Skip to content

Commit 25713ed

Browse files
spallspall
authored
[HLSL] Add additional overloads for min and max to allow for mixed scalar and vector arguments (#129334)
Add additional overloads for min and max to support min(vector<T,N>, T) and min(T, vector<T,N>) max(vector<T,N>, T) and max(T, vector<T,N>) Add tests Closes #128231
1 parent dc1ff41 commit 25713ed

File tree

3 files changed

+102
-0
lines changed

3 files changed

+102
-0
lines changed

‎clang/lib/Headers/hlsl/hlsl_intrinsics.h‎

Lines changed: 38 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -37,6 +37,26 @@ namespace hlsl {
3737
#define _HLSL_16BIT_AVAILABILITY_STAGE(environment, version, stage)
3838
#endif
3939

40+
#define GEN_VEC_SCALAR_OVERLOADS(FUNC_NAME, BASE_TYPE, AVAIL) \
41+
GEN_BOTH_OVERLOADS(FUNC_NAME, BASE_TYPE, BASE_TYPE##2, AVAIL) \
42+
GEN_BOTH_OVERLOADS(FUNC_NAME, BASE_TYPE, BASE_TYPE##3, AVAIL) \
43+
GEN_BOTH_OVERLOADS(FUNC_NAME, BASE_TYPE, BASE_TYPE##4, AVAIL)
44+
45+
#define GEN_BOTH_OVERLOADS(FUNC_NAME, BASE_TYPE, VECTOR_TYPE, AVAIL) \
46+
IF_TRUE_##AVAIL( \
47+
_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)) constexpr VECTOR_TYPE \
48+
FUNC_NAME(VECTOR_TYPE p0, BASE_TYPE p1) { \
49+
return __builtin_elementwise_##FUNC_NAME(p0, (VECTOR_TYPE)p1); \
50+
} \
51+
IF_TRUE_##AVAIL( \
52+
_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)) constexpr VECTOR_TYPE \
53+
FUNC_NAME(BASE_TYPE p0, VECTOR_TYPE p1) { \
54+
return __builtin_elementwise_##FUNC_NAME((VECTOR_TYPE)p0, p1); \
55+
}
56+
57+
#define IF_TRUE_0(EXPR)
58+
#define IF_TRUE_1(EXPR) EXPR
59+
4060
//===----------------------------------------------------------------------===//
4161
// abs builtins
4262
//===----------------------------------------------------------------------===//
@@ -1700,6 +1720,7 @@ half3 max(half3, half3);
17001720
_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
17011721
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17021722
half4 max(half4, half4);
1723+
GEN_VEC_SCALAR_OVERLOADS(max, half, 1)
17031724

17041725
#ifdef __HLSL_ENABLE_16_BIT
17051726
_HLSL_AVAILABILITY(shadermodel, 6.2)
@@ -1714,6 +1735,7 @@ int16_t3 max(int16_t3, int16_t3);
17141735
_HLSL_AVAILABILITY(shadermodel, 6.2)
17151736
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17161737
int16_t4 max(int16_t4, int16_t4);
1738+
GEN_VEC_SCALAR_OVERLOADS(max, int16_t, 1)
17171739

17181740
_HLSL_AVAILABILITY(shadermodel, 6.2)
17191741
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
@@ -1727,6 +1749,7 @@ uint16_t3 max(uint16_t3, uint16_t3);
17271749
_HLSL_AVAILABILITY(shadermodel, 6.2)
17281750
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17291751
uint16_t4 max(uint16_t4, uint16_t4);
1752+
GEN_VEC_SCALAR_OVERLOADS(max, uint16_t, 1)
17301753
#endif
17311754

17321755
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
@@ -1737,6 +1760,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17371760
int3 max(int3, int3);
17381761
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17391762
int4 max(int4, int4);
1763+
GEN_VEC_SCALAR_OVERLOADS(max, int, 0)
17401764

17411765
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17421766
uint max(uint, uint);
@@ -1746,6 +1770,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17461770
uint3 max(uint3, uint3);
17471771
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17481772
uint4 max(uint4, uint4);
1773+
GEN_VEC_SCALAR_OVERLOADS(max, uint, 0)
17491774

17501775
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17511776
int64_t max(int64_t, int64_t);
@@ -1755,6 +1780,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17551780
int64_t3 max(int64_t3, int64_t3);
17561781
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17571782
int64_t4 max(int64_t4, int64_t4);
1783+
GEN_VEC_SCALAR_OVERLOADS(max, int64_t, 0)
17581784

