Merge pull request #2811 from colinleroy/OptTosLoadPop

Optimize jsr ldax0sp/incsp2

src/cc65/codeent.h

@@ -244,6 +244,12 @@ static inline int CE_IsCallTo (const CodeEntry* E, const char* Name)
     return (E->OPC == OP65_JSR && strcmp (E->Arg, Name) == 0);
 }
 
+static inline int CE_IsJumpTo (const CodeEntry* E, const char* Name)
+/* Check if this is a jump to the given function */
+{
+    return (E->OPC == OP65_JMP && strcmp (E->Arg, Name) == 0);
+}
+
 int CE_UseLoadFlags (CodeEntry* E);
 /* Return true if the instruction uses any flags that are set by a load of
 ** a register (N and Z).

src/cc65/codeopt.c

@@ -217,6 +217,8 @@ 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 DOptTosPushPop   = { OptTosPushPop,   "OptTosPushPop",    33, 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 };
@@ -343,6 +345,8 @@ static OptFunc* OptFuncs[] = {
     &DOptSub3,
     &DOptTest1,
     &DOptTest2,
+    &DOptTosLoadPop,
+    &DOptTosPushPop,
     &DOptTransfers1,
     &DOptTransfers2,
     &DOptTransfers3,
@@ -635,6 +639,7 @@ 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 +925,8 @@ 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);
+    C += RunOptFunc (S, &DOptTosPushPop, 5);
 
     Changes += C;
     /* If we had changes, we must run dead code elimination again,

src/cc65/coptmisc.c

@@ -486,14 +486,14 @@ static unsigned OptIncDecOps (CodeSeg* S, const char* dec, const char* inc, cons
 
 unsigned OptStackPtrOps (CodeSeg* S)
 {
-    return OptIncDecOps(S, "decsp", "incsp", "subysp", "addysp");
+    return OptIncDecOps (S, "decsp", "incsp", "subysp", "addysp");
 }
 
 
 
 unsigned OptAXOps (CodeSeg* S)
 {
-    return OptIncDecOps(S, "decax", "incax", "decaxy", "incaxy");
+    return OptIncDecOps (S, "decax", "incax", "decaxy", "incaxy");
 }
 
 
@@ -557,6 +557,52 @@ 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 (CE_IsCallTo (E, "ldax0sp")                               &&
+            (N = CS_GetNextEntry (S, I)) != 0                        &&
+            (CE_IsCallTo (N, "incsp2") || CE_IsJumpTo (N, "incsp2")) &&
+            !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);
+
+            /* 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 */
 {
@@ -1150,3 +1196,49 @@ unsigned OptBinOps2 (CodeSeg* S)
     /* Return the number of changes made */
     return Changes;
 }
+
+
+unsigned OptTosPushPop (CodeSeg* S)
+/* Merge jsr pushax/j?? 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 decspn, incspn, subysp or addysp */
+        if (CE_IsCallTo (E, "pushax")                              &&
+            (N = CS_GetNextEntry (S, I)) != 0                      &&
+            (CE_IsCallTo (N, "popax") || CE_IsJumpTo (N, "popax")) &&
+            !CE_HasLabel (N)) {
+
+            /* Insert an rts if jmp popax */
+            if (N->OPC == OP65_JMP) {
+              CodeEntry* X = NewCodeEntry (OP65_RTS, AM65_IMP, 0, 0, E->LI);
+              CS_InsertEntry (S, X, I);
+            }
+
+            /* Delete the old code */
+            CS_DelEntries (S, I+1, 2);
+
+            /* Remember we had changes */
+            ++Changes;
+
+        } else {
+
+            /* Next entry */
+            ++I;
+        }
+
+    }
+
+    /* Return the number of changes made */
+    return Changes;
+}

src/cc65/coptmisc.h

@@ -131,7 +131,11 @@ unsigned OptBinOps2 (CodeSeg* S);
 ** by something simpler.
 */
 
+unsigned OptTosLoadPop (CodeSeg* S);
+/* Merge jsr ldax0sp / jsr|jmp incsp2 into jsr|jmp popax */
 
+unsigned OptTosPushPop (CodeSeg* S);
+/* Merge jsr pushax/j?? popax */
 
 /* End of coptmisc.h */