aco/wave32: Use lane mask regclass for exec/vcc.

Currently all usages of exec and vcc are hardcoded to use s2 regclass.
This commit makes it possible to use s1 in wave32 mode and
s2 in wave64 mode.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>

12 files changed, 250 insertions, 209 deletions

diff --git a/src/amd/compiler/aco_builder_h.py b/src/amd/compiler/aco_builder_h.py
index e70d9317b3f..ada0806f6a9 100644
--- a/src/amd/compiler/aco_builder_h.py
+++ b/src/amd/compiler/aco_builder_h.py
@@ -135,13 +135,14 @@ public:
    Program *program;
    bool use_iterator;
    bool start; // only when use_iterator == false
+   RegClass lm;
 
    std::vector<aco_ptr<Instruction>> *instructions;
    std::vector<aco_ptr<Instruction>>::iterator it;
 
-   Builder(Program *pgm) : program(pgm), use_iterator(false), start(false), instructions(NULL) {}
-   Builder(Program *pgm, Block *block) : program(pgm), use_iterator(false), start(false), instructions(&block->instructions) {}
-   Builder(Program *pgm, std::vector<aco_ptr<Instruction>> *instrs) : program(pgm), use_iterator(false), start(false), instructions(instrs) {}
+   Builder(Program *pgm) : program(pgm), use_iterator(false), start(false), lm(pgm->lane_mask), instructions(NULL) {}
+   Builder(Program *pgm, Block *block) : program(pgm), use_iterator(false), start(false), lm(pgm ? pgm->lane_mask : s2), instructions(&block->instructions) {}
+   Builder(Program *pgm, std::vector<aco_ptr<Instruction>> *instrs) : program(pgm), use_iterator(false), start(false), lm(pgm ? pgm->lane_mask : s2), instructions(instrs) {}
 
    void moveEnd(Block *block) {
       instructions = &block->instructions;
@@ -265,17 +266,26 @@ public:
 
 % for fixed in ['m0', 'vcc', 'exec', 'scc']:
    Operand ${fixed}(Temp tmp) {
+       % if fixed == 'vcc' or fixed == 'exec':
+          assert(tmp.regClass() == lm);
+       % endif
        Operand op(tmp);
        op.setFixed(aco::${fixed});
        return op;
    }
 
    Definition ${fixed}(Definition def) {
+       % if fixed == 'vcc' or fixed == 'exec':
+          assert(def.regClass() == lm);
+       % endif
        def.setFixed(aco::${fixed});
        return def;
    }
 
    Definition hint_${fixed}(Definition def) {
+       % if fixed == 'vcc' or fixed == 'exec':
+          assert(def.regClass() == lm);
+       % endif
        def.setHint(aco::${fixed});
        return def;
    }
@@ -350,11 +360,11 @@ public:
       assert((post_ra || b.op.hasRegClass()) && b.op.regClass().type() == RegType::vgpr);
 
       if (!carry_in.op.isUndefined())
-         return vop2(aco_opcode::v_addc_co_u32, Definition(dst), hint_vcc(def(s2)), a, b, carry_in);
+         return vop2(aco_opcode::v_addc_co_u32, Definition(dst), hint_vcc(def(lm)), a, b, carry_in);
       else if (program->chip_class >= GFX10 && carry_out)
          return vop3(aco_opcode::v_add_co_u32_e64, Definition(dst), def(s2), a, b);
       else if (program->chip_class < GFX9 || carry_out)
-         return vop2(aco_opcode::v_add_co_u32, Definition(dst), hint_vcc(def(s2)), a, b);
+         return vop2(aco_opcode::v_add_co_u32, Definition(dst), hint_vcc(def(lm)), a, b);
       else
          return vop2(aco_opcode::v_add_u32, Definition(dst), a, b);
    }
@@ -407,6 +417,7 @@ public:
       }
       return insert(std::move(sub));
    }
+
 <%
 import itertools
 formats = [("pseudo", [Format.PSEUDO], 'Pseudo_instruction', list(itertools.product(range(5), range(5))) + [(8, 1), (1, 8)]),
diff --git a/src/amd/compiler/aco_insert_exec_mask.cpp b/src/amd/compiler/aco_insert_exec_mask.cpp
index 31ae5ca658c..cbc0698096b 100644
--- a/src/amd/compiler/aco_insert_exec_mask.cpp
+++ b/src/amd/compiler/aco_insert_exec_mask.cpp
@@ -302,14 +302,15 @@ 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;
-      exec_mask = bld.sop1(aco_opcode::s_wqm_b64, bld.def(s2, exec), bld.def(s1, scc), 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;
    }
    /* otherwise, the WQM mask should be one below the current mask */
    ctx.info[idx].exec.pop_back();
    assert(ctx.info[idx].exec.back().second & mask_type_wqm);
-   ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(s2, exec),
+   assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
+   ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
                                                 ctx.info[idx].exec.back().first);
 }
 
@@ -324,14 +325,15 @@ void transition_to_Exact(exec_ctx& ctx, Builder bld, unsigned idx)
        !(ctx.info[idx].exec.back().second & mask_type_loop)) {
       ctx.info[idx].exec.pop_back();
       assert(ctx.info[idx].exec.back().second & mask_type_exact);
-      ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(s2, exec),
+      assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
+      ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
                                                    ctx.info[idx].exec.back().first);
       return;
    }
    /* otherwise, we create an exact mask and push to the stack */
    Temp wqm = ctx.info[idx].exec.back().first;
-   Temp exact = bld.tmp(s2);
-   wqm = bld.sop1(aco_opcode::s_and_saveexec_b64, bld.def(s2), bld.def(s1, scc),
+   Temp exact = bld.tmp(bld.lm);
+   wqm = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
                   bld.exec(Definition(exact)), ctx.info[idx].exec[0].first, bld.exec(wqm));
    ctx.info[idx].exec.back().first = wqm;
    ctx.info[idx].exec.emplace_back(exact, mask_type_exact);
@@ -359,7 +361,7 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
       } else {
          uint8_t mask = mask_type_global;
          if (ctx.program->needs_wqm) {
-            exec_mask = bld.sop1(aco_opcode::s_wqm_b64, bld.def(s2, exec), bld.def(s1, scc), bld.exec(exec_mask));
+            exec_mask = bld.sop1(Builder::s_wqm, bld.def(bld.lm, exec), bld.def(s1, scc), bld.exec(exec_mask));
             mask |= mask_type_wqm;
          } else {
             mask |= mask_type_exact;
@@ -383,7 +385,7 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
          aco_ptr<Pseudo_instruction> phi;
          for (int i = 0; i < info.num_exec_masks - 1; i++) {
             phi.reset(create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1));
-            phi->definitions[0] = bld.def(s2);
+            phi->definitions[0] = bld.def(bld.lm);
             phi->operands[0] = Operand(ctx.info[preds[0]].exec[i].first);
             ctx.info[idx].exec[i].first = bld.insert(std::move(phi));
          }
@@ -393,7 +395,7 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
       if (info.has_divergent_break) {
          /* this phi might be trivial but ensures a parallelcopy on the loop header */
          aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
-         phi->definitions[0] = bld.def(s2);
+         phi->definitions[0] = bld.def(bld.lm);
          phi->operands[0] = Operand(ctx.info[preds[0]].exec[info.num_exec_masks - 1].first);
          ctx.info[idx].exec.back().first = bld.insert(std::move(phi));
       }
@@ -401,9 +403,9 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
       /* create ssa name for loop active mask */
       aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
       if (info.has_divergent_continue)
-         phi->definitions[0] = bld.def(s2);
+         phi->definitions[0] = bld.def(bld.lm);
       else
-         phi->definitions[0] = bld.def(s2, exec);
+         phi->definitions[0] = bld.def(bld.lm, exec);
       phi->operands[0] = Operand(ctx.info[preds[0]].exec.back().first);
       Temp loop_active = bld.insert(std::move(phi));
 
@@ -423,7 +425,8 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
             i++;
          }
          uint8_t mask_type = ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact);
-         ctx.info[idx].exec.emplace_back(bld.pseudo(aco_opcode::p_parallelcopy, bld.def(s2, exec),
+         assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
+         ctx.info[idx].exec.emplace_back(bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
                                                     ctx.info[idx].exec.back().first), mask_type);
       }
 
@@ -480,7 +483,7 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
          } else {
             /* 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(s2);
+            phi->definitions[0] = bld.def(bld.lm);
             for (unsigned i = 0; i < phi->operands.size(); i++)
                phi->operands[i] = Operand(ctx.info[preds[i]].exec[k].first);
             ctx.info[idx].exec.emplace_back(bld.insert(std::move(phi)), type);
@@ -510,7 +513,8 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
             transition_to_Exact(ctx, bld, idx);
       }
 
-      ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(s2, exec),
+      assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
+      ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
                                                    ctx.info[idx].exec.back().first);
 
       ctx.loop.pop_back();
@@ -536,7 +540,7 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
             continue;
          }
 
-         Temp phi = bld.pseudo(aco_opcode::p_linear_phi, in_exec ? bld.def(s2, exec) : bld.def(s2),
+         Temp phi = bld.pseudo(aco_opcode::p_linear_phi, in_exec ? bld.def(bld.lm, exec) : bld.def(bld.lm),
                                ctx.info[preds[0]].exec[i].first,
                                ctx.info[preds[1]].exec[i].first);
          uint8_t mask_type = ctx.info[preds[0]].exec[i].second & ctx.info[preds[1]].exec[i].second;
@@ -578,7 +582,8 @@ unsigned add_coupling_code(exec_ctx& ctx, Block* block,
 
    if (block->kind & block_kind_merge) {
       Temp restore = ctx.info[idx].exec.back().first;
-      ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(s2, exec), restore);
+      assert(restore.size() == bld.lm.size());
+      ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec), restore);
    }
 
    return i;
@@ -589,7 +594,7 @@ void lower_fs_buffer_store_smem(Builder& bld, bool need_check, aco_ptr<Instructi
    Operand offset = instr->operands[1];
    if (need_check) {
       /* if exec is zero, then use UINT32_MAX as an offset and make this store a no-op */
-      Temp nonempty = bld.sopc(aco_opcode::s_cmp_lg_u64, bld.def(s1, scc), cur_exec, Operand(0u));
+      Temp nonempty = bld.sopc(Builder::s_cmp_lg, bld.def(s1, scc), cur_exec, Operand(0u));
 
       if (offset.isLiteral())
          offset = bld.sop1(aco_opcode::s_mov_b32, bld.def(s1), offset);
@@ -665,7 +670,7 @@ void process_instructions(exec_ctx& ctx, Block* block,
          assert(num);
          Operand cond = instr->operands[0];
          for (int i = num - 1; i >= 0; i--) {
-            Instruction *andn2 = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.def(s1, scc),
+            Instruction *andn2 = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc),
                                           ctx.info[block->index].exec[i].first, cond);
             if (i == num - 1) {
                andn2->operands[0].setFixed(exec);
@@ -689,8 +694,9 @@ void process_instructions(exec_ctx& ctx, Block* block,
 
       if (instr->opcode == aco_opcode::p_is_helper || instr->opcode == aco_opcode::p_load_helper) {
          Definition dst = instr->definitions[0];
+         assert(dst.size() == bld.lm.size());
          if (state == Exact) {
-            instr.reset(create_instruction<SOP1_instruction>(aco_opcode::s_mov_b64, Format::SOP1, 1, 1));
+            instr.reset(create_instruction<SOP1_instruction>(bld.w64or32(Builder::s_mov), Format::SOP1, 1, 1));
             instr->operands[0] = Operand(0u);
             instr->definitions[0] = dst;
          } else {
@@ -710,7 +716,7 @@ void process_instructions(exec_ctx& ctx, Block* block,
             assert(instr->opcode == aco_opcode::p_is_helper || exact_mask.second & mask_type_initial);
             assert(exact_mask.second & mask_type_exact);
 
