aco: don't emit s_endpgm for tcs with epilog

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24442>

src/amd/compiler/aco_assembler.cpp

@@ -1279,7 +1279,8 @@ align_block(asm_context& ctx, std::vector<uint32_t>& code, Block& block)
 }
 
 unsigned
-emit_program(Program* program, std::vector<uint32_t>& code, std::vector<struct aco_symbol>* symbols)
+emit_program(Program* program, std::vector<uint32_t>& code, std::vector<struct aco_symbol>* symbols,
+             bool append_endpgm)
 {
    asm_context ctx(program, symbols);
 
@@ -1298,7 +1299,8 @@ emit_program(Program* program, std::vector<uint32_t>& code, std::vector<struct a
    unsigned exec_size = code.size() * sizeof(uint32_t);
 
    /* Add end-of-code markers for the UMR disassembler. */
-   code.resize(code.size() + 5, 0xbf9f0000u);
+   if (append_endpgm)
+      code.resize(code.size() + 5, 0xbf9f0000u);
 
    fix_constaddrs(ctx, code);
 

src/amd/compiler/aco_instruction_selection.cpp

@@ -11518,8 +11518,11 @@ select_program(Program* program, unsigned shader_count, struct nir_shader* const
 
    append_logical_end(ctx.block);
    ctx.block->kind |= block_kind_uniform;
-   Builder bld(ctx.program, ctx.block);
-   bld.sopp(aco_opcode::s_endpgm);
+
+   if (!ctx.program->info.has_epilog) {
+      Builder bld(ctx.program, ctx.block);
+      bld.sopp(aco_opcode::s_endpgm);
+   }
 
    cleanup_cfg(program);
 }

src/amd/compiler/aco_interface.cpp

@@ -236,7 +236,11 @@ aco_compile_shader(const struct aco_compiler_options* options, const struct aco_
    /* assembly */
    std::vector<uint32_t> code;
    std::vector<struct aco_symbol> symbols;
-   unsigned exec_size = aco::emit_program(program.get(), code, &symbols);
+   /* OpenGL combine multi shader parts into one continous code block,
+    * so only last part need the s_endpgm instruction.
+    */
+   bool append_endpgm = !(options->is_opengl && info->has_epilog);
+   unsigned exec_size = aco::emit_program(program.get(), code, &symbols, append_endpgm);
 
    if (program->collect_statistics)
       aco::collect_postasm_stats(program.get(), code);
@@ -284,7 +288,7 @@ aco_compile_rt_prolog(const struct aco_compiler_options* options,
    /* assembly */
    std::vector<uint32_t> code;
    code.reserve(align(program->blocks[0].instructions.size() * 2, 16));
-   unsigned exec_size = aco::emit_program(program.get(), code, NULL);
+   unsigned exec_size = aco::emit_program(program.get(), code);
 
    bool get_disasm = options->dump_shader || options->record_ir;
 
@@ -322,7 +326,7 @@ aco_compile_vs_prolog(const struct aco_compiler_options* options,
    /* assembly */
    std::vector<uint32_t> code;
    code.reserve(align(program->blocks[0].instructions.size() * 2, 16));
-   unsigned exec_size = aco::emit_program(program.get(), code, NULL);
+   unsigned exec_size = aco::emit_program(program.get(), code);
 
    bool get_disasm = options->dump_shader || options->record_ir;
 
@@ -365,7 +369,7 @@ aco_compile_shader_part(const struct aco_compiler_options* options,
 
    /* assembly */
    std::vector<uint32_t> code;
-   unsigned exec_size = aco::emit_program(program.get(), code, NULL);
+   unsigned exec_size = aco::emit_program(program.get(), code);
 
    bool get_disasm = options->dump_shader || options->record_ir;
 

src/amd/compiler/aco_ir.h

@@ -2262,7 +2262,7 @@ bool dealloc_vgprs(Program* program);
 void insert_NOPs(Program* program);
 void form_hard_clauses(Program* program);
 unsigned emit_program(Program* program, std::vector<uint32_t>& code,
-                      std::vector<struct aco_symbol>* symbols);
+                      std::vector<struct aco_symbol>* symbols = NULL, bool append_endpgm = true);
 /**
  * Returns true if print_asm can disassemble the given program for the current build/runtime
  * configuration

src/amd/compiler/tests/helpers.cpp

@@ -269,7 +269,7 @@ finish_assembler_test()
 {
    finish_program(program.get());
    std::vector<uint32_t> binary;
-   unsigned exec_size = emit_program(program.get(), binary, NULL);
+   unsigned exec_size = emit_program(program.get(), binary);
 
    /* we could use CLRX for disassembly but that would require it to be
     * installed */