.maxnum.f16(
@@ -105,6 +123,11 @@ half3 test_max_half3(half3 p0, half3 p1) { return max(p0,
p1); }
// NO_HALF-LABEL: define noundef nofpclass(nan inf) <4 x float>
@_Z14test_max_half4
// NO_HALF: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.maxnum.v4f32(
half4 test_max_half4(half4 p0, half4 p1) { return max(p0, p1); }
+// NATIVE_HALF-LABEL: define noundef nofpclass(nan inf) <4 x half>
{{.*}}test_max_half4_mismatch
+// NATIVE_HALF: call reassoc nnan ninf nsz arcp afn <4 x half>
@llvm.maxnum.v4f16
+// NO_HALF-LABEL: define noundef nofpclass(nan inf) <4 x float>
{{.*}}test_max_half4_mismatch
+// NO_HALF: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.maxnum.v4f32(
+half4 test_max_half4_mismatch(half4 p0, half p1) { return max(p0, p1); }
// CHECK-LABEL: define noundef nofpclass(nan inf) float @_Z14test_max_float
// CHECK: call reassoc nnan ninf nsz arcp afn float @llvm.maxnum.f32(
@@ -118,6 +141,9 @@ float3 test_max_float3(float3 p0, float3 p1) { return
max(p0, p1); }
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x float>
@_Z15test_max_float4
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.maxnum.v4f32
float4 test_max_float4(float4 p0, float4 p1) { return max(p0, p1); }
+// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x float>
{{.*}}test_max_float4_mismatch
+// CHECK: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.maxnum.v4f32
+float4 test_max_float4_mismatch(float4 p0, float p1) { return max(p0, p1); }
// CHECK-LABEL: define noundef nofpclass(nan inf) double @_Z15test_max_double
// CHECK: call reassoc nnan ninf nsz arcp afn double @llvm.maxnum.f64(
@@ -131,3 +157,9 @@ double3 test_max_double3(double3 p0, double3 p1) { return
max(p0, p1); }
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double>
@_Z16test_max_double4
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.maxnum.v4f64
double4 test_max_double4(double4 p0, double4 p1) { return max(p0, p1); }
+// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double>
{{.*}}test_max_double4_mismatch
+// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.maxnum.v4f64
+double4 test_max_double4_mismatch(double4 p0, double p1) { return max(p0, p1);
}
+// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double>
{{.*}}test_max_double4_mismatch2
+// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.maxnum.v4f64
+double4 test_max_double4_mismatch2(double4 p0, double p1) { return max(p1,
p0); }
diff --git a/clang/test/CodeGenHLSL/builtins/min.hlsl
b/clang/test/CodeGenHLSL/builtins/min.hlsl
index a352c72f29542..551db52878e37 100644
--- a/clang/test/CodeGenHLSL/builtins/min.hlsl
+++ b/clang/test/CodeGenHLSL/builtins/min.hlsl
@@ -18,6 +18,9 @@ int16_t3 test_min_short3(int16_t3 p0, int16_t3 p1) { return
min(p0, p1); }
// NATIVE_HALF-LABEL: define noundef <4 x i16> @_Z15test_min_short4
// NATIVE_HALF: call <4 x i16> @llvm.smin.v4i16
int16_t4 test_min_short4(int16_t4 p0, int16_t4 p1) { return min(p0, p1); }
+// NATIVE_HALF-LABEL: define noundef <4 x i16> {{.*}}test_min_short4_mismatch
+// NATIVE_HALF: call <4 x i16> @llvm.smin.v4i16
+int16_t4 test_min_short4_mismatch(int16_t4 p0, int16_t p1) { return min(p0,
p1); }
// NATIVE_HALF-LABEL: define noundef i16 @_Z15test_min_ushort
// NATIVE_HALF: call i16 @llvm.umin.i16(
@@ -31,6 +34,9 @@ uint16_t3 test_min_ushort3(uint16_t3 p0, uint16_t3 p1) {
return min(p0, p1); }
// NATIVE_HALF-LABEL: define noundef <4 x i16> @_Z16test_min_ushort4
// NATIVE_HALF: call <4 x i16> @llvm.umin.v4i16
uint16_t4 test_min_ushort4(uint16_t4 p0, uint16_t4 p1) { return min(p0, p1); }
+// NATIVE_HALF-LABEL: define noundef <4 x i16> {{.*}}test_min_ushort4_mismatch
+// NATIVE_HALF: call <4 x i16> @llvm.umin.v4i16
+uint16_t4 test_min_ushort4_mismatch(uint16_t4 p0, uint16_t p1) { return
min(p0, p1); }
#endif
// CHECK-LABEL: define noundef i32 @_Z12test_min_int
