aco: add ACO_DEBUG=force-waitcnt to emit wait-states

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

diff --git a/src/amd/compiler/aco_insert_waitcnt.cpp b/src/amd/compiler/aco_insert_waitcnt.cpp
index db0c7c0d68c38457a98a857e4632bcc178734ad2..751892e44368a476d1f1ff44261826d536037691 100644 (file)
--- a/src/amd/compiler/aco_insert_waitcnt.cpp
+++ b/src/amd/compiler/aco_insert_waitcnt.cpp
@@ -473,9 +473,32 @@ wait_imm perform_barrier(wait_ctx& ctx, memory_sync_info sync, unsigned semantic
    return imm;
 }
 
+void force_waitcnt(wait_ctx& ctx, wait_imm& imm)
+{
+   if (ctx.vm_cnt)
+      imm.vm = 0;
+   if (ctx.exp_cnt)
+      imm.exp = 0;
+   if (ctx.lgkm_cnt)
+      imm.lgkm = 0;
+
+   if (ctx.chip_class >= GFX10) {
+      if (ctx.vs_cnt)
+         imm.vs = 0;
+   }
+}
+
 wait_imm kill(Instruction* instr, wait_ctx& ctx, memory_sync_info sync_info)
 {
    wait_imm imm;
+
+   if (debug_flags & DEBUG_FORCE_WAITCNT) {
+      /* Force emitting waitcnt states right after the instruction if there is
+       * something to wait for.
+       */
+      force_waitcnt(ctx, imm);
+   }
+
    if (ctx.exp_cnt || ctx.vm_cnt || ctx.lgkm_cnt)
       imm.combine(check_instr(instr, ctx));