+/*****************************************************************************/
+/* Optimize shifts */
+/*****************************************************************************/
+
+
+
+static unsigned OptShift1 (CodeSeg* S)
+/* A call to the shlaxN routine may get replaced by one or more asl insns
+ * if the value of X is not used later.
+ */
+{
+ unsigned Changes = 0;
+
+ /* Walk over the entries */
+ unsigned I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ /* Get next entry */
+ CodeEntry* E = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (E->OPC == OP65_JSR &&
+ (strncmp (E->Arg, "shlax", 5) == 0 ||
+ strncmp (E->Arg, "aslax", 5) == 0) &&
+ strlen (E->Arg) == 6 &&
+ IsDigit (E->Arg[5]) &&
+ !RegXUsed (S, I+1)) {
+
+ /* Insert shift insns */
+ unsigned Count = E->Arg[5] - '0';
+ while (Count--) {
+ CodeEntry* X = NewCodeEntry (OP65_ASL, AM65_ACC, "a", 0, E->LI);
+ CS_InsertEntry (S, X, I+1);
+ }
+
+ /* Delete the call to shlax */
+ CS_DelEntry (S, I);
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;
+
+ }
+
+ /* Return the number of changes made */
+ return Changes;
+}
+
+
+
+static unsigned OptShift2 (CodeSeg* S)
+/* A call to the shraxN routine may get replaced by one or more lsr insns
+ * if the value of X is zero.
+ */
+{
+ unsigned Changes = 0;
+ unsigned I;
+
+ /* Generate register info */
+ CS_GenRegInfo (S);
+
+ /* Walk over the entries */
+ I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ /* Get next entry */
+ CodeEntry* E = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (E->OPC == OP65_JSR &&
+ strncmp (E->Arg, "shrax", 5) == 0 &&
+ strlen (E->Arg) == 6 &&
+ IsDigit (E->Arg[5]) &&
+ E->RI->In.RegX == 0) {
+
+ /* Insert shift insns */
+ unsigned Count = E->Arg[5] - '0';
+ while (Count--) {
+ CodeEntry* X = NewCodeEntry (OP65_LSR, AM65_ACC, "a", 0, E->LI);
+ CS_InsertEntry (S, X, I+1);
+ }
+
+ /* Delete the call to shlax */
+ CS_DelEntry (S, I);
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;
+
+ }
+
+ /* Free the register info */
+ CS_FreeRegInfo (S);
+
+ /* Return the number of changes made */
+ return Changes;
+}
+
+
+
+static unsigned GetShiftType (const char* Sub)
+/* Helper function for OptShift3 */
+{
+ if (*Sub == 'a') {
+ if (strcmp (Sub+1, "slax1") == 0) {
+ return SHIFT_ASL_1;
+ } else if (strcmp (Sub+1, "srax1") == 0) {
+ return SHIFT_ASR_1;
+ }
+ } else if (*Sub == 's') {
+ if (strcmp (Sub+1, "hlax1") == 0) {
+ return SHIFT_LSL_1;
+ } else if (strcmp (Sub+1, "hrax1") == 0) {
+ return SHIFT_LSR_1;
+ }
+ }
+ return SHIFT_NONE;
+}
+
+
+
+static unsigned OptShift3 (CodeSeg* S)
+/* Search for the sequence
+ *
+ * lda xxx
+ * ldx yyy
+ * jsr aslax1/asrax1/shlax1/shrax1
+ * sta aaa
+ * stx bbb
+ *
+ * and replace it by
+ *
+ * lda xxx
+ * asl a
+ * sta aaa
+ * lda yyy
+ * rol a
+ * sta bbb
+ *
+ * or similar, provided that a/x is not used later
+ */
+{
+ unsigned Changes = 0;
+
+ /* Walk over the entries */
+ unsigned I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ unsigned ShiftType;
+ CodeEntry* L[5];
+
+ /* Get next entry */
