fix cg & tiny opt pass

build.c

@@ -35,6 +35,7 @@ struct CompileData target_lib_files[] = {
     SP(CT_C, "ir/opt/dce.c"),
     SP(CT_C, "ir/opt/join_compute.c"),
     SP(CT_C, "ir/opt/ll_jumps.c"),
+    SP(CT_C, "ir/opt/ll_binary.c"),
 
     DIR("build/ir/transform"),
     SP(CT_C, "ir/transform/single_assign_conditional.c"),

cg/x86_stupid/cg.c

@@ -452,18 +452,16 @@ static void emiti_binary(Location* a, Location* b, Location* o, const char * bin
 
     if (a != o) {
         if (b == o) {
-            Location* scratch = start_scratch_reg(a->bytesWidth, out);
+            Location* scratch = start_scratch_reg(size, out);
             emiti_move(a, scratch, false, out);
             emiti_binary(scratch, b, scratch, binary, out);
             emiti_move(scratch, o, false, out);
             end_scratch_reg(scratch, out);
-            return;
         } else {
             emiti_move(a, o, false, out);
             emiti_binary(o, b, o, binary, out);
-            return;
         }
-        assert(false);
+        return;
     }
 
     fprintf(out, "%s ", binary);
@@ -583,6 +581,23 @@ static void emit_condmove(vx_IrOp* op, const char *cc, FILE* file) {
     emiti_cmove(as_loc(out->bytesWidth, vthen), out, cc, file);
 }
 
+static size_t var_used_after_and_before_overwrite(vx_IrOp* after, vx_IrVar var) {
+    size_t i = 0;
+    for (vx_IrOp *op = after->next; op; op = op->next) {
+        for (size_t o = 0; o < op->outs_len; o ++)
+            if (op->outs[o].var == var)
+                break;
+
+        if (vx_IrOp_var_used(op, var))
+            i ++;
+    }
+    return i;
+}
+
+static bool ops_after_and_before_usage_or_overwrite(vx_IrOp* after, vx_IrVar var) {
+    return after->next && !vx_IrOp_var_used(after->next, var);
+}
+
 static vx_IrOp* emiti(vx_IrOp *prev, vx_IrOp* op, FILE* file) {
     switch (op->id) {
         case VX_IR_OP_REINTERPRET: // "val"
@@ -714,7 +729,7 @@ static vx_IrOp* emiti(vx_IrOp *prev, vx_IrOp* op, FILE* file) {
                 default: assert(false); break;
                 }
 
-                if (varData[ov].heat > 2) {
+                if (!(op->next && vx_IrOp_var_used(op->next, ov) && (op->next->id == VX_IR_OP_CMOV || op->next->id == VX_IR_OP_CONDTAILCALL || op->next->id == VX_LIR_COND))) {
                     emiti_storecond(o, cc, file);
                 }
 
@@ -999,7 +1014,7 @@ void vx_cg_x86stupid_gen(vx_IrBlock* block, FILE* out) {
                 break;
             }
         }
-        printf("var %zu: type %p , loc %s , size %zu\n", i, varData[i].type, str, loc ? loc->bytesWidth : 0);
+        printf("var %zu: type %p , loc %s , size %zu , heat %zu\n", i, varData[i].type, str, loc ? loc->bytesWidth : 0, varData[i].heat);
     }
 
     if (anyPlaced)
@@ -1024,7 +1039,7 @@ void vx_cg_x86stupid_gen(vx_IrBlock* block, FILE* out) {
 
 static Location* start_scratch_reg(size_t size, FILE* out) {
     assert(RegLut[SCRATCH_REG]->stored == NULL); // TODO 
-    Location* loc = gen_reg_var(8, SCRATCH_REG);
+    Location* loc = gen_reg_var(size, SCRATCH_REG);
     RegLut[SCRATCH_REG]->stored = loc;
     return loc;
 }

ir/llir.h

@@ -1,5 +1,7 @@
 #include "ir.h"
 
+OPT_PASS void vx_opt_ll_binary(vx_IrBlock *block);
+
 void vx_IrBlock_llir_lower(vx_IrBlock *block);
 
 void vx_IrBlock_llir_fix_decl(vx_IrBlock *root);
@@ -17,4 +19,5 @@ static void llir_prep_lower(vx_IrBlock *block) {
     vx_IrBlock_llir_compact(block);
     vx_IrBlock_lifetimes(block);
     vx_IrBlock_ll_share_slots(block);
+    vx_opt_ll_binary(block);
 }

ir/opt.h

@@ -22,10 +22,10 @@ OPT_PASS void vx_opt_reduce_loops(vx_IrBlock *block);
 OPT_PASS void vx_opt_cmov(vx_IrBlock *block);
 OPT_PASS void vx_opt_tailcall(vx_IrBlock *block);
 OPT_PASS void vx_opt_vars(vx_IrBlock *block);
+OPT_PASS void vx_opt_ll_jumps(vx_IrBlock *block);
 
 OPT_PASS void vx_opt_ll_dce(vx_IrBlock *block);
 OPT_PASS void vx_opt_ll_condtailcall(vx_IrBlock *block);
-OPT_PASS void vx_opt_ll_jumps(vx_IrBlock *block);
 
 void opt(vx_IrBlock *block);
 void opt_ll(vx_IrBlock *block);

ir/opt/ll_binary.c

@@ -0,0 +1,31 @@
+#include "../opt.h"
+
+OPT_PASS void vx_opt_ll_binary(vx_IrBlock *block) {
+    for (vx_IrOp* op = block->first; op; op = op->next) {
+        switch (op->id) {
+        case VX_IR_OP_EQ:
+        case VX_IR_OP_NEQ:
+        case VX_IR_OP_AND:
+        case VX_IR_OP_OR:
+        case VX_IR_OP_BITWISE_AND:
+        case VX_IR_OP_BITIWSE_OR:
+
+        case VX_IR_OP_ADD: // TODO: this might break programs ; add safeadd and safemul that don't have this problem
+        case VX_IR_OP_MUL:
+            {
+                vx_IrValue *a = vx_IrOp_param(op, VX_IR_NAME_OPERAND_A);
+                vx_IrValue *b = vx_IrOp_param(op, VX_IR_NAME_OPERAND_B);
+                vx_IrVar ov = op->outs[0].var;
+
+                if (b->type == VX_IR_VAL_VAR && b->var == ov) {
+                    vx_IrValue temp = *b;
+                    *b = *a;
+                    *a = temp;
+                }
+            } break;
+
+        default:
+            break;
+        }
+    }
+}