amd/compiler/aco_assembler.cpp

/*
 * Copyright © 2018 Valve Corporation
 *
 * SPDX-License-Identifier: MIT
 */

#include "aco_builder.h"
#include "aco_ir.h"

#include "common/sid.h"

#include "util/memstream.h"

#include "ac_shader_util.h"
#include <algorithm>
#include <cstdint>
#include <map>
#include <vector>

namespace aco {

struct constaddr_info {
   unsigned getpc_end;
   unsigned add_literal;
};

struct branch_info {
   unsigned pos;
   unsigned target;
};

struct asm_context {
   Program* program;
   enum amd_gfx_level gfx_level;
   std::vector<branch_info> branches;
   std::map<unsigned, constaddr_info> constaddrs;
   std::map<unsigned, constaddr_info> resumeaddrs;
   std::vector<struct aco_symbol>* symbols;
   uint32_t loop_header = -1u;
   uint32_t loop_exit = 0u;
   const int16_t* opcode;
   // TODO: keep track of branch instructions referring blocks
   // and, when emitting the block, correct the offset in instr
   asm_context(Program* program_, std::vector<struct aco_symbol>* symbols_)
       : program(program_), gfx_level(program->gfx_level), symbols(symbols_)
   {
      if (gfx_level <= GFX7)
         opcode = &instr_info.opcode_gfx7[0];
      else if (gfx_level <= GFX9)
         opcode = &instr_info.opcode_gfx9[0];
      else if (gfx_level <= GFX10_3)
         opcode = &instr_info.opcode_gfx10[0];
      else if (gfx_level <= GFX11_5)
         opcode = &instr_info.opcode_gfx11[0];
      else
         opcode = &instr_info.opcode_gfx12[0];
   }

   int subvector_begin_pos = -1;
};

unsigned
get_mimg_nsa_dwords(const Instruction* instr)
{
   unsigned addr_dwords = instr->operands.size() - 3;
   for (unsigned i = 1; i < addr_dwords; i++) {
      if (instr->operands[3 + i].physReg() !=
          instr->operands[3 + (i - 1)].physReg().advance(instr->operands[3 + (i - 1)].bytes()))
         return DIV_ROUND_UP(addr_dwords - 1, 4);
   }
   return 0;
}

unsigned
get_vopd_opy_start(const Instruction* instr)
{
   switch (instr->opcode) {
   case aco_opcode::v_dual_fmac_f32:
   case aco_opcode::v_dual_fmaak_f32:
   case aco_opcode::v_dual_fmamk_f32:
   case aco_opcode::v_dual_cndmask_b32:
   case aco_opcode::v_dual_dot2acc_f32_f16:
   case aco_opcode::v_dual_dot2acc_f32_bf16: return 3;
   case aco_opcode::v_dual_mov_b32: return 1;
   default: return 2;
   }
}

uint32_t
reg(asm_context& ctx, PhysReg reg)
{
   if (ctx.gfx_level >= GFX11) {
      if (reg == m0)
         return sgpr_null.reg();
      else if (reg == sgpr_null)
         return m0.reg();
   }
   return reg.reg();
}

ALWAYS_INLINE uint32_t
reg(asm_context& ctx, Operand op, unsigned width = 32)
{
   return reg(ctx, op.physReg()) & BITFIELD_MASK(width);
}

ALWAYS_INLINE uint32_t
reg(asm_context& ctx, Definition def, unsigned width = 32)
{
   return reg(ctx, def.physReg()) & BITFIELD_MASK(width);
}

bool
needs_vop3_gfx11(asm_context& ctx, Instruction* instr)
{
   if (ctx.gfx_level <= GFX10_3)
      return false;

   uint8_t mask = get_gfx11_true16_mask(instr->opcode);
   if (!mask)
      return false;

   u_foreach_bit (i, mask & 0x3) {
      if (instr->operands[i].physReg().reg() >= (256 + 128))
         return true;
   }
   if ((mask & 0x8) && instr->definitions[0].physReg().reg() >= (256 + 128))
      return true;
   return false;
}

template <typename T>
uint32_t
get_gfx12_cpol(const T& instr)
{
   uint32_t scope = instr.cache.gfx12.scope;
   uint32_t th = instr.cache.gfx12.temporal_hint;
   return scope | (th << 2);
}

void
emit_sop2_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];

   uint32_t encoding = (0b10 << 30);
   encoding |= opcode << 23;
   encoding |= !instr->definitions.empty() ? reg(ctx, instr->definitions[0]) << 16 : 0;
   encoding |= instr->operands.size() >= 2 ? reg(ctx, instr->operands[1]) << 8 : 0;
   encoding |= !instr->operands.empty() ? reg(ctx, instr->operands[0]) : 0;
   out.push_back(encoding);
}

void
emit_sopk_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const SALU_instruction& sopk = instr->salu();
   assert(sopk.imm <= UINT16_MAX);
   uint16_t imm = sopk.imm;

   if (instr->opcode == aco_opcode::s_subvector_loop_begin) {
      assert(ctx.gfx_level >= GFX10);
      assert(ctx.subvector_begin_pos == -1);
      ctx.subvector_begin_pos = out.size();
   } else if (instr->opcode == aco_opcode::s_subvector_loop_end) {
      assert(ctx.gfx_level >= GFX10);
      assert(ctx.subvector_begin_pos != -1);
      /* Adjust s_subvector_loop_begin instruction to the address after the end  */
      out[ctx.subvector_begin_pos] |= (out.size() - ctx.subvector_begin_pos);
      /* Adjust s_subvector_loop_end instruction to the address after the beginning  */
      imm = (uint16_t)(ctx.subvector_begin_pos - (int)out.size());
      ctx.subvector_begin_pos = -1;
   }

   uint32_t encoding = (0b1011 << 28);
   encoding |= opcode << 23;
   encoding |= !instr->definitions.empty() && !(instr->definitions[0].physReg() == scc)
                  ? reg(ctx, instr->definitions[0]) << 16
               : !instr->operands.empty() && instr->operands[0].physReg() <= 127
                  ? reg(ctx, instr->operands[0]) << 16
                  : 0;
   encoding |= imm;
   out.push_back(encoding);
}

void
emit_sop1_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];

   uint32_t encoding = (0b101111101 << 23);
   encoding |= !instr->definitions.empty() ? reg(ctx, instr->definitions[0]) << 16 : 0;
   encoding |= opcode << 8;
   encoding |= !instr->operands.empty() ? reg(ctx, instr->operands[0]) : 0;
   out.push_back(encoding);
}

void
emit_sopc_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];

   uint32_t encoding = (0b101111110 << 23);
   encoding |= opcode << 16;
   encoding |= instr->operands.size() == 2 ? reg(ctx, instr->operands[1]) << 8 : 0;
   encoding |= !instr->operands.empty() ? reg(ctx, instr->operands[0]) : 0;
   out.push_back(encoding);
}

void
emit_sopp_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr,
                      bool force_imm = false)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const SALU_instruction& sopp = instr->salu();

   uint32_t encoding = (0b101111111 << 23);
   encoding |= opcode << 16;

   if (!force_imm && instr_info.classes[(int)instr->opcode] == instr_class::branch) {
      ctx.branches.push_back({(unsigned)out.size(), sopp.imm});
   } else {
      assert(sopp.imm <= UINT16_MAX);
      encoding |= (uint16_t)sopp.imm;
   }
   out.push_back(encoding);
}

void
emit_smem_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const SMEM_instruction& smem = instr->smem();
   bool glc = smem.cache.value & ac_glc;
   bool dlc = smem.cache.value & ac_dlc;

   bool soe = instr->operands.size() >= (!instr->definitions.empty() ? 3 : 4);
   bool is_load = !instr->definitions.empty();
   uint32_t encoding = 0;

   if (ctx.gfx_level <= GFX7) {
      encoding = (0b11000 << 27);
      encoding |= opcode << 22;
      encoding |= instr->definitions.size() ? reg(ctx, instr->definitions[0]) << 15 : 0;
      encoding |= instr->operands.size() ? (reg(ctx, instr->operands[0]) >> 1) << 9 : 0;
      if (instr->operands.size() >= 2) {
         if (!instr->operands[1].isConstant()) {
            encoding |= reg(ctx, instr->operands[1]);
         } else if (instr->operands[1].constantValue() >= 1024) {
            encoding |= 255; /* SQ_SRC_LITERAL */
         } else {
            encoding |= instr->operands[1].constantValue() >> 2;
            encoding |= 1 << 8;
         }
      }
      out.push_back(encoding);
      /* SMRD instructions can take a literal on GFX7 */
      if (instr->operands.size() >= 2 && instr->operands[1].isConstant() &&
          instr->operands[1].constantValue() >= 1024)
         out.push_back(instr->operands[1].constantValue() >> 2);
      return;
   }