@@ -45,6 +51,9 @@ int3 test_min_int3(int3 p0, int3 p1) { return min(p0, p1); }
// CHECK-LABEL: define noundef <4 x i32> @_Z13test_min_int4
// CHECK: call <4 x i32> @llvm.smin.v4i32
int4 test_min_int4(int4 p0, int4 p1) { return min(p0, p1); }
+// CHECK-LABEL: define noundef <4 x i32> {{.*}}test_min_int4_mismatch
+// CHECK: call <4 x i32> @llvm.smin.v4i32
+int4 test_min_int4_mismatch(int4 p0, int p1) { return min(p0, p1); }
// CHECK-LABEL: define noundef i32 @_Z13test_min_uint
// CHECK: call i32 @llvm.umin.i32(
@@ -58,6 +67,9 @@ uint3 test_min_uint3(uint3 p0, uint3 p1) { return min(p0,
p1); }
// CHECK-LABEL: define noundef <4 x i32> @_Z14test_min_uint4
// CHECK: call <4 x i32> @llvm.umin.v4i32
uint4 test_min_uint4(uint4 p0, uint4 p1) { return min(p0, p1); }
+// CHECK-LABEL: define noundef <4 x i32> {{.*}}test_min_uint4_mismatch
+// CHECK: call <4 x i32> @llvm.umin.v4i32
+uint4 test_min_uint4_mismatch(uint4 p0, uint p1) { return min(p0, p1); }
// CHECK-LABEL: define noundef i64 @_Z13test_min_long
// CHECK: call i64 @llvm.smin.i64(
@@ -71,6 +83,9 @@ int64_t3 test_min_long3(int64_t3 p0, int64_t3 p1) { return
min(p0, p1); }
// CHECK-LABEL: define noundef <4 x i64> @_Z14test_min_long4
// CHECK: call <4 x i64> @llvm.smin.v4i64
int64_t4 test_min_long4(int64_t4 p0, int64_t4 p1) { return min(p0, p1); }
+// CHECK-LABEL: define noundef <4 x i64> {{.*}}test_min_long4_mismatch
+// CHECK: call <4 x i64> @llvm.smin.v4i64
+int64_t4 test_min_long4_mismatch(int64_t4 p0, int64_t p1) { return min(p0,
p1); }
// CHECK-LABEL: define noundef i64 @_Z14test_min_ulong
// CHECK: call i64 @llvm.umin.i64(
@@ -84,6 +99,9 @@ uint64_t3 test_min_ulong3(uint64_t3 p0, uint64_t3 p1) {
return min(p0, p1); }
// CHECK-LABEL: define noundef <4 x i64> @_Z15test_min_ulong4
// CHECK: call <4 x i64> @llvm.umin.v4i64
uint64_t4 test_min_ulong4(uint64_t4 p0, uint64_t4 p1) { return min(p0, p1); }
+// CHECK-LABEL: define noundef <4 x i64> {{.*}}test_min_ulong4_mismatch
+// CHECK: call <4 x i64> @llvm.umin.v4i64
+uint64_t4 test_min_ulong4_mismatch(uint64_t4 p0, uint64_t p1) { return min(p0,
p1); }
// NATIVE_HALF-LABEL: define noundef nofpclass(nan inf) half @_Z13test_min_half
// NATIVE_HALF: call reassoc nnan ninf nsz arcp afn half @llvm.minnum.f16(
@@ -105,6 +123,11 @@ half3 test_min_half3(half3 p0, half3 p1) { return min(p0,
p1); }
// NO_HALF-LABEL: define noundef nofpclass(nan inf) <4 x float>
@_Z14test_min_half4
// NO_HALF: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.minnum.v4f32(
half4 test_min_half4(half4 p0, half4 p1) { return min(p0, p1); }
+// NATIVE_HALF-LABEL: define noundef nofpclass(nan inf) <4 x half>
{{.*}}test_min_half4_mismatch
+// NATIVE_HALF: call reassoc nnan ninf nsz arcp afn <4 x half>
@llvm.minnum.v4f16
+// NO_HALF-LABEL: define noundef nofpclass(nan inf) <4 x float>
{{.*}}test_min_half4_mismatch
+// NO_HALF: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.minnum.v4f32(
+half4 test_min_half4_mismatch(half4 p0, half p1) { return min(p0, p1); }
// CHECK-LABEL: define noundef nofpclass(nan inf) float @_Z14test_min_float
// CHECK: call reassoc nnan ninf nsz arcp afn float @llvm.minnum.f32(
@@ -118,6 +141,9 @@ float3 test_min_float3(float3 p0, float3 p1) { return
min(p0, p1); }
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x float>
@_Z15test_min_float4
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.minnum.v4f32
float4 test_min_float4(float4 p0, float4 p1) { return min(p0, p1); }
+// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x float>
{{.*}}test_min_float4_mismatch
+// CHECK: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.minnum.v4f32
+float4 test_min_float4_mismatch(float4 p0, float p1) { return min(p0, p1); }
// CHECK-LABEL: define noundef nofpclass(nan inf) double @_Z15test_min_double
// CHECK: call reassoc nnan ninf nsz arcp afn double @llvm.minnum.f64(
@@ -131,3 +157,9 @@ double3 test_min_double3(double3 p0, double3 p1) { return