17591785
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17601786
uint64_t max(uint64_t, uint64_t);
@@ -1764,6 +1790,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17641790
uint64_t3 max(uint64_t3, uint64_t3);
17651791
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17661792
uint64_t4 max(uint64_t4, uint64_t4);
1793+
GEN_VEC_SCALAR_OVERLOADS(max, uint64_t, 0)
17671794

17681795
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17691796
float max(float, float);
@@ -1773,6 +1800,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17731800
float3 max(float3, float3);
17741801
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17751802
float4 max(float4, float4);
1803+
GEN_VEC_SCALAR_OVERLOADS(max, float, 0)
17761804

17771805
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17781806
double max(double, double);
@@ -1782,6 +1810,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17821810
double3 max(double3, double3);
17831811
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_max)
17841812
double4 max(double4, double4);
1813+
GEN_VEC_SCALAR_OVERLOADS(max, double, 0)
17851814

17861815
//===----------------------------------------------------------------------===//
17871816
// min builtins
@@ -1804,6 +1833,7 @@ half3 min(half3, half3);
18041833
_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
18051834
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18061835
half4 min(half4, half4);
1836+
GEN_VEC_SCALAR_OVERLOADS(min, half, 1)
18071837

18081838
#ifdef __HLSL_ENABLE_16_BIT
18091839
_HLSL_AVAILABILITY(shadermodel, 6.2)
@@ -1818,6 +1848,7 @@ int16_t3 min(int16_t3, int16_t3);
18181848
_HLSL_AVAILABILITY(shadermodel, 6.2)
18191849
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18201850
int16_t4 min(int16_t4, int16_t4);
1851+
GEN_VEC_SCALAR_OVERLOADS(min, int16_t, 1)
18211852

18221853
_HLSL_AVAILABILITY(shadermodel, 6.2)
18231854
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
@@ -1831,6 +1862,7 @@ uint16_t3 min(uint16_t3, uint16_t3);
18311862
_HLSL_AVAILABILITY(shadermodel, 6.2)
18321863
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18331864
uint16_t4 min(uint16_t4, uint16_t4);
1865+
GEN_VEC_SCALAR_OVERLOADS(min, uint16_t, 1)
18341866
#endif
18351867

18361868
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
@@ -1841,6 +1873,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18411873
int3 min(int3, int3);
18421874
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18431875
int4 min(int4, int4);
1876+
GEN_VEC_SCALAR_OVERLOADS(min, int, 0)
18441877

18451878
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18461879
uint min(uint, uint);
@@ -1850,6 +1883,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18501883
uint3 min(uint3, uint3);
18511884
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18521885
uint4 min(uint4, uint4);
1886+
GEN_VEC_SCALAR_OVERLOADS(min, uint, 0)
18531887

18541888
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18551889
float min(float, float);
@@ -1859,6 +1893,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18591893
float3 min(float3, float3);
18601894
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18611895
float4 min(float4, float4);
1896+
GEN_VEC_SCALAR_OVERLOADS(min, float, 0)
18621897

18631898
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18641899
int64_t min(int64_t, int64_t);
@@ -1868,6 +1903,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18681903
int64_t3 min(int64_t3, int64_t3);
18691904
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18701905
int64_t4 min(int64_t4, int64_t4);
1906+
GEN_VEC_SCALAR_OVERLOADS(min, int64_t, 0)
18711907

18721908
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18731909
uint64_t min(uint64_t, uint64_t);
@@ -1877,6 +1913,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18771913
uint64_t3 min(uint64_t3, uint64_t3);
18781914
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18791915
uint64_t4 min(uint64_t4, uint64_t4);
1916+
GEN_VEC_SCALAR_OVERLOADS(min, uint64_t, 0)
18801917

18811918
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18821919
double min(double, double);
@@ -1886,6 +1923,7 @@ _HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18861923
double3 min(double3, double3);
18871924
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_min)
18881925
double4 min(double4, double4);
1926+
GEN_VEC_SCALAR_OVERLOADS(min, double, 0)
18891927

18901928
//===----------------------------------------------------------------------===//
18911929
// normalize builtins

‎clang/test/CodeGenHLSL/builtins/max.hlsl‎

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -18,6 +18,9 @@ int16_t3 test_max_short3(int16_t3 p0, int16_t3 p1) { return max(p0, p1); }
1818
// NATIVE_HALF-LABEL: define noundef <4 x i16> @_Z15test_max_short4
1919
// NATIVE_HALF: call <4 x i16> @llvm.smax.v4i16
2020
int16_t4 test_max_short4(int16_t4 p0, int16_t4 p1) { return max(p0, p1); }
21+
// NATIVE_HALF-LABEL: define noundef <4 x i16> {{.*}}test_max_short4_mismatch
22+
// NATIVE_HALF: call <4 x i16> @llvm.smax.v4i16
23+
int16_t4 test_max_short4_mismatch(int16_t4 p0, int16_t p1) { return max(p0, p1); }
2124