   if (ctx.gfx_level <= GFX9) {
      encoding = (0b110000 << 26);
      assert(!dlc); /* Device-level coherent is not supported on GFX9 and lower */
      /* We don't use the NV bit. */
   } else {
      encoding = (0b111101 << 26);
      if (ctx.gfx_level <= GFX11_5)
         encoding |= dlc ? 1 << (ctx.gfx_level >= GFX11 ? 13 : 14) : 0;
   }

   if (ctx.gfx_level <= GFX11_5) {
      encoding |= opcode << 18;
      encoding |= glc ? 1 << (ctx.gfx_level >= GFX11 ? 14 : 16) : 0;
   } else {
      encoding |= opcode << 13;
      encoding |= get_gfx12_cpol(smem) << 21;
   }

   if (ctx.gfx_level <= GFX9) {
      if (instr->operands.size() >= 2)
         encoding |= instr->operands[1].isConstant() ? 1 << 17 : 0; /* IMM - immediate enable */
   }
   if (ctx.gfx_level == GFX9) {
      encoding |= soe ? 1 << 14 : 0;
   }

   if (is_load || instr->operands.size() >= 3) { /* SDATA */
      encoding |= (is_load ? reg(ctx, instr->definitions[0]) : reg(ctx, instr->operands[2])) << 6;
   }
   if (instr->operands.size() >= 1) { /* SBASE */
      encoding |= reg(ctx, instr->operands[0]) >> 1;
   }

   out.push_back(encoding);
   encoding = 0;

   int32_t offset = 0;
   uint32_t soffset =
      ctx.gfx_level >= GFX10
         ? reg(ctx, sgpr_null) /* On GFX10 this is disabled by specifying SGPR_NULL */
         : 0;                  /* On GFX9, it is disabled by the SOE bit (and it's not present on
                                  GFX8 and below) */
   if (instr->operands.size() >= 2) {
      const Operand& op_off1 = instr->operands[1];
      if (ctx.gfx_level <= GFX9) {
         offset = op_off1.isConstant() ? op_off1.constantValue() : reg(ctx, op_off1);
      } else {
         /* GFX10 only supports constants in OFFSET, so put the operand in SOFFSET if it's an
          * SGPR */
         if (op_off1.isConstant()) {
            offset = op_off1.constantValue();
         } else {
            soffset = reg(ctx, op_off1);
            assert(!soe); /* There is no place to put the other SGPR offset, if any */
         }
      }

      if (soe) {
         const Operand& op_off2 = instr->operands.back();
         assert(ctx.gfx_level >= GFX9); /* GFX8 and below don't support specifying a constant
                                            and an SGPR at the same time */
         assert(!op_off2.isConstant());
         soffset = reg(ctx, op_off2);
      }
   }
   encoding |= offset;
   encoding |= soffset << 25;

   out.push_back(encoding);
}

void
emit_vop2_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const VALU_instruction& valu = instr->valu();

   uint32_t encoding = 0;
   encoding |= opcode << 25;
   encoding |= reg(ctx, instr->definitions[0], 8) << 17;
   encoding |= (valu.opsel[3] ? 128 : 0) << 17;
   encoding |= reg(ctx, instr->operands[1], 8) << 9;
   encoding |= (valu.opsel[1] ? 128 : 0) << 9;
   encoding |= reg(ctx, instr->operands[0]);
   encoding |= valu.opsel[0] ? 128 : 0;
   out.push_back(encoding);
}

void
emit_vop1_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const VALU_instruction& valu = instr->valu();

   uint32_t encoding = (0b0111111 << 25);
   if (!instr->definitions.empty()) {
      encoding |= reg(ctx, instr->definitions[0], 8) << 17;
      encoding |= (valu.opsel[3] ? 128 : 0) << 17;
   }
   encoding |= opcode << 9;
   if (!instr->operands.empty()) {
      encoding |= reg(ctx, instr->operands[0]);
      encoding |= valu.opsel[0] ? 128 : 0;
   }
   out.push_back(encoding);
}

void
emit_vopc_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const VALU_instruction& valu = instr->valu();

   uint32_t encoding = (0b0111110 << 25);
   encoding |= opcode << 17;
   encoding |= reg(ctx, instr->operands[1], 8) << 9;
   encoding |= (valu.opsel[1] ? 128 : 0) << 9;
   encoding |= reg(ctx, instr->operands[0]);
   encoding |= valu.opsel[0] ? 128 : 0;
   out.push_back(encoding);
}

void
emit_vintrp_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const VINTRP_instruction& interp = instr->vintrp();

   uint32_t encoding = 0;
   if (instr->opcode == aco_opcode::v_interp_p1ll_f16 ||
       instr->opcode == aco_opcode::v_interp_p1lv_f16 ||
       instr->opcode == aco_opcode::v_interp_p2_legacy_f16 ||
       instr->opcode == aco_opcode::v_interp_p2_f16 ||
       instr->opcode == aco_opcode::v_interp_p2_hi_f16) {
      if (ctx.gfx_level == GFX8 || ctx.gfx_level == GFX9) {
         encoding = (0b110100 << 26);
      } else if (ctx.gfx_level >= GFX10) {
         encoding = (0b110101 << 26);
      } else {
         unreachable("Unknown gfx_level.");
      }

      unsigned opsel = instr->opcode == aco_opcode::v_interp_p2_hi_f16 ? 0x8 : 0;

      encoding |= opcode << 16;
      encoding |= opsel << 11;
      encoding |= reg(ctx, instr->definitions[0], 8);
      out.push_back(encoding);

      encoding = 0;
      encoding |= interp.attribute;
      encoding |= interp.component << 6;
      encoding |= interp.high_16bits << 8;
      encoding |= reg(ctx, instr->operands[0]) << 9;
      if (instr->opcode == aco_opcode::v_interp_p2_f16 ||
          instr->opcode == aco_opcode::v_interp_p2_hi_f16 ||
          instr->opcode == aco_opcode::v_interp_p2_legacy_f16 ||
          instr->opcode == aco_opcode::v_interp_p1lv_f16) {
         encoding |= reg(ctx, instr->operands[2]) << 18;
      }
      out.push_back(encoding);
   } else {
      if (ctx.gfx_level == GFX8 || ctx.gfx_level == GFX9) {
         encoding = (0b110101 << 26); /* Vega ISA doc says 110010 but it's wrong */
      } else {
         encoding = (0b110010 << 26);
      }

      assert(encoding);
      encoding |= reg(ctx, instr->definitions[0], 8) << 18;
      encoding |= opcode << 16;
      encoding |= interp.attribute << 10;
      encoding |= interp.component << 8;
      if (instr->opcode == aco_opcode::v_interp_mov_f32)
         encoding |= (0x3 & instr->operands[0].constantValue());
      else
         encoding |= reg(ctx, instr->operands[0], 8);
      out.push_back(encoding);
   }
}

void
emit_vinterp_inreg_instruction(asm_context& ctx, std::vector<uint32_t>& out,
                               const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const VINTERP_inreg_instruction& interp = instr->vinterp_inreg();

   uint32_t encoding = (0b11001101 << 24);
   encoding |= reg(ctx, instr->definitions[0], 8);
   encoding |= (uint32_t)interp.wait_exp << 8;
   encoding |= (uint32_t)interp.opsel << 11;
   encoding |= (uint32_t)interp.clamp << 15;
   encoding |= opcode << 16;
   out.push_back(encoding);

   encoding = 0;
   for (unsigned i = 0; i < instr->operands.size(); i++)
      encoding |= reg(ctx, instr->operands[i]) << (i * 9);
   for (unsigned i = 0; i < 3; i++)
      encoding |= interp.neg[i] << (29 + i);
   out.push_back(encoding);
}

void
emit_vopd_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const VOPD_instruction& vopd = instr->vopd();

   uint32_t encoding = (0b110010 << 26);
   encoding |= reg(ctx, instr->operands[0]);
   if (instr->opcode != aco_opcode::v_dual_mov_b32)
      encoding |= reg(ctx, instr->operands[1], 8) << 9;
   encoding |= (uint32_t)ctx.opcode[(int)vopd.opy] << 17;
   encoding |= opcode << 22;
   out.push_back(encoding);

   unsigned opy_start = get_vopd_opy_start(instr);

   encoding = reg(ctx, instr->operands[opy_start]);
   if (vopd.opy != aco_opcode::v_dual_mov_b32)
      encoding |= reg(ctx, instr->operands[opy_start + 1], 8) << 9;
   encoding |= (reg(ctx, instr->definitions[1], 8) >> 1) << 17;
   encoding |= reg(ctx, instr->definitions[0], 8) << 24;
   out.push_back(encoding);
}

void
emit_ds_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const DS_instruction& ds = instr->ds();