+ L[0] = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (L[0]->OPC == OP65_LDA &&
+ (L[0]->AM == AM65_ABS || L[0]->AM == AM65_ZP) &&
+ CS_GetEntries (S, L+1, I+1, 4) &&
+ !CS_RangeHasLabel (S, I+1, 4) &&
+ L[1]->OPC == OP65_LDX &&
+ (L[1]->AM == AM65_ABS || L[1]->AM == AM65_ZP) &&
+ L[2]->OPC == OP65_JSR &&
+ (ShiftType = GetShiftType (L[2]->Arg)) != SHIFT_NONE&&
+ L[3]->OPC == OP65_STA &&
+ (L[3]->AM == AM65_ABS || L[3]->AM == AM65_ZP) &&
+ L[4]->OPC == OP65_STX &&
+ (L[4]->AM == AM65_ABS || L[4]->AM == AM65_ZP) &&
+ !RegAXUsed (S, I+5)) {
+
+ CodeEntry* X;
+
+ /* Handle the four shift types differently */
+ switch (ShiftType) {
+
+ case SHIFT_ASR_1:
+ X = NewCodeEntry (OP65_LDA, L[1]->AM, L[1]->Arg, 0, L[1]->LI);
+ CS_InsertEntry (S, X, I+5);
+ X = NewCodeEntry (OP65_CMP, AM65_IMM, "$80", 0, L[2]->LI);
+ CS_InsertEntry (S, X, I+6);
+ X = NewCodeEntry (OP65_ROR, AM65_ACC, "a", 0, L[2]->LI);
+ CS_InsertEntry (S, X, I+7);
+ X = NewCodeEntry (OP65_STA, L[4]->AM, L[4]->Arg, 0, L[4]->LI);
+ CS_InsertEntry (S, X, I+8);
+ X = NewCodeEntry (OP65_LDA, L[0]->AM, L[0]->Arg, 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+9);
+ X = NewCodeEntry (OP65_ROR, AM65_ACC, "a", 0, L[2]->LI);
+ CS_InsertEntry (S, X, I+10);
+ X = NewCodeEntry (OP65_STA, L[3]->AM, L[3]->Arg, 0, L[3]->LI);
+ CS_InsertEntry (S, X, I+11);
+ CS_DelEntries (S, I, 5);
+ break;
+
+ case SHIFT_LSR_1:
+ X = NewCodeEntry (OP65_LDA, L[1]->AM, L[1]->Arg, 0, L[1]->LI);
+ CS_InsertEntry (S, X, I+5);
+ X = NewCodeEntry (OP65_LSR, AM65_ACC, "a", 0, L[2]->LI);
+ CS_InsertEntry (S, X, I+6);
+ X = NewCodeEntry (OP65_STA, L[4]->AM, L[4]->Arg, 0, L[4]->LI);
+ CS_InsertEntry (S, X, I+7);
+ X = NewCodeEntry (OP65_LDA, L[0]->AM, L[0]->Arg, 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+8);
+ X = NewCodeEntry (OP65_ROR, AM65_ACC, "a", 0, L[2]->LI);
+ CS_InsertEntry (S, X, I+9);
+ X = NewCodeEntry (OP65_STA, L[3]->AM, L[3]->Arg, 0, L[3]->LI);
+ CS_InsertEntry (S, X, I+10);
+ CS_DelEntries (S, I, 5);
+ break;
+
+ case SHIFT_LSL_1:
+ case SHIFT_ASL_1:
+ /* These two are identical */
+ X = NewCodeEntry (OP65_ASL, AM65_ACC, "a", 0, L[2]->LI);
+ CS_InsertEntry (S, X, I+1);
+ X = NewCodeEntry (OP65_STA, L[3]->AM, L[3]->Arg, 0, L[3]->LI);
+ CS_InsertEntry (S, X, I+2);
+ X = NewCodeEntry (OP65_LDA, L[1]->AM, L[1]->Arg, 0, L[1]->LI);
+ CS_InsertEntry (S, X, I+3);
+ X = NewCodeEntry (OP65_ROL, AM65_ACC, "a", 0, L[2]->LI);
+ CS_InsertEntry (S, X, I+4);
+ X = NewCodeEntry (OP65_STA, L[4]->AM, L[4]->Arg, 0, L[4]->LI);
+ CS_InsertEntry (S, X, I+5);
+ CS_DelEntries (S, I+6, 4);
+ break;
+
+ }
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;
+
+ }
+
+ /* Return the number of changes made */
+ return Changes;
+}
+
+
+
+/*****************************************************************************/
+/* Optimize loads */
+/*****************************************************************************/
+
+
+
+static unsigned OptLoad1 (CodeSeg* S)
+/* Search for a call to ldaxysp where X is not used later and replace it by
+ * a load of just the A register.