-            instr.reset(create_instruction<SOP2_instruction>(aco_opcode::s_andn2_b64, Format::SOP2, 2, 2));
+            instr.reset(create_instruction<SOP2_instruction>(bld.w64or32(Builder::s_andn2), Format::SOP2, 2, 2));
             instr->operands[0] = Operand(ctx.info[block->index].exec.back().first); /* current exec */
             instr->operands[1] = Operand(exact_mask.first);
             instr->definitions[0] = dst;
@@ -726,8 +732,8 @@ void process_instructions(exec_ctx& ctx, Block* block,
          if (instr->operands.empty()) {
             /* transition to exact and set exec to zero */
             Temp old_exec = ctx.info[block->index].exec.back().first;
-            Temp new_exec = bld.tmp(s2);
-            cond = bld.sop1(aco_opcode::s_and_saveexec_b64, bld.def(s2), bld.def(s1, scc),
+            Temp new_exec = bld.tmp(bld.lm);
+            cond = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
                             bld.exec(Definition(new_exec)), Operand(0u), bld.exec(old_exec));
             if (ctx.info[block->index].exec.back().second & mask_type_exact) {
                ctx.info[block->index].exec.back().first = new_exec;
@@ -746,7 +752,7 @@ void process_instructions(exec_ctx& ctx, Block* block,
          num += ctx.info[block->index].exec.size() - 1;
          for (int i = num - 1; i >= 0; i--) {
             if (ctx.info[block->index].exec[i].second & mask_type_exact) {
-               Instruction *andn2 = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.def(s1, scc),
+               Instruction *andn2 = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc),
                                              ctx.info[block->index].exec[i].first, cond);
                if (i == num - 1) {
                   andn2->operands[0].setFixed(exec);
@@ -878,13 +884,13 @@ void add_branch_code(exec_ctx& ctx, Block* block)
       }
 
       Temp old_exec = ctx.info[idx].exec.back().first;
-      Temp new_exec = bld.tmp(s2);
-      Temp cond = bld.sop1(aco_opcode::s_and_saveexec_b64, bld.def(s2), bld.def(s1, scc),
+      Temp new_exec = bld.tmp(bld.lm);
+      Temp cond = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
                            bld.exec(Definition(new_exec)), Operand(0u), bld.exec(old_exec));
       ctx.info[idx].exec.back().first = new_exec;
 
       for (int i = num - 1; i >= 0; i--) {
-         Instruction *andn2 = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.def(s1, scc),
+         Instruction *andn2 = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc),
                                        ctx.info[block->index].exec[i].first, cond);
          if (i == (int)ctx.info[idx].exec.size() - 1)
             andn2->definitions[0].setFixed(exec);
@@ -912,7 +918,7 @@ void add_branch_code(exec_ctx& ctx, Block* block)
          Temp cond = Temp();
          for (int exec_idx = ctx.info[idx].exec.size() - 1; exec_idx >= 0; exec_idx--) {
             if (ctx.info[idx].exec[exec_idx].second & mask_type_loop) {
-               cond = bld.sopc(aco_opcode::s_cmp_lg_u64, bld.def(s1, scc), ctx.info[idx].exec[exec_idx].first, Operand(0u));
+               cond = bld.sopc(Builder::s_cmp_lg, bld.def(s1, scc), ctx.info[idx].exec[exec_idx].first, Operand(0u));
                break;
             }
          }
@@ -957,8 +963,8 @@ void add_branch_code(exec_ctx& ctx, Block* block)
       Temp current_exec = ctx.info[idx].exec.back().first;
       uint8_t mask_type = ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact);
 
-      Temp then_mask = bld.tmp(s2);
-      Temp old_exec = bld.sop1(aco_opcode::s_and_saveexec_b64, bld.def(s2), bld.def(s1, scc),
+      Temp then_mask = bld.tmp(bld.lm);
+      Temp old_exec = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
                                bld.exec(Definition(then_mask)), cond, bld.exec(current_exec));
 
       ctx.info[idx].exec.back().first = old_exec;
@@ -978,7 +984,7 @@ void add_branch_code(exec_ctx& ctx, Block* block)
       uint8_t mask_type = ctx.info[idx].exec.back().second;
       ctx.info[idx].exec.pop_back();
       Temp orig_exec = ctx.info[idx].exec.back().first;
-      Temp else_mask = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2, exec),
+      Temp else_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm, exec),
                                 bld.def(s1, scc), orig_exec, bld.exec(then_mask));
 
       /* add next current exec to the stack */
@@ -998,7 +1004,7 @@ void add_branch_code(exec_ctx& ctx, Block* block)
       for (int exec_idx = ctx.info[idx].exec.size() - 2; exec_idx >= 0; exec_idx--) {
          cond = bld.tmp(s1);
          Temp exec_mask = ctx.info[idx].exec[exec_idx].first;
-         exec_mask = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.scc(Definition(cond)),
+         exec_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.scc(Definition(cond)),
                               exec_mask, current_exec);
          ctx.info[idx].exec[exec_idx].first = exec_mask;
          if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
@@ -1010,7 +1016,7 @@ void add_branch_code(exec_ctx& ctx, Block* block)
       unsigned succ_idx = ctx.program->blocks[block->linear_succs[1]].linear_succs[0];
       Block& succ = ctx.program->blocks[succ_idx];
       if (!(succ.kind & block_kind_invert || succ.kind & block_kind_merge)) {
-         ctx.info[idx].exec.back().first = bld.sop1(aco_opcode::s_mov_b64, bld.def(s2, exec), Operand(0u));
+         ctx.info[idx].exec.back().first = bld.sop1(Builder::s_mov, bld.def(bld.lm, exec), Operand(0u));
       }
 
       bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
@@ -1028,7 +1034,7 @@ void add_branch_code(exec_ctx& ctx, Block* block)
             break;
          cond = bld.tmp(s1);
          Temp exec_mask = ctx.info[idx].exec[exec_idx].first;
-         exec_mask = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.scc(Definition(cond)),
+         exec_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.scc(Definition(cond)),
                               exec_mask, bld.exec(current_exec));
          ctx.info[idx].exec[exec_idx].first = exec_mask;
       }
@@ -1039,7 +1045,7 @@ void add_branch_code(exec_ctx& ctx, Block* block)
       unsigned succ_idx = ctx.program->blocks[block->linear_succs[1]].linear_succs[0];
       Block& succ = ctx.program->blocks[succ_idx];
       if (!(succ.kind & block_kind_invert || succ.kind & block_kind_merge)) {
-         ctx.info[idx].exec.back().first = bld.sop1(aco_opcode::s_mov_b64, bld.def(s2, exec), Operand(0u));
+         ctx.info[idx].exec.back().first = bld.sop1(Builder::s_mov, bld.def(bld.lm, exec), Operand(0u));
       }
 
       bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index a2b2c21170c..9de9d5dec14 100644
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -351,12 +351,12 @@ Temp bool_to_vector_condition(isel_context *ctx, Temp val, Temp dst = Temp(0, s2
 {
    Builder bld(ctx->program, ctx->block);
    if (!dst.id())
-      dst = bld.tmp(s2);
+      dst = bld.tmp(bld.lm);
 
    assert(val.regClass() == s1);
-   assert(dst.regClass() == s2);
+   assert(dst.regClass() == bld.lm);
 
-   return bld.sop2(aco_opcode::s_cselect_b64, bld.hint_vcc(Definition(dst)), Operand((uint32_t) -1), Operand(0u), bld.scc(val));
+   return bld.sop2(Builder::s_cselect, bld.hint_vcc(Definition(dst)), Operand((uint32_t) -1), Operand(0u), bld.scc(val));
 }
 
 Temp bool_to_scalar_condition(isel_context *ctx, Temp val, Temp dst = Temp(0, s1))
@@ -365,12 +365,12 @@ Temp bool_to_scalar_condition(isel_context *ctx, Temp val, Temp dst = Temp(0, s1
    if (!dst.id())
       dst = bld.tmp(s1);
 
-   assert(val.regClass() == s2);
+   assert(val.regClass() == bld.lm);
    assert(dst.regClass() == s1);
 
    /* if we're currently in WQM mode, ensure that the source is also computed in WQM */
    Temp tmp = bld.tmp(s1);
-   bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.scc(Definition(tmp)), val, Operand(exec, s2));
+   bld.sop2(Builder::s_and, bld.def(bld.lm), bld.scc(Definition(tmp)), val, Operand(exec, bld.lm));
    return emit_wqm(ctx, tmp, dst);
 }
 
@@ -489,6 +489,8 @@ void emit_vopc_instruction(isel_context *ctx, nir_alu_instr *instr, aco_opcode o
 {
    Temp src0 = get_alu_src(ctx, instr->src[0]);
    Temp src1 = get_alu_src(ctx, instr->src[1]);
+   assert(src0.size() == src1.size());
+
    aco_ptr<Instruction> vopc;
    if (src1.type() == RegType::sgpr) {
       if (src0.type() == RegType::vgpr) {
@@ -549,12 +551,13 @@ void emit_sopc_instruction(isel_context *ctx, nir_alu_instr *instr, aco_opcode o
 {
    Temp src0 = get_alu_src(ctx, instr->src[0]);
    Temp src1 = get_alu_src(ctx, instr->src[1]);
+   Builder bld(ctx->program, ctx->block);
 
-   assert(dst.regClass() == s2);
+   assert(dst.regClass() == bld.lm);
    assert(src0.type() == RegType::sgpr);
    assert(src1.type() == RegType::sgpr);
+   assert(src0.regClass() == src1.regClass());
 
-   Builder bld(ctx->program, ctx->block);
    /* Emit the SALU comparison instruction */
    Temp cmp = bld.sopc(op, bld.scc(bld.def(s1)), src0, src1);
    /* Turn the result into a per-lane bool */
@@ -580,17 +583,17 @@ void emit_comparison(isel_context *ctx, nir_alu_instr *instr, Temp dst,
       emit_sopc_instruction(ctx, instr, op, dst);
 }
 
-void emit_boolean_logic(isel_context *ctx, nir_alu_instr *instr, aco_opcode op64, Temp dst)
+void emit_boolean_logic(isel_context *ctx, nir_alu_instr *instr, Builder::WaveSpecificOpcode op, Temp dst)
 {
    Builder bld(ctx->program, ctx->block);
    Temp src0 = get_alu_src(ctx, instr->src[0]);
    Temp src1 = get_alu_src(ctx, instr->src[1]);
 
-   assert(dst.regClass() == s2);
-   assert(src0.regClass() == s2);
-   assert(src1.regClass() == s2);
+   assert(dst.regClass() == bld.lm);
+   assert(src0.regClass() == bld.lm);
+   assert(src1.regClass() == bld.lm);
 
-   bld.sop2(op64, Definition(dst), bld.def(s1, scc), src0, src1);
+   bld.sop2(op, Definition(dst), bld.def(s1, scc), src0, src1);
 }
 
 void emit_bcsel(isel_context *ctx, nir_alu_instr *instr, Temp dst)
@@ -600,7 +603,7 @@ void emit_bcsel(isel_context *ctx, nir_alu_instr *instr, Temp dst)
    Temp then = get_alu_src(ctx, instr->src[1]);
    Temp els = get_alu_src(ctx, instr->src[2]);
 