   uint32_t encoding = (0b110110 << 26);
   if (ctx.gfx_level == GFX8 || ctx.gfx_level == GFX9) {
      encoding |= opcode << 17;
      encoding |= (ds.gds ? 1 : 0) << 16;
   } else {
      encoding |= opcode << 18;
      encoding |= (ds.gds ? 1 : 0) << 17;
   }
   encoding |= ((0xFF & ds.offset1) << 8);
   encoding |= (0xFFFF & ds.offset0);
   out.push_back(encoding);
   encoding = 0;
   if (!instr->definitions.empty())
      encoding |= reg(ctx, instr->definitions[0], 8) << 24;
   for (unsigned i = 0; i < MIN2(instr->operands.size(), 3); i++) {
      const Operand& op = instr->operands[i];
      if (op.physReg() != m0 && !op.isUndefined())
         encoding |= reg(ctx, op, 8) << (8 * i);
   }
   out.push_back(encoding);
}

void
emit_ldsdir_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const LDSDIR_instruction& dir = instr->ldsdir();

   uint32_t encoding = (0b11001110 << 24);
   encoding |= opcode << 20;
   encoding |= (uint32_t)dir.wait_vdst << 16;
   if (ctx.gfx_level >= GFX12)
      encoding |= (uint32_t)dir.wait_vsrc << 23;
   encoding |= (uint32_t)dir.attr << 10;
   encoding |= (uint32_t)dir.attr_chan << 8;
   encoding |= reg(ctx, instr->definitions[0], 8);
   out.push_back(encoding);
}

void
emit_mubuf_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const MUBUF_instruction& mubuf = instr->mubuf();
   bool glc = mubuf.cache.value & ac_glc;
   bool slc = mubuf.cache.value & ac_slc;
   bool dlc = mubuf.cache.value & ac_dlc;

   uint32_t encoding = (0b111000 << 26);
   if (ctx.gfx_level >= GFX11 && mubuf.lds) /* GFX11 has separate opcodes for LDS loads */
      opcode = opcode == 0 ? 0x32 : (opcode + 0x1d);
   else
      encoding |= (mubuf.lds ? 1 : 0) << 16;
   encoding |= opcode << 18;
   encoding |= (glc ? 1 : 0) << 14;
   if (ctx.gfx_level <= GFX10_3)
      encoding |= (mubuf.idxen ? 1 : 0) << 13;
   assert(!mubuf.addr64 || ctx.gfx_level <= GFX7);
   if (ctx.gfx_level == GFX6 || ctx.gfx_level == GFX7)
      encoding |= (mubuf.addr64 ? 1 : 0) << 15;
   if (ctx.gfx_level <= GFX10_3)
      encoding |= (mubuf.offen ? 1 : 0) << 12;
   if (ctx.gfx_level == GFX8 || ctx.gfx_level == GFX9) {
      assert(!dlc); /* Device-level coherent is not supported on GFX9 and lower */
      encoding |= (slc ? 1 : 0) << 17;
   } else if (ctx.gfx_level >= GFX11) {
      encoding |= (slc ? 1 : 0) << 12;
      encoding |= (dlc ? 1 : 0) << 13;
   } else if (ctx.gfx_level >= GFX10) {
      encoding |= (dlc ? 1 : 0) << 15;
   }
   encoding |= 0x0FFF & mubuf.offset;
   out.push_back(encoding);
   encoding = 0;
   if (ctx.gfx_level <= GFX7 || (ctx.gfx_level >= GFX10 && ctx.gfx_level <= GFX10_3)) {
      encoding |= (slc ? 1 : 0) << 22;
   }
   encoding |= reg(ctx, instr->operands[2]) << 24;
   if (ctx.gfx_level >= GFX11) {
      encoding |= (mubuf.tfe ? 1 : 0) << 21;
      encoding |= (mubuf.offen ? 1 : 0) << 22;
      encoding |= (mubuf.idxen ? 1 : 0) << 23;
   } else {
      encoding |= (mubuf.tfe ? 1 : 0) << 23;
   }
   encoding |= (reg(ctx, instr->operands[0]) >> 2) << 16;
   if (instr->operands.size() > 3 && !mubuf.lds)
      encoding |= reg(ctx, instr->operands[3], 8) << 8;
   else if (!mubuf.lds)
      encoding |= reg(ctx, instr->definitions[0], 8) << 8;
   encoding |= reg(ctx, instr->operands[1], 8);
   out.push_back(encoding);
}

void
emit_mubuf_instruction_gfx12(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const MUBUF_instruction& mubuf = instr->mubuf();
   assert(!mubuf.lds);

   uint32_t encoding = 0b110001 << 26;
   encoding |= opcode << 14;
   if (instr->operands[2].isConstant()) {
      assert(instr->operands[2].constantValue() == 0);
      encoding |= reg(ctx, sgpr_null);
   } else {
      encoding |= reg(ctx, instr->operands[2]);
   }
   encoding |= (mubuf.tfe ? 1 : 0) << 22;
   out.push_back(encoding);

   encoding = 0;
   if (instr->operands.size() > 3)
      encoding |= reg(ctx, instr->operands[3], 8);
   else
      encoding |= reg(ctx, instr->definitions[0], 8);
   encoding |= reg(ctx, instr->operands[0]) << 9;
   encoding |= (mubuf.offen ? 1 : 0) << 30;
   encoding |= (mubuf.idxen ? 1 : 0) << 31;
   encoding |= get_gfx12_cpol(mubuf) << 18;
   encoding |= 1 << 23;
   out.push_back(encoding);

   encoding = 0;
   if (!instr->operands[1].isUndefined())
      encoding |= reg(ctx, instr->operands[1], 8);
   encoding |= (mubuf.offset & 0x00ffffff) << 8;
   out.push_back(encoding);
}

void
emit_mtbuf_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const MTBUF_instruction& mtbuf = instr->mtbuf();
   bool glc = mtbuf.cache.value & ac_glc;
   bool slc = mtbuf.cache.value & ac_slc;
   bool dlc = mtbuf.cache.value & ac_dlc;
   uint32_t img_format = ac_get_tbuffer_format(ctx.gfx_level, mtbuf.dfmt, mtbuf.nfmt);
   assert(img_format <= 0x7F);
   assert(!dlc || ctx.gfx_level >= GFX10);

   uint32_t encoding = (0b111010 << 26);
   encoding |= (img_format << 19); /* Handles both the GFX10 FORMAT and the old NFMT+DFMT */
   if (ctx.gfx_level < GFX8) {
      encoding |= opcode << 16;
      /* ADDR64 is unused */
   } else if (ctx.gfx_level >= GFX10 && ctx.gfx_level < GFX11) {
      /* DLC bit replaces one bit of the OPCODE on GFX10 */
      encoding |= (opcode & 0x07) << 16; /* 3 LSBs of 4-bit OPCODE */
      encoding |= (dlc ? 1 : 0) << 15;
   } else {
      encoding |= opcode << 15;
   }
   encoding |= (glc ? 1 : 0) << 14;
   if (ctx.gfx_level >= GFX11) {
      encoding |= (dlc ? 1 : 0) << 13;
      encoding |= (slc ? 1 : 0) << 12;
   } else {
      encoding |= (mtbuf.idxen ? 1 : 0) << 13;
      encoding |= (mtbuf.offen ? 1 : 0) << 12;
   }
   encoding |= 0x0FFF & mtbuf.offset;
   out.push_back(encoding);

   encoding = 0;
   encoding |= reg(ctx, instr->operands[2]) << 24;
   if (ctx.gfx_level >= GFX11) {
      encoding |= (mtbuf.idxen ? 1 : 0) << 23;
      encoding |= (mtbuf.offen ? 1 : 0) << 22;
      encoding |= (mtbuf.tfe ? 1 : 0) << 21;
   } else {
      encoding |= (mtbuf.tfe ? 1 : 0) << 23;
      encoding |= (slc ? 1 : 0) << 22;
      if (ctx.gfx_level >= GFX10)
         encoding |= (((opcode & 0x08) >> 3) << 21); /* MSB of 4-bit OPCODE */
   }
   encoding |= (reg(ctx, instr->operands[0]) >> 2) << 16;
   if (instr->operands.size() > 3)
      encoding |= reg(ctx, instr->operands[3], 8) << 8;
   else
      encoding |= reg(ctx, instr->definitions[0], 8) << 8;
   encoding |= reg(ctx, instr->operands[1], 8);
   out.push_back(encoding);
}

void
emit_mtbuf_instruction_gfx12(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const MTBUF_instruction& mtbuf = instr->mtbuf();

   uint32_t img_format = ac_get_tbuffer_format(ctx.gfx_level, mtbuf.dfmt, mtbuf.nfmt);

   uint32_t encoding = 0b110001 << 26;
   encoding |= 0b1000 << 18;
   encoding |= opcode << 14;
   if (instr->operands[2].isConstant()) {
      assert(instr->operands[2].constantValue() == 0);
      encoding |= reg(ctx, sgpr_null);
   } else {
      encoding |= reg(ctx, instr->operands[2]);
   }
   encoding |= (mtbuf.tfe ? 1 : 0) << 22;
   out.push_back(encoding);