+ */
+{
+ unsigned I;
+ unsigned Changes = 0;
+
+ /* Generate register info */
+ CS_GenRegInfo (S);
+
+ /* Walk over the entries */
+ I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ CodeEntry* E;
+
+ /* Get next entry */
+ E = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (CE_IsCallTo (E, "ldaxysp") &&
+ RegValIsKnown (E->RI->In.RegY) &&
+ !RegXUsed (S, I+1)) {
+
+ CodeEntry* X;
+
+ /* Reload the Y register */
+ const char* Arg = MakeHexArg (E->RI->In.RegY - 1);
+ X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, E->LI);
+ CS_InsertEntry (S, X, I+1);
+
+ /* Load from stack */
+ X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, "sp", 0, E->LI);
+ CS_InsertEntry (S, X, I+2);
+
+ /* Now remove the call to the subroutine */
+ CS_DelEntry (S, I);
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;
+
+ }
+
+ /* Free the register info */
+ CS_FreeRegInfo (S);
+
+ /* Return the number of changes made */
+ return Changes;
+}
+
+
+
+/*****************************************************************************/
+/* Optimize stores through pointers */
+/*****************************************************************************/
+
+
+
+static unsigned OptPtrStore1Sub (CodeSeg* S, unsigned I, CodeEntry** const L)
+/* Check if this is one of the allowed suboperation for OptPtrStore1 */
+{
+ /* Check for a label attached to the entry */
+ if (CE_HasLabel (L[0])) {
+ return 0;
+ }
+
+ /* Check for single insn sub ops */
+ if (L[0]->OPC == OP65_AND ||
+ L[0]->OPC == OP65_EOR ||
+ L[0]->OPC == OP65_ORA ||
+ (L[0]->OPC == OP65_JSR && strncmp (L[0]->Arg, "shlax", 5) == 0) ||
+ (L[0]->OPC == OP65_JSR && strncmp (L[0]->Arg, "shrax", 5) == 0)) {
+
+ /* One insn */
+ return 1;
+
+ } else if (L[0]->OPC == OP65_CLC &&
+ (L[1] = CS_GetNextEntry (S, I)) != 0 &&
+ L[1]->OPC == OP65_ADC &&
+ !CE_HasLabel (L[1])) {
+ return 2;
+ } else if (L[0]->OPC == OP65_SEC &&
+ (L[1] = CS_GetNextEntry (S, I)) != 0 &&
+ L[1]->OPC == OP65_SBC &&
+ !CE_HasLabel (L[1])) {
+ return 2;
+ }
+
+
+
+ /* Not found */
+ return 0;
+}
+
+
+
+static unsigned OptPtrStore1 (CodeSeg* S)
+/* Search for the sequence:
+ *
+ * jsr pushax
+ * ldy xxx
+ * jsr ldauidx
+ * subop
+ * ldy yyy
+ * jsr staspidx
+ *
+ * and replace it by:
+ *
+ * sta ptr1
+ * stx ptr1+1
+ * ldy xxx
+ * ldx #$00
+ * lda (ptr1),y
+ * subop
+ * ldy yyy
+ * sta (ptr1),y
+ *
+ * In case a/x is loaded from the register bank before the pushax, we can even
+ * use the register bank instead of ptr1.
+ */
+/*
+ * jsr pushax
+ * ldy xxx
+ * jsr ldauidx
+ * ldx #$00
+ * lda (zp),y
+ * subop
+ * ldy yyy
+ * sta (zp),y
+ * jsr staspidx
+ */
+{
+ unsigned Changes = 0;
+
+ /* Walk over the entries */
+ unsigned I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ unsigned K;
+ CodeEntry* L[10];
+
+ /* Get next entry */
+ L[0] = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (CE_IsCallTo (L[0], "pushax") &&
+ CS_GetEntries (S, L+1, I+1, 3) &&
+ L[1]->OPC == OP65_LDY &&
+ CE_KnownImm (L[1]) &&
+ !CE_HasLabel (L[1]) &&
+ CE_IsCallTo (L[2], "ldauidx") &&
+ !CE_HasLabel (L[2]) &&
+ (K = OptPtrStore1Sub (S, I+3, L+3)) > 0 &&
+ CS_GetEntries (S, L+3+K, I+3+K, 2) &&
+ L[3+K]->OPC == OP65_LDY &&
+ CE_KnownImm (L[3+K]) &&
+ !CE_HasLabel (L[3+K]) &&
+ CE_IsCallTo (L[4+K], "staspidx") &&
+ !CE_HasLabel (L[4+K])) {
+
+
+ const char* RegBank = 0;
+ const char* ZPLoc = "ptr1";
+ CodeEntry* X;
