aco: treat VINTERP_INREG as VALU

It's just v_fma with fixed DPP8 and builtin s_waitcnt_expcnt, so it can mostly
be handled as a pure VALU instruction.

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21023>

diff --git a/src/amd/compiler/aco_insert_NOPs.cpp b/src/amd/compiler/aco_insert_NOPs.cpp
index e9a0c32..243388a 100644 (file)
--- a/src/amd/compiler/aco_insert_NOPs.cpp
+++ b/src/amd/compiler/aco_insert_NOPs.cpp
@@ -1054,7 +1054,7 @@ handle_lds_direct_valu_hazard_instr(LdsDirectVALUHazardGlobalState& global_state
                                     LdsDirectVALUHazardBlockState& block_state,
                                     aco_ptr<Instruction>& instr)
 {
-   if (instr->isVALU() || instr->isVINTERP_INREG()) {
+   if (instr->isVALU()) {
       block_state.has_trans |= instr->isTrans();
 
       bool uses_vgpr = false;
@@ -1153,7 +1153,7 @@ handle_valu_partial_forwarding_hazard_instr(VALUPartialForwardingHazardGlobalSta
    if (instr->isSALU() && !instr->definitions.empty()) {
       if (block_state.state == written_after_exec_write && instr_writes_exec(instr))
          block_state.state = exec_written;
-   } else if (instr->isVALU() || instr->isVINTERP_INREG()) {
+   } else if (instr->isVALU()) {
       bool vgpr_write = false;
       for (Definition& def : instr->definitions) {
          if (def.physReg().reg() < 256)
@@ -1239,7 +1239,7 @@ handle_valu_partial_forwarding_hazard(State& state, aco_ptr<Instruction>& instr)
     * For the hazard, there must be less than 3 VALU between the first and second VGPR writes.
     * There also must be less than 5 VALU between the second VGPR write and the current instruction.
     */
-   if (state.program->wave_size != 64 || (!instr->isVALU() && !instr->isVINTERP_INREG()))
+   if (state.program->wave_size != 64 || !instr->isVALU())
       return false;
 
    unsigned num_vgprs = 0;
@@ -1319,7 +1319,7 @@ handle_instruction_gfx11(State& state, NOP_ctx_gfx11& ctx, aco_ptr<Instruction>&
     * VALU reads VGPR written by transcendental instruction without 6+ VALU or 2+ transcendental
     * in-between.
     */
-   if (va_vdst > 0 && (instr->isVALU() || instr->isVINTERP_INREG())) {
+   if (va_vdst > 0 && instr->isVALU()) {
       uint8_t num_valu = 15;
       uint8_t num_trans = 15;
       for (Operand& op : instr->operands) {
@@ -1362,7 +1362,7 @@ handle_instruction_gfx11(State& state, NOP_ctx_gfx11& ctx, aco_ptr<Instruction>&
    if (sa_sdst == 0)
       ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_salu.reset();
 
-   if (instr->isVALU() || instr->isVINTERP_INREG()) {
+   if (instr->isVALU()) {
       bool is_trans = instr->isTrans();
 
       ctx.valu_since_wr_by_trans.inc();
@@ -1419,7 +1419,7 @@ handle_instruction_gfx11(State& state, NOP_ctx_gfx11& ctx, aco_ptr<Instruction>&
       for (Operand& op : instr->operands)
          fill_vgpr_bitset(ctx.vgpr_used_by_ds, op.physReg(), op.bytes());
    }
-   if (instr->isVALU() || instr->isVINTERP_INREG() || instr->isEXP() || vm_vsrc == 0) {
+   if (instr->isVALU() || instr->isEXP() || vm_vsrc == 0) {
       ctx.vgpr_used_by_vmem_load.reset();
       ctx.vgpr_used_by_vmem_store.reset();
       ctx.vgpr_used_by_ds.reset();