min(p0, p1); }
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double>
@_Z16test_min_double4
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.minnum.v4f64
double4 test_min_double4(double4 p0, double4 p1) { return min(p0, p1); }
+// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double>
{{.*}}test_min_double4_mismatch
+// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.minnum.v4f64
+double4 test_min_double4_mismatch(double4 p0, double p1) { return min(p0, p1);
}
+// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double>
{{.*}}test_min_double4_mismatch2
+// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.minnum.v4f64
+double4 test_min_double4_mismatch2(double4 p0, double p1) { return min(p1,
p0); }
>From dabf38402b6ecf903a7065bb7bff74a13ce6fc1c Mon Sep 17 00:00:00 2001
From: Sarah Spall <sarahsp...@microsoft.com>
Date: Fri, 28 Feb 2025 16:46:27 -0800
Subject: [PATCH 3/3] use macro to make header file shorter
---
clang/lib/Headers/hlsl/hlsl_intrinsics.h | 373 +++--------------------
1 file changed, 37 insertions(+), 336 deletions(-)
diff --git a/clang/lib/Headers/hlsl/hlsl_intrinsics.h
b/clang/lib/Headers/hlsl/hlsl_intrinsics.h
index 68acdf1720318..7a75a31505376 100644
--- a/clang/lib/Headers/hlsl/hlsl_intrinsics.h
+++ b/clang/lib/Headers/hlsl/hlsl_intrinsics.h
@@ -37,6 +37,26 @@ namespace hlsl {
#define _HLSL_16BIT_AVAILABILITY_STAGE(environment, version, stage)
#endif
+#define GEN_VEC_SCALAR_OVERLOADS(FUNC_NAME, BASE_TYPE, AVAIL)
\
+ GEN_BOTH_OVERLOADS(FUNC_NAME, BASE_TYPE, BASE_TYPE##2, AVAIL)
\
+ GEN_BOTH_OVERLOADS(FUNC_NAME, BASE_TYPE, BASE_TYPE##3, AVAIL)
\
+ GEN_BOTH_OVERLOADS(FUNC_NAME, BASE_TYPE, BASE_TYPE##4, AVAIL)
+
+#define GEN_BOTH_OVERLOADS(FUNC_NAME, BASE_TYPE, VECTOR_TYPE, AVAIL)
\
+ IF_TRUE_##AVAIL(
\
+ _HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)) constexpr VECTOR_TYPE
\
+ FUNC_NAME(VECTOR_TYPE p0, BASE_TYPE p1) {
\
+ return __builtin_elementwise_##FUNC_NAME(p0, (VECTOR_TYPE)p1);
\
+ }
\
+ IF_TRUE_##AVAIL(
\
+ _HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)) constexpr VECTOR_TYPE
\
+ FUNC_NAME(BASE_TYPE p0, VECTOR_TYPE p1) {
\
+ return __builtin_elementwise_##FUNC_NAME((VECTOR_TYPE)p0, p1);
\
+ }
+
+#define IF_TRUE_0(EXPR)
+#define IF_TRUE_1(EXPR) EXPR
+
//===----------------------------------------------------------------------===//
// abs builtins
//===----------------------------------------------------------------------===//
@@ -1716,30 +1736,7 @@ int16_t3 max(int16_t3, int16_t3);
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int16_t4 max(int16_t4, int16_t4);
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t2 max(int16_t2 p0, int16_t p1) {
- return __builtin_elementwise_max(p0, (int16_t2)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t2 max(int16_t p0, int16_t2 p1) {
- return __builtin_elementwise_max((int16_t2)p0, p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t3 max(int16_t3 p0, int16_t p1) {
- return __builtin_elementwise_max(p0, (int16_t3)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t3 max(int16_t p0, int16_t3 p1) {
- return __builtin_elementwise_max((int16_t3)p0, p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t4 max(int16_t4 p0, int16_t p1) {
- return __builtin_elementwise_max(p0, (int16_t4)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t4 max(int16_t p0, int16_t4 p1) {
- return __builtin_elementwise_max((int16_t4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(max, int16_t, 1)
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
@@ -1753,30 +1750,7 @@ uint16_t3 max(uint16_t3, uint16_t3);
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint16_t4 max(uint16_t4, uint16_t4);
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t2 max(uint16_t2 p0, uint16_t p1) {
