aco/isel: Move add_startpgm to aco_instruction_selection.cpp
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6504>
diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index fe1e9b7..eeb2035 100644
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -10391,6 +10391,65 @@
end_divergent_if(ctx, &ic);
}
+Pseudo_instruction *add_startpgm(struct isel_context *ctx)
+{
+ unsigned arg_count = ctx->args->ac.arg_count;
+ if (ctx->stage == fragment_fs) {
+ /* LLVM optimizes away unused FS inputs and computes spi_ps_input_addr
+ * itself and then communicates the results back via the ELF binary.
+ * Mirror what LLVM does by re-mapping the VGPR arguments here.
+ *
+ * TODO: If we made the FS input scanning code into a separate pass that
+ * could run before argument setup, then this wouldn't be necessary
+ * anymore.
+ */
+ struct ac_shader_args *args = &ctx->args->ac;
+ arg_count = 0;
+ for (unsigned i = 0, vgpr_arg = 0, vgpr_reg = 0; i < args->arg_count; i++) {
+ if (args->args[i].file != AC_ARG_VGPR) {
+ arg_count++;
+ continue;
+ }
+
+ if (!(ctx->program->config->spi_ps_input_addr & (1 << vgpr_arg))) {
+ args->args[i].skip = true;
+ } else {
+ args->args[i].offset = vgpr_reg;
+ vgpr_reg += args->args[i].size;
+ arg_count++;
+ }
+ vgpr_arg++;
+ }
+ }
+
+ aco_ptr<Pseudo_instruction> startpgm{create_instruction<Pseudo_instruction>(aco_opcode::p_startpgm, Format::PSEUDO, 0, arg_count + 1)};
+ for (unsigned i = 0, arg = 0; i < ctx->args->ac.arg_count; i++) {
+ if (ctx->args->ac.args[i].skip)
+ continue;
+
+ enum ac_arg_regfile file = ctx->args->ac.args[i].file;
+ unsigned size = ctx->args->ac.args[i].size;
+ unsigned reg = ctx->args->ac.args[i].offset;
+ RegClass type = RegClass(file == AC_ARG_SGPR ? RegType::sgpr : RegType::vgpr, size);
+ Temp dst = Temp{ctx->program->allocateId(), type};
+ 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_count] = Definition{ctx->program->allocateId(), exec, ctx->program->lane_mask};
+ Pseudo_instruction *instr = startpgm.get();
+ ctx->block->instructions.push_back(std::move(startpgm));
+
+ /* Stash these in the program so that they can be accessed later when
+ * handling spilling.
+ */
+ ctx->program->private_segment_buffer = get_arg(ctx, ctx->args->ring_offsets);
+ ctx->program->scratch_offset = get_arg(ctx, ctx->args->scratch_offset);
+
+ return instr;
+}
+
} /* end namespace */
void fix_ls_vgpr_init_bug(isel_context *ctx, Pseudo_instruction *startpgm)
diff --git a/src/amd/compiler/aco_instruction_selection.h b/src/amd/compiler/aco_instruction_selection.h
index aff969d..a9249a88 100644
--- a/src/amd/compiler/aco_instruction_selection.h
+++ b/src/amd/compiler/aco_instruction_selection.h
@@ -201,65 +201,6 @@
void init_context(isel_context *ctx, nir_shader *shader);
-inline Pseudo_instruction *add_startpgm(struct isel_context *ctx)
-{
- unsigned arg_count = ctx->args->ac.arg_count;
- if (ctx->stage == fragment_fs) {
- /* LLVM optimizes away unused FS inputs and computes spi_ps_input_addr
- * itself and then communicates the results back via the ELF binary.
- * Mirror what LLVM does by re-mapping the VGPR arguments here.
- *
- * TODO: If we made the FS input scanning code into a separate pass that
- * could run before argument setup, then this wouldn't be necessary
- * anymore.
- */
- struct ac_shader_args *args = &ctx->args->ac;
- arg_count = 0;
- for (unsigned i = 0, vgpr_arg = 0, vgpr_reg = 0; i < args->arg_count; i++) {
- if (args->args[i].file != AC_ARG_VGPR) {
- arg_count++;
- continue;
- }
-
- if (!(ctx->program->config->spi_ps_input_addr & (1 << vgpr_arg))) {
- args->args[i].skip = true;
- } else {
- args->args[i].offset = vgpr_reg;
- vgpr_reg += args->args[i].size;
- arg_count++;
- }
- vgpr_arg++;
- }
- }
-
- aco_ptr<Pseudo_instruction> startpgm{create_instruction<Pseudo_instruction>(aco_opcode::p_startpgm, Format::PSEUDO, 0, arg_count + 1)};
- for (unsigned i = 0, arg = 0; i < ctx->args->ac.arg_count; i++) {
- if (ctx->args->ac.args[i].skip)
- continue;
-
- enum ac_arg_regfile file = ctx->args->ac.args[i].file;
- unsigned size = ctx->args->ac.args[i].size;
- unsigned reg = ctx->args->ac.args[i].offset;
- RegClass type = RegClass(file == AC_ARG_SGPR ? RegType::sgpr : RegType::vgpr, size);
- Temp dst = Temp{ctx->program->allocateId(), type};
- 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_count] = Definition{ctx->program->allocateId(), exec, ctx->program->lane_mask};
- Pseudo_instruction *instr = startpgm.get();
- ctx->block->instructions.push_back(std::move(startpgm));
-
- /* Stash these in the program so that they can be accessed later when
- * handling spilling.
- */
- ctx->program->private_segment_buffer = get_arg(ctx, ctx->args->ring_offsets);
- ctx->program->scratch_offset = get_arg(ctx, ctx->args->scratch_offset);
-
- return instr;
-}
-
isel_context
setup_isel_context(Program* program,
unsigned shader_count,