-   assert(cond.regClass() == s2);
+   assert(cond.regClass() == bld.lm);
 
    if (dst.type() == RegType::vgpr) {
       aco_ptr<Instruction> bcsel;
@@ -628,14 +631,15 @@ void emit_bcsel(isel_context *ctx, nir_alu_instr *instr, Temp dst)
    }
 
    if (instr->dest.dest.ssa.bit_size == 1) {
-      assert(dst.regClass() == s2);
-      assert(then.regClass() == s2);
-      assert(els.regClass() == s2);
+      assert(dst.regClass() == bld.lm);
+      assert(then.regClass() == bld.lm);
+      assert(els.regClass() == bld.lm);
    }
 
    if (!ctx->divergent_vals[instr->src[0].src.ssa->index]) { /* uniform condition and values in sgpr */
       if (dst.regClass() == s1 || dst.regClass() == s2) {
          assert((then.regClass() == s1 || then.regClass() == s2) && els.regClass() == then.regClass());
+         assert(dst.size() == then.size());
          aco_opcode op = dst.regClass() == s1 ? aco_opcode::s_cselect_b32 : aco_opcode::s_cselect_b64;
          bld.sop2(op, Definition(dst), then, els, bld.scc(bool_to_scalar_condition(ctx, cond)));
       } else {
@@ -652,20 +656,20 @@ void emit_bcsel(isel_context *ctx, nir_alu_instr *instr, Temp dst)
    assert(instr->dest.dest.ssa.bit_size == 1);
 
    if (cond.id() != then.id())
-      then = bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), cond, then);
+      then = bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), cond, then);
 
    if (cond.id() == els.id())
-      bld.sop1(aco_opcode::s_mov_b64, Definition(dst), then);
+      bld.sop1(Builder::s_mov, Definition(dst), then);
    else
-      bld.sop2(aco_opcode::s_or_b64, Definition(dst), bld.def(s1, scc), then,
-               bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.def(s1, scc), els, cond));
+      bld.sop2(Builder::s_or, Definition(dst), bld.def(s1, scc), then,
+               bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc), els, cond));
 }
 
 void emit_scaled_op(isel_context *ctx, Builder& bld, Definition dst, Temp val,
                     aco_opcode op, uint32_t undo)
 {
    /* multiply by 16777216 to handle denormals */
-   Temp is_denormal = bld.vopc(aco_opcode::v_cmp_class_f32, bld.hint_vcc(bld.def(s2)),
+   Temp is_denormal = bld.vopc(aco_opcode::v_cmp_class_f32, bld.hint_vcc(bld.def(bld.lm)),
                                as_vgpr(ctx, val), bld.copy(bld.def(v1), Operand((1u << 7) | (1u << 4))));
    Temp scaled = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), Operand(0x4b800000u), val);
    scaled = bld.vop1(op, bld.def(v1), scaled);
@@ -766,9 +770,9 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    case nir_op_inot: {
       Temp src = get_alu_src(ctx, instr->src[0]);
       if (instr->dest.dest.ssa.bit_size == 1) {
-         assert(src.regClass() == s2);
-         assert(dst.regClass() == s2);
-         bld.sop2(aco_opcode::s_andn2_b64, Definition(dst), bld.def(s1, scc), Operand(exec, s2), src);
+         assert(src.regClass() == bld.lm);
+         assert(dst.regClass() == bld.lm);
+         bld.sop2(Builder::s_andn2, Definition(dst), bld.def(s1, scc), Operand(exec, bld.lm), src);
       } else if (dst.regClass() == v1) {
          emit_vop1_instruction(ctx, instr, aco_opcode::v_not_b32, dst);
       } else if (dst.type() == RegType::sgpr) {
@@ -835,12 +839,12 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
          bld.sop2(aco_opcode::s_or_b64, Definition(dst), bld.def(s1, scc), neg, neqz);
       } else if (dst.regClass() == v1) {
          Temp tmp = bld.vop2(aco_opcode::v_ashrrev_i32, bld.def(v1), Operand(31u), src);
-         Temp gtz = bld.vopc(aco_opcode::v_cmp_ge_i32, bld.hint_vcc(bld.def(s2)), Operand(0u), src);
+         Temp gtz = bld.vopc(aco_opcode::v_cmp_ge_i32, bld.hint_vcc(bld.def(bld.lm)), Operand(0u), src);
          bld.vop2(aco_opcode::v_cndmask_b32, Definition(dst), Operand(1u), tmp, gtz);
       } else if (dst.regClass() == v2) {
          Temp upper = emit_extract_vector(ctx, src, 1, v1);
          Temp neg = bld.vop2(aco_opcode::v_ashrrev_i32, bld.def(v1), Operand(31u), upper);
-         Temp gtz = bld.vopc(aco_opcode::v_cmp_ge_i64, bld.hint_vcc(bld.def(s2)), Operand(0u), src);
+         Temp gtz = bld.vopc(aco_opcode::v_cmp_ge_i64, bld.hint_vcc(bld.def(bld.lm)), Operand(0u), src);
          Temp lower = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), Operand(1u), neg, gtz);
          upper = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), Operand(0u), neg, gtz);
          bld.pseudo(aco_opcode::p_create_vector, Definition(dst), lower, upper);
@@ -901,7 +905,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    }
    case nir_op_ior: {
       if (instr->dest.dest.ssa.bit_size == 1) {
-         emit_boolean_logic(ctx, instr, aco_opcode::s_or_b64, dst);
+         emit_boolean_logic(ctx, instr, Builder::s_or, dst);
       } else if (dst.regClass() == v1) {
          emit_vop2_instruction(ctx, instr, aco_opcode::v_or_b32, dst, true);
       } else if (dst.regClass() == s1) {
@@ -917,7 +921,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    }
    case nir_op_iand: {
       if (instr->dest.dest.ssa.bit_size == 1) {
-         emit_boolean_logic(ctx, instr, aco_opcode::s_and_b64, dst);
+         emit_boolean_logic(ctx, instr, Builder::s_and, dst);
       } else if (dst.regClass() == v1) {
          emit_vop2_instruction(ctx, instr, aco_opcode::v_and_b32, dst, true);
       } else if (dst.regClass() == s1) {
@@ -933,7 +937,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    }
    case nir_op_ixor: {
       if (instr->dest.dest.ssa.bit_size == 1) {
-         emit_boolean_logic(ctx, instr, aco_opcode::s_xor_b64, dst);
+         emit_boolean_logic(ctx, instr, Builder::s_xor, dst);
       } else if (dst.regClass() == v1) {
          emit_vop2_instruction(ctx, instr, aco_opcode::v_xor_b32, dst, true);
       } else if (dst.regClass() == s1) {
@@ -1709,16 +1713,16 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    case nir_op_fsign: {
       Temp src = as_vgpr(ctx, get_alu_src(ctx, instr->src[0]));
       if (dst.size() == 1) {
-         Temp cond = bld.vopc(aco_opcode::v_cmp_nlt_f32, bld.hint_vcc(bld.def(s2)), Operand(0u), src);
+         Temp cond = bld.vopc(aco_opcode::v_cmp_nlt_f32, bld.hint_vcc(bld.def(bld.lm)), Operand(0u), src);
          src = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), Operand(0x3f800000u), src, cond);
-         cond = bld.vopc(aco_opcode::v_cmp_le_f32, bld.hint_vcc(bld.def(s2)), Operand(0u), src);
+         cond = bld.vopc(aco_opcode::v_cmp_le_f32, bld.hint_vcc(bld.def(bld.lm)), Operand(0u), src);
          bld.vop2(aco_opcode::v_cndmask_b32, Definition(dst), Operand(0xbf800000u), src, cond);
       } else if (dst.size() == 2) {
-         Temp cond = bld.vopc(aco_opcode::v_cmp_nlt_f64, bld.hint_vcc(bld.def(s2)), Operand(0u), src);
+         Temp cond = bld.vopc(aco_opcode::v_cmp_nlt_f64, bld.hint_vcc(bld.def(bld.lm)), Operand(0u), src);
          Temp tmp = bld.vop1(aco_opcode::v_mov_b32, bld.def(v1), Operand(0x3FF00000u));
          Temp upper = bld.vop2_e64(aco_opcode::v_cndmask_b32, bld.def(v1), tmp, emit_extract_vector(ctx, src, 1, v1), cond);
 
-         cond = bld.vopc(aco_opcode::v_cmp_le_f64, bld.hint_vcc(bld.def(s2)), Operand(0u), src);
+         cond = bld.vopc(aco_opcode::v_cmp_le_f64, bld.hint_vcc(bld.def(bld.lm)), Operand(0u), src);
          tmp = bld.vop1(aco_opcode::v_mov_b32, bld.def(v1), Operand(0xBFF00000u));
          upper = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), tmp, upper, cond);
 
@@ -1922,7 +1926,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
       Temp src = get_alu_src(ctx, instr->src[0]);
       if (instr->src[0].src.ssa->bit_size == 32 && dst.type() == RegType::vgpr) {
          Temp exponent = bld.vop1(aco_opcode::v_frexp_exp_i32_f32, bld.def(v1), src);
-         Temp exponent_in_range = bld.vopc(aco_opcode::v_cmp_ge_i32, bld.hint_vcc(bld.def(s2)), Operand(64u), exponent);
+         Temp exponent_in_range = bld.vopc(aco_opcode::v_cmp_ge_i32, bld.hint_vcc(bld.def(bld.lm)), Operand(64u), exponent);
          exponent = bld.vop2(aco_opcode::v_max_i32, bld.def(v1), Operand(0x0u), exponent);
          Temp mantissa = bld.vop2(aco_opcode::v_and_b32, bld.def(v1), Operand(0x7fffffu), src);
          mantissa = bld.vop2(aco_opcode::v_or_b32, bld.def(v1), Operand(0x800000u), mantissa);
@@ -1986,7 +1990,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    }
    case nir_op_b2f32: {
       Temp src = get_alu_src(ctx, instr->src[0]);
-      assert(src.regClass() == s2);
+      assert(src.regClass() == bld.lm);
 
       if (dst.regClass() == s1) {
          src = bool_to_scalar_condition(ctx, src);
@@ -2000,7 +2004,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    }
    case nir_op_b2f64: {
       Temp src = get_alu_src(ctx, instr->src[0]);
-      assert(src.regClass() == s2);
+      assert(src.regClass() == bld.lm);
 
       if (dst.regClass() == s2) {
          src = bool_to_scalar_condition(ctx, src);
@@ -2073,7 +2077,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    }
    case nir_op_b2i32: {
       Temp src = get_alu_src(ctx, instr->src[0]);
-      assert(src.regClass() == s2);
+      assert(src.regClass() == bld.lm);
 
       if (dst.regClass() == s1) {
          // TODO: in a post-RA optimization, we can check if src is in VCC, and directly use VCCNZ
@@ -2087,7 +2091,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    }
    case nir_op_i2b1: {
       Temp src = get_alu_src(ctx, instr->src[0]);
-      assert(dst.regClass() == s2);
+      assert(dst.regClass() == bld.lm);
 
       if (src.type() == RegType::vgpr) {
          assert(src.regClass() == v1 || src.regClass() == v2);
@@ -2164,7 +2168,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
 
       Temp mask = bld.copy(bld.def(s1), Operand(0x36Fu)); /* value is NOT negative/positive denormal value */
 
-      Temp cmp_res = bld.tmp(s2);
+      Temp cmp_res = bld.tmp(bld.lm);
       bld.vopc_e64(aco_opcode::v_cmp_class_f16, Definition(cmp_res), f16, mask).def(0).setHint(vcc);
 
