Optimize jsr ldax0sp/incsp2

src/cc65/codeopt.c

@@ -217,6 +217,7 @@ static OptFunc DOptSub2         = { OptSub2,         "OptSub2",         100, 0,
 static OptFunc DOptSub3         = { OptSub3,         "OptSub3",         100, 0, 0, 0, 0, 0 };
 static OptFunc DOptTest1        = { OptTest1,        "OptTest1",         65, 0, 0, 0, 0, 0 };
 static OptFunc DOptTest2        = { OptTest2,        "OptTest2",         50, 0, 0, 0, 0, 0 };
+static OptFunc DOptTosLoadPop   = { OptTosLoadPop,   "OptTosLoadPop",    50, 0, 0, 0, 0, 0 };
 static OptFunc DOptTransfers1   = { OptTransfers1,   "OptTransfers1",     0, 0, 0, 0, 0, 0 };
 static OptFunc DOptTransfers2   = { OptTransfers2,   "OptTransfers2",    60, 0, 0, 0, 0, 0 };
 static OptFunc DOptTransfers3   = { OptTransfers3,   "OptTransfers3",    65, 0, 0, 0, 0, 0 };
@@ -347,6 +348,7 @@ static OptFunc* OptFuncs[] = {
     &DOptTransfers2,
     &DOptTransfers3,
     &DOptTransfers4,
+    &DOptTosLoadPop,
     &DOptUnusedLoads,
     &DOptUnusedStores,
 /* END SORTED_CODEOPT.SH */
@@ -635,6 +637,7 @@ static unsigned RunOptGroup1 (CodeSeg* S)
 
     Changes += RunOptFunc (S, &DOptGotoSPAdj, 1);
     Changes += RunOptFunc (S, &DOptStackPtrOps, 5);
+    Changes += RunOptFunc (S, &DOptTosLoadPop, 5);
     Changes += RunOptFunc (S, &DOptAXOps, 5);
     Changes += RunOptFunc (S, &DOptAdd3, 1);    /* Before OptPtrLoad5! */
     Changes += RunOptFunc (S, &DOptPtrStore1, 1);
@@ -920,6 +923,7 @@ static unsigned RunOptGroup7 (CodeSeg* S)
     C += RunOptFunc (S, &DOptStackPtrOps, 5);
     /* Re-optimize JSR/RTS that may now be grouped */
     C += RunOptFunc (S, &DOptRTS, 1);
+    C += RunOptFunc (S, &DOptTosLoadPop, 5);
 
     Changes += C;
     /* If we had changes, we must run dead code elimination again,

src/cc65/coptmisc.c

@@ -557,6 +557,57 @@ unsigned OptAXLoad (CodeSeg* S)
 
 
 
+unsigned OptTosLoadPop (CodeSeg* S)
+/* Merge jsr ldax0sp / jsr|jmp incsp2 into jsr|jmp popax */
+{
+    unsigned Changes = 0;
+    unsigned I;
+
+    /* Walk over the entries */
+    I = 0;
+    while (I < CS_GetEntryCount (S)) {
+
+        const CodeEntry* N;
+
+        /* Get the next entry */
+        const CodeEntry* E = CS_GetEntry (S, I);
+
+        /* Check for the sequence */
+        if (E->OPC == OP65_JSR                            &&
+            strncmp (E->Arg, "ldax0sp", 5) == 0           &&
+            (N = CS_GetNextEntry (S, I)) != 0             &&
+            (N->OPC == OP65_JSR || N->OPC == OP65_JMP)    &&
+            strcmp (N->Arg, "incsp2") == 0                &&
+            !CE_HasLabel (N)) {
+
+            CodeEntry* X;
+
+            X = NewCodeEntry (N->OPC, AM65_ABS, "popax", 0, N->LI);
+            CS_InsertEntry (S, X, I+2);
+
+            /* Delete the old code */
+            CS_DelEntries (S, I, 2);
+
+            /* Regenerate register info */
+            CS_GenRegInfo (S);
+
+            /* Remember we had changes */
+            ++Changes;
+
+        } else {
+
+            /* Next entry */
+            ++I;
+        }
+
+    }
+
+    /* Return the number of changes made */
+    return Changes;
+}
+
+
+
 unsigned OptAXLoad2 (CodeSeg* S)
 /* Merge ldy/jsr incaxy/jsr ldaxi into ldy/jsr ldaxidx */
 {

src/cc65/coptmisc.h

@@ -131,6 +131,9 @@ unsigned OptBinOps2 (CodeSeg* S);
 ** by something simpler.
 */
 
+unsigned OptTosLoadPop (CodeSeg* S);
+/* Merge jsr ldax0sp / jsr|jmp incsp2 into jsr|jmp popax */
+
 
 
 /* End of coptmisc.h */