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 (('fmul', a
, 0.0), 0.0),
61 (('fmul', a
, 1.0), a
),
63 (('fmul', a
, -1.0), ('fneg', a
)),
64 (('imul', a
, -1), ('ineg', a
)),
65 (('ffma', 0.0, a
, b
), b
),
66 (('ffma', a
, 0.0, b
), b
),
67 (('ffma', a
, b
, 0.0), ('fmul', a
, b
)),
68 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
69 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
70 (('flrp', a
, b
, 0.0), a
),
71 (('flrp', a
, b
, 1.0), b
),
72 (('flrp', a
, a
, b
), a
),
73 (('flrp', 0.0, a
, b
), ('fmul', a
, b
)),
74 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
75 (('fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), '!options->lower_ffma'),
76 # Comparison simplifications
77 (('inot', ('flt', a
, b
)), ('fge', a
, b
)),
78 (('inot', ('fge', a
, b
)), ('flt', a
, b
)),
79 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
80 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
81 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
82 (('fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
83 (('feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
84 (('fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
85 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
86 (('bcsel', ('flt', a
, b
), a
, b
), ('fmin', a
, b
)),
87 (('bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
88 (('fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
)),
89 # Comparison with the same args. Note that these are not done for
90 # the float versions because NaN always returns false on float
92 (('ilt', a
, a
), False),
93 (('ige', a
, a
), True),
94 (('ieq', a
, a
), True),
95 (('ine', a
, a
), False),
96 (('ult', a
, a
), False),
97 (('uge', a
, a
), True),
98 # Logical and bit operations
99 (('fand', a
, 0.0), 0.0),
104 (('fxor', a
, a
), 0.0),
106 (('inot', ('inot', a
)), a
),
108 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
109 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
110 # Shift optimizations
117 # Exponential/logarithmic identities
118 (('fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
119 (('fexp', ('flog', a
)), a
), # e^ln(a) = a
120 (('flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
121 (('flog', ('fexp', a
)), a
), # ln(e^a) = a
122 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
123 (('fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
124 (('fexp', ('fmul', ('flog', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # e^(ln(a)*b) = a^b
125 (('fpow', a
, 1.0), a
),
126 (('fpow', a
, 2.0), ('fmul', a
, a
)),
127 (('fpow', 2.0, a
), ('fexp2', a
)),
128 # Division and reciprocal
129 (('fdiv', 1.0, a
), ('frcp', a
)),
130 (('frcp', ('frcp', a
)), a
),
131 (('frcp', ('fsqrt', a
)), ('frsq', a
)),
132 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
133 (('frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
134 # Boolean simplifications
135 (('ine', 'a@bool', 0), 'a'),
136 (('ieq', 'a@bool', 0), ('inot', 'a')),
137 (('bcsel', a
, True, False), ('ine', a
, 0)),
138 (('bcsel', a
, False, True), ('ieq', a
, 0)),
139 (('bcsel', True, b
, c
), b
),
140 (('bcsel', False, b
, c
), c
),
141 # The result of this should be hit by constant propagation and, in the
142 # next round of opt_algebraic, get picked up by one of the above two.
143 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
146 (('fsub', 0.0, ('fsub', 0.0, a
)), a
),
147 (('isub', 0, ('isub', 0, a
)), a
),
148 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
149 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
150 (('fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
151 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
152 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
153 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
155 # This one may not be exact
156 (('feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
159 # Add optimizations to handle the case where the result of a ternary is
160 # compared to a constant. This way we can take things like
166 # a ? (0 > 0) : (1 > 0)
168 # which constant folding will eat for lunch. The resulting ternary will
169 # further get cleaned up by the boolean reductions above and we will be
170 # left with just the original variable "a".
171 for op
in ['flt', 'fge', 'feq', 'fne',
172 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
174 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
175 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
176 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
177 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
180 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render()