nir/algebraic: Add better lowering of ldexp

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index ed21c5d5c8ccc4bedc91ae0699820f0287bd11a9..60cd73132f720a1aa46f7edba09c9e4d271d5754 100644 (file)
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -276,8 +276,6 @@ optimizations = [
    (('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod'),
    (('uadd_carry', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'),
    (('usub_borrow', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'),
-   (('ldexp', 'x', 'exp'),
-    ('fmul', 'x', ('ishl', ('imin', ('imax', ('iadd', 'exp', 0x7f), 0), 0xff), 23))),
 
    (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
     ('bcsel', ('ilt', 31, 'bits'), 'insert',
@@ -359,6 +357,33 @@ optimizations = [
      'options->lower_unpack_snorm_4x8'),
 ]
 
+def fexp2i(exp):
+   # We assume that exp is already in range.
+   return ('ishl', ('iadd', exp, 127), 23)
+
+def ldexp32(f, exp):
+   # First, we clamp exp to a reasonable range.  The maximum range that we
+   # need is the largest range for an exponent, ([-127, 128] if you include
+   # inf and 0) plus the number of mantissa bits in either direction to
+   # account for denormals.  This means that we need at least a range of
+   # [-150, 151].  For our implementation, however, what we really care
+   # about is that neither exp/2 nor exp-exp/2 go out of the regular range
+   # for floating-point exponents.
+   exp = ('imin', ('imax', exp, -252), 254)
+
+   # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
+   # While the spec technically defines ldexp as f * 2.0^exp, simply
+   # multiplying once doesn't work when denormals are involved because
+   # 2.0^exp may not be representable even though ldexp(f, exp) is (see
+   # comments above about range).  Instead, we create two powers of two and
+   # multiply by them each in turn.  That way the effective range of our
+   # exponent is doubled.
+   pow2_1 = fexp2i(('ishr', exp, 1))
+   pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)))
+   return ('fmul', ('fmul', f, pow2_1), pow2_2)
+
+optimizations += [(('ldexp', 'x', 'exp'), ldexp32('x', 'exp'))]
+
 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
 def bitfield_reverse(u):
     step1 = ('ior', ('ishl', u, 16), ('ushr', u, 16))