       Temp f32 = bld.vop1(aco_opcode::v_cvt_f32_f16, bld.def(v1), f16);
@@ -2338,14 +2342,14 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    }
    case nir_op_ieq: {
       if (instr->src[0].src.ssa->bit_size == 1)
-         emit_boolean_logic(ctx, instr, aco_opcode::s_xnor_b64, dst);
+         emit_boolean_logic(ctx, instr, Builder::s_xnor, dst);
       else
          emit_comparison(ctx, instr, dst, aco_opcode::v_cmp_eq_i32, aco_opcode::v_cmp_eq_i64, aco_opcode::s_cmp_eq_i32, aco_opcode::s_cmp_eq_u64);
       break;
    }
    case nir_op_ine: {
       if (instr->src[0].src.ssa->bit_size == 1)
-         emit_boolean_logic(ctx, instr, aco_opcode::s_xor_b64, dst);
+         emit_boolean_logic(ctx, instr, Builder::s_xor, dst);
       else
          emit_comparison(ctx, instr, dst, aco_opcode::v_cmp_lg_i32, aco_opcode::v_cmp_lg_i64, aco_opcode::s_cmp_lg_i32, aco_opcode::s_cmp_lg_u64);
       break;
@@ -2405,8 +2409,10 @@ void visit_load_const(isel_context *ctx, nir_load_const_instr *instr)
    Builder bld(ctx->program, ctx->block);
 
    if (instr->def.bit_size == 1) {
-      assert(dst.regClass() == s2);
-      bld.sop1(aco_opcode::s_mov_b64, Definition(dst), Operand((uint64_t)(instr->value[0].b ? -1 : 0)));
+      assert(dst.regClass() == bld.lm);
+      int val = instr->value[0].b ? -1 : 0;
+      Operand op = bld.lm.size() == 1 ? Operand((uint32_t) val) : Operand((uint64_t) val);
+      bld.sop1(Builder::s_mov, Definition(dst), op);
    } else if (dst.size() == 1) {
       bld.copy(Definition(dst), Operand(instr->value[0].u32));
    } else {
@@ -3033,7 +3039,7 @@ Temp adjust_vertex_fetch_alpha(isel_context *ctx, unsigned adjustment, Temp alph
    /* Convert back to the right type. */
    if (adjustment == RADV_ALPHA_ADJUST_SNORM) {
       alpha = bld.vop1(aco_opcode::v_cvt_f32_i32, bld.def(v1), alpha);
-      Temp clamp = bld.vopc(aco_opcode::v_cmp_le_f32, bld.hint_vcc(bld.def(s2)), Operand(0xbf800000u), alpha);
+      Temp clamp = bld.vopc(aco_opcode::v_cmp_le_f32, bld.hint_vcc(bld.def(bld.lm)), Operand(0xbf800000u), alpha);
       alpha = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), Operand(0xbf800000u), alpha, clamp);
    } else if (adjustment == RADV_ALPHA_ADJUST_SSCALED) {
       alpha = bld.vop1(aco_opcode::v_cvt_f32_i32, bld.def(v1), alpha);
@@ -3599,8 +3605,8 @@ void visit_discard_if(isel_context *ctx, nir_intrinsic_instr *instr)
    // TODO: optimize uniform conditions
    Builder bld(ctx->program, ctx->block);
    Temp src = get_ssa_temp(ctx, instr->src[0].ssa);
-   assert(src.regClass() == s2);
-   src = bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), src, Operand(exec, s2));
+   assert(src.regClass() == bld.lm);
+   src = bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), src, Operand(exec, bld.lm));
    bld.pseudo(aco_opcode::p_discard_if, src);
    ctx->block->kind |= block_kind_uses_discard_if;
    return;
@@ -3663,7 +3669,7 @@ void visit_discard(isel_context* ctx, nir_intrinsic_instr *instr)
       ctx->program->needs_exact = true;
       /* save exec somewhere temporarily so that it doesn't get
        * overwritten before the discard from outer exec masks */
-      Temp cond = bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), Operand(0xFFFFFFFF), Operand(exec, s2));
+      Temp cond = bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), Operand(0xFFFFFFFF), Operand(exec, bld.lm));
       bld.pseudo(aco_opcode::p_discard_if, cond);
       ctx->block->kind |= block_kind_uses_discard_if;
       return;
@@ -3950,7 +3956,7 @@ static Temp adjust_sample_index_using_fmask(isel_context *ctx, bool da, Temp coo
    /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
     * resource descriptor is 0 (invalid),
     */
-   Temp compare = bld.tmp(s2);
+   Temp compare = bld.tmp(bld.lm);
    bld.vopc_e64(aco_opcode::v_cmp_lg_u32, Definition(compare),
                 Operand(0u), emit_extract_vector(ctx, fmask_desc_ptr, 1, s1)).def(0).setHint(vcc);
 
@@ -4739,12 +4745,12 @@ void visit_store_global(isel_context *ctx, nir_intrinsic_instr *instr)
       if (offset > 0 && ctx->options->chip_class < GFX9) {
          Temp addr0 = bld.tmp(v1), addr1 = bld.tmp(v1);
          Temp new_addr0 = bld.tmp(v1), new_addr1 = bld.tmp(v1);
-         Temp carry = bld.tmp(s2);
+         Temp carry = bld.tmp(bld.lm);
          bld.pseudo(aco_opcode::p_split_vector, Definition(addr0), Definition(addr1), addr);
 
          bld.vop2(aco_opcode::v_add_co_u32, Definition(new_addr0), bld.hint_vcc(Definition(carry)),
                   Operand(offset), addr0);
-         bld.vop2(aco_opcode::v_addc_co_u32, Definition(new_addr1), bld.def(s2),
+         bld.vop2(aco_opcode::v_addc_co_u32, Definition(new_addr1), bld.def(bld.lm),
                   Operand(0u), addr1,
                   carry).def(1).setHint(vcc);
 
@@ -5219,25 +5225,25 @@ Temp emit_boolean_reduce(isel_context *ctx, nir_op op, unsigned cluster_size, Te
       return src;
    } if (op == nir_op_iand && cluster_size == 4) {
       //subgroupClusteredAnd(val, 4) -> ~wqm(exec & ~val)
-      Temp tmp = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.def(s1, scc), Operand(exec, s2), src);
-      return bld.sop1(aco_opcode::s_not_b64, bld.def(s2), bld.def(s1, scc),
-                      bld.sop1(aco_opcode::s_wqm_b64, bld.def(s2), bld.def(s1, scc), tmp));
+      Temp tmp = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc), Operand(exec, bld.lm), src);
+      return bld.sop1(Builder::s_not, bld.def(bld.lm), bld.def(s1, scc),
+                      bld.sop1(Builder::s_wqm, bld.def(bld.lm), bld.def(s1, scc), tmp));
    } else if (op == nir_op_ior && cluster_size == 4) {
       //subgroupClusteredOr(val, 4) -> wqm(val & exec)
-      return bld.sop1(aco_opcode::s_wqm_b64, bld.def(s2), bld.def(s1, scc),
-                      bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), src, Operand(exec, s2)));
+      return bld.sop1(Builder::s_wqm, bld.def(bld.lm), bld.def(s1, scc),
+                      bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), src, Operand(exec, bld.lm)));
    } else if (op == nir_op_iand && cluster_size == 64) {
       //subgroupAnd(val) -> (exec & ~val) == 0
-      Temp tmp = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.def(s1, scc), Operand(exec, s2), src).def(1).getTemp();
-      return bld.sop2(aco_opcode::s_cselect_b64, bld.def(s2), Operand(0u), Operand(-1u), bld.scc(tmp));
+      Temp tmp = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc), Operand(exec, bld.lm), src).def(1).getTemp();
+      return bld.sop2(Builder::s_cselect, bld.def(bld.lm), Operand(0u), Operand(-1u), bld.scc(tmp));
    } else if (op == nir_op_ior && cluster_size == 64) {
       //subgroupOr(val) -> (val & exec) != 0
-      Temp tmp = bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), src, Operand(exec, s2)).def(1).getTemp();
+      Temp tmp = bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), src, Operand(exec, bld.lm)).def(1).getTemp();
       return bool_to_vector_condition(ctx, tmp);
    } else if (op == nir_op_ixor && cluster_size == 64) {
       //subgroupXor(val) -> s_bcnt1_i32_b64(val & exec) & 1
-      Temp tmp = bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), src, Operand(exec, s2));
-      tmp = bld.sop1(aco_opcode::s_bcnt1_i32_b64, bld.def(s1), bld.def(s1, scc), tmp);
+      Temp tmp = bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), src, Operand(exec, bld.lm));
+      tmp = bld.sop1(Builder::s_bcnt1_i32, bld.def(s1), bld.def(s1, scc), tmp);
       tmp = bld.sop2(aco_opcode::s_and_b32, bld.def(s1), bld.def(s1, scc), tmp, Operand(1u)).def(1).getTemp();
       return bool_to_vector_condition(ctx, tmp);
    } else {
@@ -5256,25 +5262,28 @@ Temp emit_boolean_reduce(isel_context *ctx, nir_op op, unsigned cluster_size, Te
 
       Temp tmp;
       if (op == nir_op_iand)
-         tmp = bld.sop2(aco_opcode::s_orn2_b64, bld.def(s2), bld.def(s1, scc), src, Operand(exec, s2));
+         tmp = bld.sop2(Builder::s_orn2, bld.def(bld.lm), bld.def(s1, scc), src, Operand(exec, bld.lm));
       else
-         tmp = bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), src, Operand(exec, s2));
+         tmp = bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), src, Operand(exec, bld.lm));
 
       uint32_t cluster_mask = cluster_size == 32 ? -1 : (1u << cluster_size) - 1u;
-      tmp = bld.vop3(aco_opcode::v_lshrrev_b64, bld.def(v2), cluster_offset, tmp);
+      if (ctx->program->wave_size == 64)
+         tmp = bld.vop3(aco_opcode::v_lshrrev_b64, bld.def(v2), cluster_offset, tmp);
+      else
+         tmp = bld.vop2_e64(aco_opcode::v_lshrrev_b32, bld.def(v1), cluster_offset, tmp);
       tmp = emit_extract_vector(ctx, tmp, 0, v1);
       if (cluster_mask != 0xffffffff)
          tmp = bld.vop2(aco_opcode::v_and_b32, bld.def(v1), Operand(cluster_mask), tmp);
 
       Definition cmp_def = Definition();
       if (op == nir_op_iand) {
-         cmp_def = bld.vopc(aco_opcode::v_cmp_eq_u32, bld.def(s2), Operand(cluster_mask), tmp).def(0);
+         cmp_def = bld.vopc(aco_opcode::v_cmp_eq_u32, bld.def(bld.lm), Operand(cluster_mask), tmp).def(0);
       } else if (op == nir_op_ior) {
-         cmp_def = bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(s2), Operand(0u), tmp).def(0);
+         cmp_def = bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(bld.lm), Operand(0u), tmp).def(0);
       } else if (op == nir_op_ixor) {
          tmp = bld.vop2(aco_opcode::v_and_b32, bld.def(v1), Operand(1u),
                         bld.vop3(aco_opcode::v_bcnt_u32_b32, bld.def(v1), tmp, Operand(0u)));
-         cmp_def = bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(s2), Operand(0u), tmp).def(0);
+         cmp_def = bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(bld.lm), Operand(0u), tmp).def(0);
       }
       cmp_def.setHint(vcc);
       return cmp_def.getTemp();
@@ -5290,9 +5299,9 @@ Temp emit_boolean_exclusive_scan(isel_context *ctx, nir_op op, Temp src)
    //subgroupExclusiveXor(val) -> mbcnt(val & exec) & 1 != 0
    Temp tmp;
    if (op == nir_op_iand)
-      tmp = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.def(s1, scc), Operand(exec, s2), src);
+      tmp = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc), Operand(exec, bld.lm), src);
    else
-      tmp = bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), src, Operand(exec, s2));
+      tmp = bld.sop2(Builder::s_and, bld.def(s2), bld.def(s1, scc), src, Operand(exec, bld.lm));
 
    Builder::Result lohi = bld.pseudo(aco_opcode::p_split_vector, bld.def(s1), bld.def(s1), tmp);
    Temp lo = lohi.def(0).getTemp();
@@ -5301,11 +5310,11 @@ Temp emit_boolean_exclusive_scan(isel_context *ctx, nir_op op, Temp src)
 
