Skip to content

Commit e4106de

Browse files
committed
[naga wgsl-out] Improve representation of minimum int64 literal
#7424 fixed a bug where the representation of the minimum int64 literal generated by naga was invalid WGSL. It changed us from expressing it as `-9223372036854775808` which was invalid, to `-9223372036854775807li - 1li`. This is valid WGSL. However, as the values are concrete i64 types if the shader is parsed again by naga the expression does not get const evaluated away, leading to suboptimal code generated by the backends. This patch makes us perform the subtraction using abstract integers before casting to i64, solving this problem. Additionally the input WGSL test is updated to use the same construct.
1 parent 90afc88 commit e4106de

File tree

6 files changed

+230
-230
lines changed

6 files changed

+230
-230
lines changed

naga/src/back/wgsl/writer.rs

Lines changed: 5 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -1147,12 +1147,12 @@ impl<W: Write> Writer<W> {
11471147
crate::Literal::Bool(value) => write!(self.out, "{value}")?,
11481148
crate::Literal::F64(value) => write!(self.out, "{value:?}lf")?,
11491149
crate::Literal::I64(value) => {
1150-
// `-9223372036854775808li` is not valid WGSL. Nor can we use the AbstractInt
1151-
// trick above, as AbstractInt also cannot represent `9223372036854775808`.
1152-
// The most negative `i64` value can only be expressed in WGSL using
1153-
// subtracting 1 from the second most negative value.
1150+
// `-9223372036854775808li` is not valid WGSL. Nor can we simply use the
1151+
// AbstractInt trick above, as AbstractInt also cannot represent
1152+
// `9223372036854775808`. Instead construct the second most negative
1153+
// AbstractInt, subtract one from it, then cast to i64.
11541154
if value == i64::MIN {
1155-
write!(self.out, "{}li - 1li", value + 1)?;
1155+
write!(self.out, "i64({} - 1)", value + 1)?;
11561156
} else {
11571157
write!(self.out, "{value}li")?;
11581158
}

naga/tests/in/wgsl/int64.wgsl

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -56,7 +56,7 @@ fn int64_function(x: i64) -> i64 {
5656
val += bitcast<vec3<i64>>(input_uniform.val_u64_3).z;
5757
val += bitcast<vec4<i64>>(input_uniform.val_u64_4).w;
5858
// Most negative i64
59-
val += -9223372036854775807li - 1li;
59+
val += i64(-9223372036854775807 - 1);
6060

6161
// Reading/writing to a uniform/storage buffer
6262
output.val_i64 = input_uniform.val_i64 + input_storage.val_i64;