diff --git a/src/amd/compiler/aco_insert_waitcnt.cpp b/src/amd/compiler/aco_insert_waitcnt.cpp
index b7edf04..948799f 100644 (file)
--- a/src/amd/compiler/aco_insert_waitcnt.cpp
+++ b/src/amd/compiler/aco_insert_waitcnt.cpp
@@ -372,7 +372,7 @@ check_instr(wait_ctx& ctx, wait_imm& wait, alu_delay_info& delay, Instruction* i
             continue;
 
          wait.combine(it->second.imm);
-         if (instr->isVALU() || instr->isSALU() || instr->isVINTERP_INREG())
+         if (instr->isVALU() || instr->isSALU())
             delay.combine(it->second.delay);
       }
    }
@@ -788,7 +788,7 @@ void
 gen_alu(Instruction* instr, wait_ctx& ctx)
 {
    Instruction_cycle_info cycle_info = get_cycle_info(*ctx.program, *instr);
-   bool is_valu = instr->isVALU() || instr->isVINTERP_INREG();
+   bool is_valu = instr->isVALU();
    bool is_trans = instr->isTrans();
    bool clear = instr->isEXP() || instr->isDS() || instr->isMIMG() || instr->isFlatLike() ||
                 instr->isMUBUF() || instr->isMTBUF();

diff --git a/src/amd/compiler/aco_ir.cpp b/src/amd/compiler/aco_ir.cpp
index 4652af8..c4f3367 100644 (file)
--- a/src/amd/compiler/aco_ir.cpp
+++ b/src/amd/compiler/aco_ir.cpp
@@ -343,7 +343,7 @@ can_use_DPP(const aco_ptr<Instruction>& instr, bool pre_ra, bool dpp8)
    if (instr->operands.size() && instr->operands[0].isLiteral())
       return false;
 
-   if (instr->isSDWA() || instr->isVOP3P())
+   if (instr->isSDWA() || instr->isVINTERP_INREG() || instr->isVOP3P())
       return false;
 
    if (!pre_ra && (instr->isVOPC() || instr->definitions.size() > 1) &&

diff --git a/src/amd/compiler/aco_ir.h b/src/amd/compiler/aco_ir.h
index f57ba4a..c3ad28a 100644 (file)
--- a/src/amd/compiler/aco_ir.h
+++ b/src/amd/compiler/aco_ir.h
@@ -1372,7 +1372,7 @@ struct Instruction {
 
    constexpr bool isVALU() const noexcept
    {
-      return isVOP1() || isVOP2() || isVOPC() || isVOP3() || isVOP3P();
+      return isVOP1() || isVOP2() || isVOPC() || isVOP3() || isVOP3P() || isVINTERP_INREG();
    }
 
    constexpr bool isSALU() const noexcept

diff --git a/src/amd/compiler/aco_lower_to_hw_instr.cpp b/src/amd/compiler/aco_lower_to_hw_instr.cpp
index b0be28a..3477563 100644 (file)
--- a/src/amd/compiler/aco_lower_to_hw_instr.cpp
+++ b/src/amd/compiler/aco_lower_to_hw_instr.cpp
@@ -2675,7 +2675,7 @@ lower_to_hw_instr(Program* program)
                         can_remove = false;
                   } else if (inst->isSALU()) {
                      num_scalar++;
-                  } else if (inst->isVALU() || inst->isVINTRP() || inst->isVINTERP_INREG()) {
+                  } else if (inst->isVALU() || inst->isVINTRP()) {
                      num_vector++;
                      /* VALU which writes SGPRs are always executed on GFX10+ */
                      if (ctx.program->gfx_level >= GFX10) {

diff --git a/src/amd/compiler/aco_optimizer_postRA.cpp b/src/amd/compiler/aco_optimizer_postRA.cpp
index 6435a54..b39af7b 100644 (file)
--- a/src/amd/compiler/aco_optimizer_postRA.cpp
+++ b/src/amd/compiler/aco_optimizer_postRA.cpp
@@ -568,7 +568,7 @@ num_encoded_alu_operands(const aco_ptr<Instruction>& instr)
       else if (instr->opcode == aco_opcode::v_writelane_b32_e64 ||
                instr->opcode == aco_opcode::v_writelane_b32)
          return 2; /* potentially VOP3, but reads VDST as SRC2 */
-      else if (instr->isVOP3() || instr->isVOP3P())
+      else if (instr->isVOP3() || instr->isVOP3P() || instr->isVINTERP_INREG())
          return instr->operands.size();
    }
 

