nir: Saturating integer arithmetic is not associative

In 8-bits,

    iadd_sat(iadd_sat(0x7f, 0x7f), -1) =
    iadd_sat(0x7f, -1) =
    0x7e

but,

    iadd_sat(0x7f, iadd_sat(0x7f, -1)) =
    iadd_sat(0x7f, 0x7e) =
    0x7f

Fixes: 272e927d0e ("nir/spirv: initial handling of OpenCL.std extension opcodes")
Reviewed-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This commit is contained in:
Ian Romanick 2019-04-30 08:00:32 -07:00
parent 70da00ffd6
commit 85e6865ff6

View file

@ -482,7 +482,7 @@ def binop_reduce(name, output_size, output_type, src_type, prereduce_expr,
binop("fadd", tfloat, commutative + associative, "src0 + src1")
binop("iadd", tint, commutative + associative, "src0 + src1")
binop("iadd_sat", tint, commutative + associative, """
binop("iadd_sat", tint, commutative, """
src1 > 0 ?
(src0 + src1 < src0 ? (1ull << (bit_size - 1)) - 1 : src0 + src1) :
(src0 < src0 + src1 ? (1ull << (bit_size - 1)) : src0 + src1)