- return __builtin_elementwise_max(p0, (uint16_t2)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t2 max(uint16_t p0, uint16_t2 p1) {
- return __builtin_elementwise_max((uint16_t2)p0, p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t3 max(uint16_t3 p0, uint16_t p1) {
- return __builtin_elementwise_max(p0, (uint16_t3)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t3 max(uint16_t p0, uint16_t3 p1) {
- return __builtin_elementwise_max((uint16_t3)p0, p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t4 max(uint16_t4 p0, uint16_t p1) {
- return __builtin_elementwise_max(p0, (uint16_t4)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t4 max(uint16_t p0, uint16_t4 p1) {
- return __builtin_elementwise_max((uint16_t4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(max, uint16_t, 1)
#endif
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
@@ -1787,24 +1761,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int3 max(int3, int3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int4 max(int4, int4);
-constexpr int2 max(int2 p0, int p1) {
- return __builtin_elementwise_max(p0, (int2)p1);
-}
-constexpr int2 max(int p0, int2 p1) {
- return __builtin_elementwise_max((int2)p0, p1);
-}
-constexpr int3 max(int3 p0, int p1) {
- return __builtin_elementwise_max(p0, (int3)p1);
-}
-constexpr int3 max(int p0, int3 p1) {
- return __builtin_elementwise_max((int3)p0, p1);
-}
-constexpr int4 max(int4 p0, int p1) {
- return __builtin_elementwise_max(p0, (int4)p1);
-}
-constexpr int4 max(int p0, int4 p1) {
- return __builtin_elementwise_max((int4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(max, int, 0)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint max(uint, uint);
@@ -1814,24 +1771,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint3 max(uint3, uint3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint4 max(uint4, uint4);
-constexpr uint2 max(uint2 p0, uint p1) {
- return __builtin_elementwise_max(p0, (uint2)p1);
-}
-constexpr uint2 max(uint p0, uint2 p1) {
- return __builtin_elementwise_max((uint2)p0, p1);
-}
-constexpr uint3 max(uint3 p0, uint p1) {
- return __builtin_elementwise_max(p0, (uint3)p1);
-}
-constexpr uint3 max(uint p0, uint3 p1) {
- return __builtin_elementwise_max((uint3)p0, p1);
-}
-constexpr uint4 max(uint4 p0, uint p1) {
- return __builtin_elementwise_max(p0, (uint4)p1);
-}
-constexpr uint4 max(uint p0, uint4 p1) {
- return __builtin_elementwise_max((uint4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(max, uint, 0)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int64_t max(int64_t, int64_t);
@@ -1841,24 +1781,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int64_t3 max(int64_t3, int64_t3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
int64_t4 max(int64_t4, int64_t4);
-constexpr int64_t2 max(int64_t2 p0, int64_t p1) {
- return __builtin_elementwise_max(p0, (int64_t2)p1);
-}
-constexpr int64_t2 max(int64_t p0, int64_t2 p1) {
- return __builtin_elementwise_max((int64_t2)p0, p1);
-}
-constexpr int64_t3 max(int64_t3 p0, int64_t p1) {
- return __builtin_elementwise_max(p0, (int64_t3)p1);
-}
-constexpr int64_t3 max(int64_t p0, int64_t3 p1) {
- return __builtin_elementwise_max((int64_t3)p0, p1);
-}
-constexpr int64_t4 max(int64_t4 p0, int64_t p1) {
- return __builtin_elementwise_max(p0, (int64_t4)p1);
-}
-constexpr int64_t4 max(int64_t p0, int64_t4 p1) {
- return __builtin_elementwise_max((int64_t4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(max, int64_t, 0)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint64_t max(uint64_t, uint64_t);
@@ -1868,24 +1791,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint64_t3 max(uint64_t3, uint64_t3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
uint64_t4 max(uint64_t4, uint64_t4);
-constexpr uint64_t2 max(uint64_t2 p0, uint64_t p1) {
- return __builtin_elementwise_max(p0, (uint64_t2)p1);
-}
