X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fcompiler%2Fnir%2Fnir_opt_algebraic.py;h=472db765026cab36a0f095933f0b4a5f35ea80a7;hb=a18c4ee7b07cb0c78b7d93005cc76eded4e8001c;hp=7d9775950a4d596d519df681ce6f86e605583173;hpb=4e3d69ad07d9c439fc99b7c1b1e999ce556d260f;p=mesa.git

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 7d9775950a4..472db765026 100644
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -629,6 +629,17 @@ optimizations.extend([
    (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', a, b), 0), '!options->lower_bitops'),
    (('ior',  ('ine', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('ior', a, b), 0), '!options->lower_bitops'),
 
+   # This pattern occurs coutresy of __flt64_nonnan in the soft-fp64 code.
+   # The first part of the iand comes from the !__feq64_nonnan.
+   #
+   # The second pattern is a reformulation of the first based on the relation
+   # (a == 0 || y == 0) <=> umin(a, y) == 0, where b in the first equation
+   # happens to be y == 0.
+   (('iand', ('inot', ('iand', ('ior', ('ieq', a, 0),  b), c)), ('ilt', a, 0)),
+    ('iand', ('inot', ('iand',                         b , c)), ('ilt', a, 0))),
+   (('iand', ('inot', ('iand', ('ieq', ('umin', a, b), 0), c)), ('ilt', a, 0)),
+    ('iand', ('inot', ('iand', ('ieq',             b , 0), c)), ('ilt', a, 0))),
+
    # These patterns can result when (a < b || a < c) => (a < min(b, c))
    # transformations occur before constant propagation and loop-unrolling.
    (('~flt', a, ('fmax', b, a)), ('flt', a, b)),
@@ -776,6 +787,7 @@ optimizations.extend([
    (('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
    (('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2', c), d))),
     ('~fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
+   (('~fexp2', ('fmul', ('flog2', a), 0.5)), ('fsqrt', a)),
    (('~fexp2', ('fmul', ('flog2', a), 2.0)), ('fmul', a, a)),
    (('~fexp2', ('fmul', ('flog2', a), 4.0)), ('fmul', ('fmul', a, a), ('fmul', a, a))),
    (('~fpow', a, 1.0), a),
@@ -1347,7 +1359,7 @@ for x, y in itertools.product(['f', 'u', 'i'], ['f', 'u', 'i']):
    optimizations.append(((x2yN, (b2x, a)), (b2y, a)))
 
 # Optimize away x2xN(a@N)
-for t in ['int', 'uint', 'float']:
+for t in ['int', 'uint', 'float', 'bool']:
    for N in type_sizes(t):
       x2xN = '{0}2{0}{1}'.format(t[0], N)
       aN = 'a@{0}'.format(N)
@@ -1388,6 +1400,15 @@ for N, M in itertools.product(type_sizes('uint'), type_sizes('uint')):
       # The N == M case is handled by other optimizations
       pass
 
+# Downcast operations should be able to see through pack
+for t in ['i', 'u']:
+    for N in [8, 16, 32]:
+        x2xN = '{0}2{0}{1}'.format(t, N)
+        optimizations += [
+            ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)),
+            ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)),
+        ]
+
 # Optimize comparisons with up-casts
 for t in ['int', 'uint', 'float']:
     for N, M in itertools.product(type_sizes(t), repeat=2):