Optimize lda/sta/lda and friends

src/cc65/codeopt.c

@@ -197,6 +197,7 @@ static OptFunc DOptStore3       = { OptStore3,       "OptStore3",       120, 0,
 static OptFunc DOptStore4       = { OptStore4,       "OptStore4",        50, 0, 0, 0, 0, 0 };
 static OptFunc DOptStore5       = { OptStore5,       "OptStore5",       100, 0, 0, 0, 0, 0 };
 static OptFunc DOptStoreLoad    = { OptStoreLoad,    "OptStoreLoad",      0, 0, 0, 0, 0, 0 };
+static OptFunc DOptLoadStoreLoad= { OptLoadStoreLoad,"OptLoadStoreLoad",  0, 0, 0, 0, 0, 0 };
 static OptFunc DOptSub1         = { OptSub1,         "OptSub1",         100, 0, 0, 0, 0, 0 };
 static OptFunc DOptSub2         = { OptSub2,         "OptSub2",         100, 0, 0, 0, 0, 0 };
 static OptFunc DOptSub3         = { OptSub3,         "OptSub3",         100, 0, 0, 0, 0, 0 };
@@ -307,6 +308,7 @@ static OptFunc* OptFuncs[] = {
     &DOptStore4,
     &DOptStore5,
     &DOptStoreLoad,
+    &DOptLoadStoreLoad,
     &DOptSub1,
     &DOptSub2,
     &DOptSub3,
@@ -723,6 +725,7 @@ static unsigned RunOptGroup3 (CodeSeg* S)
         C += RunOptFunc (S, &DOptUnusedStores, 1);
         C += RunOptFunc (S, &DOptDupLoads, 1);
         C += RunOptFunc (S, &DOptStoreLoad, 1);
+        C += RunOptFunc (S, &DOptLoadStoreLoad, 1);
         C += RunOptFunc (S, &DOptTransfers1, 1);
         C += RunOptFunc (S, &DOptTransfers3, 1);
         C += RunOptFunc (S, &DOptTransfers4, 1);

src/cc65/coptind.c

@@ -581,6 +581,60 @@ unsigned OptStoreLoad (CodeSeg* S)
 
 
 
+unsigned OptLoadStoreLoad (CodeSeg* S)
+/* Search for the sequence
+**
+**      ld.     xx
+**      st.     yy
+**      ld.     xx
+**
+** and remove the useless load.
+*/
+{
+    unsigned Changes = 0;
+
+    /* Walk over the entries */
+    unsigned I = 0;
+    while (I < CS_GetEntryCount (S)) {
+
+        CodeEntry* L[3];
+
+        /* Get next entry */
+        L[0] = CS_GetEntry (S, I);
+
+        /* Check for the sequence */
+        if ((L[0]->OPC == OP65_LDA ||
+             L[0]->OPC == OP65_LDX ||
+             L[0]->OPC == OP65_LDY)                         &&
+            (L[0]->AM == AM65_ABS || L[0]->AM == AM65_ZP)   &&
+            !CS_RangeHasLabel (S, I+1, 3)                   &&
+            CS_GetEntries (S, L+1, I+1, 2)                  &&
+            (L[1]->OPC == OP65_STA ||
+             L[1]->OPC == OP65_STX ||
+             L[1]->OPC == OP65_STY)                         &&
+            L[2]->OPC == L[0]->OPC                          &&
+            L[2]->AM == L[0]->AM                            &&
+            strcmp (L[0]->Arg, L[2]->Arg) == 0) {
+
+            /* Remove the second load */
+            CS_DelEntries (S, I+2, 1);
+
+            /* Remember, we had changes */
+            ++Changes;
+
+        }
+
+        /* Next entry */
+        ++I;
+
+    }
+
+    /* Return the number of changes made */
+    return Changes;
+}
+
+
+
 unsigned OptTransfers1 (CodeSeg* S)
 /* Remove transfers from one register to another and back */
 {

src/cc65/coptind.h

@@ -64,6 +64,9 @@ unsigned OptDupLoads (CodeSeg* S);
 unsigned OptStoreLoad (CodeSeg* S);
 /* Remove a store followed by a load from the same location. */
 
+unsigned OptLoadStoreLoad (CodeSeg* S);
+/* Remove a load, store followed by a reload of the same location. */
+
 unsigned OptTransfers1 (CodeSeg* S);
 /* Remove transfers from one register to another and back */