-constexpr uint64_t2 max(uint64_t p0, uint64_t2 p1) {
- return __builtin_elementwise_max((uint64_t2)p0, p1);
-}
-constexpr uint64_t3 max(uint64_t3 p0, uint64_t p1) {
- return __builtin_elementwise_max(p0, (uint64_t3)p1);
-}
-constexpr uint64_t3 max(uint64_t p0, uint64_t3 p1) {
- return __builtin_elementwise_max((uint64_t3)p0, p1);
-}
-constexpr uint64_t4 max(uint64_t4 p0, uint64_t p1) {
- return __builtin_elementwise_max(p0, (uint64_t4)p1);
-}
-constexpr uint64_t4 max(uint64_t p0, uint64_t4 p1) {
- return __builtin_elementwise_max((uint64_t4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(max, uint64_t, 0)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
float max(float, float);
@@ -1895,24 +1801,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
float3 max(float3, float3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
float4 max(float4, float4);
-constexpr float2 max(float2 p0, float p1) {
- return __builtin_elementwise_max(p0, (float2)p1);
-}
-constexpr float2 max(float p0, float2 p1) {
- return __builtin_elementwise_max((float2)p0, p1);
-}
-constexpr float3 max(float3 p0, float p1) {
- return __builtin_elementwise_max(p0, (float3)p1);
-}
-constexpr float3 max(float p0, float3 p1) {
- return __builtin_elementwise_max((float3)p0, p1);
-}
-constexpr float4 max(float4 p0, float p1) {
- return __builtin_elementwise_max(p0, (float4)p1);
-}
-constexpr float4 max(float p0, float4 p1) {
- return __builtin_elementwise_max((float4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(max, float, 0)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
double max(double, double);
@@ -1922,24 +1811,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
double3 max(double3, double3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
double4 max(double4, double4);
-constexpr double2 max(double2 p0, double p1) {
- return __builtin_elementwise_max(p0, (double2)p1);
-}
-constexpr double2 max(double p0, double2 p1) {
- return __builtin_elementwise_max((double2)p0, p1);
-}
-constexpr double3 max(double3 p0, double p1) {
- return __builtin_elementwise_max(p0, (double3)p1);
-}
-constexpr double3 max(double p0, double3 p1) {
- return __builtin_elementwise_max((double3)p0, p1);
-}
-constexpr double4 max(double4 p0, double p1) {
- return __builtin_elementwise_max(p0, (double4)p1);
-}
-constexpr double4 max(double p0, double4 p1) {
- return __builtin_elementwise_max((double4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(max, double, 0)
//===----------------------------------------------------------------------===//
// min builtins
@@ -1962,30 +1834,7 @@ half3 min(half3, half3);
_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
half4 min(half4, half4);
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
-constexpr half2 min(half2 p0, half p1) {
- return __builtin_elementwise_min(p0, (half2)p1);
-}
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
-constexpr half2 min(half p0, half2 p1) {
- return __builtin_elementwise_min((half2)p0, p1);
-}
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
-constexpr half3 min(half3 p0, half p1) {
- return __builtin_elementwise_min(p0, (half3)p1);
-}
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
-constexpr half3 min(half p0, half3 p1) {
- return __builtin_elementwise_min((half3)p0, p1);
-}
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
-constexpr half4 min(half4 p0, half p1) {
- return __builtin_elementwise_min(p0, (half4)p1);
-}
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
-constexpr half4 min(half p0, half4 p1) {
- return __builtin_elementwise_min((half4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(min, half, 1)
#ifdef __HLSL_ENABLE_16_BIT
_HLSL_AVAILABILITY(shadermodel, 6.2)
@@ -2000,30 +1849,7 @@ int16_t3 min(int16_t3, int16_t3);
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int16_t4 min(int16_t4, int16_t4);
