nir/algebraic: Optimize some 64-bit integer comparisons involving zero

I noticed that we can do better for these kinds of comparisons while
working on the lowering for iadd_sat@64 and isub_sat@64.  This
eliminated 11 instruction from the fs-addSaturate-int64.shader_test.

My hope is that this will improve the run-time of int64 tests on Ice
Lake.  I have no data to support or refute this.

Unsurprisingly, no changes on shader-db.

v2: Condition the min and max patterns with nir_lower_minmax64.
Suggested by Caio.  Very long discussion in the MR. :)

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/767>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/767>

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index c52522379c63cf6d5d2b0f78445d73d4dcd0ff80..7b9a6a8e45d3ee3915f3a97372337d928a2f8c35 100644 (file)
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -1063,6 +1063,24 @@ optimizations.extend([
                              0x7fffffffffffffff)),
     '(options->lower_int64_options & nir_lower_iadd64) != 0'),
 
+   # These are done here instead of in the backend because the int64 lowering
+   # pass will make a mess of the patterns.  The first patterns are
+   # conditioned on nir_lower_minmax64 because it was not clear that it was
+   # always an improvement on platforms that have real int64 support.  No
+   # shaders in shader-db hit this, so it was hard to say one way or the
+   # other.
+   (('ilt', ('imax(is_used_once)', 'a@64', 'b@64'), 0), ('ilt', ('imax', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
+   (('ilt', ('imin(is_used_once)', 'a@64', 'b@64'), 0), ('ilt', ('imin', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
+   (('ige', ('imax(is_used_once)', 'a@64', 'b@64'), 0), ('ige', ('imax', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
+   (('ige', ('imin(is_used_once)', 'a@64', 'b@64'), 0), ('ige', ('imin', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
+   (('ilt', 'a@64', 0), ('ilt', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
+   (('ige', 'a@64', 0), ('ige', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
+
+   (('ine', 'a@64', 0), ('ine', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
+   (('ieq', 'a@64', 0), ('ieq', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
+   # 0u < uint(a) <=> uint(a) != 0u
+   (('ult', 0, 'a@64'), ('ine', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
+
    # Alternative lowering that doesn't rely on bfi.
    (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
     ('bcsel', ('ult', 31, 'bits'),