    Definition cmp_def = Definition();
    if (op == nir_op_iand)
-      cmp_def = bld.vopc(aco_opcode::v_cmp_eq_u32, bld.def(s2), Operand(0u), mbcnt).def(0);
+      cmp_def = bld.vopc(aco_opcode::v_cmp_eq_u32, bld.def(bld.lm), Operand(0u), mbcnt).def(0);
    else if (op == nir_op_ior)
-      cmp_def = bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(s2), Operand(0u), mbcnt).def(0);
+      cmp_def = bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(bld.lm), Operand(0u), mbcnt).def(0);
    else if (op == nir_op_ixor)
-      cmp_def = bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(s2), Operand(0u),
+      cmp_def = bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(bld.lm), Operand(0u),
                          bld.vop2(aco_opcode::v_and_b32, bld.def(v1), Operand(1u), mbcnt)).def(0);
    cmp_def.setHint(vcc);
    return cmp_def.getTemp();
@@ -5320,11 +5329,11 @@ Temp emit_boolean_inclusive_scan(isel_context *ctx, nir_op op, Temp src)
    //subgroupInclusiveXor(val) -> subgroupExclusiveXor(val) ^^ val
    Temp tmp = emit_boolean_exclusive_scan(ctx, op, src);
    if (op == nir_op_iand)
-      return bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), tmp, src);
+      return bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), tmp, src);
    else if (op == nir_op_ior)
-      return bld.sop2(aco_opcode::s_or_b64, bld.def(s2), bld.def(s1, scc), tmp, src);
+      return bld.sop2(Builder::s_or, bld.def(bld.lm), bld.def(s1, scc), tmp, src);
    else if (op == nir_op_ixor)
-      return bld.sop2(aco_opcode::s_xor_b64, bld.def(s2), bld.def(s1, scc), tmp, src);
+      return bld.sop2(Builder::s_xor, bld.def(bld.lm), bld.def(s1, scc), tmp, src);
 
    assert(false);
    return Temp();
@@ -5453,7 +5462,7 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
          Temp pck0 = bld.tmp(v1);
          Temp carry = bld.vadd32(Definition(pck0), tmp0, addr, true).def(1).getTemp();
          tmp1 = as_vgpr(ctx, tmp1);
-         Temp pck1 = bld.vop2_e64(aco_opcode::v_addc_co_u32, bld.def(v1), bld.hint_vcc(bld.def(s2)), tmp1, Operand(0u), carry);
+         Temp pck1 = bld.vop2_e64(aco_opcode::v_addc_co_u32, bld.def(v1), bld.hint_vcc(bld.def(bld.lm)), tmp1, Operand(0u), carry);
          addr = bld.pseudo(aco_opcode::p_create_vector, bld.def(v2), pck0, pck1);
 
          /* sample_pos = flat_load_dwordx2 addr */
@@ -5685,11 +5694,12 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
       break;
    }
    case nir_intrinsic_ballot: {
-      Definition tmp = bld.def(s2);
       Temp src = get_ssa_temp(ctx, instr->src[0].ssa);
+      Temp dst = get_ssa_temp(ctx, &instr->dest.ssa);
+      Definition tmp = bld.def(dst.regClass());
       if (instr->src[0].ssa->bit_size == 1) {
-         assert(src.regClass() == s2);
-         bld.sop2(aco_opcode::s_and_b64, tmp, bld.def(s1, scc), Operand(exec, s2), src);
+         assert(src.regClass() == bld.lm);
+         bld.sop2(Builder::s_and, tmp, bld.def(s1, scc), Operand(exec, bld.lm), src);
       } else if (instr->src[0].ssa->bit_size == 32 && src.regClass() == v1) {
          bld.vopc(aco_opcode::v_cmp_lg_u32, tmp, Operand(0u), src);
       } else if (instr->src[0].ssa->bit_size == 64 && src.regClass() == v2) {
@@ -5699,7 +5709,7 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
          nir_print_instr(&instr->instr, stderr);
          fprintf(stderr, "\n");
       }
-      emit_wqm(ctx, tmp.getTemp(), get_ssa_temp(ctx, &instr->dest.ssa));
+      emit_wqm(ctx, tmp.getTemp(), dst);
       break;
    }
    case nir_intrinsic_shuffle:
@@ -5722,15 +5732,19 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
             bld.pseudo(aco_opcode::p_create_vector, Definition(dst), lo, hi);
             emit_split_vector(ctx, dst, 2);
          } else if (instr->dest.ssa.bit_size == 1 && tid.regClass() == s1) {
-            assert(src.regClass() == s2);
-            Temp tmp = bld.sopc(aco_opcode::s_bitcmp1_b64, bld.def(s1, scc), src, tid);
+            assert(src.regClass() == bld.lm);
+            Temp tmp = bld.sopc(Builder::s_bitcmp1, bld.def(s1, scc), src, tid);
             bool_to_vector_condition(ctx, emit_wqm(ctx, tmp), dst);
          } else if (instr->dest.ssa.bit_size == 1 && tid.regClass() == v1) {
-            assert(src.regClass() == s2);
-            Temp tmp = bld.vop3(aco_opcode::v_lshrrev_b64, bld.def(v2), tid, src);
+            assert(src.regClass() == bld.lm);
+            Temp tmp;
+            if (ctx->program->wave_size == 64)
+               tmp = bld.vop3(aco_opcode::v_lshrrev_b64, bld.def(v2), tid, src);
+            else
+               tmp = bld.vop2_e64(aco_opcode::v_lshrrev_b32, bld.def(v1), tid, src);
             tmp = emit_extract_vector(ctx, tmp, 0, v1);
             tmp = bld.vop2(aco_opcode::v_and_b32, bld.def(v1), Operand(1u), tmp);
-            emit_wqm(ctx, bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(s2), Operand(0u), tmp), dst);
+            emit_wqm(ctx, bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(bld.lm), Operand(0u), tmp), dst);
          } else {
             fprintf(stderr, "Unimplemented NIR instr bit size: ");
             nir_print_instr(&instr->instr, stderr);
@@ -5763,9 +5777,9 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
          bld.pseudo(aco_opcode::p_create_vector, Definition(dst), lo, hi);
          emit_split_vector(ctx, dst, 2);
       } else if (instr->dest.ssa.bit_size == 1) {
-         assert(src.regClass() == s2);
-         Temp tmp = bld.sopc(aco_opcode::s_bitcmp1_b64, bld.def(s1, scc), src,
-                             bld.sop1(aco_opcode::s_ff1_i32_b64, bld.def(s1), Operand(exec, s2)));
+         assert(src.regClass() == bld.lm);
+         Temp tmp = bld.sopc(Builder::s_bitcmp1, bld.def(s1, scc), src,
+                             bld.sop1(Builder::s_ff1_i32, bld.def(s1), Operand(exec, bld.lm)));
          bool_to_vector_condition(ctx, emit_wqm(ctx, tmp), dst);
       } else if (src.regClass() == s1) {
          bld.sop1(aco_opcode::s_mov_b32, Definition(dst), src);
@@ -5781,22 +5795,22 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
    case nir_intrinsic_vote_all: {
       Temp src = get_ssa_temp(ctx, instr->src[0].ssa);
       Temp dst = get_ssa_temp(ctx, &instr->dest.ssa);
-      assert(src.regClass() == s2);
-      assert(dst.regClass() == s2);
+      assert(src.regClass() == bld.lm);
+      assert(dst.regClass() == bld.lm);
 
-      Temp tmp = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.def(s1, scc), Operand(exec, s2), src).def(1).getTemp();
-      Temp val = bld.sop2(aco_opcode::s_cselect_b64, bld.def(s2), Operand(0u), Operand(-1u), bld.scc(tmp));
+      Temp tmp = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc), Operand(exec, bld.lm), src).def(1).getTemp();
+      Temp val = bld.sop2(Builder::s_cselect, bld.def(bld.lm), Operand(0u), Operand(-1u), bld.scc(tmp));
       emit_wqm(ctx, val, dst);
       break;
    }
    case nir_intrinsic_vote_any: {
       Temp src = get_ssa_temp(ctx, instr->src[0].ssa);
       Temp dst = get_ssa_temp(ctx, &instr->dest.ssa);
-      assert(src.regClass() == s2);
-      assert(dst.regClass() == s2);
+      assert(src.regClass() == bld.lm);
+      assert(dst.regClass() == bld.lm);
 
-      Temp tmp = bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), Operand(exec, s2), src).def(1).getTemp();
-      Temp val = bld.sop2(aco_opcode::s_cselect_b64, bld.def(s2), Operand(-1u), Operand(0u), bld.scc(tmp));
+      Temp tmp = bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), Operand(exec, bld.lm), src).def(1).getTemp();
+      Temp val = bld.sop2(Builder::s_cselect, bld.def(bld.lm), Operand(-1u), Operand(0u), bld.scc(tmp));
       emit_wqm(ctx, val, dst);
       break;
    }
@@ -5879,7 +5893,7 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
 
          Temp tmp_dst = bld.tmp(dst.regClass());
          reduce->definitions[0] = Definition(tmp_dst);
-         reduce->definitions[1] = bld.def(s2); // used internally
+         reduce->definitions[1] = bld.def(ctx->program->lane_mask); // used internally
          reduce->definitions[2] = Definition();
          reduce->definitions[3] = Definition(scc, s1);
          reduce->definitions[4] = Definition();
@@ -5899,13 +5913,14 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
          Temp dst = get_ssa_temp(ctx, &instr->dest.ssa);
          unsigned lane = nir_src_as_const_value(instr->src[1])->u32;
          if (instr->dest.ssa.bit_size == 1) {
-            assert(src.regClass() == s2);
+            assert(src.regClass() == bld.lm);
+            assert(dst.regClass() == bld.lm);
             uint32_t half_mask = 0x11111111u << lane;
             Temp mask_tmp = bld.pseudo(aco_opcode::p_create_vector, bld.def(s2), Operand(half_mask), Operand(half_mask));
-            Temp tmp = bld.tmp(s2);
-            bld.sop1(aco_opcode::s_wqm_b64, Definition(tmp),
-                     bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), mask_tmp,
-                              bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), src, Operand(exec, s2))));
+            Temp tmp = bld.tmp(bld.lm);
+            bld.sop1(Builder::s_wqm, Definition(tmp),
+                     bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), mask_tmp,
+                              bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), src, Operand(exec, bld.lm))));
             emit_wqm(ctx, tmp, dst);
          } else if (instr->dest.ssa.bit_size == 32) {
             emit_wqm(ctx,
@@ -5957,10 +5972,10 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
 
       Temp dst = get_ssa_temp(ctx, &instr->dest.ssa);
       if (instr->dest.ssa.bit_size == 1) {
-         assert(src.regClass() == s2);
+         assert(src.regClass() == bld.lm);
          src = bld.vop2_e64(aco_opcode::v_cndmask_b32, bld.def(v1), Operand(0u), Operand((uint32_t)-1), src);
          src = bld.vop1_dpp(aco_opcode::v_mov_b32, bld.def(v1), src, dpp_ctrl);
-         Temp tmp = bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(s2), Operand(0u), src);
+         Temp tmp = bld.vopc(aco_opcode::v_cmp_lg_u32, bld.def(bld.lm), Operand(0u), src);
          emit_wqm(ctx, tmp, dst);
       } else if (instr->dest.ssa.bit_size == 32) {
          Temp tmp = bld.vop1_dpp(aco_opcode::v_mov_b32, bld.def(v1), src, dpp_ctrl);
@@ -6060,15 +6075,15 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
       break;
    case nir_intrinsic_demote_if: {
       Temp src = get_ssa_temp(ctx, instr->src[0].ssa);
-      assert(src.regClass() == s2);
-      Temp cond = bld.sop2(aco_opcode::s_and_b64, bld.def(s2), bld.def(s1, scc), src, Operand(exec, s2));
+      assert(src.regClass() == bld.lm);
+      Temp cond = bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc), src, Operand(exec, bld.lm));
       bld.pseudo(aco_opcode::p_demote_to_helper, cond);
       ctx->block->kind |= block_kind_uses_demote;
       ctx->program->needs_exact = true;
       break;
    }
    case nir_intrinsic_first_invocation: {
-      emit_wqm(ctx, bld.sop1(aco_opcode::s_ff1_i32_b64, bld.def(s1), Operand(exec, s2)),
+      emit_wqm(ctx, bld.sop1(Builder::s_ff1_i32, bld.def(s1), Operand(exec, bld.lm)),
                get_ssa_temp(ctx, &instr->dest.ssa));
       break;
    }
@@ -6180,14 +6195,14 @@ void build_cube_select(isel_context *ctx, Temp ma, Temp id, Temp deriv,
    Operand two(0x40000000u);
    Operand four(0x40800000u);
 