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t2 min(int16_t2 p0, int16_t p1) {
- return __builtin_elementwise_min(p0, (int16_t2)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t2 min(int16_t p0, int16_t2 p1) {
- return __builtin_elementwise_min((int16_t2)p0, p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t3 min(int16_t3 p0, int16_t p1) {
- return __builtin_elementwise_min(p0, (int16_t3)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t3 min(int16_t p0, int16_t3 p1) {
- return __builtin_elementwise_min((int16_t3)p0, p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t4 min(int16_t4 p0, int16_t p1) {
- return __builtin_elementwise_min(p0, (int16_t4)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr int16_t4 min(int16_t p0, int16_t4 p1) {
- return __builtin_elementwise_min((int16_t4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(min, int16_t, 1)
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
@@ -2037,30 +1863,7 @@ uint16_t3 min(uint16_t3, uint16_t3);
_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint16_t4 min(uint16_t4, uint16_t4);
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t2 min(uint16_t2 p0, uint16_t p1) {
- return __builtin_elementwise_min(p0, (uint16_t2)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t2 min(uint16_t p0, uint16_t2 p1) {
- return __builtin_elementwise_min((uint16_t2)p0, p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t3 min(uint16_t3 p0, uint16_t p1) {
- return __builtin_elementwise_min(p0, (uint16_t3)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t3 min(uint16_t p0, uint16_t3 p1) {
- return __builtin_elementwise_min((uint16_t3)p0, p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t4 min(uint16_t4 p0, uint16_t p1) {
- return __builtin_elementwise_min(p0, (uint16_t4)p1);
-}
-_HLSL_AVAILABILITY(shadermodel, 6.2)
-constexpr uint16_t4 min(uint16_t p0, uint16_t4 p1) {
- return __builtin_elementwise_min((uint16_t4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(min, uint16_t, 1)
#endif
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
@@ -2071,24 +1874,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int3 min(int3, int3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int4 min(int4, int4);
-constexpr int2 min(int2 p0, int p1) {
- return __builtin_elementwise_min(p0, (int2)p1);
-}
-constexpr int2 min(int p0, int2 p1) {
- return __builtin_elementwise_min((int2)p0, p1);
-}
-constexpr int3 min(int3 p0, int p1) {
- return __builtin_elementwise_min(p0, (int3)p1);
-}
-constexpr int3 min(int p0, int3 p1) {
- return __builtin_elementwise_min((int3)p0, p1);
-}
-constexpr int4 min(int4 p0, int p1) {
- return __builtin_elementwise_min(p0, (int4)p1);
-}
-constexpr int4 min(int p0, int4 p1) {
- return __builtin_elementwise_min((int4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(min, int, 0)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint min(uint, uint);
@@ -2098,24 +1884,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint3 min(uint3, uint3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint4 min(uint4, uint4);
-constexpr uint2 min(uint2 p0, uint p1) {
- return __builtin_elementwise_min(p0, (uint2)p1);
-}
-constexpr uint2 min(uint p0, uint2 p1) {
- return __builtin_elementwise_min((uint2)p0, p1);
-}
-constexpr uint3 min(uint3 p0, uint p1) {
- return __builtin_elementwise_min(p0, (uint3)p1);
-}
-constexpr uint3 min(uint p0, uint3 p1) {
- return __builtin_elementwise_min((uint3)p0, p1);
-}
-constexpr uint4 min(uint4 p0, uint p1) {
- return __builtin_elementwise_min(p0, (uint4)p1);
-}
-constexpr uint4 min(uint p0, uint4 p1) {
- return __builtin_elementwise_min((uint4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(min, uint, 0)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
float min(float, float);
@@ -2125,24 +1894,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