2225
// NATIVE_HALF-LABEL: define noundef i16 @_Z15test_max_ushort
2326
// 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); }
3134
// NATIVE_HALF-LABEL: define noundef <4 x i16> @_Z16test_max_ushort4
3235
// NATIVE_HALF: call <4 x i16> @llvm.umax.v4i16
3336
uint16_t4 test_max_ushort4(uint16_t4 p0, uint16_t4 p1) { return max(p0, p1); }
37+
// NATIVE_HALF-LABEL: define noundef <4 x i16> {{.*}}test_max_ushort4_mismatch
38+
// NATIVE_HALF: call <4 x i16> @llvm.umax.v4i16
39+
uint16_t4 test_max_ushort4_mismatch(uint16_t4 p0, uint16_t p1) { return max(p0, p1); }
3440
#endif
3541

3642
// CHECK-LABEL: define noundef i32 @_Z12test_max_int
@@ -45,6 +51,9 @@ int3 test_max_int3(int3 p0, int3 p1) { return max(p0, p1); }
4551
// CHECK-LABEL: define noundef <4 x i32> @_Z13test_max_int4
4652
// CHECK: call <4 x i32> @llvm.smax.v4i32
4753
int4 test_max_int4(int4 p0, int4 p1) { return max(p0, p1); }
54+
// CHECK-LABEL: define noundef <4 x i32> {{.*}}test_max_int4_mismatch
55+
// CHECK: call <4 x i32> @llvm.smax.v4i32
56+
int4 test_max_int4_mismatch(int4 p0, int p1) { return max(p0, p1); }
4857

4958
// CHECK-LABEL: define noundef i32 @_Z13test_max_uint
5059
// CHECK: call i32 @llvm.umax.i32(
@@ -58,6 +67,9 @@ uint3 test_max_uint3(uint3 p0, uint3 p1) { return max(p0, p1); }
5867
// CHECK-LABEL: define noundef <4 x i32> @_Z14test_max_uint4
5968
// CHECK: call <4 x i32> @llvm.umax.v4i32
6069
uint4 test_max_uint4(uint4 p0, uint4 p1) { return max(p0, p1); }
70+
// CHECK-LABEL: define noundef <4 x i32> {{.*}}test_max_uint4_mismatch
71+
// CHECK: call <4 x i32> @llvm.umax.v4i32
72+
uint4 test_max_uint4_mismatch(uint4 p0, uint p1) { return max(p0, p1); }
6173

6274
// CHECK-LABEL: define noundef i64 @_Z13test_max_long
6375
// 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); }
7183
// CHECK-LABEL: define noundef <4 x i64> @_Z14test_max_long4
7284
// CHECK: call <4 x i64> @llvm.smax.v4i64
7385
int64_t4 test_max_long4(int64_t4 p0, int64_t4 p1) { return max(p0, p1); }
86+
// CHECK-LABEL: define noundef <4 x i64> {{.*}}test_max_long4_mismatch
87+
// CHECK: call <4 x i64> @llvm.smax.v4i64
88+
int64_t4 test_max_long4_mismatch(int64_t4 p0, int64_t p1) { return max(p0, p1); }
7489

7590
// CHECK-LABEL: define noundef i64 @_Z14test_max_ulong
7691
// 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); }
8499
// CHECK-LABEL: define noundef <4 x i64> @_Z15test_max_ulong4
85100
// CHECK: call <4 x i64> @llvm.umax.v4i64
86101
uint64_t4 test_max_ulong4(uint64_t4 p0, uint64_t4 p1) { return max(p0, p1); }
102+
// CHECK-LABEL: define noundef <4 x i64> {{.*}}test_max_ulong4_mismatch
103+
// CHECK: call <4 x i64> @llvm.umax.v4i64
104+
uint64_t4 test_max_ulong4_mismatch(uint64_t4 p0, uint64_t p1) { return max(p0, p1); }
87105