-   Temp is_ma_positive = bld.vopc(aco_opcode::v_cmp_le_f32, bld.hint_vcc(bld.def(s2)), Operand(0u), ma);
+   Temp is_ma_positive = bld.vopc(aco_opcode::v_cmp_le_f32, bld.hint_vcc(bld.def(bld.lm)), Operand(0u), ma);
    Temp sgn_ma = bld.vop2_e64(aco_opcode::v_cndmask_b32, bld.def(v1), neg_one, one, is_ma_positive);
    Temp neg_sgn_ma = bld.vop2(aco_opcode::v_sub_f32, bld.def(v1), Operand(0u), sgn_ma);
 
-   Temp is_ma_z = bld.vopc(aco_opcode::v_cmp_le_f32, bld.hint_vcc(bld.def(s2)), four, id);
+   Temp is_ma_z = bld.vopc(aco_opcode::v_cmp_le_f32, bld.hint_vcc(bld.def(bld.lm)), four, id);
    Temp is_ma_y = bld.vopc(aco_opcode::v_cmp_le_f32, bld.def(s2), two, id);
-   is_ma_y = bld.sop2(aco_opcode::s_andn2_b64, bld.hint_vcc(bld.def(s2)), is_ma_y, is_ma_z);
-   Temp is_not_ma_x = bld.sop2(aco_opcode::s_or_b64, bld.hint_vcc(bld.def(s2)), bld.def(s1, scc), is_ma_z, is_ma_y);
+   is_ma_y = bld.sop2(Builder::s_andn2, bld.hint_vcc(bld.def(bld.lm)), is_ma_y, is_ma_z);
+   Temp is_not_ma_x = bld.sop2(aco_opcode::s_or_b64, bld.hint_vcc(bld.def(bld.lm)), bld.def(s1, scc), is_ma_z, is_ma_y);
 
    // select sc
    Temp tmp = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), deriv_z, deriv_x, is_not_ma_x);
@@ -6667,7 +6682,7 @@ void visit_tex(isel_context *ctx, nir_tex_instr *instr)
                             Operand((uint32_t)V_008F14_IMG_NUM_FORMAT_SINT),
                             bld.scc(compare_cube_wa));
          }
-         tg4_compare_cube_wa64 = bld.tmp(s2);
+         tg4_compare_cube_wa64 = bld.tmp(bld.lm);
          bool_to_vector_condition(ctx, compare_cube_wa, tg4_compare_cube_wa64);
 
          nfmt = bld.sop2(aco_opcode::s_lshl_b32, bld.def(s1), bld.def(s1, scc), nfmt, Operand(26u));
@@ -6800,7 +6815,7 @@ void visit_tex(isel_context *ctx, nir_tex_instr *instr)
          assert(dmask == 1 && dst.regClass() == v1);
          assert(dst.id() != tmp_dst.id());
 
-         Temp tmp = bld.tmp(s2);
+         Temp tmp = bld.tmp(bld.lm);
          bld.vopc(aco_opcode::v_cmp_eq_u32, Definition(tmp), Operand(0u), tmp_dst).def(0).setHint(vcc);
          bld.vop2_e64(aco_opcode::v_cndmask_b32, Definition(dst), Operand(0u), Operand((uint32_t)-1), tmp);
 
@@ -6921,7 +6936,7 @@ void visit_phi(isel_context *ctx, nir_phi_instr *instr)
 {
    aco_ptr<Pseudo_instruction> phi;
    Temp dst = get_ssa_temp(ctx, &instr->dest.ssa);
-   assert(instr->dest.ssa.bit_size != 1 || dst.regClass() == s2);
+   assert(instr->dest.ssa.bit_size != 1 || dst.regClass() == ctx->program->lane_mask);
 
    bool logical = !dst.is_linear() || ctx->divergent_vals[instr->dest.ssa.index];
    logical |= ctx->block->kind & block_kind_merge;
@@ -7295,7 +7310,7 @@ static void begin_divergent_if_then(isel_context *ctx, if_context *ic, Temp cond
    ctx->block->kind |= block_kind_branch;
 
    /* branch to linear then block */
-   assert(cond.regClass() == s2);
+   assert(cond.regClass() == ctx->program->lane_mask);
    aco_ptr<Pseudo_branch_instruction> branch;
    branch.reset(create_instruction<Pseudo_branch_instruction>(aco_opcode::p_cbranch_z, Format::PSEUDO_BRANCH, 1, 0));
    branch->operands[0] = Operand(cond);
@@ -7439,7 +7454,7 @@ static void visit_if(isel_context *ctx, nir_if *if_stmt)
       ctx->block->kind |= block_kind_uniform;
 
       /* emit branch */
-      assert(cond.regClass() == s2);
+      assert(cond.regClass() == bld.lm);
       // TODO: in a post-RA optimizer, we could check if the condition is in VCC and omit this instruction
       cond = bool_to_scalar_condition(ctx, cond);
 
@@ -7825,7 +7840,7 @@ void handle_bc_optimize(isel_context *ctx)
    ctx->persp_centroid = get_arg(ctx, ctx->args->ac.persp_centroid);
    ctx->linear_centroid = get_arg(ctx, ctx->args->ac.linear_centroid);
    if (uses_center && uses_centroid) {
-      Temp sel = bld.vopc_e64(aco_opcode::v_cmp_lt_i32, bld.hint_vcc(bld.def(s2)),
+      Temp sel = bld.vopc_e64(aco_opcode::v_cmp_lt_i32, bld.hint_vcc(bld.def(bld.lm)),
                               get_arg(ctx, ctx->args->ac.prim_mask), Operand(0u));
 
       if (G_0286CC_PERSP_CENTROID_ENA(spi_ps_input_ena)) {
@@ -7934,7 +7949,7 @@ void select_program(Program *program,
          Builder bld(ctx.program, ctx.block);
          Temp count = bld.sop2(aco_opcode::s_bfe_u32, bld.def(s1), bld.def(s1, scc), ctx.merged_wave_info, Operand((8u << 16) | (i * 8u)));
          Temp thread_id = emit_mbcnt(&ctx, bld.def(v1));
-         Temp cond = bld.vopc(aco_opcode::v_cmp_gt_u32, bld.hint_vcc(bld.def(s2)), count, thread_id);
+         Temp cond = bld.vopc(aco_opcode::v_cmp_gt_u32, bld.hint_vcc(bld.def(bld.lm)), count, thread_id);
 
          begin_divergent_if_then(&ctx, &ic, cond);
       }
diff --git a/src/amd/compiler/aco_instruction_selection_setup.cpp b/src/amd/compiler/aco_instruction_selection_setup.cpp
index ab96a4507cf..a7446c6c058 100644
--- a/src/amd/compiler/aco_instruction_selection_setup.cpp
+++ b/src/amd/compiler/aco_instruction_selection_setup.cpp
@@ -126,6 +126,7 @@ unsigned get_interp_input(nir_intrinsic_op intrin, enum glsl_interp_mode interp)
 void init_context(isel_context *ctx, nir_shader *shader)
 {
    nir_function_impl *impl = nir_shader_get_entrypoint(shader);
+   unsigned lane_mask_size = ctx->program->lane_mask.size();
 
    ctx->shader = shader;
    ctx->divergent_vals = nir_divergence_analysis(shader, nir_divergence_view_index_uniform);
@@ -207,7 +208,7 @@ void init_context(isel_context *ctx, nir_shader *shader)
                   case nir_op_ieq:
                   case nir_op_ine:
                   case nir_op_i2b1:
-                     size = 2;
+                     size = lane_mask_size;
                      break;
                   case nir_op_f2i64:
                   case nir_op_f2u64:
@@ -219,7 +220,7 @@ void init_context(isel_context *ctx, nir_shader *shader)
                      break;
                   case nir_op_bcsel:
                      if (alu_instr->dest.dest.ssa.bit_size == 1) {
-                        size = 2;
+                        size = lane_mask_size;
                      } else {
                         if (ctx->divergent_vals[alu_instr->dest.dest.ssa.index]) {
                            type = RegType::vgpr;
@@ -237,14 +238,14 @@ void init_context(isel_context *ctx, nir_shader *shader)
                      break;
                   case nir_op_mov:
                      if (alu_instr->dest.dest.ssa.bit_size == 1) {
-                        size = 2;
+                        size = lane_mask_size;
                      } else {
                         type = ctx->divergent_vals[alu_instr->dest.dest.ssa.index] ? RegType::vgpr : RegType::sgpr;
                      }
                      break;
                   default:
                      if (alu_instr->dest.dest.ssa.bit_size == 1) {
-                        size = 2;
+                        size = lane_mask_size;
                      } else {
                         for (unsigned i = 0; i < nir_op_infos[alu_instr->op].num_inputs; i++) {
                            if (allocated[alu_instr->src[i].src.ssa->index].type() == RegType::vgpr)
@@ -261,7 +262,7 @@ void init_context(isel_context *ctx, nir_shader *shader)
                if (nir_instr_as_load_const(instr)->def.bit_size == 64)
                   size *= 2;
                else if (nir_instr_as_load_const(instr)->def.bit_size == 1)
-                  size *= 2;
+                  size *= lane_mask_size;
                allocated[nir_instr_as_load_const(instr)->def.index] = Temp(0, RegClass(RegType::sgpr, size));
                break;
             }
@@ -289,11 +290,11 @@ void init_context(isel_context *ctx, nir_shader *shader)
                   case nir_intrinsic_first_invocation:
                      type = RegType::sgpr;
                      if (intrinsic->dest.ssa.bit_size == 1)
-                        size = 2;
+                        size = lane_mask_size;
                      break;
                   case nir_intrinsic_ballot:
                      type = RegType::sgpr;
-                     size = 2;
+                     size = lane_mask_size;
                      break;
                   case nir_intrinsic_load_sample_id:
                   case nir_intrinsic_load_sample_mask_in:
@@ -369,7 +370,7 @@ void init_context(isel_context *ctx, nir_shader *shader)
                   case nir_intrinsic_inclusive_scan:
                   case nir_intrinsic_exclusive_scan:
                      if (intrinsic->dest.ssa.bit_size == 1) {
-                        size = 2;
+                        size = lane_mask_size;
                         type = RegType::sgpr;
                      } else if (!ctx->divergent_vals[intrinsic->dest.ssa.index]) {
                         type = RegType::sgpr;
@@ -384,11 +385,11 @@ void init_context(isel_context *ctx, nir_shader *shader)
                   case nir_intrinsic_load_helper_invocation:
                   case nir_intrinsic_is_helper_invocation:
                      type = RegType::sgpr;
-                     size = 2;
+                     size = lane_mask_size;
                      break;
                   case nir_intrinsic_reduce:
                      if (intrinsic->dest.ssa.bit_size == 1) {
-                        size = 2;
+                        size = lane_mask_size;
                         type = RegType::sgpr;
                      } else if (nir_intrinsic_cluster_size(intrinsic) == 0 ||
                          !ctx->divergent_vals[intrinsic->dest.ssa.index]) {
@@ -489,7 +490,7 @@ void init_context(isel_context *ctx, nir_shader *shader)
                if (phi->dest.ssa.bit_size == 1) {
                   assert(size == 1 && "multiple components not yet supported on boolean phis.");
                   type = RegType::sgpr;
-                  size *= 2;
+                  size *= lane_mask_size;
                   allocated[phi->dest.ssa.index] = Temp(0, RegClass(type, size));
                   break;
                }
@@ -590,7 +591,7 @@ Pseudo_instruction *add_startpgm(struct isel_context *ctx)
       startpgm->definitions[arg].setFixed(PhysReg{file == AC_ARG_SGPR ? reg : reg + 256});
       arg++;
    }
-   startpgm->definitions[arg_count] = Definition{ctx->program->allocateId(), exec, s2};
+   startpgm->definitions[arg_count] = Definition{ctx->program->allocateId(), exec, ctx->program->lane_mask};
    Pseudo_instruction *instr = startpgm.get();
    ctx->block->instructions.push_back(std::move(startpgm));
 
