nir, ir3: Add lower_fmulz_with_abs_min backend option

This commits adds the `lower_fmulz_with_abs_min` which lowers
`fmulz` -> `min(abs(a), abs(b)) == 0.0 ? 0.0 : a * b`
`ffmaz` -> `min(abs(a), abs(b)) == 0.0 ? c : ffma(a, b, c)

This is useful for ISAs which have `abs` for free on `min` such as
ir3.

Adreno A750 Benchmark of 10 runs of 5 DX9 single frame trimmed
captures looped 2048 times using u_trace measuring
`start_render_pass` to `end_render_pass` results:

sysmem:
-1.91156%, -2.21791%, -2.02533%, -2.21666%, -2.33272%,
-2.67349%, -1.75278%, -2.05923%, -2.26892%, -2.10506%
Avg:  ~ -2.16%
ST.S: ~  0.25%

gmem:
-3.61496%, -3.66682%, -3.80901%, -3.51198%, -3.72950%,
-3.71413%, -3.64467%, -3.67092%, -3.90640%, -3.83888%
Avg:  ~ -3.71%
ST.S: ~  0.12%

Signed-off-by: Karmjit Mahil <karmjit.mahil@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31479>

src/compiler/nir/nir_opt_algebraic.py

@@ -3954,6 +3954,14 @@ for op in ['ffma', 'ffmaz']:
         (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, e, d)), (op, b, ('bcsel', a, c, e), d)),
     ]
 
+# Lower fmulz using min(abs(a), abs(b))==0.0
+late_optimizations += [
+   (('fmulz@32', a, b),
+    ('bcsel', ('feq', ('fmin', ('fabs', a), ('fabs', b)), 0.0), 0.0, ('fmul', a, b)), 'options->lower_fmulz_with_abs_min'),
+   (('ffmaz@32', a, b, c),
+    ('bcsel', ('feq', ('fmin', ('fabs', a), ('fabs', b)), 0.0), c, ('ffma@32', a, b, c)), 'options->lower_fmulz_with_abs_min')
+]
+
 # mediump: If an opcode is surrounded by conversions, remove the conversions.
 # The rationale is that type conversions + the low precision opcode are more
 # expensive that the same arithmetic opcode at higher precision.

src/compiler/nir/nir_shader_compiler_options.h

@@ -712,6 +712,11 @@ typedef struct nir_shader_compiler_options {
    uint8_t support_indirect_inputs;
    uint8_t support_indirect_outputs;
 
+   /**
+    * Lower fmulz to `min(abs(a), abs(b)) == 0.0 ? 0.0 : a * b`.
+    */
+   bool lower_fmulz_with_abs_min;
+
    /** store the variable offset into the instrinsic range_base instead
     *  of adding it to the image index.
     */

src/freedreno/ir3/ir3_compiler.c

@@ -113,6 +113,12 @@ static const nir_shader_compiler_options ir3_base_options = {
    .lower_uniforms_to_ubo = true,
    .max_unroll_iterations = 32,
 
+   /* Not actually supported but we want fmulz to be produced and then be
+    * lowered with the abs min pattern since we have free abs on min.
+    */
+   .has_fmulz = true,
+   .lower_fmulz_with_abs_min = true,
+
    .lower_cs_local_index_to_id = true,
    .lower_wpos_pntc = true,