88106
// NATIVE_HALF-LABEL: define noundef nofpclass(nan inf) half @_Z13test_max_half
89107
// 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); }
105123
// NO_HALF-LABEL: define noundef nofpclass(nan inf) <4 x float> @_Z14test_max_half4
106124
// NO_HALF: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.maxnum.v4f32(
107125
half4 test_max_half4(half4 p0, half4 p1) { return max(p0, p1); }
126+
// NATIVE_HALF-LABEL: define noundef nofpclass(nan inf) <4 x half> {{.*}}test_max_half4_mismatch
127+
// NATIVE_HALF: call reassoc nnan ninf nsz arcp afn <4 x half> @llvm.maxnum.v4f16
128+
// NO_HALF-LABEL: define noundef nofpclass(nan inf) <4 x float> {{.*}}test_max_half4_mismatch
129+
// NO_HALF: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.maxnum.v4f32(
130+
half4 test_max_half4_mismatch(half4 p0, half p1) { return max(p0, p1); }
108131

109132
// CHECK-LABEL: define noundef nofpclass(nan inf) float @_Z14test_max_float
110133
// 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); }
118141
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x float> @_Z15test_max_float4
119142
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.maxnum.v4f32
120143
float4 test_max_float4(float4 p0, float4 p1) { return max(p0, p1); }
144+
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x float> {{.*}}test_max_float4_mismatch
145+
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.maxnum.v4f32
146+
float4 test_max_float4_mismatch(float4 p0, float p1) { return max(p0, p1); }
121147

122148
// CHECK-LABEL: define noundef nofpclass(nan inf) double @_Z15test_max_double
123149
// 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); }
131157
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double> @_Z16test_max_double4
132158
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.maxnum.v4f64
133159
double4 test_max_double4(double4 p0, double4 p1) { return max(p0, p1); }
160+
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double> {{.*}}test_max_double4_mismatch
161+
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.maxnum.v4f64
162+
double4 test_max_double4_mismatch(double4 p0, double p1) { return max(p0, p1); }
163+
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double> {{.*}}test_max_double4_mismatch2
164+
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.maxnum.v4f64
165+
double4 test_max_double4_mismatch2(double4 p0, double p1) { return max(p1, p0); }

‎clang/test/CodeGenHLSL/builtins/min.hlsl‎

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -18,6 +18,9 @@ int16_t3 test_min_short3(int16_t3 p0, int16_t3 p1) { return min(p0, p1); }
1818
// NATIVE_HALF-LABEL: define noundef <4 x i16> @_Z15test_min_short4
1919
// NATIVE_HALF: call <4 x i16> @llvm.smin.v4i16
2020
int16_t4 test_min_short4(int16_t4 p0, int16_t4 p1) { return min(p0, p1); }
21+
// NATIVE_HALF-LABEL: define noundef <4 x i16> {{.*}}test_min_short4_mismatch
22+
// NATIVE_HALF: call <4 x i16> @llvm.smin.v4i16
23+
int16_t4 test_min_short4_mismatch(int16_t4 p0, int16_t p1) { return min(p0, p1); }
2124

2225
// NATIVE_HALF-LABEL: define noundef i16 @_Z15test_min_ushort
2326
// 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); }
3134
// NATIVE_HALF-LABEL: define noundef <4 x i16> @_Z16test_min_ushort4
3235
// NATIVE_HALF: call <4 x i16> @llvm.umin.v4i16
3336
uint16_t4 test_min_ushort4(uint16_t4 p0, uint16_t4 p1) { return min(p0, p1); }
37+
// NATIVE_HALF-LABEL: define noundef <4 x i16> {{.*}}test_min_ushort4_mismatch
38+
// NATIVE_HALF: call <4 x i16> @llvm.umin.v4i16
39+
uint16_t4 test_min_ushort4_mismatch(uint16_t4 p0, uint16_t p1) { return min(p0, p1); }
3440
#endif
3541

3642
// CHECK-LABEL: define noundef i32 @_Z12test_min_int
@@ -45,6 +51,9 @@ int3 test_min_int3(int3 p0, int3 p1) { return min(p0, p1); }
4551
// CHECK-LABEL: define noundef <4 x i32> @_Z13test_min_int4
4652
// CHECK: call <4 x i32> @llvm.smin.v4i32
4753
int4 test_min_int4(int4 p0, int4 p1) { return min(p0, p1); }
54+
// CHECK-LABEL: define noundef <4 x i32> {{.*}}test_min_int4_mismatch
55+
// CHECK: call <4 x i32> @llvm.smin.v4i32
56+
int4 test_min_int4_mismatch(int4 p0, int p1) { return min(p0, p1); }
4857