naga/tests/out/hlsl/int64.hlsl

Lines changed: 28 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -95,51 +95,51 @@ int64_t int64_function(int64_t x)
9595
uint64_t4 _e71 = input_uniform.val_u64_4_;
9696
int64_t _e74 = val;
9797
val = (_e74 + _e71.w);
98-
int64_t _e79 = val;
99-
val = (_e79 + (-9223372036854775807L - 1L));
100-
int64_t _e85 = input_uniform.val_i64_;
101-
int64_t _e88 = input_storage.Load<int64_t>(128);
102-
output.Store(128, (_e85 + _e88));
103-
int64_t2 _e94 = input_uniform.val_i64_2_;
104-
int64_t2 _e97 = input_storage.Load<int64_t2>(144);
105-
output.Store(144, (_e94 + _e97));
106-
int64_t3 _e103 = input_uniform.val_i64_3_;
107-
int64_t3 _e106 = input_storage.Load<int64_t3>(160);
108-
output.Store(160, (_e103 + _e106));
109-
int64_t4 _e112 = input_uniform.val_i64_4_;
110-
int64_t4 _e115 = input_storage.Load<int64_t4>(192);
111-
output.Store(192, (_e112 + _e115));
112-
int64_t _e121[2] = Constructarray2_int64_t_(input_arrays.Load<int64_t>(16+0), input_arrays.Load<int64_t>(16+8));
98+
int64_t _e77 = val;
99+
val = (_e77 + -9223372036854775808L);
100+
int64_t _e83 = input_uniform.val_i64_;
101+
int64_t _e86 = input_storage.Load<int64_t>(128);
102+
output.Store(128, (_e83 + _e86));
103+
int64_t2 _e92 = input_uniform.val_i64_2_;
104+
int64_t2 _e95 = input_storage.Load<int64_t2>(144);
105+
output.Store(144, (_e92 + _e95));
106+
int64_t3 _e101 = input_uniform.val_i64_3_;
107+
int64_t3 _e104 = input_storage.Load<int64_t3>(160);
108+
output.Store(160, (_e101 + _e104));
109+
int64_t4 _e110 = input_uniform.val_i64_4_;
110+
int64_t4 _e113 = input_storage.Load<int64_t4>(192);
111+
output.Store(192, (_e110 + _e113));
112+
int64_t _e119[2] = Constructarray2_int64_t_(input_arrays.Load<int64_t>(16+0), input_arrays.Load<int64_t>(16+8));
113113
{
114-
int64_t _value2[2] = _e121;
114+
int64_t _value2[2] = _e119;
115115
output_arrays.Store(16+0, _value2[0]);
116116
output_arrays.Store(16+8, _value2[1]);
117117
}
118+
int64_t _e120 = val;
118119
int64_t _e122 = val;
120+
val = (_e122 + abs(_e120));
119121
int64_t _e124 = val;
120-
val = (_e124 + abs(_e122));
122+
int64_t _e125 = val;
121123
int64_t _e126 = val;
122-
int64_t _e127 = val;
123124
int64_t _e128 = val;
125+
val = (_e128 + clamp(_e124, _e125, _e126));
124126
int64_t _e130 = val;
125-
val = (_e130 + clamp(_e126, _e127, _e128));
126127
int64_t _e132 = val;
127-
int64_t _e134 = val;
128+
int64_t _e135 = val;
129+
val = (_e135 + dot((_e130).xx, (_e132).xx));
128130
int64_t _e137 = val;
129-
val = (_e137 + dot((_e132).xx, (_e134).xx));
130-
int64_t _e139 = val;
131+
int64_t _e138 = val;
131132
int64_t _e140 = val;
133+
val = (_e140 + max(_e137, _e138));
132134
int64_t _e142 = val;
133-
val = (_e142 + max(_e139, _e140));
134-
int64_t _e144 = val;
135+
int64_t _e143 = val;
135136
int64_t _e145 = val;
137+
val = (_e145 + min(_e142, _e143));
136138
int64_t _e147 = val;
137-
val = (_e147 + min(_e144, _e145));
138139
int64_t _e149 = val;
140+
val = (_e149 + sign(_e147));
139141
int64_t _e151 = val;
140-
val = (_e151 + sign(_e149));
141-
int64_t _e153 = val;
142-
return _e153;
142+
return _e151;
143143
}
144144

145145
typedef uint64_t ret_Constructarray2_uint64_t_[2];

