radv: don't lower subdword phis to scalar

Totals from 193 (0.24% of 79839) affected shaders: (Navi48)

MaxWaves: 6004 -> 6024 (+0.33%)
Instrs: 169276 -> 166784 (-1.47%); split: -3.01%, +1.53%
CodeSize: 940608 -> 915768 (-2.64%); split: -4.29%, +1.64%
VGPRs: 8012 -> 7716 (-3.69%); split: -3.99%, +0.30%
SpillVGPRs: 185 -> 0 (-inf%)
Scratch: 13568 -> 0 (-inf%)
Latency: 2159787 -> 2147084 (-0.59%); split: -2.86%, +2.28%
InvThroughput: 664022 -> 395859 (-40.38%); split: -42.59%, +2.21%
VClause: 2998 -> 2880 (-3.94%); split: -4.27%, +0.33%
SClause: 3117 -> 3120 (+0.10%)
Copies: 21290 -> 16278 (-23.54%); split: -24.74%, +1.20%
Branches: 4757 -> 4760 (+0.06%); split: -0.34%, +0.40%
PreSGPRs: 7369 -> 7378 (+0.12%); split: -0.11%, +0.23%
PreVGPRs: 4257 -> 3859 (-9.35%); split: -9.94%, +0.59%
VALU: 83173 -> 79804 (-4.05%); split: -5.68%, +1.63%
SALU: 36672 -> 37318 (+1.76%); split: -0.02%, +1.78%
VMEM: 4012 -> 3762 (-6.23%); split: -6.83%, +0.60%
SMEM: 4300 -> 4303 (+0.07%)
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35784>

src/amd/common/nir/ac_nir.c

@@ -879,3 +879,13 @@ ac_nir_repack_invocations_in_workgroup(nir_builder *b, nir_def **input_bool,
          nir_mbcnt_amd(b, input_mask[i], wg_repacked_index_base);
    }
 }
+
+uint8_t
+ac_nir_lower_phis_to_scalar_cb(const nir_instr *instr, const void *_)
+{
+   nir_phi_instr *phi = nir_instr_as_phi(instr);
+   if (phi->def.bit_size == 1 || phi->def.bit_size >= 32)
+      return 1;
+
+   return 32 / phi->def.bit_size;
+}

src/amd/common/nir/ac_nir.h

@@ -434,6 +434,9 @@ ac_nir_scalarize_overfetching_loads_callback(const nir_instr *instr, const void
 enum gl_access_qualifier
 ac_nir_get_mem_access_flags(const nir_intrinsic_instr *instr);
 
+uint8_t
+ac_nir_lower_phis_to_scalar_cb(const nir_instr *instr, const void *_);
+
 #ifdef __cplusplus
 }
 #endif

src/amd/common/nir/ac_nir_lower_ngg.c

@@ -1797,7 +1797,7 @@ ac_nir_lower_ngg_nogs(nir_shader *shader, const ac_nir_lower_ngg_options *option
    nir_lower_vars_to_ssa(shader);
    nir_remove_dead_variables(shader, nir_var_function_temp, NULL);
    nir_lower_alu_to_scalar(shader, NULL, NULL);
-   nir_lower_all_phis_to_scalar(shader);
+   nir_lower_phis_to_scalar(shader, ac_nir_lower_phis_to_scalar_cb, NULL);
 
    if (options->can_cull) {
       /* It's beneficial to redo these opts after splitting the shader. */

src/amd/common/nir/ac_nir_lower_ngg_mesh.c

@@ -1429,7 +1429,7 @@ ac_nir_lower_ngg_mesh(nir_shader *shader,
    nir_lower_vars_to_ssa(shader);
    nir_remove_dead_variables(shader, nir_var_function_temp, NULL);
    nir_lower_alu_to_scalar(shader, NULL, NULL);
-   nir_lower_all_phis_to_scalar(shader);
+   nir_lower_phis_to_scalar(shader, ac_nir_lower_phis_to_scalar_cb, NULL);
 
    /* Optimize load_local_invocation_index. When the API workgroup is smaller than the HW workgroup,
     * local_invocation_id isn't initialized for all lanes and we can't perform this optimization for

src/amd/common/nir/ac_nir_lower_tess_io_to_mem.c

@@ -1624,7 +1624,7 @@ ac_nir_lower_hs_outputs_to_mem(nir_shader *shader, const nir_tcs_info *info,
    NIR_PASS(_, shader, nir_lower_vars_to_ssa);
    NIR_PASS(_, shader, nir_remove_dead_variables, nir_var_function_temp, NULL);
    NIR_PASS(_, shader, nir_lower_alu_to_scalar, NULL, NULL);
-   NIR_PASS(_, shader, nir_lower_all_phis_to_scalar);
+   NIR_PASS(_, shader, nir_lower_phis_to_scalar, ac_nir_lower_phis_to_scalar_cb, NULL);
 
    return true;
 }

src/amd/compiler/instruction_selection/aco_isel_setup.cpp

@@ -249,7 +249,7 @@ void
 setup_nir(isel_context* ctx, nir_shader* nir)
 {
    nir_convert_to_lcssa(nir, true, false);
-   if (nir_lower_all_phis_to_scalar(nir)) {
+   if (nir_lower_phis_to_scalar(nir, ac_nir_lower_phis_to_scalar_cb, NULL)) {
       nir_copy_prop(nir);
       nir_opt_dce(nir);
    }

src/amd/vulkan/radv_shader.c

@@ -181,7 +181,7 @@ radv_optimize_nir(struct nir_shader *shader, bool optimize_conservatively)
       NIR_LOOP_PASS(_, skip, shader, nir_lower_vars_to_ssa);
 
       NIR_LOOP_PASS(_, skip, shader, nir_lower_alu_width, vectorize_vec2_16bit, NULL);
-      NIR_LOOP_PASS(_, skip, shader, nir_lower_all_phis_to_scalar);
+      NIR_LOOP_PASS(_, skip, shader, nir_lower_phis_to_scalar, ac_nir_lower_phis_to_scalar_cb, NULL);
 
       NIR_LOOP_PASS(progress, skip, shader, nir_copy_prop);
       NIR_LOOP_PASS(progress, skip, shader, nir_opt_remove_phis);