aco: fix regclass checks when fixing to vcc/exec with Builder

[mesa.git] / src / amd / compiler / aco_builder_h.py

diff --git a/src/amd/compiler/aco_builder_h.py b/src/amd/compiler/aco_builder_h.py
index 0296653efdaab664fac5e222ac0d68e0dd899796..8d541cbd72f2a15a4d2856498728d8e8189f01e0 100644 (file)
--- a/src/amd/compiler/aco_builder_h.py
+++ b/src/amd/compiler/aco_builder_h.py
@@ -69,6 +69,13 @@ dpp_row_sr(unsigned amount)
     return (dpp_ctrl)(((unsigned) _dpp_row_sr) | amount);
 }
 
+inline dpp_ctrl
+dpp_row_rr(unsigned amount)
+{
+    assert(amount > 0 && amount < 16);
+    return (dpp_ctrl)(((unsigned) _dpp_row_rr) | amount);
+}
+
 inline unsigned
 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
 {
@@ -169,8 +176,9 @@ public:
    std::vector<aco_ptr<Instruction>> *instructions;
    std::vector<aco_ptr<Instruction>>::iterator it;
    bool is_precise = false;
+   bool is_nuw = false;
 
-   Builder(Program *pgm) : program(pgm), use_iterator(false), start(false), lm(pgm->lane_mask), instructions(NULL) {}
+   Builder(Program *pgm) : program(pgm), use_iterator(false), start(false), lm(pgm ? pgm->lane_mask : s2), 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) {}
 
@@ -180,6 +188,12 @@ public:
       return res;
    };
 
+   Builder nuw() const {
+      Builder res = *this;
+      res.is_nuw = true;
+      return res;
+   }
+
    void moveEnd(Block *block) {
       instructions = &block->instructions;
    }
@@ -303,7 +317,8 @@ public:
 % for fixed in ['m0', 'vcc', 'exec', 'scc']:
    Operand ${fixed}(Temp tmp) {
        % if fixed == 'vcc' or fixed == 'exec':
-          assert(tmp.regClass() == lm);
+          //vcc_hi and exec_hi can still be used in wave32
+          assert(tmp.type() == RegType::sgpr && tmp.bytes() <= 8);
        % endif
        Operand op(tmp);
        op.setFixed(aco::${fixed});
@@ -312,7 +327,8 @@ public:
 
    Definition ${fixed}(Definition def) {
        % if fixed == 'vcc' or fixed == 'exec':
-          assert(def.regClass() == lm);
+          //vcc_hi and exec_hi can still be used in wave32
+          assert(def.regClass().type() == RegType::sgpr && def.bytes() <= 8);
        % endif
        def.setFixed(aco::${fixed});
        return def;
@@ -320,7 +336,8 @@ public:
 
    Definition hint_${fixed}(Definition def) {
        % if fixed == 'vcc' or fixed == 'exec':
-          assert(def.regClass() == lm);
+          //vcc_hi and exec_hi can still be used in wave32
+          assert(def.regClass().type() == RegType::sgpr && def.bytes() <= 8);
        % endif
        def.setHint(aco::${fixed});
        return def;
@@ -545,8 +562,9 @@ formats = [("pseudo", [Format.PSEUDO], 'Pseudo_instruction', list(itertools.prod
            ("vopc_e64", [Format.VOPC, Format.VOP3A], 'VOP3A_instruction', itertools.product([1, 2], [2])),
            ("flat", [Format.FLAT], 'FLAT_instruction', [(0, 3), (1, 2)]),
            ("global", [Format.GLOBAL], 'FLAT_instruction', [(0, 3), (1, 2)])]
+formats = [(f if len(f) == 5 else f + ('',)) for f in formats]
 %>\\
-% for name, formats, struct, shapes in formats:
+% for name, formats, struct, shapes, extra_field_setup in formats:
     % for num_definitions, num_operands in shapes:
         <%
         args = ['aco_opcode opcode']
@@ -564,6 +582,7 @@ formats = [("pseudo", [Format.PSEUDO], 'Pseudo_instruction', list(itertools.prod
         % for i in range(num_definitions):
             instr->definitions[${i}] = def${i};
             instr->definitions[${i}].setPrecise(is_precise);
+            instr->definitions[${i}].setNUW(is_nuw);
         % endfor
         % for i in range(num_operands):
             instr->operands[${i}] = op${i}.op;
@@ -574,6 +593,7 @@ formats = [("pseudo", [Format.PSEUDO], 'Pseudo_instruction', list(itertools.prod
             % endfor
             ${f.get_builder_initialization(num_operands)}
         % endfor
+       ${extra_field_setup}
       return insert(instr);
    }