naga/tests/out/msl/int64.msl

Lines changed: 29 additions & 29 deletions
Original file line numberDiff line numberDiff line change
@@ -79,49 +79,49 @@ long int64_function(
7979
metal::ulong4 _e71 = input_uniform.val_u64_4_;
8080
long _e74 = val;
8181
val = as_type<long>(as_type<ulong>(_e74) + as_type<ulong>(as_type<metal::long4>(_e71).w));
82-
long _e79 = val;
83-
val = as_type<long>(as_type<ulong>(_e79) + as_type<ulong>(as_type<long>(as_type<ulong>(-9223372036854775807L) - as_type<ulong>(1L))));
84-
long _e85 = input_uniform.val_i64_;
85-
long _e88 = input_storage.val_i64_;
86-
output.val_i64_ = as_type<long>(as_type<ulong>(_e85) + as_type<ulong>(_e88));
87-
metal::long2 _e94 = input_uniform.val_i64_2_;
88-
metal::long2 _e97 = input_storage.val_i64_2_;
89-
output.val_i64_2_ = as_type<metal::long2>(as_type<metal::ulong2>(_e94) + as_type<metal::ulong2>(_e97));
90-
metal::long3 _e103 = input_uniform.val_i64_3_;
91-
metal::long3 _e106 = input_storage.val_i64_3_;
92-
output.val_i64_3_ = as_type<metal::long3>(as_type<metal::ulong3>(_e103) + as_type<metal::ulong3>(_e106));
93-
metal::long4 _e112 = input_uniform.val_i64_4_;
94-
metal::long4 _e115 = input_storage.val_i64_4_;
95-
output.val_i64_4_ = as_type<metal::long4>(as_type<metal::ulong4>(_e112) + as_type<metal::ulong4>(_e115));
96-
type_12 _e121 = input_arrays.val_i64_array_2_;
97-
output_arrays.val_i64_array_2_ = _e121;
82+
long _e77 = val;
83+
val = as_type<long>(as_type<ulong>(_e77) + as_type<ulong>(-9223372036854775808L));
84+
long _e83 = input_uniform.val_i64_;
85+
long _e86 = input_storage.val_i64_;
86+
output.val_i64_ = as_type<long>(as_type<ulong>(_e83) + as_type<ulong>(_e86));
87+
metal::long2 _e92 = input_uniform.val_i64_2_;
88+
metal::long2 _e95 = input_storage.val_i64_2_;
89+
output.val_i64_2_ = as_type<metal::long2>(as_type<metal::ulong2>(_e92) + as_type<metal::ulong2>(_e95));
90+
metal::long3 _e101 = input_uniform.val_i64_3_;
91+
metal::long3 _e104 = input_storage.val_i64_3_;
92+
output.val_i64_3_ = as_type<metal::long3>(as_type<metal::ulong3>(_e101) + as_type<metal::ulong3>(_e104));
93+
metal::long4 _e110 = input_uniform.val_i64_4_;
94+
metal::long4 _e113 = input_storage.val_i64_4_;
95+
output.val_i64_4_ = as_type<metal::long4>(as_type<metal::ulong4>(_e110) + as_type<metal::ulong4>(_e113));
96+
type_12 _e119 = input_arrays.val_i64_array_2_;
97+
output_arrays.val_i64_array_2_ = _e119;
98+
long _e120 = val;
9899
long _e122 = val;
100+
val = as_type<long>(as_type<ulong>(_e122) + as_type<ulong>(naga_abs(_e120)));
99101
long _e124 = val;
100-
val = as_type<long>(as_type<ulong>(_e124) + as_type<ulong>(naga_abs(_e122)));
102+
long _e125 = val;
101103
long _e126 = val;
102-
long _e127 = val;
103104
long _e128 = val;
105+
val = as_type<long>(as_type<ulong>(_e128) + as_type<ulong>(metal::clamp(_e124, _e125, _e126)));
104106
long _e130 = val;
105-
val = as_type<long>(as_type<ulong>(_e130) + as_type<ulong>(metal::clamp(_e126, _e127, _e128)));
107+
metal::long2 _e131 = metal::long2(_e130);
106108
long _e132 = val;
107109
metal::long2 _e133 = metal::long2(_e132);
108-
long _e134 = val;
109-
metal::long2 _e135 = metal::long2(_e134);
110+
long _e135 = val;
111+
val = as_type<long>(as_type<ulong>(_e135) + as_type<ulong>(( + _e131.x * _e133.x + _e131.y * _e133.y)));
110112
long _e137 = val;
111-
val = as_type<long>(as_type<ulong>(_e137) + as_type<ulong>(( + _e133.x * _e135.x + _e133.y * _e135.y)));
112-
long _e139 = val;
113+
long _e138 = val;
113114
long _e140 = val;
115+
val = as_type<long>(as_type<ulong>(_e140) + as_type<ulong>(metal::max(_e137, _e138)));
114116
long _e142 = val;
115-
val = as_type<long>(as_type<ulong>(_e142) + as_type<ulong>(metal::max(_e139, _e140)));
116-
long _e144 = val;
117+
long _e143 = val;
117118
long _e145 = val;
119+
val = as_type<long>(as_type<ulong>(_e145) + as_type<ulong>(metal::min(_e142, _e143)));
118120
long _e147 = val;
119-
val = as_type<long>(as_type<ulong>(_e147) + as_type<ulong>(metal::min(_e144, _e145)));
120121
long _e149 = val;
122+
val = as_type<long>(as_type<ulong>(_e149) + as_type<ulong>(metal::select(metal::select(long(-1), long(1), (_e147 > 0)), long(0), (_e147 == 0))));
121123
long _e151 = val;
122-
val = as_type<long>(as_type<ulong>(_e151) + as_type<ulong>(metal::select(metal::select(long(-1), long(1), (_e149 > 0)), long(0), (_e149 == 0))));
123-
long _e153 = val;
124-
return _e153;
124+
return _e151;
125125
}
126126

127127
ulong uint64_function(

0 commit comments

Comments
 (0)