   encoding = 0;
   if (instr->operands.size() > 3)
      encoding |= reg(ctx, instr->operands[3], 8);
   else
      encoding |= reg(ctx, instr->definitions[0], 8);
   encoding |= reg(ctx, instr->operands[0]) << 9;
   encoding |= (mtbuf.offen ? 1 : 0) << 30;
   encoding |= (mtbuf.idxen ? 1 : 0) << 31;
   encoding |= get_gfx12_cpol(mtbuf) << 18;
   encoding |= img_format << 23;
   out.push_back(encoding);

   encoding = 0;
   encoding |= reg(ctx, instr->operands[1], 8);
   encoding |= (mtbuf.offset & 0x00ffffff) << 8;
   out.push_back(encoding);
}

void
emit_mimg_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const MIMG_instruction& mimg = instr->mimg();
   bool glc = mimg.cache.value & ac_glc;
   bool slc = mimg.cache.value & ac_slc;
   bool dlc = mimg.cache.value & ac_dlc;

   unsigned nsa_dwords = get_mimg_nsa_dwords(instr);
   assert(!nsa_dwords || ctx.gfx_level >= GFX10);

   uint32_t encoding = (0b111100 << 26);
   if (ctx.gfx_level >= GFX11) { /* GFX11: rearranges most fields */
      assert(nsa_dwords <= 1);
      encoding |= nsa_dwords;
      encoding |= mimg.dim << 2;
      encoding |= mimg.unrm ? 1 << 7 : 0;
      encoding |= (0xF & mimg.dmask) << 8;
      encoding |= slc ? 1 << 12 : 0;
      encoding |= dlc ? 1 << 13 : 0;
      encoding |= glc ? 1 << 14 : 0;
      encoding |= mimg.r128 ? 1 << 15 : 0;
      encoding |= mimg.a16 ? 1 << 16 : 0;
      encoding |= mimg.d16 ? 1 << 17 : 0;
      encoding |= (opcode & 0xFF) << 18;
   } else {
      encoding |= slc ? 1 << 25 : 0;
      encoding |= (opcode & 0x7f) << 18;
      encoding |= (opcode >> 7) & 1;
      encoding |= mimg.lwe ? 1 << 17 : 0;
      encoding |= mimg.tfe ? 1 << 16 : 0;
      encoding |= glc ? 1 << 13 : 0;
      encoding |= mimg.unrm ? 1 << 12 : 0;
      if (ctx.gfx_level <= GFX9) {
         assert(!dlc); /* Device-level coherent is not supported on GFX9 and lower */
         assert(!mimg.r128);
         encoding |= mimg.a16 ? 1 << 15 : 0;
         encoding |= mimg.da ? 1 << 14 : 0;
      } else {
         encoding |= mimg.r128 ? 1 << 15
                               : 0; /* GFX10: A16 moved to 2nd word, R128 replaces it in 1st word */
         encoding |= nsa_dwords << 1;
         encoding |= mimg.dim << 3; /* GFX10: dimensionality instead of declare array */
         encoding |= dlc ? 1 << 7 : 0;
      }
      encoding |= (0xF & mimg.dmask) << 8;
   }
   out.push_back(encoding);

   encoding = reg(ctx, instr->operands[3], 8); /* VADDR */
   if (!instr->definitions.empty()) {
      encoding |= reg(ctx, instr->definitions[0], 8) << 8; /* VDATA */
   } else if (!instr->operands[2].isUndefined()) {
      encoding |= reg(ctx, instr->operands[2], 8) << 8; /* VDATA */
   }
   encoding |= (0x1F & (reg(ctx, instr->operands[0]) >> 2)) << 16; /* T# (resource) */

   assert(!mimg.d16 || ctx.gfx_level >= GFX9);
   if (ctx.gfx_level >= GFX11) {
      if (!instr->operands[1].isUndefined())
         encoding |= (0x1F & (reg(ctx, instr->operands[1]) >> 2)) << 26; /* sampler */

      encoding |= mimg.tfe ? 1 << 21 : 0;
      encoding |= mimg.lwe ? 1 << 22 : 0;
   } else {
      if (!instr->operands[1].isUndefined())
         encoding |= (0x1F & (reg(ctx, instr->operands[1]) >> 2)) << 21; /* sampler */

      encoding |= mimg.d16 ? 1 << 31 : 0;
      if (ctx.gfx_level >= GFX10) {
         /* GFX10: A16 still exists, but is in a different place */
         encoding |= mimg.a16 ? 1 << 30 : 0;
      }
   }

   out.push_back(encoding);

   if (nsa_dwords) {
      out.resize(out.size() + nsa_dwords);
      std::vector<uint32_t>::iterator nsa = std::prev(out.end(), nsa_dwords);
      for (unsigned i = 0; i < instr->operands.size() - 4u; i++)
         nsa[i / 4] |= reg(ctx, instr->operands[4 + i], 8) << (i % 4 * 8);
   }
}

void
emit_mimg_instruction_gfx12(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const MIMG_instruction& mimg = instr->mimg();

   bool vsample = !instr->operands[1].isUndefined() || instr->opcode == aco_opcode::image_msaa_load;
   uint32_t encoding = opcode << 14;
   if (vsample) {
      encoding |= 0b111001 << 26;
      encoding |= mimg.tfe << 3;
      encoding |= mimg.unrm << 13;
   } else {
      encoding |= 0b110100 << 26;
   }
   encoding |= mimg.dim;
   encoding |= mimg.r128 << 4;
   encoding |= mimg.d16 << 5;
   encoding |= mimg.a16 << 6;
   encoding |= (mimg.dmask & 0xf) << 22;
   out.push_back(encoding);

   uint8_t vaddr[5] = {0, 0, 0, 0, 0};
   for (unsigned i = 3; i < instr->operands.size(); i++)
      vaddr[i - 3] = reg(ctx, instr->operands[i], 8);
   unsigned num_vaddr = instr->operands.size() - 3;
   for (unsigned i = 0; i < MIN2(instr->operands.back().size() - 1, 5 - num_vaddr); i++)
      vaddr[num_vaddr + i] = reg(ctx, instr->operands.back(), 8) + i + 1;

   encoding = 0;
   if (!instr->definitions.empty())
      encoding |= reg(ctx, instr->definitions[0], 8); /* VDATA */
   else if (!instr->operands[2].isUndefined())
      encoding |= reg(ctx, instr->operands[2], 8); /* VDATA */
   encoding |= reg(ctx, instr->operands[0]) << 9;  /* T# (resource) */
   if (vsample) {
      encoding |= mimg.lwe << 8;
      if (instr->opcode != aco_opcode::image_msaa_load)
         encoding |= reg(ctx, instr->operands[1]) << 23; /* sampler */
   } else {
      encoding |= mimg.tfe << 23;
      encoding |= vaddr[4] << 24;
   }
   encoding |= get_gfx12_cpol(mimg) << 18;
   out.push_back(encoding);

   encoding = 0;
   for (unsigned i = 0; i < 4; i++)
      encoding |= vaddr[i] << (i * 8);
   out.push_back(encoding);
}

void
emit_flatlike_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const FLAT_instruction& flat = instr->flatlike();
   bool glc = flat.cache.value & ac_glc;
   bool slc = flat.cache.value & ac_slc;
   bool dlc = flat.cache.value & ac_dlc;

   uint32_t encoding = (0b110111 << 26);
   encoding |= opcode << 18;
   if (ctx.gfx_level == GFX9 || ctx.gfx_level >= GFX11) {
      if (instr->isFlat())
         assert(flat.offset <= 0xfff);
      else
         assert(flat.offset >= -4096 && flat.offset < 4096);
      encoding |= flat.offset & 0x1fff;
   } else if (ctx.gfx_level <= GFX8 || instr->isFlat()) {
      /* GFX10 has a 12-bit immediate OFFSET field,
       * but it has a hw bug: it ignores the offset, called FlatSegmentOffsetBug
       */
      assert(flat.offset == 0);
   } else {
      assert(flat.offset >= -2048 && flat.offset <= 2047);
      encoding |= flat.offset & 0xfff;
   }
   if (instr->isScratch())
      encoding |= 1 << (ctx.gfx_level >= GFX11 ? 16 : 14);
   else if (instr->isGlobal())
      encoding |= 2 << (ctx.gfx_level >= GFX11 ? 16 : 14);
   encoding |= flat.lds ? 1 << 13 : 0;
   encoding |= glc ? 1 << (ctx.gfx_level >= GFX11 ? 14 : 16) : 0;
   encoding |= slc ? 1 << (ctx.gfx_level >= GFX11 ? 15 : 17) : 0;
   if (ctx.gfx_level >= GFX10) {
      assert(!flat.nv);
      encoding |= dlc ? 1 << (ctx.gfx_level >= GFX11 ? 13 : 12) : 0;
   } else {
      assert(!dlc);
   }
   out.push_back(encoding);
   encoding = reg(ctx, instr->operands[0], 8);
   if (!instr->definitions.empty())
      encoding |= reg(ctx, instr->definitions[0], 8) << 24;
   if (instr->operands.size() >= 3)
      encoding |= reg(ctx, instr->operands[2], 8) << 8;
   if (!instr->operands[1].isUndefined()) {
      assert(ctx.gfx_level >= GFX10 || instr->operands[1].physReg() != 0x7F);
      assert(instr->format != Format::FLAT);
      encoding |= reg(ctx, instr->operands[1], 8) << 16;
   } else if (instr->format != Format::FLAT ||
              ctx.gfx_level >= GFX10) { /* SADDR is actually used with FLAT on GFX10 */
      /* For GFX10.3 scratch, 0x7F disables both ADDR and SADDR, unlike sgpr_null, which only
       * disables SADDR. On GFX11, this was replaced with SVE.
       */
      if (ctx.gfx_level <= GFX9 ||
          (instr->isScratch() && instr->operands[0].isUndefined() && ctx.gfx_level < GFX11))
         encoding |= 0x7F << 16;
      else
         encoding |= reg(ctx, sgpr_null) << 16;
   }
   if (ctx.gfx_level >= GFX11 && instr->isScratch())
      encoding |= !instr->operands[0].isUndefined() ? 1 << 23 : 0;
   else
      encoding |= flat.nv ? 1 << 23 : 0;
   out.push_back(encoding);
}