diff --git a/src/amd/compiler/aco_register_allocation.cpp b/src/amd/compiler/aco_register_allocation.cpp
index e8a1095..54d4174 100644 (file)
--- a/src/amd/compiler/aco_register_allocation.cpp
+++ b/src/amd/compiler/aco_register_allocation.cpp
@@ -504,7 +504,7 @@ get_subdword_operand_stride(amd_gfx_level gfx_level, const aco_ptr<Instruction>&
    }
 
    assert(rc.bytes() <= 2);
-   if (instr->isVALU() || instr->isVINTERP_INREG()) {
+   if (instr->isVALU()) {
       if (can_use_SDWA(gfx_level, instr, false))
          return rc.bytes();
       if (can_use_opsel(gfx_level, instr->opcode, idx))
@@ -539,7 +539,7 @@ add_subdword_operand(ra_ctx& ctx, aco_ptr<Instruction>& instr, unsigned idx, uns
       return;
 
    assert(rc.bytes() <= 2);
-   if (instr->isVALU() || instr->isVINTERP_INREG()) {
+   if (instr->isVALU()) {
       /* check if we can use opsel */
       if (instr->format == Format::VOP3) {
          assert(byte == 2);
@@ -616,7 +616,7 @@ get_subdword_definition_info(Program* program, const aco_ptr<Instruction>& instr
          return std::make_pair(4, rc.size() * 4u);
    }
 
-   if (instr->isVALU() || instr->isVINTRP() || instr->isVINTERP_INREG()) {
+   if (instr->isVALU() || instr->isVINTRP()) {
       assert(rc.bytes() <= 2);
 
       if (can_use_SDWA(gfx_level, instr, false))
@@ -684,7 +684,7 @@ add_subdword_definition(Program* program, aco_ptr<Instruction>& instr, PhysReg r
    if (instr->isPseudo())
       return;
 
-   if (instr->isVALU() || instr->isVINTERP_INREG()) {
+   if (instr->isVALU()) {
       amd_gfx_level gfx_level = program->gfx_level;
       assert(instr->definitions[0].bytes() <= 2);
 

diff --git a/src/amd/compiler/aco_validate.cpp b/src/amd/compiler/aco_validate.cpp
index 0494f4c..599eb31 100644 (file)
--- a/src/amd/compiler/aco_validate.cpp
+++ b/src/amd/compiler/aco_validate.cpp
@@ -284,7 +284,7 @@ validate_ir(Program* program)
                      instr.get());
          }
 
-         if (instr->isSALU() || instr->isVALU() || instr->isVINTERP_INREG()) {
+         if (instr->isSALU() || instr->isVALU()) {
             /* check literals */
             Operand literal(s1);
             for (unsigned i = 0; i < instr->operands.size(); i++) {
@@ -306,7 +306,7 @@ validate_ir(Program* program)
             }
 
             /* check num sgprs for VALU */
-            if (instr->isVALU() || instr->isVINTERP_INREG()) {
+            if (instr->isVALU()) {
                bool is_shift64 = instr->opcode == aco_opcode::v_lshlrev_b64 ||
                                  instr->opcode == aco_opcode::v_lshrrev_b64 ||
                                  instr->opcode == aco_opcode::v_ashrrev_i64;
@@ -929,7 +929,7 @@ get_subdword_bytes_written(Program* program, const aco_ptr<Instruction>& instr,
 
    if (instr->isPseudo())
       return gfx_level >= GFX8 ? def.bytes() : def.size() * 4u;
-   if (instr->isVALU() || instr->isVINTERP_INREG()) {
+   if (instr->isVALU()) {
       assert(def.bytes() <= 2);
       if (instr->isSDWA())
          return instr->sdwa().dst_sel.size();