+
+
+ /* Get the preceeding two instructions and check them. We check
+ * for:
+ * lda regbank+n
+ * ldx regbank+n+1
+ */
+ if (I > 1) {
+ CodeEntry* P[2];
+ P[0] = CS_GetEntry (S, I-2);
+ P[1] = CS_GetEntry (S, I-1);
+ if (P[0]->OPC == OP65_LDA &&
+ P[0]->AM == AM65_ZP &&
+ P[1]->OPC == OP65_LDX &&
+ P[1]->AM == AM65_ZP &&
+ !CE_HasLabel (P[1]) &&
+ strncmp (P[0]->Arg, "regbank+", 8) == 0) {
+
+ unsigned Len = strlen (P[0]->Arg);
+
+ if (strncmp (P[0]->Arg, P[1]->Arg, Len) == 0 &&
+ P[1]->Arg[Len+0] == '+' &&
+ P[1]->Arg[Len+1] == '1' &&
+ P[1]->Arg[Len+2] == '\0') {
+
+ /* Ok, found. Use the name of the register bank */
+ RegBank = ZPLoc = P[0]->Arg;
+ }
+ }
+ }
+
+ /* Insert the load via the zp pointer */
+ X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[3]->LI);
+ CS_InsertEntry (S, X, I+3);
+ X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, ZPLoc, 0, L[2]->LI);
+ CS_InsertEntry (S, X, I+4);
+
+ /* Insert the store through the zp pointer */
+ X = NewCodeEntry (OP65_STA, AM65_ZP_INDY, ZPLoc, 0, L[3]->LI);
+ CS_InsertEntry (S, X, I+6+K);
+
+ /* Delete the old code */
+ CS_DelEntry (S, I+7+K); /* jsr spaspidx */
+ CS_DelEntry (S, I+2); /* jsr ldauidx */
+
+ /* Create and insert the stores into the zp pointer if needed */
+ if (RegBank == 0) {
+ X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1", 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+1);
+ X = NewCodeEntry (OP65_STX, AM65_ZP, "ptr1+1", 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+2);
+ }
+
+ /* Delete more old code. Do it here to keep a label attached to
+ * entry I in place.
+ */
+ CS_DelEntry (S, I); /* jsr pushax */
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;
+
+ }
+
+ /* Return the number of changes made */
+ return Changes;
+}
+
+
+
+static unsigned OptPtrStore2 (CodeSeg* S)
+/* Search for the sequence:
+ *
+ * lda #<(label+0)
+ * ldx #>(label+0)
+ * clc
+ * adc xxx
+ * bcc L
+ * inx
+ * L: jsr pushax
+ * ldx #$00
+ * lda yyy
+ * ldy #$00
+ * jsr staspidx
+ *
+ * and replace it by:
+ *
+ * ldy xxx
+ * ldx #$00
+ * lda yyy
+ * sta label,y
+ */
+{
+ unsigned Changes = 0;
+
+ /* Walk over the entries */
+ unsigned I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ CodeEntry* L[11];
+ unsigned Len;
+
+ /* Get next entry */
+ L[0] = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (L[0]->OPC == OP65_LDA &&
+ L[0]->AM == AM65_IMM &&
+ CS_GetEntries (S, L+1, I+1, 10) &&
+ L[1]->OPC == OP65_LDX &&
+ L[1]->AM == AM65_IMM &&
+ L[2]->OPC == OP65_CLC &&
+ L[3]->OPC == OP65_ADC &&
+ (L[3]->AM == AM65_ABS || L[3]->AM == AM65_ZP) &&
+ (L[4]->OPC == OP65_BCC || L[4]->OPC == OP65_JCC) &&
+ L[4]->JumpTo != 0 &&
+ L[4]->JumpTo->Owner == L[6] &&
+ L[5]->OPC == OP65_INX &&
+ CE_IsCallTo (L[6], "pushax") &&
+ L[7]->OPC == OP65_LDX &&
+ L[8]->OPC == OP65_LDA &&
+ L[9]->OPC == OP65_LDY &&
+ CE_KnownImm (L[9]) &&
+ L[9]->Num == 0 &&
+ CE_IsCallTo (L[10], "staspidx") &&
+ !CS_RangeHasLabel (S, I+1, 5) &&
+ !CS_RangeHasLabel (S, I+7, 4) &&
+ /* Check the label last because this is quite costly */
+ (Len = strlen (L[0]->Arg)) > 3 &&
+ L[0]->Arg[0] == '<' &&
+ L[0]->Arg[1] == '(' &&
+ strlen (L[1]->Arg) == Len &&
+ L[1]->Arg[0] == '>' &&
+ memcmp (L[0]->Arg+1, L[1]->Arg+1, Len-1) == 0) {
+
+ CodeEntry* X;
+ char* Label;
+
+ /* We will create all the new stuff behind the current one so
+ * we keep the line references.