void
emit_flatlike_instruction_gfx12(asm_context& ctx, std::vector<uint32_t>& out,
                                const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const FLAT_instruction& flat = instr->flatlike();
   assert(!flat.lds);

   uint32_t encoding = opcode << 14;
   encoding |= 0b111011 << 26;
   if (!instr->operands[1].isUndefined()) {
      assert(!instr->isFlat());
      encoding |= reg(ctx, instr->operands[1]);
   } else {
      encoding |= reg(ctx, sgpr_null);
   }
   if (instr->isScratch())
      encoding |= 1 << 24;
   else if (instr->isGlobal())
      encoding |= 2 << 24;
   out.push_back(encoding);

   encoding = 0;
   if (!instr->definitions.empty())
      encoding |= reg(ctx, instr->definitions[0], 8);
   if (instr->isScratch())
      encoding |= !instr->operands[0].isUndefined() ? 1 << 17 : 0;
   encoding |= get_gfx12_cpol(flat) << 18;
   if (instr->operands.size() >= 3)
      encoding |= reg(ctx, instr->operands[2], 8) << 23;
   out.push_back(encoding);

   encoding = 0;
   if (!instr->operands[0].isUndefined())
      encoding |= reg(ctx, instr->operands[0], 8);
   encoding |= (flat.offset & 0x00ffffff) << 8;
   out.push_back(encoding);
}

void
emit_exp_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   const Export_instruction& exp = instr->exp();
   uint32_t encoding;
   if (ctx.gfx_level == GFX8 || ctx.gfx_level == GFX9) {
      encoding = (0b110001 << 26);
   } else {
      encoding = (0b111110 << 26);
   }

   if (ctx.gfx_level >= GFX11) {
      encoding |= exp.row_en ? 0b1 << 13 : 0;
   } else {
      encoding |= exp.valid_mask ? 0b1 << 12 : 0;
      encoding |= exp.compressed ? 0b1 << 10 : 0;
   }
   encoding |= exp.done ? 0b1 << 11 : 0;
   encoding |= exp.dest << 4;
   encoding |= exp.enabled_mask;
   out.push_back(encoding);
   encoding = reg(ctx, exp.operands[0], 8);
   encoding |= reg(ctx, exp.operands[1], 8) << 8;
   encoding |= reg(ctx, exp.operands[2], 8) << 16;
   encoding |= reg(ctx, exp.operands[3], 8) << 24;
   out.push_back(encoding);
}

void emit_instruction(asm_context& ctx, std::vector<uint32_t>& out, Instruction* instr);

void
emit_dpp16_instruction(asm_context& ctx, std::vector<uint32_t>& out, Instruction* instr)
{
   assert(ctx.gfx_level >= GFX8);
   DPP16_instruction& dpp = instr->dpp16();

   /* first emit the instruction without the DPP operand */
   Operand dpp_op = instr->operands[0];
   instr->operands[0] = Operand(PhysReg{250}, v1);
   instr->format = (Format)((uint16_t)instr->format & ~(uint16_t)Format::DPP16);
   emit_instruction(ctx, out, instr);
   instr->format = (Format)((uint16_t)instr->format | (uint16_t)Format::DPP16);
   instr->operands[0] = dpp_op;

   uint32_t encoding = (0xF & dpp.row_mask) << 28;
   encoding |= (0xF & dpp.bank_mask) << 24;
   encoding |= dpp.abs[1] << 23;
   encoding |= dpp.neg[1] << 22;
   encoding |= dpp.abs[0] << 21;
   encoding |= dpp.neg[0] << 20;
   encoding |= dpp.fetch_inactive << 18;
   encoding |= dpp.bound_ctrl << 19;
   encoding |= dpp.dpp_ctrl << 8;
   encoding |= reg(ctx, dpp_op, 8);
   encoding |= dpp.opsel[0] && !instr->isVOP3() ? 128 : 0;
   out.push_back(encoding);
}

void
emit_dpp8_instruction(asm_context& ctx, std::vector<uint32_t>& out, Instruction* instr)
{
   assert(ctx.gfx_level >= GFX10);
   DPP8_instruction& dpp = instr->dpp8();

   /* first emit the instruction without the DPP operand */
   Operand dpp_op = instr->operands[0];
   instr->operands[0] = Operand(PhysReg{233u + dpp.fetch_inactive}, v1);
   instr->format = (Format)((uint16_t)instr->format & ~(uint16_t)Format::DPP8);
   emit_instruction(ctx, out, instr);
   instr->format = (Format)((uint16_t)instr->format | (uint16_t)Format::DPP8);
   instr->operands[0] = dpp_op;

   uint32_t encoding = reg(ctx, dpp_op, 8);
   encoding |= dpp.opsel[0] && !instr->isVOP3() ? 128 : 0;
   encoding |= dpp.lane_sel << 8;
   out.push_back(encoding);
}

void
emit_vop3_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const VALU_instruction& vop3 = instr->valu();

   if (instr->isVOP2()) {
      opcode = opcode + 0x100;
   } else if (instr->isVOP1()) {
      if (ctx.gfx_level == GFX8 || ctx.gfx_level == GFX9)
         opcode = opcode + 0x140;
      else
         opcode = opcode + 0x180;
   } else if (instr->isVOPC()) {
      opcode = opcode + 0x0;
   } else if (instr->isVINTRP()) {
      opcode = opcode + 0x270;
   }

   uint32_t encoding;
   if (ctx.gfx_level <= GFX9) {
      encoding = (0b110100 << 26);
   } else if (ctx.gfx_level >= GFX10) {
      encoding = (0b110101 << 26);
   } else {
      unreachable("Unknown gfx_level.");
   }

   if (ctx.gfx_level <= GFX7) {
      encoding |= opcode << 17;
      encoding |= (vop3.clamp ? 1 : 0) << 11;
   } else {
      encoding |= opcode << 16;
      encoding |= (vop3.clamp ? 1 : 0) << 15;
   }
   encoding |= vop3.opsel << 11;
   for (unsigned i = 0; i < 3; i++)
      encoding |= vop3.abs[i] << (8 + i);
   /* On GFX9 and older, v_cmpx implicitly writes exec besides writing an SGPR pair.
    * On GFX10 and newer, v_cmpx always writes just exec.
    */
   if (instr->definitions.size() == 2 && instr->isVOPC())
      assert(ctx.gfx_level <= GFX9 && instr->definitions[1].physReg() == exec);
   else if (instr->definitions.size() == 2 && instr->opcode != aco_opcode::v_swap_b16)
      encoding |= reg(ctx, instr->definitions[1]) << 8;
   encoding |= reg(ctx, instr->definitions[0], 8);
   out.push_back(encoding);
   encoding = 0;

   unsigned num_ops = instr->operands.size();
   /* Encoding implicit sources works fine with hardware but breaks some disassemblers. */
   if (instr->opcode == aco_opcode::v_writelane_b32_e64)
      num_ops = 2;
   else if (instr->opcode == aco_opcode::v_swap_b16)
      num_ops = 1;

   for (unsigned i = 0; i < num_ops; i++)
      encoding |= reg(ctx, instr->operands[i]) << (i * 9);
   encoding |= vop3.omod << 27;
   for (unsigned i = 0; i < 3; i++)
      encoding |= vop3.neg[i] << (29 + i);
   out.push_back(encoding);
}

void
emit_vop3p_instruction(asm_context& ctx, std::vector<uint32_t>& out, const Instruction* instr)
{
   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   const VALU_instruction& vop3 = instr->valu();

   uint32_t encoding;
   if (ctx.gfx_level == GFX9) {
      encoding = (0b110100111 << 23);
   } else if (ctx.gfx_level >= GFX10) {
      encoding = (0b110011 << 26);
   } else {
      unreachable("Unknown gfx_level.");
   }

   encoding |= opcode << 16;
   encoding |= (vop3.clamp ? 1 : 0) << 15;
   encoding |= vop3.opsel_lo << 11;
   encoding |= ((vop3.opsel_hi & 0x4) ? 1 : 0) << 14;
   for (unsigned i = 0; i < 3; i++)
      encoding |= vop3.neg_hi[i] << (8 + i);
   encoding |= reg(ctx, instr->definitions[0], 8);
   out.push_back(encoding);
   encoding = 0;
   for (unsigned i = 0; i < instr->operands.size(); i++)
      encoding |= reg(ctx, instr->operands[i]) << (i * 9);
   encoding |= (vop3.opsel_hi & 0x3) << 27;
   for (unsigned i = 0; i < 3; i++)
      encoding |= vop3.neg_lo[i] << (29 + i);
   out.push_back(encoding);
}

void
emit_sdwa_instruction(asm_context& ctx, std::vector<uint32_t>& out, Instruction* instr)
{
   assert(ctx.gfx_level >= GFX8 && ctx.gfx_level < GFX11);
   SDWA_instruction& sdwa = instr->sdwa();