@@ -796,6 +797,7 @@ setup_isel_context(Program* program,
    program->chip_class = args->options->chip_class;
    program->family = args->options->family;
    program->wave_size = args->shader_info->wave_size;
+   program->lane_mask = program->wave_size == 32 ? s1 : s2;
 
    program->lds_alloc_granule = args->options->chip_class >= GFX7 ? 512 : 256;
    program->lds_limit = args->options->chip_class >= GFX7 ? 65536 : 32768;
diff --git a/src/amd/compiler/aco_ir.h b/src/amd/compiler/aco_ir.h
index 4073086662a..1f4721f5ffd 100644
--- a/src/amd/compiler/aco_ir.h
+++ b/src/amd/compiler/aco_ir.h
@@ -1149,6 +1149,7 @@ public:
    enum chip_class chip_class;
    enum radeon_family family;
    unsigned wave_size;
+   RegClass lane_mask;
    Stage stage; /* Stage */
    bool needs_exact = false; /* there exists an instruction with disable_wqm = true */
    bool needs_wqm = false; /* there exists a p_wqm instruction */
diff --git a/src/amd/compiler/aco_live_var_analysis.cpp b/src/amd/compiler/aco_live_var_analysis.cpp
index 05ddb7bc68a..4255d56173b 100644
--- a/src/amd/compiler/aco_live_var_analysis.cpp
+++ b/src/amd/compiler/aco_live_var_analysis.cpp
@@ -54,7 +54,7 @@ void process_live_temps_per_block(Program *program, live& lives, Block* block,
    bool exec_live = false;
    if (block->live_out_exec != Temp()) {
       live_sgprs.insert(block->live_out_exec);
-      new_demand.sgpr += 2;
+      new_demand.sgpr += program->lane_mask.size();
       exec_live = true;
    }
 
@@ -77,10 +77,10 @@ void process_live_temps_per_block(Program *program, live& lives, Block* block,
       if (is_phi(insn))
          break;
 
-      /* substract the 2 sgprs from exec */
+      /* substract the 1 or 2 sgprs from exec */
       if (exec_live)
-         assert(new_demand.sgpr >= 2);
-      register_demand[idx] = RegisterDemand(new_demand.vgpr, new_demand.sgpr - (exec_live ? 2 : 0));
+         assert(new_demand.sgpr >= (int16_t) program->lane_mask.size());
+      register_demand[idx] = RegisterDemand(new_demand.vgpr, new_demand.sgpr - (exec_live ? program->lane_mask.size() : 0));
 
       /* KILL */
       for (Definition& definition : insn->definitions) {
@@ -144,8 +144,8 @@ void process_live_temps_per_block(Program *program, live& lives, Block* block,
 
    /* update block's register demand for a last time */
    if (exec_live)
-      assert(new_demand.sgpr >= 2);
-   new_demand.sgpr -= exec_live ? 2 : 0;
+      assert(new_demand.sgpr >= (int16_t) program->lane_mask.size());
+   new_demand.sgpr -= exec_live ? program->lane_mask.size() : 0;
    block->register_demand.update(new_demand);
 
    /* handle phi definitions */
diff --git a/src/amd/compiler/aco_lower_bool_phis.cpp b/src/amd/compiler/aco_lower_bool_phis.cpp
index dc64f0133b5..988f753c82d 100644
--- a/src/amd/compiler/aco_lower_bool_phis.cpp
+++ b/src/amd/compiler/aco_lower_bool_phis.cpp
@@ -54,12 +54,12 @@ Operand get_ssa(Program *program, unsigned block_idx, ssa_state *state)
    while (true) {
       auto pos = state->latest.find(block_idx);
       if (pos != state->latest.end())
-         return Operand({pos->second, s2});
+         return Operand({pos->second, program->lane_mask});
 
       Block& block = program->blocks[block_idx];
       size_t pred = block.linear_preds.size();
       if (pred == 0) {
-         return Operand(s2);
+         return Operand(program->lane_mask);
       } else if (pred == 1) {
          block_idx = block.linear_preds[0];
          continue;
@@ -75,10 +75,10 @@ Operand get_ssa(Program *program, unsigned block_idx, ssa_state *state)
                state->phis[phi->operands[i].tempId()][(phi_use){&block, res}] |= (uint64_t)1 << i;
             }
          }
-         phi->definitions[0] = Definition(Temp{res, s2});
+         phi->definitions[0] = Definition(Temp{res, program->lane_mask});
          block.instructions.emplace(block.instructions.begin(), std::move(phi));
 
-         return Operand({res, s2});
+         return Operand({res, program->lane_mask});
       }
    }
 }
@@ -118,7 +118,7 @@ Temp write_ssa(Program *program, Block *block, ssa_state *state, unsigned previo
          update_phi(program, state, phi.first.block, phi.first.phi_def, phi.second);
    }
 
-   return {id, s2};
+   return {id, program->lane_mask};
 }
 
 void insert_before_logical_end(Block *block, aco_ptr<Instruction> instr)
@@ -150,23 +150,25 @@ void lower_divergent_bool_phi(Program *program, Block *block, aco_ptr<Instructio
 
       assert(phi->operands[i].isTemp());
       Temp phi_src = phi->operands[i].getTemp();
-      assert(phi_src.regClass() == s2);
+      assert(phi_src.regClass() == bld.lm);
 
       Operand cur = get_ssa(program, pred->index, &state);
+      assert(cur.regClass() == bld.lm);
       Temp new_cur = write_ssa(program, pred, &state, cur.isTemp() ? cur.tempId() : 0);
+      assert(new_cur.regClass() == bld.lm);
 
       if (cur.isUndefined()) {
          insert_before_logical_end(pred, bld.sop1(aco_opcode::s_mov_b64, Definition(new_cur), phi_src).get_ptr());
       } else {
-         Temp tmp1 = bld.tmp(s2), tmp2 = bld.tmp(s2);
+         Temp tmp1 = bld.tmp(bld.lm), tmp2 = bld.tmp(bld.lm);
          insert_before_logical_end(pred,
-            bld.sop2(aco_opcode::s_andn2_b64, Definition(tmp1), bld.def(s1, scc),
-                     cur, Operand(exec, s2)).get_ptr());
+            bld.sop2(Builder::s_andn2, Definition(tmp1), bld.def(s1, scc),
+                     cur, Operand(exec, bld.lm)).get_ptr());
          insert_before_logical_end(pred,
-            bld.sop2(aco_opcode::s_and_b64, Definition(tmp2), bld.def(s1, scc),
-                     phi_src, Operand(exec, s2)).get_ptr());
+            bld.sop2(Builder::s_and, Definition(tmp2), bld.def(s1, scc),
+                     phi_src, Operand(exec, bld.lm)).get_ptr());
          insert_before_logical_end(pred,
-            bld.sop2(aco_opcode::s_or_b64, Definition(new_cur), bld.def(s1, scc),
+            bld.sop2(Builder::s_or, Definition(new_cur), bld.def(s1, scc),
                      tmp1, tmp2).get_ptr());
       }
    }
@@ -192,8 +194,8 @@ void lower_bool_phis(Program* program)
    for (Block& block : program->blocks) {
       for (aco_ptr<Instruction>& phi : block.instructions) {
          if (phi->opcode == aco_opcode::p_phi) {
-            assert(phi->definitions[0].regClass() != s1);
-            if (phi->definitions[0].regClass() == s2)
+            assert(program->wave_size == 64 ? phi->definitions[0].regClass() != s1 : phi->definitions[0].regClass() != s2);
+            if (phi->definitions[0].regClass() == program->lane_mask)
                lower_divergent_bool_phi(program, &block, phi);
          } else if (!is_phi(phi)) {
             break;
diff --git a/src/amd/compiler/aco_lower_to_hw_instr.cpp b/src/amd/compiler/aco_lower_to_hw_instr.cpp
index cbb3b55179c..e9c2d66d823 100644
--- a/src/amd/compiler/aco_lower_to_hw_instr.cpp
+++ b/src/amd/compiler/aco_lower_to_hw_instr.cpp
@@ -76,8 +76,10 @@ aco_opcode get_reduce_opcode(chip_class chip, ReduceOp op) {
 void emit_vadd32(Builder& bld, Definition def, Operand src0, Operand src1)
 {
    Instruction *instr = bld.vadd32(def, src0, src1, false, Operand(s2), true);
-   if (instr->definitions.size() >= 2)
+   if (instr->definitions.size() >= 2) {
+      assert(instr->definitions[1].regClass() == bld.lm);
       instr->definitions[1].setFixed(vcc);
+   }
 }
 
 void emit_int64_dpp_op(lower_context *ctx, PhysReg dst_reg, PhysReg src0_reg, PhysReg src1_reg,
@@ -99,12 +101,12 @@ void emit_int64_dpp_op(lower_context *ctx, PhysReg dst_reg, PhysReg src0_reg, Ph
             bld.vop1(aco_opcode::v_mov_b32, vtmp_def[0], identity[0]);
          bld.vop1_dpp(aco_opcode::v_mov_b32, vtmp_def[0], src0[0],
                       dpp_ctrl, row_mask, bank_mask, bound_ctrl);
-         bld.vop3(aco_opcode::v_add_co_u32_e64, dst[0], bld.def(s2, vcc), vtmp_op[0], src1[0]);
+         bld.vop3(aco_opcode::v_add_co_u32_e64, dst[0], bld.def(bld.lm, vcc), vtmp_op[0], src1[0]);
       } else {
-         bld.vop2_dpp(aco_opcode::v_add_co_u32, dst[0], bld.def(s2, vcc), src0[0], src1[0],
+         bld.vop2_dpp(aco_opcode::v_add_co_u32, dst[0], bld.def(bld.lm, vcc), src0[0], src1[0],
                       dpp_ctrl, row_mask, bank_mask, bound_ctrl);
       }
-      bld.vop2_dpp(aco_opcode::v_addc_co_u32, dst[1], bld.def(s2, vcc), src0[1], src1[1], Operand(vcc, s2),
+      bld.vop2_dpp(aco_opcode::v_addc_co_u32, dst[1], bld.def(bld.lm, vcc), src0[1], src1[1], Operand(vcc, bld.lm),
                    dpp_ctrl, row_mask, bank_mask, bound_ctrl);
    } else if (op == iand64) {
       bld.vop2_dpp(aco_opcode::v_and_b32, dst[0], src0[0], src1[0],
@@ -149,9 +151,9 @@ void emit_int64_dpp_op(lower_context *ctx, PhysReg dst_reg, PhysReg src0_reg, Ph
       bld.vop1_dpp(aco_opcode::v_mov_b32, vtmp_def[1], src0[1],
                    dpp_ctrl, row_mask, bank_mask, bound_ctrl);
 