+ */
+ X = NewCodeEntry (OP65_LDY, L[3]->AM, L[3]->Arg, 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+11);
+
+ X = NewCodeEntry (OP65_LDX, L[7]->AM, L[7]->Arg, 0, L[7]->LI);
+ CS_InsertEntry (S, X, I+12);
+
+ X = NewCodeEntry (OP65_LDA, L[8]->AM, L[8]->Arg, 0, L[8]->LI);
+ CS_InsertEntry (S, X, I+13);
+
+ Label = memcpy (xmalloc (Len-2), L[0]->Arg+2, Len-3);
+ Label[Len-3] = '\0';
+ X = NewCodeEntry (OP65_STA, AM65_ABSY, Label, 0, L[10]->LI);
+ CS_InsertEntry (S, X, I+14);
+ xfree (Label);
+
+ /* Remove the old code */
+ CS_DelEntries (S, I, 11);
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;
+
+ }
+
+ /* Return the number of changes made */
+ return Changes;
+}
+
+
+
+/*****************************************************************************/
+/* Optimize loads through pointers */
+/*****************************************************************************/
+
+
+
+static unsigned OptPtrLoad1 (CodeSeg* S)
+/* Search for the sequence:
+ *
+ * clc
+ * adc xxx
+ * tay
+ * txa
+ * adc yyy
+ * tax
+ * tya
+ * ldy
+ * jsr ldauidx
+ *
+ * and replace it by:
+ *
+ * clc
+ * adc xxx
+ * sta ptr1
+ * txa
+ * adc yyy
+ * sta ptr1+1
+ * ldy
+ * ldx #$00
+ * lda (ptr1),y
+ */
+{
+ unsigned Changes = 0;
+
+ /* Walk over the entries */
+ unsigned I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ CodeEntry* L[9];
+
+ /* Get next entry */
+ L[0] = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (L[0]->OPC == OP65_CLC &&
+ CS_GetEntries (S, L+1, I+1, 8) &&
+ L[1]->OPC == OP65_ADC &&
+ L[2]->OPC == OP65_TAY &&
+ L[3]->OPC == OP65_TXA &&
+ L[4]->OPC == OP65_ADC &&
+ L[5]->OPC == OP65_TAX &&
+ L[6]->OPC == OP65_TYA &&
+ L[7]->OPC == OP65_LDY &&
+ CE_IsCallTo (L[8], "ldauidx") &&
+ !CS_RangeHasLabel (S, I+1, 8)) {
+
+ CodeEntry* X;
+ CodeEntry* P;
+
+ /* Track the insertion point */
+ unsigned IP = I+2;
+
+ /* sta ptr1 */
+ X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1", 0, L[2]->LI);
+ CS_InsertEntry (S, X, IP++);
+
+ /* If the instruction before the clc is a ldx, replace the
+ * txa by an lda with the same location of the ldx. Otherwise
+ * transfer the value in X to A.