   /* first emit the instruction without the SDWA operand */
   Operand sdwa_op = instr->operands[0];
   instr->operands[0] = Operand(PhysReg{249}, v1);
   instr->format = (Format)((uint16_t)instr->format & ~(uint16_t)Format::SDWA);
   emit_instruction(ctx, out, instr);
   instr->format = (Format)((uint16_t)instr->format | (uint16_t)Format::SDWA);
   instr->operands[0] = sdwa_op;

   uint32_t encoding = 0;

   if (instr->isVOPC()) {
      if (instr->definitions[0].physReg() !=
          (ctx.gfx_level >= GFX10 && is_cmpx(instr->opcode) ? exec : vcc)) {
         encoding |= reg(ctx, instr->definitions[0]) << 8;
         encoding |= 1 << 15;
      }
      encoding |= (sdwa.clamp ? 1 : 0) << 13;
   } else {
      encoding |= sdwa.dst_sel.to_sdwa_sel(instr->definitions[0].physReg().byte()) << 8;
      uint32_t dst_u = sdwa.dst_sel.sign_extend() ? 1 : 0;
      if (instr->definitions[0].bytes() < 4) /* dst_preserve */
         dst_u = 2;
      encoding |= dst_u << 11;
      encoding |= (sdwa.clamp ? 1 : 0) << 13;
      encoding |= sdwa.omod << 14;
   }

   encoding |= sdwa.sel[0].to_sdwa_sel(sdwa_op.physReg().byte()) << 16;
   encoding |= sdwa.sel[0].sign_extend() ? 1 << 19 : 0;
   encoding |= sdwa.abs[0] << 21;
   encoding |= sdwa.neg[0] << 20;

   if (instr->operands.size() >= 2) {
      encoding |= sdwa.sel[1].to_sdwa_sel(instr->operands[1].physReg().byte()) << 24;
      encoding |= sdwa.sel[1].sign_extend() ? 1 << 27 : 0;
      encoding |= sdwa.abs[1] << 29;
      encoding |= sdwa.neg[1] << 28;
   }

   encoding |= reg(ctx, sdwa_op, 8);
   encoding |= (sdwa_op.physReg() < 256) << 23;
   if (instr->operands.size() >= 2)
      encoding |= (instr->operands[1].physReg() < 256) << 31;
   out.push_back(encoding);
}

void
emit_instruction(asm_context& ctx, std::vector<uint32_t>& out, Instruction* instr)
{
   /* lower remaining pseudo-instructions */
   if (instr->opcode == aco_opcode::p_constaddr_getpc) {
      ctx.constaddrs[instr->operands[0].constantValue()].getpc_end = out.size() + 1;

      instr->opcode = aco_opcode::s_getpc_b64;
      instr->operands.pop_back();
   } else if (instr->opcode == aco_opcode::p_constaddr_addlo) {
      ctx.constaddrs[instr->operands[2].constantValue()].add_literal = out.size() + 1;

      instr->opcode = aco_opcode::s_add_u32;
      instr->operands.pop_back();
      assert(instr->operands[1].isConstant());
      /* in case it's an inline constant, make it a literal */
      instr->operands[1] = Operand::literal32(instr->operands[1].constantValue());
   } else if (instr->opcode == aco_opcode::p_resumeaddr_getpc) {
      ctx.resumeaddrs[instr->operands[0].constantValue()].getpc_end = out.size() + 1;

      instr->opcode = aco_opcode::s_getpc_b64;
      instr->operands.pop_back();
   } else if (instr->opcode == aco_opcode::p_resumeaddr_addlo) {
      ctx.resumeaddrs[instr->operands[2].constantValue()].add_literal = out.size() + 1;

      instr->opcode = aco_opcode::s_add_u32;
      instr->operands.pop_back();
      assert(instr->operands[1].isConstant());
      /* in case it's an inline constant, make it a literal */
      instr->operands[1] = Operand::literal32(instr->operands[1].constantValue());
   } else if (instr->opcode == aco_opcode::p_load_symbol) {
      assert(instr->operands[0].isConstant());
      assert(ctx.symbols);

      struct aco_symbol info;
      info.id = (enum aco_symbol_id)instr->operands[0].constantValue();
      info.offset = out.size() + 1;
      ctx.symbols->push_back(info);

      instr->opcode = aco_opcode::s_mov_b32;
      /* in case it's an inline constant, make it a literal */
      instr->operands[0] = Operand::literal32(0);
   } else if (instr->opcode == aco_opcode::p_debug_info) {
      assert(instr->operands[0].isConstant());
      uint32_t index = instr->operands[0].constantValue();
      ctx.program->debug_info[index].offset = (out.size() - 1) * 4;
      return;
   }

   /* Promote VOP12C to VOP3 if necessary. */
   if ((instr->isVOP1() || instr->isVOP2() || instr->isVOPC()) && !instr->isVOP3() &&
       needs_vop3_gfx11(ctx, instr)) {
      instr->format = asVOP3(instr->format);
      if (instr->opcode == aco_opcode::v_fmaak_f16) {
         instr->opcode = aco_opcode::v_fma_f16;
         instr->format = (Format)((uint32_t)instr->format & ~(uint32_t)Format::VOP2);
      } else if (instr->opcode == aco_opcode::v_fmamk_f16) {
         instr->valu().swapOperands(1, 2);
         instr->opcode = aco_opcode::v_fma_f16;
         instr->format = (Format)((uint32_t)instr->format & ~(uint32_t)Format::VOP2);
      }
   }

   uint32_t opcode = ctx.opcode[(int)instr->opcode];
   if (opcode == (uint32_t)-1) {
      char* outmem;
      size_t outsize;
      struct u_memstream mem;
      u_memstream_open(&mem, &outmem, &outsize);
      FILE* const memf = u_memstream_get(&mem);

      fprintf(memf, "Unsupported opcode: ");
      aco_print_instr(ctx.gfx_level, instr, memf);
      u_memstream_close(&mem);

      aco_err(ctx.program, outmem);
      free(outmem);

      abort();
   }

   switch (instr->format) {
   case Format::SOP2: {
      emit_sop2_instruction(ctx, out, instr);
      break;
   }
   case Format::SOPK: {
      emit_sopk_instruction(ctx, out, instr);
      break;
   }
   case Format::SOP1: {
      emit_sop1_instruction(ctx, out, instr);
      break;
   }
   case Format::SOPC: {
      emit_sopc_instruction(ctx, out, instr);
      break;
   }
   case Format::SOPP: {
      emit_sopp_instruction(ctx, out, instr);
      break;
   }
   case Format::SMEM: {
      emit_smem_instruction(ctx, out, instr);
      return;
   }
   case Format::VOP2: {
      emit_vop2_instruction(ctx, out, instr);
      break;
   }
   case Format::VOP1: {
      emit_vop1_instruction(ctx, out, instr);
      break;
   }
   case Format::VOPC: {
      emit_vopc_instruction(ctx, out, instr);
      break;
   }
   case Format::VINTRP: {
      emit_vintrp_instruction(ctx, out, instr);
      break;
   }
   case Format::VINTERP_INREG: {
      emit_vinterp_inreg_instruction(ctx, out, instr);
      break;
   }
   case Format::VOPD: {
      emit_vopd_instruction(ctx, out, instr);
      break;
   }
   case Format::DS: {
      emit_ds_instruction(ctx, out, instr);
      break;
   }
   case Format::LDSDIR: {
      emit_ldsdir_instruction(ctx, out, instr);
      break;
   }
   case Format::MUBUF: {
      if (ctx.gfx_level >= GFX12)
         emit_mubuf_instruction_gfx12(ctx, out, instr);
      else
         emit_mubuf_instruction(ctx, out, instr);
      break;
   }
   case Format::MTBUF: {
      if (ctx.gfx_level >= GFX12)
         emit_mtbuf_instruction_gfx12(ctx, out, instr);
      else
         emit_mtbuf_instruction(ctx, out, instr);
      break;
   }
   case Format::MIMG: {
      if (ctx.gfx_level >= GFX12)
         emit_mimg_instruction_gfx12(ctx, out, instr);
      else
         emit_mimg_instruction(ctx, out, instr);
      break;
   }
   case Format::FLAT:
   case Format::SCRATCH:
   case Format::GLOBAL: {
      if (ctx.gfx_level >= GFX12)
         emit_flatlike_instruction_gfx12(ctx, out, instr);
      else
         emit_flatlike_instruction(ctx, out, instr);
      break;
   }
   case Format::EXP: {
      emit_exp_instruction(ctx, out, instr);
      break;
   }
   case Format::PSEUDO:
   case Format::PSEUDO_BARRIER:
      if (instr->opcode != aco_opcode::p_unit_test)
         unreachable("Pseudo instructions should be lowered before assembly.");
      break;
   default:
      if (instr->isDPP16()) {
         emit_dpp16_instruction(ctx, out, instr);
         return;
      } else if (instr->isDPP8()) {
         emit_dpp8_instruction(ctx, out, instr);
         return;
      } else if (instr->isVOP3()) {
         emit_vop3_instruction(ctx, out, instr);
      } else if (instr->isVOP3P()) {
         emit_vop3p_instruction(ctx, out, instr);
      } else if (instr->isSDWA()) {
         emit_sdwa_instruction(ctx, out, instr);
      } else {
         unreachable("unimplemented instruction format");
      }
      break;
   }

