aco/gfx11.5: workaround export priority issue

https://github.com/llvm/llvm-project/pull/99273

fossil-db (gfx1150):
Totals from 73996 (93.20% of 79395) affected shaders:
Instrs: 36015357 -> 36807177 (+2.20%)
CodeSize: 189072544 -> 192238748 (+1.67%)
Latency: 245845181 -> 246790550 (+0.38%); split: -0.00%, +0.38%
InvThroughput: 45068018 -> 45116177 (+0.11%); split: -0.00%, +0.11%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Backport-to: 24.2
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30241>
(cherry picked from commit 0919ce1ac4)

.pick_status.json

@@ -234,7 +234,7 @@
         "description": "aco/gfx11.5: workaround export priority issue",
         "nominated": true,
         "nomination_type": 4,
-        "resolution": 0,
+        "resolution": 1,
         "main_sha": null,
         "because_sha": null,
         "notes": null

src/amd/compiler/aco_insert_NOPs.cpp

@@ -1772,6 +1772,70 @@ mitigate_hazards(Program* program)
    }
 }
 
+/* FeatureRequiredExportPriority in LLVM */
+void
+required_export_priority(Program* program)
+{
+   /* Skip callees, assuming that the caller has already increased the priority. */
+   bool increase_priority = !program->is_epilog && !program->info.vs.has_prolog &&
+                            (!program->info.merged_shader_compiled_separately ||
+                             program->stage.sw == SWStage::VS || program->stage.sw == SWStage::TES);
+   increase_priority |= program->is_prolog;
+
+   for (Block& block : program->blocks) {
+      std::vector<aco_ptr<Instruction>> new_instructions;
+      new_instructions.reserve(block.instructions.size() + 6);
+
+      Builder bld(program, &new_instructions);
+
+      if (increase_priority && block.index == 0) {
+         if (!block.instructions.empty() && block.instructions[0]->opcode == aco_opcode::s_setprio)
+            block.instructions[0]->salu().imm = MAX2(block.instructions[0]->salu().imm, 2);
+         else
+            bld.sopp(aco_opcode::s_setprio, 2);
+      }
+
+      for (unsigned i = 0; i < block.instructions.size(); i++) {
+         Instruction* instr = block.instructions[i].get();
+         new_instructions.push_back(std::move(block.instructions[i]));
+
+         if (instr->opcode == aco_opcode::s_setprio) {
+            instr->salu().imm = MAX2(instr->salu().imm, 2);
+            continue;
+         }
+
+         bool end_of_export_sequence = instr->isEXP() && (i == block.instructions.size() - 1 ||
+                                                          !block.instructions[i + 1]->isEXP());
+         if (!end_of_export_sequence)
+            continue;
+
+         bool before_endpgm = false;
+         if (i != block.instructions.size() - 1) {
+            before_endpgm = block.instructions[i + 1]->opcode == aco_opcode::s_endpgm;
+         } else {
+            /* Does this fallthrough to a s_endpgm? */
+            for (unsigned j = block.index + 1; j < program->blocks.size(); j++) {
+               if (program->blocks[j].instructions.size() == 1 &&
+                   program->blocks[j].instructions[0]->opcode == aco_opcode::s_endpgm)
+                  before_endpgm = true;
+               if (!program->blocks[j].instructions.empty())
+                  break;
+            }
+         }
+
+         bld.sopp(aco_opcode::s_setprio, 0);
+         if (!before_endpgm)
+            bld.sopk(aco_opcode::s_waitcnt_expcnt, Operand(sgpr_null, s1), 0);
+         bld.sopp(aco_opcode::s_nop, 0);
+         bld.sopp(aco_opcode::s_nop, 0);
+         if (!before_endpgm)
+            bld.sopp(aco_opcode::s_setprio, 2);
+      }
+
+      block.instructions = std::move(new_instructions);
+   }
+}
+
 } /* end namespace */
 
 void
@@ -1785,6 +1849,10 @@ insert_NOPs(Program* program)
       mitigate_hazards<NOP_ctx_gfx10, handle_instruction_gfx10, resolve_all_gfx10>(program);
    else
       mitigate_hazards<NOP_ctx_gfx6, handle_instruction_gfx6, resolve_all_gfx6>(program);
+
+   if (program->gfx_level == GFX11_5 && (program->stage.hw == AC_HW_NEXT_GEN_GEOMETRY_SHADER ||
+                                         program->stage.hw == AC_HW_PIXEL_SHADER))
+      required_export_priority(program);
 }
 
 } // namespace aco

src/amd/compiler/aco_instruction_selection.cpp

@@ -12912,7 +12912,9 @@ select_vs_prolog(Program* program, const struct aco_vs_prolog_info* pinfo, ac_sh
 
    block->instructions.reserve(16 + pinfo->num_attributes * 4);
 
-   bld.sopp(aco_opcode::s_setprio, 0x3u);
+   /* Besides performance, the purpose of this is also for the FeatureRequiredExportPriority GFX11.5
+    * issue. */
+   bld.sopp(aco_opcode::s_setprio, 3);
 
    uint32_t attrib_mask = BITFIELD_MASK(pinfo->num_attributes);
    bool has_nontrivial_divisors = pinfo->nontrivial_divisors;

src/amd/compiler/aco_interface.cpp

@@ -222,6 +222,7 @@ aco_compile_shader_part(const struct aco_compiler_options* options,
    program->debug.private_data = options->debug.private_data;
 
    program->is_prolog = is_prolog;
+   program->is_epilog = !is_prolog;
 
    /* Instruction selection */
    select_shader_part(program.get(), pinfo, &config, options, info, args);

src/amd/compiler/aco_ir.h

@@ -2035,6 +2035,7 @@ public:
    bool has_pops_overlapped_waves_wait = false;
    bool has_color_exports = false;
    bool is_prolog = false;
+   bool is_epilog = false;
 
    std::vector<uint8_t> constant_data;
    Temp private_segment_buffer;