nir/algebraic: Remove spurious conversions from inside logic ops

Not only does this eliminate a bunch of unnecessary type converting
MOVs, but it can also enable some SWAR.  The
dEQP-VK.spirv_assembly.type.vec3.i8.bitwise_xor_frag test does
something about like:

    c = a.x ^ b.x;
    d = a.y ^ b.y;
    e = a.z ^ b.z;

After this change, it looks more like:

    uint t = i8vec3AsUint(a) ^ i8vec3AsUint(b);
    c = extract_u8(t, 0);
    d = extract_u8(t, 1);
    e = extract_u8(t, 2);

On Ice Lake, this results in:

SIMD8 shader: 41 instructions. 1 loops. 3804 cycles. 0:0 spills:fills, 5 sends
SIMD8 shader: 31 instructions. 1 loops. 2844 cycles. 0:0 spills:fills, 5 sends

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9025>

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 1bf08ee..264d49a 100644 (file)
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -829,6 +829,33 @@ optimizations.extend([
     (('i2f16', ('i2imp', 'a@32')), ('i2f16', a)),
 ])
 
+# Clean up junk left from 8-bit integer to 16-bit integer lowering.
+optimizations.extend([
+    # The u2u16(u2u8(X)) just masks off the upper 8-bits of X.  This can be
+    # accomplished by mask the upper 8-bit of the immediate operand to the
+    # iand instruction.  Often times, both patterns will end up being applied
+    # to the same original expression tree.
+    (('iand', ('u2u16', ('u2u8', 'a@16')), '#b'),               ('iand', a, ('iand', b, 0xff))),
+    (('u2u16', ('u2u8(is_used_once)', ('iand', 'a@16', '#b'))), ('iand', a, ('iand', b, 0xff))),
+])
+
+for op in ['iand', 'ior', 'ixor']:
+    optimizations.extend([
+        (('u2u8', (op, ('u2u16', ('u2u8', 'a@16')), ('u2u16', ('u2u8', 'b@16')))), ('u2u8', (op, a, b))),
+        (('u2u8', (op, ('u2u16', ('u2u8', 'a@32')), ('u2u16', ('u2u8', 'b@32')))), ('u2u8', (op, a, b))),
+
+        # Undistribute extract from a logic op
+        ((op, ('extract_i8', a, '#b'), ('extract_i8', c, b)), ('extract_i8', (op, a, c), b)),
+        ((op, ('extract_u8', a, '#b'), ('extract_u8', c, b)), ('extract_u8', (op, a, c), b)),
+        ((op, ('extract_i16', a, '#b'), ('extract_i16', c, b)), ('extract_i16', (op, a, c), b)),
+        ((op, ('extract_u16', a, '#b'), ('extract_u16', c, b)), ('extract_u16', (op, a, c), b)),
+
+        # Undistribute shifts from a logic op
+        ((op, ('ushr(is_used_once)', a, '#b'), ('ushr', c, b)), ('ushr', (op, a, c), b)),
+        ((op, ('ishr(is_used_once)', a, '#b'), ('ishr', c, b)), ('ishr', (op, a, c), b)),
+        ((op, ('ishl(is_used_once)', a, '#b'), ('ishl', c, b)), ('ishl', (op, a, c), b)),
+    ])
+
 # Integer sizes
 for s in [8, 16, 32, 64]:
     optimizations.extend([