-      bld.vopc(cmp, bld.def(s2, vcc), vtmp_op64, src1_64);
-      bld.vop2(aco_opcode::v_cndmask_b32, dst[0], vtmp_op[0], src1[0], Operand(vcc, s2));
-      bld.vop2(aco_opcode::v_cndmask_b32, dst[1], vtmp_op[1], src1[1], Operand(vcc, s2));
+      bld.vopc(cmp, bld.def(bld.lm, vcc), vtmp_op64, src1_64);
+      bld.vop2(aco_opcode::v_cndmask_b32, dst[0], vtmp_op[0], src1[0], Operand(vcc, bld.lm));
+      bld.vop2(aco_opcode::v_cndmask_b32, dst[1], vtmp_op[1], src1[1], Operand(vcc, bld.lm));
    } else if (op == imul64) {
       /* t4 = dpp(x_hi)
        * t1 = umul_lo(t4, y_lo)
@@ -216,11 +218,11 @@ void emit_int64_op(lower_context *ctx, PhysReg dst_reg, PhysReg src0_reg, PhysRe
 
    if (op == iadd64) {
       if (ctx->program->chip_class >= GFX10) {
-         bld.vop3(aco_opcode::v_add_co_u32_e64, dst[0], bld.def(s2, vcc), src0[0], src1[0]);
+         bld.vop3(aco_opcode::v_add_co_u32_e64, dst[0], bld.def(bld.lm, vcc), src0[0], src1[0]);
       } else {
-         bld.vop2(aco_opcode::v_add_co_u32, dst[0], bld.def(s2, vcc), src0[0], src1[0]);
+         bld.vop2(aco_opcode::v_add_co_u32, dst[0], bld.def(bld.lm, vcc), src0[0], src1[0]);
       }
-      bld.vop2(aco_opcode::v_addc_co_u32, dst[1], bld.def(s2, vcc), src0[1], src1[1], Operand(vcc, s2));
+      bld.vop2(aco_opcode::v_addc_co_u32, dst[1], bld.def(bld.lm, vcc), src0[1], src1[1], Operand(vcc, bld.lm));
    } else if (op == iand64) {
       bld.vop2(aco_opcode::v_and_b32, dst[0], src0[0], src1[0]);
       bld.vop2(aco_opcode::v_and_b32, dst[1], src0[1], src1[1]);
@@ -249,9 +251,9 @@ void emit_int64_op(lower_context *ctx, PhysReg dst_reg, PhysReg src0_reg, PhysRe
          break;
       }
 
-      bld.vopc(cmp, bld.def(s2, vcc), src0_64, src1_64);
-      bld.vop2(aco_opcode::v_cndmask_b32, dst[0], src0[0], src1[0], Operand(vcc, s2));
-      bld.vop2(aco_opcode::v_cndmask_b32, dst[1], src0[1], src1[1], Operand(vcc, s2));
+      bld.vopc(cmp, bld.def(bld.lm, vcc), src0_64, src1_64);
+      bld.vop2(aco_opcode::v_cndmask_b32, dst[0], src0[0], src1[0], Operand(vcc, bld.lm));
+      bld.vop2(aco_opcode::v_cndmask_b32, dst[1], src0[1], src1[1], Operand(vcc, bld.lm));
    } else if (op == imul64) {
       if (src1_reg == dst_reg) {
          /* it's fine if src0==dst but not if src1==dst */
@@ -298,7 +300,7 @@ void emit_dpp_op(lower_context *ctx, PhysReg dst_reg, PhysReg src0_reg, PhysReg
 
    if (!vop3) {
       if (opcode == aco_opcode::v_add_co_u32)
-         bld.vop2_dpp(opcode, dst, bld.def(s2, vcc), src0, src1, dpp_ctrl, row_mask, bank_mask, bound_ctrl);
+         bld.vop2_dpp(opcode, dst, bld.def(bld.lm, vcc), src0, src1, dpp_ctrl, row_mask, bank_mask, bound_ctrl);
       else
          bld.vop2_dpp(opcode, dst, src0, src1, dpp_ctrl, row_mask, bank_mask, bound_ctrl);
       return;
@@ -342,7 +344,7 @@ void emit_op(lower_context *ctx, PhysReg dst_reg, PhysReg src0_reg, PhysReg src1
    if (vop3) {
       bld.vop3(opcode, dst, src0, src1);
    } else if (opcode == aco_opcode::v_add_co_u32) {
-      bld.vop2(opcode, dst, bld.def(s2, vcc), src0, src1);
+      bld.vop2(opcode, dst, bld.def(bld.lm, vcc), src0, src1);
    } else {
       bld.vop2(opcode, dst, src0, src1);
    }
@@ -420,7 +422,7 @@ void emit_reduction(lower_context *ctx, aco_opcode op, ReduceOp reduce_op, unsig
    Operand vcndmask_identity[2] = {identity[0], identity[1]};
 
    /* First, copy the source to tmp and set inactive lanes to the identity */
-   bld.sop1(aco_opcode::s_or_saveexec_b64, Definition(stmp, s2), Definition(scc, s1), Definition(exec, s2), Operand(UINT64_MAX), Operand(exec, s2));
+   bld.sop1(Builder::s_or_saveexec, Definition(stmp, bld.lm), Definition(scc, s1), Definition(exec, bld.lm), Operand(UINT64_MAX), Operand(exec, bld.lm));
 
    for (unsigned i = 0; i < src.size(); i++) {
       /* p_exclusive_scan needs it to be a sgpr or inline constant for the v_writelane_b32
@@ -440,7 +442,7 @@ void emit_reduction(lower_context *ctx, aco_opcode op, ReduceOp reduce_op, unsig
    for (unsigned i = 0; i < src.size(); i++) {
       bld.vop2_e64(aco_opcode::v_cndmask_b32, Definition(PhysReg{tmp + i}, v1),
                    vcndmask_identity[i], Operand(PhysReg{src.physReg() + i}, v1),
-                   Operand(stmp, s2));
+                   Operand(stmp, bld.lm));
    }
 
    bool exec_restored = false;
@@ -463,7 +465,7 @@ void emit_reduction(lower_context *ctx, aco_opcode op, ReduceOp reduce_op, unsig
       if (cluster_size == 32) {
          for (unsigned i = 0; i < src.size(); i++)
             bld.ds(aco_opcode::ds_swizzle_b32, Definition(PhysReg{vtmp+i}, v1), Operand(PhysReg{tmp+i}, s1), ds_pattern_bitmode(0x1f, 0, 0x10));
-         bld.sop1(aco_opcode::s_mov_b64, Definition(exec, s2), Operand(stmp, s2));
+         bld.sop1(Builder::s_mov, Definition(exec, bld.lm), Operand(stmp, bld.lm));
          exec_restored = true;
          emit_op(ctx, dst.physReg(), vtmp, tmp, PhysReg{0}, reduce_op, src.size());
          dst_written = true;
@@ -500,7 +502,7 @@ void emit_reduction(lower_context *ctx, aco_opcode op, ReduceOp reduce_op, unsig
                                          Operand(0xffffffffu), Operand(0xffffffffu)).instr;
             static_cast<VOP3A_instruction*>(perm)->opsel[0] = true; /* FI (Fetch Inactive) */
          }
-         bld.sop1(aco_opcode::s_mov_b64, Definition(exec, s2), Operand(UINT64_MAX));
+         bld.sop1(Builder::s_mov, Definition(exec, bld.lm), Operand(UINT64_MAX));
 
          /* fill in the gap in row 2 */
          for (unsigned i = 0; i < src.size(); i++) {
@@ -559,7 +561,7 @@ void emit_reduction(lower_context *ctx, aco_opcode op, ReduceOp reduce_op, unsig
    }
 
    if (!exec_restored)
-      bld.sop1(aco_opcode::s_mov_b64, Definition(exec, s2), Operand(stmp, s2));
+      bld.sop1(Builder::s_mov, Definition(exec, bld.lm), Operand(stmp, bld.lm));
 
    if (op == aco_opcode::p_reduce && cluster_size == 64) {
       for (unsigned k = 0; k < src.size(); k++) {
diff --git a/src/amd/compiler/aco_reduce_assign.cpp b/src/amd/compiler/aco_reduce_assign.cpp
index 28a779580a2..68a0dc15761 100644
--- a/src/amd/compiler/aco_reduce_assign.cpp
+++ b/src/amd/compiler/aco_reduce_assign.cpp
@@ -172,7 +172,7 @@ void setup_reduce_temp(Program* program)
             clobber_vcc = true;
 
          if (clobber_vcc)
-            instr->definitions[4] = Definition(vcc, s2);
+            instr->definitions[4] = Definition(vcc, bld.lm);
       }
    }
 }
diff --git a/src/amd/compiler/aco_register_allocation.cpp b/src/amd/compiler/aco_register_allocation.cpp
index c4144cc42f0..504ad015746 100644
--- a/src/amd/compiler/aco_register_allocation.cpp
+++ b/src/amd/compiler/aco_register_allocation.cpp
@@ -1719,6 +1719,7 @@ void register_allocation(Program *program, std::vector<std::set<Temp>> live_out_
 
                pc->operands[i] = parallelcopy[i].first;
                pc->definitions[i] = parallelcopy[i].second;
+               assert(pc->operands[i].size() == pc->definitions[i].size());
 
                /* it might happen that the operand is already renamed. we have to restore the original name. */
                std::map<unsigned, Temp>::iterator it = ctx.orig_names.find(pc->operands[i].tempId());
diff --git a/src/amd/compiler/aco_ssa_elimination.cpp b/src/amd/compiler/aco_ssa_elimination.cpp
index 3d76dcd8867..54e691ba476 100644
--- a/src/amd/compiler/aco_ssa_elimination.cpp
+++ b/src/amd/compiler/aco_ssa_elimination.cpp
@@ -58,6 +58,7 @@ void collect_phi_info(ssa_elimination_ctx& ctx)
             std::vector<unsigned>& preds = phi->opcode == aco_opcode::p_phi ? block.logical_preds : block.linear_preds;
             phi_info& info = phi->opcode == aco_opcode::p_phi ? ctx.logical_phi_info : ctx.linear_phi_info;
             const auto result = info.emplace(preds[i], std::vector<std::pair<Definition, Operand>>());
+            assert(phi->definitions[0].size() == phi->operands[i].size());
             result.first->second.emplace_back(phi->definitions[0], phi->operands[i]);
             ctx.empty_blocks[preds[i]] = false;
          }
diff --git a/src/amd/compiler/aco_validate.cpp b/src/amd/compiler/aco_validate.cpp
index 8d2bf8449db..8282d7e27e3 100644
--- a/src/amd/compiler/aco_validate.cpp
+++ b/src/amd/compiler/aco_validate.cpp
@@ -190,7 +190,7 @@ void validate(Program* program, FILE * output)
                }
             } else if (instr->opcode == aco_opcode::p_phi) {
                check(instr->operands.size() == block.logical_preds.size(), "Number of Operands does not match number of predecessors", instr.get());
-               check(instr->definitions[0].getTemp().type() == RegType::vgpr || instr->definitions[0].getTemp().regClass() == s2, "Logical Phi Definition must be vgpr or divergent boolean", instr.get());
+               check(instr->definitions[0].getTemp().type() == RegType::vgpr || instr->definitions[0].getTemp().regClass() == program->lane_mask, "Logical Phi Definition must be vgpr or divergent boolean", instr.get());
             } else if (instr->opcode == aco_opcode::p_linear_phi) {
                for (const Operand& op : instr->operands)
                   check(!op.isTemp() || op.getTemp().is_linear(), "Wrong Operand type", instr.get());