From: Ian Romanick Date: Thu, 24 Oct 2019 21:41:20 +0000 (-0700) Subject: nir/algebraic: Mark other comparison exact when removing a == a X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=9be4a422a055d1e829d56c3cc91e1fc2f6e8fb31;p=mesa.git nir/algebraic: Mark other comparison exact when removing a == a This prevents some additional optimizations that would change the original result. This includes things like (b < a && b < c) => b < min(a, c) and !(a < b) => b >= a. Both of these optimizations were specifically observed in the piglit tests added in piglit!160. This was discovered while investigating https://gitlab.freedesktop.org/mesa/mesa/issues/1958. However, the problem in that issue was Chrome or Angle is replacing calls to isnan() with some stuff that we (correctly) optimize to false. If they had left the calls to isnan() alone, everything would have just worked. No shader-db changes on any Intel platform. I also tried marking the comparison generated by the isnan() function precise. The precise marker "infects" every computation involved in calculating the parameter to the isnan() function, and this severely hurt all of the (few) shaders in shader-db that use isnan(). I also considered adding a new ir_unop_isnan opcode that would implement the functionality. During GLSL IR-to-NIR translation, the resulting comparison operation would be marked exact (and the samething would need to happen in SPIR-V translation). This approach taken by this patch seemed easier, but we may want to do the ir_unop_isnan thing anyway. Fixes: d55835b8bdf ("nir/algebraic: Add optimizations for "a == a && a CMP b"") Reviewed-by: Matt Turner Reviewed-by: Connor Abbott --- diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py index c706eed5258..e4295d7a9e9 100644 --- a/src/compiler/nir/nir_opt_algebraic.py +++ b/src/compiler/nir/nir_opt_algebraic.py @@ -1354,8 +1354,8 @@ optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!option # and, if a is a NaN then the second comparison will fail anyway. for op in ['flt', 'fge', 'feq']: optimizations += [ - (('iand', ('feq', a, a), (op, a, b)), (op, a, b)), - (('iand', ('feq', a, a), (op, b, a)), (op, b, a)), + (('iand', ('feq', a, a), (op, a, b)), ('!' + op, a, b)), + (('iand', ('feq', a, a), (op, b, a)), ('!' + op, b, a)), ] # Add optimizations to handle the case where the result of a ternary is