4958
// CHECK-LABEL: define noundef i32 @_Z13test_min_uint
5059
// CHECK: call i32 @llvm.umin.i32(
@@ -58,6 +67,9 @@ uint3 test_min_uint3(uint3 p0, uint3 p1) { return min(p0, p1); }
5867
// CHECK-LABEL: define noundef <4 x i32> @_Z14test_min_uint4
5968
// CHECK: call <4 x i32> @llvm.umin.v4i32
6069
uint4 test_min_uint4(uint4 p0, uint4 p1) { return min(p0, p1); }
70+
// CHECK-LABEL: define noundef <4 x i32> {{.*}}test_min_uint4_mismatch
71+
// CHECK: call <4 x i32> @llvm.umin.v4i32
72+
uint4 test_min_uint4_mismatch(uint4 p0, uint p1) { return min(p0, p1); }
6173

6274
// CHECK-LABEL: define noundef i64 @_Z13test_min_long
6375
// 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); }
7183
// CHECK-LABEL: define noundef <4 x i64> @_Z14test_min_long4
7284
// CHECK: call <4 x i64> @llvm.smin.v4i64
7385
int64_t4 test_min_long4(int64_t4 p0, int64_t4 p1) { return min(p0, p1); }
86+
// CHECK-LABEL: define noundef <4 x i64> {{.*}}test_min_long4_mismatch
87+
// CHECK: call <4 x i64> @llvm.smin.v4i64
88+
int64_t4 test_min_long4_mismatch(int64_t4 p0, int64_t p1) { return min(p0, p1); }
7489

7590
// CHECK-LABEL: define noundef i64 @_Z14test_min_ulong
7691
// 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); }
8499
// CHECK-LABEL: define noundef <4 x i64> @_Z15test_min_ulong4
85100
// CHECK: call <4 x i64> @llvm.umin.v4i64
86101
uint64_t4 test_min_ulong4(uint64_t4 p0, uint64_t4 p1) { return min(p0, p1); }
102+
// CHECK-LABEL: define noundef <4 x i64> {{.*}}test_min_ulong4_mismatch
103+
// CHECK: call <4 x i64> @llvm.umin.v4i64
104+
uint64_t4 test_min_ulong4_mismatch(uint64_t4 p0, uint64_t p1) { return min(p0, p1); }
87105

88106
// NATIVE_HALF-LABEL: define noundef nofpclass(nan inf) half @_Z13test_min_half
89107
// 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); }
105123
// NO_HALF-LABEL: define noundef nofpclass(nan inf) <4 x float> @_Z14test_min_half4
106124
// NO_HALF: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.minnum.v4f32(
107125
half4 test_min_half4(half4 p0, half4 p1) { return min(p0, p1); }
126+
// NATIVE_HALF-LABEL: define noundef nofpclass(nan inf) <4 x half> {{.*}}test_min_half4_mismatch
127+
// NATIVE_HALF: call reassoc nnan ninf nsz arcp afn <4 x half> @llvm.minnum.v4f16
128+
// NO_HALF-LABEL: define noundef nofpclass(nan inf) <4 x float> {{.*}}test_min_half4_mismatch
129+
// NO_HALF: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.minnum.v4f32(
130+
half4 test_min_half4_mismatch(half4 p0, half p1) { return min(p0, p1); }
108131

109132
// CHECK-LABEL: define noundef nofpclass(nan inf) float @_Z14test_min_float
110133
// 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); }
118141
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x float> @_Z15test_min_float4
119142
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.minnum.v4f32
120143
float4 test_min_float4(float4 p0, float4 p1) { return min(p0, p1); }
144+
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x float> {{.*}}test_min_float4_mismatch
145+
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.minnum.v4f32
146+
float4 test_min_float4_mismatch(float4 p0, float p1) { return min(p0, p1); }
121147

122148
// CHECK-LABEL: define noundef nofpclass(nan inf) double @_Z15test_min_double
123149
// 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); }
131157
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double> @_Z16test_min_double4
132158
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.minnum.v4f64
133159
double4 test_min_double4(double4 p0, double4 p1) { return min(p0, p1); }
160+
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double> {{.*}}test_min_double4_mismatch
161+
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.minnum.v4f64
162+
double4 test_min_double4_mismatch(double4 p0, double p1) { return min(p0, p1); }
163+
// CHECK-LABEL: define noundef nofpclass(nan inf) <4 x double> {{.*}}test_min_double4_mismatch2
164+
// CHECK: call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.minnum.v4f64
165+
double4 test_min_double4_mismatch2(double4 p0, double p1) { return min(p1, p0); }

0 commit comments

Comments
 (0)