   /* append literal dword */
   for (const Operand& op : instr->operands) {
      if (op.isLiteral()) {
         out.push_back(op.constantValue());
         break;
      }
   }
}

void
emit_block(asm_context& ctx, std::vector<uint32_t>& out, Block& block)
{
   for (aco_ptr<Instruction>& instr : block.instructions) {
#if 0
      int start_idx = out.size();
      std::cerr << "Encoding:\t" << std::endl;
      aco_print_instr(&*instr, stderr);
      std::cerr << std::endl;
#endif
      emit_instruction(ctx, out, instr.get());
#if 0
      for (int i = start_idx; i < out.size(); i++)
         std::cerr << "encoding: " << "0x" << std::setfill('0') << std::setw(8) << std::hex << out[i] << std::endl;
#endif
   }
}

void
fix_exports(asm_context& ctx, std::vector<uint32_t>& out, Program* program)
{
   bool exported = false;
   for (Block& block : program->blocks) {
      if (!(block.kind & block_kind_export_end))
         continue;
      std::vector<aco_ptr<Instruction>>::reverse_iterator it = block.instructions.rbegin();
      while (it != block.instructions.rend()) {
         if ((*it)->isEXP()) {
            Export_instruction& exp = (*it)->exp();
            if (program->stage.hw == AC_HW_VERTEX_SHADER ||
                program->stage.hw == AC_HW_NEXT_GEN_GEOMETRY_SHADER) {
               if (exp.dest >= V_008DFC_SQ_EXP_POS && exp.dest <= (V_008DFC_SQ_EXP_POS + 3)) {
                  exp.done = true;
                  exported = true;
                  break;
               }
            } else {
               exp.done = true;
               exp.valid_mask = true;
               exported = true;
               break;
            }
         } else if ((*it)->definitions.size() && (*it)->definitions[0].physReg() == exec) {
            break;
         }
         ++it;
      }
   }

   /* GFX10+ FS may not export anything if no discard is used. */
   bool may_skip_export = program->stage.hw == AC_HW_PIXEL_SHADER && program->gfx_level >= GFX10;

   if (!exported && !may_skip_export) {
      /* Abort in order to avoid a GPU hang. */
      bool is_vertex_or_ngg = (program->stage.hw == AC_HW_VERTEX_SHADER ||
                               program->stage.hw == AC_HW_NEXT_GEN_GEOMETRY_SHADER);
      aco_err(program,
              "Missing export in %s shader:", is_vertex_or_ngg ? "vertex or NGG" : "fragment");
      aco_print_program(program, stderr);
      abort();
   }
}

static void
insert_code(asm_context& ctx, std::vector<uint32_t>& out, unsigned insert_before,
            unsigned insert_count, const uint32_t* insert_data)
{
   out.insert(out.begin() + insert_before, insert_data, insert_data + insert_count);

   /* Update the offset of each affected block */
   for (Block& block : ctx.program->blocks) {
      if (block.offset >= insert_before)
         block.offset += insert_count;
   }

   /* Update the locations of branches */
   for (branch_info& info : ctx.branches) {
      if (info.pos >= insert_before)
         info.pos += insert_count;
   }

   /* Update the locations of p_constaddr instructions */
   for (auto& constaddr : ctx.constaddrs) {
      constaddr_info& info = constaddr.second;
      if (info.getpc_end >= insert_before)
         info.getpc_end += insert_count;
      if (info.add_literal >= insert_before)
         info.add_literal += insert_count;
   }
   for (auto& constaddr : ctx.resumeaddrs) {
      constaddr_info& info = constaddr.second;
      if (info.getpc_end >= insert_before)
         info.getpc_end += insert_count;
      if (info.add_literal >= insert_before)
         info.add_literal += insert_count;
   }

   if (ctx.symbols) {
      for (auto& symbol : *ctx.symbols) {
         if (symbol.offset >= insert_before)
            symbol.offset += insert_count;
      }
   }
}

static void
fix_branches_gfx10(asm_context& ctx, std::vector<uint32_t>& out)
{
   /* Branches with an offset of 0x3f are buggy on GFX10,
    * we workaround by inserting NOPs if needed.
    */
   bool gfx10_3f_bug = false;

   do {
      auto buggy_branch_it = std::find_if(
         ctx.branches.begin(), ctx.branches.end(), [&](const branch_info& branch) -> bool
         { return ((int)ctx.program->blocks[branch.target].offset - branch.pos - 1) == 0x3f; });
      gfx10_3f_bug = buggy_branch_it != ctx.branches.end();

      if (gfx10_3f_bug) {
         /* Insert an s_nop after the branch */
         constexpr uint32_t s_nop_0 = 0xbf800000u;
         insert_code(ctx, out, buggy_branch_it->pos + 1, 1, &s_nop_0);
      }
   } while (gfx10_3f_bug);
}

void
chain_branches(asm_context& ctx, std::vector<uint32_t>& out, branch_info& branch)
{
   /* Create an empty block in order to remember the offset of the chained branch instruction.
    * The new branch instructions are inserted into the program in source code order.
    */
   Block* new_block = ctx.program->create_and_insert_block();
   Builder bld(ctx.program);
   std::vector<uint32_t> code;
   Instruction* branch_instr;

   /* Re-direct original branch to new block (offset). */
   unsigned target = branch.target;
   branch.target = new_block->index;

   unsigned skip_branch_target = 0; /* Target of potentially inserted short jump. */

   /* Find suitable insertion point:
    * We define two offset ranges within our new branch instruction should be placed.
    * Then we try to maximize the distance from either the previous branch or the target.
    */
   const int half_dist = (INT16_MAX - 31) / 2;
   const unsigned upper_start = MIN2(ctx.program->blocks[target].offset, branch.pos) + half_dist;
   const unsigned upper_end = upper_start + half_dist;
   const unsigned lower_end = MAX2(ctx.program->blocks[target].offset, branch.pos) - half_dist;
   const unsigned lower_start = lower_end - half_dist;
   unsigned insert_at = 0;
   for (unsigned i = 0; i < ctx.program->blocks.size() - 1; i++) {
      Block& block = ctx.program->blocks[i];
      Block& next = ctx.program->blocks[i + 1];
      if (next.offset >= lower_end)
         break;
      if (next.offset < upper_start || (next.offset > upper_end && next.offset < lower_start))
         continue;

      /* If this block ends in an unconditional branch, we can insert
       * another branch right after it without additional cost for the
       * existing code.
       */
      if (!block.instructions.empty() &&
          block.instructions.back()->opcode == aco_opcode::s_branch) {
         insert_at = next.offset;
         bld.reset(&block.instructions);
         if (next.offset >= lower_start)
            break;
      }
   }

   /* If we didn't find a suitable insertion point, split the existing code. */
   if (insert_at == 0) {
      /* Find the last block that is still within reach. */
      unsigned insertion_block_idx = 0;
      while (ctx.program->blocks[insertion_block_idx + 1].offset < upper_end)
         insertion_block_idx++;

      insert_at = ctx.program->blocks[insertion_block_idx].offset;
      auto it = ctx.program->blocks[insertion_block_idx].instructions.begin();
      int skip = 0;
      if (insert_at < upper_start) {
         /* Ensure some forward progress by splitting the block if necessary. */
         while (skip-- > 0 || insert_at < upper_start) {
            Instruction* instr = (it++)->get();
            if (instr->isSOPP()) {
               if (instr->opcode == aco_opcode::s_clause)
                  skip = instr->salu().imm + 1;
               else if (instr->opcode == aco_opcode::s_delay_alu)
                  skip = ((instr->salu().imm >> 4) & 0x7) + 1;
               else if (instr->opcode == aco_opcode::s_branch)
                  skip = 1;
               insert_at++;
               continue;
            }
            emit_instruction(ctx, code, instr);
            assert(out[insert_at] == code[0]);
            insert_at += code.size();
            code.clear();
         }