+ */
+ if ((P = CS_GetPrevEntry (S, I)) != 0 &&
+ P->OPC == OP65_LDX &&
+ !CE_HasLabel (P)) {
+ X = NewCodeEntry (OP65_LDA, P->AM, P->Arg, 0, P->LI);
+ } else {
+ X = NewCodeEntry (OP65_TXA, AM65_IMP, 0, 0, L[3]->LI);
+ }
+ CS_InsertEntry (S, X, IP++);
+
+ /* adc yyy */
+ X = NewCodeEntry (OP65_ADC, L[4]->AM, L[4]->Arg, 0, L[4]->LI);
+ CS_InsertEntry (S, X, IP++);
+
+ /* sta ptr1+1 */
+ X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1+1", 0, L[5]->LI);
+ CS_InsertEntry (S, X, IP++);
+
+ /* ldy ... */
+ X = NewCodeEntry (OP65_LDY, L[7]->AM, L[7]->Arg, 0, L[7]->LI);
+ CS_InsertEntry (S, X, IP++);
+
+ /* ldx #$00 */
+ X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[8]->LI);
+ CS_InsertEntry (S, X, IP++);
+
+ /* lda (ptr1),y */
+ X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, "ptr1", 0, L[8]->LI);
+ CS_InsertEntry (S, X, IP++);
+
+ /* Remove the old instructions */
+ CS_DelEntries (S, IP, 7);
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;
+
+ }
+
+ /* Return the number of changes made */
+ return Changes;
+}
+
+
+
+static unsigned OptPtrLoad2 (CodeSeg* S)
+/* Search for the sequence:
+ *
+ * adc xxx
+ * pha
+ * txa
+ * iny
+ * adc yyy
+ * tax
+ * pla
+ * ldy
+ * jsr ldauidx
+ *
+ * and replace it by:
+ *
+ * adc xxx
+ * sta ptr1
+ * txa
+ * iny
+ * adc yyy
+ * sta ptr1+1
+ * ldy
+ * ldx #$00
+ * lda (ptr1),y
+ */
+{
+ unsigned Changes = 0;
+
+ /* Walk over the entries */
+ unsigned I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ CodeEntry* L[9];
+
+ /* Get next entry */
+ L[0] = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (L[0]->OPC == OP65_ADC &&
+ CS_GetEntries (S, L+1, I+1, 8) &&
+ L[1]->OPC == OP65_PHA &&
+ L[2]->OPC == OP65_TXA &&
+ L[3]->OPC == OP65_INY &&
+ L[4]->OPC == OP65_ADC &&
+ L[5]->OPC == OP65_TAX &&
+ L[6]->OPC == OP65_PLA &&
+ L[7]->OPC == OP65_LDY &&
+ CE_IsCallTo (L[8], "ldauidx") &&
+ !CS_RangeHasLabel (S, I+1, 8)) {
+
+ CodeEntry* X;
+
+ /* Store the low byte and remove the PHA instead */
+ X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1", 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+1);
+
+ /* Store the high byte */
+ X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1+1", 0, L[4]->LI);
+ CS_InsertEntry (S, X, I+6);
+
+ /* Load high and low byte */
+ X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[6]->LI);
+ CS_InsertEntry (S, X, I+10);
+ X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, "ptr1", 0, L[6]->LI);
+ CS_InsertEntry (S, X, I+11);
+
+ /* Delete the old code */
+ CS_DelEntry (S, I+12); /* jsr ldauidx */
+ CS_DelEntry (S, I+8); /* pla */
+ CS_DelEntry (S, I+7); /* tax */
+ CS_DelEntry (S, I+2); /* pha */
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;
+
+ }
+
+ /* Return the number of changes made */
+ return Changes;
+}
+
+
+
+static unsigned OptPtrLoad3 (CodeSeg* S)
+/* Search for the sequence:
+ *
+ * lda #<(label+0)
+ * ldx #>(label+0)
+ * clc
+ * adc xxx
+ * bcc L
+ * inx
+ * L: ldy #$00
+ * jsr ldauidx
+ *
+ * and replace it by:
+ *
+ * ldy xxx
+ * ldx #$00
+ * lda label,y
+ */
+{
+ unsigned Changes = 0;
+
+ /* Walk over the entries */
+ unsigned I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ CodeEntry* L[8];
+ unsigned Len;
+
+ /* Get next entry */
+ L[0] = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (L[0]->OPC == OP65_LDA &&
+ L[0]->AM == AM65_IMM &&
+ CS_GetEntries (S, L+1, I+1, 7) &&
+ L[1]->OPC == OP65_LDX &&
+ L[1]->AM == AM65_IMM &&
+ L[2]->OPC == OP65_CLC &&
+ L[3]->OPC == OP65_ADC &&
+ (L[3]->AM == AM65_ABS || L[3]->AM == AM65_ZP) &&
+ (L[4]->OPC == OP65_BCC || L[4]->OPC == OP65_JCC) &&
+ L[4]->JumpTo != 0 &&
+ L[4]->JumpTo->Owner == L[6] &&
+ L[5]->OPC == OP65_INX &&
+ L[6]->OPC == OP65_LDY &&
+ CE_KnownImm (L[6]) &&
+ L[6]->Num == 0 &&
+ CE_IsCallTo (L[7], "ldauidx") &&
+ !CS_RangeHasLabel (S, I+1, 5) &&
+ !CE_HasLabel (L[7]) &&
+ /* Check the label last because this is quite costly */
+ (Len = strlen (L[0]->Arg)) > 3 &&
+ L[0]->Arg[0] == '<' &&
+ L[0]->Arg[1] == '(' &&
+ strlen (L[1]->Arg) == Len &&
+ L[1]->Arg[0] == '>' &&
+ memcmp (L[0]->Arg+1, L[1]->Arg+1, Len-1) == 0) {
+
+ CodeEntry* X;
+ char* Label;
+
+ /* We will create all the new stuff behind the current one so
+ * we keep the line references.
