aco: forget valu delays after certain s_waitcnt_depctr/LDSDIR

fossil-db (navi31):
Totals from 55242 (69.58% of 79395) affected shaders:
Instrs: 40507666 -> 40138006 (-0.91%); split: -0.91%, +0.00%
CodeSize: 212516104 -> 211025880 (-0.70%); split: -0.70%, +0.00%
Latency: 281643258 -> 281628053 (-0.01%); split: -0.01%, +0.00%
InvThroughput: 46370668 -> 46369637 (-0.00%); split: -0.00%, +0.00%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Acked-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23337>

src/amd/compiler/aco_insert_NOPs.cpp

@@ -1136,19 +1136,6 @@ test_vgpr_bitset(std::bitset<256>& set, Operand op)
 }
 
 /* GFX11 */
-unsigned
-parse_vdst_wait(aco_ptr<Instruction>& instr)
-{
-   if (instr->isVMEM() || instr->isFlatLike() || instr->isDS() || instr->isEXP())
-      return 0;
-   else if (instr->isLDSDIR())
-      return instr->ldsdir().wait_vdst;
-   else if (instr->opcode == aco_opcode::s_waitcnt_depctr)
-      return (instr->salu().imm >> 12) & 0xf;
-   else
-      return 15;
-}
-
 struct LdsDirectVALUHazardGlobalState {
    unsigned wait_vdst = 15;
    PhysReg vgpr;
@@ -1188,7 +1175,7 @@ handle_lds_direct_valu_hazard_instr(LdsDirectVALUHazardGlobalState& global_state
       block_state.num_valu++;
    }
 
-   if (parse_vdst_wait(instr) == 0)
+   if (parse_vdst_wait(instr.get()) == 0)
       return true;
 
    block_state.num_instrs++;
@@ -1310,7 +1297,7 @@ handle_valu_partial_forwarding_hazard_instr(VALUPartialForwardingHazardGlobalSta
       }
 
       block_state.num_valu_since_read++;
-   } else if (parse_vdst_wait(instr) == 0) {
+   } else if (parse_vdst_wait(instr.get()) == 0) {
       return true;
    }
 
@@ -1407,7 +1394,7 @@ handle_instruction_gfx11(State& state, NOP_ctx_gfx11& ctx, aco_ptr<Instruction>&
       ctx.has_Vcmpx = false;
    }
 
-   unsigned va_vdst = parse_vdst_wait(instr);
+   unsigned va_vdst = parse_vdst_wait(instr.get());
    unsigned vm_vsrc = 7;
    unsigned sa_sdst = 1;
 
@@ -1608,7 +1595,7 @@ handle_instruction_gfx11(State& state, NOP_ctx_gfx11& ctx, aco_ptr<Instruction>&
 bool
 has_vdst0_since_valu_instr(bool& global_state, unsigned& block_state, aco_ptr<Instruction>& pred)
 {
-   if (parse_vdst_wait(pred) == 0)
+   if (parse_vdst_wait(pred.get()) == 0)
       return true;
 
    if (--block_state == 0) {

src/amd/compiler/aco_insert_delay_alu.cpp

@@ -192,6 +192,16 @@ update_alu(delay_ctx& ctx, bool is_valu, bool is_trans, int cycles)
 void
 kill_alu(alu_delay_info& delay, Instruction* instr, delay_ctx& ctx)
 {
+   if (parse_vdst_wait(instr) == 0) {
+      std::map<PhysReg, alu_delay_info>::iterator it = ctx.gpr_map.begin();
+      while (it != ctx.gpr_map.end()) {
+         alu_delay_info& entry = it->second;
+         entry.valu_instrs = alu_delay_info::valu_nop;
+         entry.trans_instrs = alu_delay_info::trans_nop;
+         it = it->second.fixup() ? ctx.gpr_map.erase(it) : std::next(it);
+      }
+   }
+
    if (instr->isVALU() || instr->isSALU())
       check_alu(ctx, delay, instr);
 
@@ -213,12 +223,6 @@ kill_alu(alu_delay_info& delay, Instruction* instr, delay_ctx& ctx)
 void
 gen_alu(Instruction* instr, delay_ctx& ctx)
 {
-   if (instr->isEXP() || instr->isDS() || instr->isMIMG() || instr->isFlatLike() ||
-       instr->isMUBUF() || instr->isMTBUF()) {
-      ctx.gpr_map.clear();
-      return;
-   }
-
    Instruction_cycle_info cycle_info = get_cycle_info(*ctx.program, *instr);
    bool is_valu = instr->isVALU();
    bool is_trans = instr->isTrans();

src/amd/compiler/aco_ir.cpp

@@ -1398,6 +1398,19 @@ get_vmem_type(enum amd_gfx_level gfx_level, Instruction* instr)
    return 0;
 }
 
+unsigned
+parse_vdst_wait(Instruction* instr)
+{
+   if (instr->isVMEM() || instr->isFlatLike() || instr->isDS() || instr->isEXP())
+      return 0;
+   else if (instr->isLDSDIR())
+      return instr->ldsdir().wait_vdst;
+   else if (instr->opcode == aco_opcode::s_waitcnt_depctr)
+      return (instr->salu().imm >> 12) & 0xf;
+   else
+      return 15;
+}
+
 bool
 dealloc_vgprs(Program* program)
 {

src/amd/compiler/aco_ir.h

@@ -1853,6 +1853,8 @@ enum vmem_type : uint8_t {
  */
 uint8_t get_vmem_type(enum amd_gfx_level gfx_level, Instruction* instr);
 
+unsigned parse_vdst_wait(Instruction* instr);
+
 enum block_kind {
    /* uniform indicates that leaving this block,
     * all actives lanes stay active */