aco: consider intrinsic access in visit_{load,store}_image

[mesa.git] / src / amd / compiler / aco_insert_exec_mask.cpp

diff --git a/src/amd/compiler/aco_insert_exec_mask.cpp b/src/amd/compiler/aco_insert_exec_mask.cpp
index 539d0a2580b880da4fe41531da2f65e5b9359ba5..7246fb74e0cf1411c4e77041857dc48a3bc3033e 100644 (file)
--- a/src/amd/compiler/aco_insert_exec_mask.cpp
+++ b/src/amd/compiler/aco_insert_exec_mask.cpp
@@ -107,10 +107,13 @@ bool pred_by_exec_mask(aco_ptr<Instruction>& instr) {
    if (instr->format == Format::PSEUDO) {
       switch (instr->opcode) {
       case aco_opcode::p_create_vector:
-         return instr->definitions[0].getTemp().type() == RegType::vgpr;
       case aco_opcode::p_extract_vector:
       case aco_opcode::p_split_vector:
-         return instr->operands[0].getTemp().type() == RegType::vgpr;
+         for (Definition def : instr->definitions) {
+            if (def.getTemp().type() == RegType::vgpr)
+               return true;
+         }
+         return false;
       case aco_opcode::p_spill:
       case aco_opcode::p_reload:
          return false;
@@ -186,11 +189,14 @@ void get_block_needs(wqm_ctx &ctx, exec_ctx &exec_ctx, Block* block)
 
    if (block->kind & block_kind_top_level) {
       if (ctx.loop && ctx.wqm) {
-         /* mark all break conditions as WQM */
          unsigned block_idx = block->index + 1;
          while (!(ctx.program->blocks[block_idx].kind & block_kind_top_level)) {
+            /* flag all break conditions as WQM:
+             * the conditions might be computed outside the nested CF */
             if (ctx.program->blocks[block_idx].kind & block_kind_break)
                mark_block_wqm(ctx, block_idx);
+            /* flag all blocks as WQM to ensure we enter all (nested) loops in WQM */
+            exec_ctx.info[block_idx].block_needs |= WQM;
             block_idx++;
          }
       } else if (ctx.loop && !ctx.wqm) {
@@ -315,6 +321,13 @@ void transition_to_WQM(exec_ctx& ctx, Builder bld, unsigned idx)
       return;
    if (ctx.info[idx].exec.back().second & mask_type_global) {
       Temp exec_mask = ctx.info[idx].exec.back().first;
+      /* TODO: we might generate better code if we pass the uncopied "exec_mask"
+       * directly to the s_wqm (we still need to keep this parallelcopy for
+       * potential later uses of exec_mask though). We currently can't do this
+       * because of a RA bug. */
+      exec_mask = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm), bld.exec(exec_mask));
+      ctx.info[idx].exec.back().first = exec_mask;
+
       exec_mask = bld.sop1(Builder::s_wqm, bld.def(bld.lm, exec), bld.def(s1, scc), exec_mask);
       ctx.info[idx].exec.emplace_back(exec_mask, mask_type_global | mask_type_wqm);
       return;
@@ -367,7 +380,7 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
       bld.insert(std::move(startpgm));
 
       /* exec seems to need to be manually initialized with combined shaders */
-      if (util_bitcount(ctx.program->stage & sw_mask) > 1) {
+      if (util_bitcount(ctx.program->stage & sw_mask) > 1 || (ctx.program->stage & hw_ngg_gs)) {
          bld.sop1(Builder::s_mov, bld.exec(Definition(exec_mask)), bld.lm == s2 ? Operand(UINT64_MAX) : Operand(UINT32_MAX));
          instructions[0]->definitions.pop_back();
       }
@@ -519,7 +532,7 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
             /* create phi for loop footer */
             aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
             phi->definitions[0] = bld.def(bld.lm);
-            if (k == info.num_exec_masks - 1) {
+            if (k == info.num_exec_masks - 1u) {
                phi->definitions[0].setFixed(exec);
                need_parallelcopy = false;
             }
@@ -961,10 +974,14 @@ void add_branch_code(exec_ctx& ctx, Block* block)
       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
       block->instructions.pop_back();
 
-      while (!(ctx.info[idx].exec.back().second & mask_type_loop))
+      bool need_parallelcopy = false;
+      while (!(ctx.info[idx].exec.back().second & mask_type_loop)) {
          ctx.info[idx].exec.pop_back();
+         need_parallelcopy = true;
+      }
 
-      ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec), ctx.info[idx].exec.back().first);
+      if (need_parallelcopy)
+         ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec), ctx.info[idx].exec.back().first);
       bld.branch(aco_opcode::p_cbranch_nz, bld.exec(ctx.info[idx].exec.back().first), block->linear_succs[1], block->linear_succs[0]);
       return;
    }
@@ -1045,7 +1062,7 @@ void add_branch_code(exec_ctx& ctx, Block* block)
          cond = bld.tmp(s1);
          Temp exec_mask = ctx.info[idx].exec[exec_idx].first;
          exec_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.scc(Definition(cond)),
-                              exec_mask, current_exec);
+                              exec_mask, bld.exec(current_exec));
          ctx.info[idx].exec[exec_idx].first = exec_mask;
          if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
             break;