3 # Copyright (C) 2014 Intel Corporation
5 # Permission is hereby granted, free of charge, to any person obtaining a
6 # copy of this software and associated documentation files (the "Software"),
7 # to deal in the Software without restriction, including without limitation
8 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 # and/or sell copies of the Software, and to permit persons to whom the
10 # Software is furnished to do so, subject to the following conditions:
12 # The above copyright notice and this permission notice (including the next
13 # paragraph) shall be included in all copies or substantial portions of the
16 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 # Jason Ekstrand (jason@jlekstrand.net)
29 # Convenience variables
35 # Written in the form (<search>, <replace>) where <search> is an expression
36 # and <replace> is either an expression or a value. An expression is
37 # defined as a tuple of the form (<op>, <src0>, <src1>, <src2>, <src3>)
38 # where each source is either an expression or a value. A value can be
39 # either a numeric constant or a string representing a variable name.
41 # Variable names are specified as "[#]name[@type]" where "#" inicates that
42 # the given variable will only match constants and the type indicates that
43 # the given variable will only match values from ALU instructions with the
46 # For constants, you have to be careful to make sure that it is the right
47 # type because python is unaware of the source and destination types of the
51 (('fneg', ('fneg', a
)), a
),
52 (('ineg', ('ineg', a
)), a
),
53 (('fabs', ('fabs', a
)), ('fabs', a
)),
54 (('fabs', ('fneg', a
)), ('fabs', a
)),
55 (('iabs', ('iabs', a
)), ('iabs', a
)),
56 (('iabs', ('ineg', a
)), ('iabs', a
)),
57 (('fadd', a
, 0.0), a
),
59 (('fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
60 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
61 (('fmul', a
, 0.0), 0.0),
63 (('fmul', a
, 1.0), a
),
65 (('fmul', a
, -1.0), ('fneg', a
)),
66 (('imul', a
, -1), ('ineg', a
)),
67 (('ffma', 0.0, a
, b
), b
),
68 (('ffma', a
, 0.0, b
), b
),
69 (('ffma', a
, b
, 0.0), ('fmul', a
, b
)),
70 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
71 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
72 (('flrp', a
, b
, 0.0), a
),
73 (('flrp', a
, b
, 1.0), b
),
74 (('flrp', a
, a
, b
), a
),
75 (('flrp', 0.0, a
, b
), ('fmul', a
, b
)),
76 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
77 (('fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), '!options->lower_ffma'),
78 # Comparison simplifications
79 (('inot', ('flt', a
, b
)), ('fge', a
, b
)),
80 (('inot', ('fge', a
, b
)), ('flt', a
, b
)),
81 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
82 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
83 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
84 (('fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
85 (('feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
86 (('fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
87 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
88 (('bcsel', ('flt', a
, b
), a
, b
), ('fmin', a
, b
)),
89 (('bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
90 (('bcsel', ('inot', 'a@bool'), b
, c
), ('bcsel', a
, c
, b
)),
91 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
92 (('fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
93 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
94 (('fsat', ('fsat', a
)), ('fsat', a
)),
95 (('fmin', ('fmax', ('fmin', ('fmax', a
, 0.0), 1.0), 0.0), 1.0), ('fmin', ('fmax', a
, 0.0), 1.0)),
96 # Comparison with the same args. Note that these are not done for
97 # the float versions because NaN always returns false on float
99 (('ilt', a
, a
), False),
100 (('ige', a
, a
), True),
101 (('ieq', a
, a
), True),
102 (('ine', a
, a
), False),
103 (('ult', a
, a
), False),
104 (('uge', a
, a
), True),
105 # Logical and bit operations
106 (('fand', a
, 0.0), 0.0),
111 (('fxor', a
, a
), 0.0),
113 (('inot', ('inot', a
)), a
),
115 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
116 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
117 # Shift optimizations
124 # Exponential/logarithmic identities
125 (('fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
126 (('fexp', ('flog', a
)), a
), # e^ln(a) = a
127 (('flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
128 (('flog', ('fexp', a
)), a
), # ln(e^a) = a
129 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
130 (('fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
131 (('fexp', ('fmul', ('flog', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # e^(ln(a)*b) = a^b
132 (('fpow', a
, 1.0), a
),
133 (('fpow', a
, 2.0), ('fmul', a
, a
)),
134 (('fpow', 2.0, a
), ('fexp2', a
)),
135 # Division and reciprocal
136 (('fdiv', 1.0, a
), ('frcp', a
)),
137 (('frcp', ('frcp', a
)), a
),
138 (('frcp', ('fsqrt', a
)), ('frsq', a
)),
139 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
140 (('frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
141 # Boolean simplifications
142 (('ine', 'a@bool', 0), 'a'),
143 (('ieq', 'a@bool', 0), ('inot', 'a')),
144 (('bcsel', a
, True, False), ('ine', a
, 0)),
145 (('bcsel', a
, False, True), ('ieq', a
, 0)),
146 (('bcsel', True, b
, c
), b
),
147 (('bcsel', False, b
, c
), c
),
148 # The result of this should be hit by constant propagation and, in the
149 # next round of opt_algebraic, get picked up by one of the above two.
150 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
152 (('bcsel', a
, b
, b
), b
),
153 (('fcsel', a
, b
, b
), b
),
156 (('fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
157 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
158 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
159 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
160 (('fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
161 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
162 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
163 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
165 # This one may not be exact
166 (('feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
169 # Add optimizations to handle the case where the result of a ternary is
170 # compared to a constant. This way we can take things like
176 # a ? (0 > 0) : (1 > 0)
178 # which constant folding will eat for lunch. The resulting ternary will
179 # further get cleaned up by the boolean reductions above and we will be
180 # left with just the original variable "a".
181 for op
in ['flt', 'fge', 'feq', 'fne',
182 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
184 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
185 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
186 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
187 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
190 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render()