+ */
+ X = NewCodeEntry (OP65_LDY, L[3]->AM, L[3]->Arg, 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+8);
+
+ X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+9);
+
+ Label = memcpy (xmalloc (Len-2), L[0]->Arg+2, Len-3);
+ Label[Len-3] = '\0';
+ X = NewCodeEntry (OP65_LDA, AM65_ABSY, Label, 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+10);
+ xfree (Label);
+
+ /* Remove the old code */
+ CS_DelEntries (S, I, 8);
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;
+
+ }
+
+ /* Return the number of changes made */
+ return Changes;
+}
+
+
+
+static unsigned OptPtrLoad4 (CodeSeg* S)
+/* Search for the sequence:
+ *
+ * lda #<(label+0)
+ * ldx #>(label+0)
+ * ldy #$xx
+ * clc
+ * adc (sp),y
+ * bcc L
+ * inx
+ * L: ldy #$00
+ * jsr ldauidx
+ *
+ * and replace it by:
+ *
+ * ldy #$xx
+ * lda (sp),y
+ * tay
+ * ldx #$00
+ * lda label,y
+ */
+{
+ unsigned Changes = 0;
+
+ /* Walk over the entries */
+ unsigned I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ CodeEntry* L[9];
+ unsigned Len;
+
+ /* Get next entry */
+ L[0] = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (L[0]->OPC == OP65_LDA &&
+ L[0]->AM == AM65_IMM &&
+ CS_GetEntries (S, L+1, I+1, 8) &&
+ L[1]->OPC == OP65_LDX &&
+ L[1]->AM == AM65_IMM &&
+ !CE_HasLabel (L[1]) &&
+ L[2]->OPC == OP65_LDY &&
+ CE_KnownImm (L[2]) &&
+ !CE_HasLabel (L[2]) &&
+ L[3]->OPC == OP65_CLC &&
+ !CE_HasLabel (L[3]) &&
+ L[4]->OPC == OP65_ADC &&
+ L[4]->AM == AM65_ZP_INDY &&
+ !CE_HasLabel (L[4]) &&
+ (L[5]->OPC == OP65_BCC || L[5]->OPC == OP65_JCC) &&
+ L[5]->JumpTo != 0 &&
+ L[5]->JumpTo->Owner == L[7] &&
+ !CE_HasLabel (L[5]) &&
+ L[6]->OPC == OP65_INX &&
+ !CE_HasLabel (L[6]) &&
+ L[7]->OPC == OP65_LDY &&
+ CE_KnownImm (L[7]) &&
+ L[7]->Num == 0 &&
+ CE_IsCallTo (L[8], "ldauidx") &&
+ !CE_HasLabel (L[8]) &&
+ /* Check the label last because this is quite costly */
+ (Len = strlen (L[0]->Arg)) > 3 &&
+ L[0]->Arg[0] == '<' &&
+ L[0]->Arg[1] == '(' &&
+ strlen (L[1]->Arg) == Len &&
+ L[1]->Arg[0] == '>' &&
+ memcmp (L[0]->Arg+1, L[1]->Arg+1, Len-1) == 0) {
+
+ CodeEntry* X;
+ char* Label;
+
+ /* Add the lda */
+ X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, L[4]->Arg, 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+3);
+
+ /* Add the tay */
+ X = NewCodeEntry (OP65_TAY, AM65_IMP, 0, 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+4);
+
+ /* Add the ldx */
+ X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+5);
+
+ /* Add the lda */
+ Label = memcpy (xmalloc (Len-2), L[0]->Arg+2, Len-3);
+ Label[Len-3] = '\0';
+ X = NewCodeEntry (OP65_LDA, AM65_ABSY, Label, 0, L[0]->LI);
+ CS_InsertEntry (S, X, I+6);
+ xfree (Label);
+
+ /* Remove the old code */
+ CS_DelEntries (S, I, 2);
+ CS_DelEntries (S, I+5, 6);
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;
+
+ }
+
+ /* Return the number of changes made */
+ return Changes;
+}
+
+
+
+static unsigned OptPtrLoad5 (CodeSeg* S)
+/* Search for the sequence:
+ *
+ * lda regbank+n
+ * ldx regbank+n+1
+ * sta regsave
+ * stx regsave+1
+ * clc
+ * adc #$01
+ * bcc L0005
+ * inx
+ * L: sta regbank+n
+ * stx regbank+n+1
+ * lda regsave
+ * ldx regsave+1
+ * ldy #$00
+ * jsr ldauidx
+ *
+ * and replace it by:
+ *
+ * ldy #$00
+ * ldx #$00
+ * lda (regbank+n),y
+ * inc regbank+n
+ * bne L1
+ * inc regbank+n+1
+ * L1: tay <- only if flags are used
+ *
+ * This function must execute before OptPtrLoad5!