float3 min(float3, float3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
float4 min(float4, float4);
-constexpr float2 min(float2 p0, float p1) {
- return __builtin_elementwise_min(p0, (float2)p1);
-}
-constexpr float2 min(float p0, float2 p1) {
- return __builtin_elementwise_min((float2)p0, p1);
-}
-constexpr float3 min(float3 p0, float p1) {
- return __builtin_elementwise_min(p0, (float3)p1);
-}
-constexpr float3 min(float p0, float3 p1) {
- return __builtin_elementwise_min((float3)p0, p1);
-}
-constexpr float4 min(float4 p0, float p1) {
- return __builtin_elementwise_min(p0, (float4)p1);
-}
-constexpr float4 min(float p0, float4 p1) {
- return __builtin_elementwise_min((float4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(min, float, 0)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int64_t min(int64_t, int64_t);
@@ -2152,24 +1904,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int64_t3 min(int64_t3, int64_t3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
int64_t4 min(int64_t4, int64_t4);
-constexpr int64_t2 min(int64_t2 p0, int64_t p1) {
- return __builtin_elementwise_min(p0, (int64_t2)p1);
-}
-constexpr int64_t2 min(int64_t p0, int64_t2 p1) {
- return __builtin_elementwise_min((int64_t2)p0, p1);
-}
-constexpr int64_t3 min(int64_t3 p0, int64_t p1) {
- return __builtin_elementwise_min(p0, (int64_t3)p1);
-}
-constexpr int64_t3 min(int64_t p0, int64_t3 p1) {
- return __builtin_elementwise_min((int64_t3)p0, p1);
-}
-constexpr int64_t4 min(int64_t4 p0, int64_t p1) {
- return __builtin_elementwise_min(p0, (int64_t4)p1);
-}
-constexpr int64_t4 min(int64_t p0, int64_t4 p1) {
- return __builtin_elementwise_min((int64_t4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(min, int64_t, 0)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint64_t min(uint64_t, uint64_t);
@@ -2179,24 +1914,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint64_t3 min(uint64_t3, uint64_t3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
uint64_t4 min(uint64_t4, uint64_t4);
-constexpr uint64_t2 min(uint64_t2 p0, uint64_t p1) {
- return __builtin_elementwise_min(p0, (uint64_t2)p1);
-}
-constexpr uint64_t2 min(uint64_t p0, uint64_t2 p1) {
- return __builtin_elementwise_min((uint64_t2)p0, p1);
-}
-constexpr uint64_t3 min(uint64_t3 p0, uint64_t p1) {
- return __builtin_elementwise_min(p0, (uint64_t3)p1);
-}
-constexpr uint64_t3 min(uint64_t p0, uint64_t3 p1) {
- return __builtin_elementwise_min((uint64_t3)p0, p1);
-}
-constexpr uint64_t4 min(uint64_t4 p0, uint64_t p1) {
- return __builtin_elementwise_min(p0, (uint64_t4)p1);
-}
-constexpr uint64_t4 min(uint64_t p0, uint64_t4 p1) {
- return __builtin_elementwise_min((uint64_t4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(min, uint64_t, 0)
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
double min(double, double);
@@ -2206,24 +1924,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
double3 min(double3, double3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
double4 min(double4, double4);
-constexpr double2 min(double2 p0, double p1) {
- return __builtin_elementwise_min(p0, (double2)p1);
-}
-constexpr double2 min(double p0, double2 p1) {
- return __builtin_elementwise_min((double2)p0, p1);
-}
-constexpr double3 min(double3 p0, double p1) {
- return __builtin_elementwise_min(p0, (double3)p1);
-}
-constexpr double3 min(double p0, double3 p1) {
- return __builtin_elementwise_min((double3)p0, p1);
-}
-constexpr double4 min(double4 p0, double p1) {
- return __builtin_elementwise_min(p0, (double4)p1);
-}
-constexpr double4 min(double p0, double4 p1) {
- return __builtin_elementwise_min((double4)p0, p1);
-}
+GEN_VEC_SCALAR_OVERLOADS(min, double, 0)
//===----------------------------------------------------------------------===//
// normalize builtins
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
