aco: Generalize vs_inputs to args_pending_vmem.

Handle arguments that need a waitcnt without relying on
RADV specific VS input information.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21696>

src/amd/compiler/aco_insert_waitcnt.cpp

@@ -1032,11 +1032,9 @@ insert_wait_states(Program* program)
    std::stack<unsigned, std::vector<unsigned>> loop_header_indices;
    unsigned loop_progress = 0;
 
-   if (program->stage.has(SWStage::VS) && program->info.vs.dynamic_inputs) {
-      for (Definition def : program->vs_inputs) {
-         update_counters(in_ctx[0], event_vmem);
-         insert_wait_entry(in_ctx[0], def, event_vmem);
-      }
+   for (Definition def : program->args_pending_vmem) {
+      update_counters(in_ctx[0], event_vmem);
+      insert_wait_entry(in_ctx[0], def, event_vmem);
    }
 
    for (unsigned i = 0; i < program->blocks.size();) {

src/amd/compiler/aco_instruction_selection.cpp

@@ -11271,10 +11271,15 @@ add_startpgm(struct isel_context* ctx)
          ctx->arg_temps[i] = create_vec_from_array(ctx, elems, size, RegType::sgpr, 4);
       } else {
          Temp dst = ctx->program->allocateTmp(type);
+         Definition def(dst);
+         def.setFixed(PhysReg{file == AC_ARG_SGPR ? reg : reg + 256});
          ctx->arg_temps[i] = dst;
-         startpgm->definitions[arg] = Definition(dst);
-         startpgm->definitions[arg].setFixed(PhysReg{file == AC_ARG_SGPR ? reg : reg + 256});
-         arg++;
+         startpgm->definitions[arg++] = def;
+
+         if (ctx->args->ac.args[i].pending_vmem) {
+            assert(file == AC_ARG_VGPR);
+            ctx->program->args_pending_vmem.push_back(def);
+         }
       }
    }
 
@@ -11294,18 +11299,6 @@ add_startpgm(struct isel_context* ctx)
       }
    }
 
-   if (ctx->stage.has(SWStage::VS) && ctx->program->info.vs.dynamic_inputs) {
-      unsigned num_attributes = util_last_bit(ctx->program->info.vs.input_slot_usage_mask);
-      for (unsigned i = 0; i < num_attributes; i++) {
-         Definition def(get_arg(ctx, ctx->args->vs_inputs[i]));
-
-         unsigned idx = ctx->args->vs_inputs[i].arg_index;
-         def.setFixed(PhysReg(256 + ctx->args->ac.args[idx].offset));
-
-         ctx->program->vs_inputs.push_back(def);
-      }
-   }
-
    return startpgm;
 }
 

src/amd/compiler/aco_ir.h

@@ -2120,7 +2120,7 @@ public:
    unsigned next_divergent_if_logical_depth = 0;
    unsigned next_uniform_if_depth = 0;
 
-   std::vector<Definition> vs_inputs;
+   std::vector<Definition> args_pending_vmem;
 
    struct {
       FILE* output = stderr;

src/amd/compiler/aco_statistics.cpp

@@ -544,13 +544,11 @@ collect_preasm_stats(Program* program)
    double usage[(int)BlockCycleEstimator::resource_count] = {0};
    std::vector<BlockCycleEstimator> blocks(program->blocks.size(), program);
 
-   if (program->stage.has(SWStage::VS) && program->info.vs.has_prolog) {
-      unsigned vs_input_latency = 320;
-      for (Definition def : program->vs_inputs) {
-         blocks[0].vm.push_back(vs_input_latency);
-         for (unsigned i = 0; i < def.size(); i++)
-            blocks[0].reg_available[def.physReg().reg() + i] = vs_input_latency;
-      }
+   constexpr const unsigned vmem_latency = 320;
+   for (const Definition def : program->args_pending_vmem) {
+      blocks[0].vm.push_back(vmem_latency);
+      for (unsigned i = 0; i < def.size(); i++)
+         blocks[0].reg_available[def.physReg().reg() + i] = vmem_latency;
    }
 
    for (Block& block : program->blocks) {