         /* If the insertion point is in the middle of the block, insert the branch instructions
          * into that block instead. */
         bld.reset(&ctx.program->blocks[insertion_block_idx].instructions, it);
      } else {
         bld.reset(&ctx.program->blocks[insertion_block_idx - 1].instructions);
         skip_branch_target = insertion_block_idx;
      }

      /* Since we insert a branch into existing code, mitigate LdsBranchVmemWARHazard on GFX10. */
      if (ctx.program->gfx_level == GFX10) {
         emit_sopk_instruction(
            ctx, code, bld.sopk(aco_opcode::s_waitcnt_vscnt, Operand(sgpr_null, s1), 0).instr);
      }

      /* For the existing code, create a short jump over the new branch. */
      branch_instr = bld.sopp(aco_opcode::s_branch, 1).instr;
      emit_sopp_instruction(ctx, code, branch_instr, true);
   }
   const unsigned block_offset = insert_at + code.size();

   branch_instr = bld.sopp(aco_opcode::s_branch, 0);
   emit_sopp_instruction(ctx, code, branch_instr, true);
   insert_code(ctx, out, insert_at, code.size(), code.data());

   new_block->offset = block_offset;
   if (skip_branch_target) {
      /* If we insert a short jump over the new branch at the end of a block,
       * ensure that it gets updated accordingly after additional changes. */
      ctx.branches.push_back({block_offset - 1, skip_branch_target});
   }
   ctx.branches.push_back({block_offset, target});
   assert(out[ctx.branches.back().pos] == code.back());
}

void
fix_branches(asm_context& ctx, std::vector<uint32_t>& out)
{
   bool repeat = false;
   do {
      repeat = false;

      if (ctx.gfx_level == GFX10)
         fix_branches_gfx10(ctx, out);

      for (branch_info& branch : ctx.branches) {
         int offset = (int)ctx.program->blocks[branch.target].offset - branch.pos - 1;
         if (offset >= INT16_MIN && offset <= INT16_MAX) {
            out[branch.pos] &= 0xffff0000u;
            out[branch.pos] |= (uint16_t)offset;
         } else {
            chain_branches(ctx, out, branch);
            repeat = true;
            break;
         }
      }
   } while (repeat);
}

void
fix_constaddrs(asm_context& ctx, std::vector<uint32_t>& out)
{
   for (auto& constaddr : ctx.constaddrs) {
      constaddr_info& info = constaddr.second;
      out[info.add_literal] += (out.size() - info.getpc_end) * 4u;

      if (ctx.symbols) {
         struct aco_symbol sym;
         sym.id = aco_symbol_const_data_addr;
         sym.offset = info.add_literal;
         ctx.symbols->push_back(sym);
      }
   }
   for (auto& addr : ctx.resumeaddrs) {
      constaddr_info& info = addr.second;
      const Block& block = ctx.program->blocks[out[info.add_literal]];
      assert(block.kind & block_kind_resume);
      out[info.add_literal] = (block.offset - info.getpc_end) * 4u;
   }
}

void
align_block(asm_context& ctx, std::vector<uint32_t>& code, Block& block)
{
   /* Align the previous loop. */
   if (ctx.loop_header != -1u &&
       block.loop_nest_depth < ctx.program->blocks[ctx.loop_header].loop_nest_depth) {
      assert(ctx.loop_exit != -1u);
      Block& loop_header = ctx.program->blocks[ctx.loop_header];
      Block& loop_exit = ctx.program->blocks[ctx.loop_exit];
      ctx.loop_header = -1u;
      ctx.loop_exit = -1u;
      std::vector<uint32_t> nops;

      const unsigned loop_num_cl = DIV_ROUND_UP(block.offset - loop_header.offset, 16);

      /* On GFX10.3+, change the prefetch mode if the loop fits into 2 or 3 cache lines.
       * Don't use the s_inst_prefetch instruction on GFX10 as it might cause hangs.
       */
      const bool change_prefetch = ctx.program->gfx_level >= GFX10_3 &&
                                   ctx.program->gfx_level <= GFX11 && loop_num_cl > 1 &&
                                   loop_num_cl <= 3;

      if (change_prefetch) {
         Builder bld(ctx.program, &ctx.program->blocks[loop_header.linear_preds[0]]);
         int16_t prefetch_mode = loop_num_cl == 3 ? 0x1 : 0x2;
         Instruction* instr = bld.sopp(aco_opcode::s_inst_prefetch, prefetch_mode);
         emit_instruction(ctx, nops, instr);
         insert_code(ctx, code, loop_header.offset, nops.size(), nops.data());

         /* Change prefetch mode back to default (0x3) at the loop exit. */
         bld.reset(&loop_exit.instructions, loop_exit.instructions.begin());
         instr = bld.sopp(aco_opcode::s_inst_prefetch, 0x3);
         if (ctx.loop_exit < block.index) {
            nops.clear();
            emit_instruction(ctx, nops, instr);
            insert_code(ctx, code, loop_exit.offset, nops.size(), nops.data());
         }
      }

      const unsigned loop_start_cl = loop_header.offset >> 4;
      const unsigned loop_end_cl = (block.offset - 1) >> 4;

      /* Align the loop if it fits into the fetched cache lines or if we can
       * reduce the number of cache lines with less than 8 NOPs.
       */
      const bool align_loop = loop_end_cl - loop_start_cl >= loop_num_cl &&
                              (loop_num_cl == 1 || change_prefetch || loop_header.offset % 16 > 8);

      if (align_loop) {
         nops.clear();
         nops.resize(16 - (loop_header.offset % 16), 0xbf800000u);
         insert_code(ctx, code, loop_header.offset, nops.size(), nops.data());
      }
   }

   if (block.kind & block_kind_loop_header) {
      /* In case of nested loops, only handle the inner-most loops in order
       * to not break the alignment of inner loops by handling outer loops.
       * Also ignore loops without back-edge.
       */
      if (block.linear_preds.size() > 1) {
         ctx.loop_header = block.index;
         ctx.loop_exit = -1u;
      }
   }

   /* Blocks with block_kind_loop_exit might be eliminated after jump threading,
    * so we instead find loop exits using the successors when in loop_nest_depth.
    * This works, because control flow always re-converges after loops.
    */
   if (ctx.loop_header != -1u && ctx.loop_exit == -1u) {
      for (uint32_t succ_idx : block.linear_succs) {
         Block& succ = ctx.program->blocks[succ_idx];
         if (succ.loop_nest_depth < ctx.program->blocks[ctx.loop_header].loop_nest_depth)
            ctx.loop_exit = succ_idx;
      }
   }

   /* align resume shaders with cache line */
   if (block.kind & block_kind_resume) {
      size_t cache_aligned = align(code.size(), 16);
      code.resize(cache_aligned, 0xbf800000u); /* s_nop 0 */
      block.offset = code.size();
   }
}

unsigned
emit_program(Program* program, std::vector<uint32_t>& code, std::vector<struct aco_symbol>* symbols,
             bool append_endpgm)
{
   asm_context ctx(program, symbols);

   bool is_separately_compiled_ngg_vs_or_es =
      (program->stage.sw == SWStage::VS || program->stage.sw == SWStage::TES) &&
      program->stage.hw == AC_HW_NEXT_GEN_GEOMETRY_SHADER &&
      program->info.merged_shader_compiled_separately;

   /* Prolog has no exports. */
   if (!program->is_prolog && !program->info.ps.has_epilog &&
       !is_separately_compiled_ngg_vs_or_es &&
       (program->stage.hw == AC_HW_VERTEX_SHADER || program->stage.hw == AC_HW_PIXEL_SHADER ||
        program->stage.hw == AC_HW_NEXT_GEN_GEOMETRY_SHADER))
      fix_exports(ctx, code, program);

   for (Block& block : program->blocks) {
      block.offset = code.size();
      align_block(ctx, code, block);
      emit_block(ctx, code, block);
   }

   fix_branches(ctx, code);

   unsigned exec_size = code.size() * sizeof(uint32_t);

   /* Add end-of-code markers for the UMR disassembler. */
   if (append_endpgm)
      code.resize(code.size() + 5, 0xbf9f0000u);

   fix_constaddrs(ctx, code);

   while (program->constant_data.size() % 4u)
      program->constant_data.push_back(0);
   /* Copy constant data */
   code.insert(code.end(), (uint32_t*)program->constant_data.data(),
               (uint32_t*)(program->constant_data.data() + program->constant_data.size()));

   program->config->scratch_bytes_per_wave =
      align(program->config->scratch_bytes_per_wave, program->dev.scratch_alloc_granule);

   return exec_size;
}

} // namespace aco