+ *
+ */
+{
+ unsigned Changes = 0;
+
+ /* Walk over the entries */
+ unsigned I = 0;
+ while (I < CS_GetEntryCount (S)) {
+
+ CodeEntry* L[15];
+ unsigned Len;
+
+ /* Get next entry */
+ L[0] = CS_GetEntry (S, I);
+
+ /* Check for the sequence */
+ if (L[0]->OPC == OP65_LDA &&
+ L[0]->AM == AM65_ZP &&
+ strncmp (L[0]->Arg, "regbank+", 8) == 0 &&
+ (Len = strlen (L[0]->Arg)) > 0 &&
+ CS_GetEntries (S, L+1, I+1, 14) &&
+ !CS_RangeHasLabel (S, I+1, 7) &&
+ !CS_RangeHasLabel (S, I+9, 5) &&
+ L[1]->OPC == OP65_LDX &&
+ L[1]->AM == AM65_ZP &&
+ strncmp (L[1]->Arg, L[0]->Arg, Len) == 0 &&
+ strcmp (L[1]->Arg+Len, "+1") == 0 &&
+ L[2]->OPC == OP65_STA &&
+ L[2]->AM == AM65_ZP &&
+ strcmp (L[2]->Arg, "regsave") == 0 &&
+ L[3]->OPC == OP65_STX &&
+ L[3]->AM == AM65_ZP &&
+ strcmp (L[3]->Arg, "regsave+1") == 0 &&
+ L[4]->OPC == OP65_CLC &&
+ L[5]->OPC == OP65_ADC &&
+ CE_KnownImm (L[5]) &&
+ L[5]->Num == 1 &&
+ L[6]->OPC == OP65_BCC &&
+ L[6]->JumpTo != 0 &&
+ L[6]->JumpTo->Owner == L[8] &&
+ L[7]->OPC == OP65_INX &&
+ L[8]->OPC == OP65_STA &&
+ L[8]->AM == AM65_ZP &&
+ strcmp (L[8]->Arg, L[0]->Arg) == 0 &&
+ L[9]->OPC == OP65_STX &&
+ L[9]->AM == AM65_ZP &&
+ strcmp (L[9]->Arg, L[1]->Arg) == 0 &&
+ L[10]->OPC == OP65_LDA &&
+ L[10]->AM == AM65_ZP &&
+ strcmp (L[10]->Arg, "regsave") == 0 &&
+ L[11]->OPC == OP65_LDX &&
+ L[11]->AM == AM65_ZP &&
+ strcmp (L[11]->Arg, "regsave+1") == 0 &&
+ L[12]->OPC == OP65_LDY &&
+ CE_KnownImm (L[12]) &&
+ CE_IsCallTo (L[13], "ldauidx")) {
+
+ CodeEntry* X;
+ CodeLabel* Label;
+
+ /* Check if the instruction following the sequence uses the flags
+ * set by the load. If so, insert a test of the value in the
+ * accumulator.
+ */
+ if (CE_UseLoadFlags (L[14])) {
+ X = NewCodeEntry (OP65_TAY, AM65_IMP, 0, 0, L[13]->LI);
+ CS_InsertEntry (S, X, I+14);
+ }
+
+ /* Attach a label to L[14]. This may be either the just inserted
+ * instruction, or the one following the sequence.
+ */
+ Label = CS_GenLabel (S, L[14]);
+
+ /* ldy #$xx */
+ X = NewCodeEntry (OP65_LDY, AM65_IMM, L[12]->Arg, 0, L[12]->LI);
+ CS_InsertEntry (S, X, I+14);
+
+ /* ldx #$xx */
+ X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[13]->LI);
+ CS_InsertEntry (S, X, I+15);
+
+ /* lda (regbank+n),y */
+ X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, L[0]->Arg, 0, L[13]->LI);
+ CS_InsertEntry (S, X, I+16);
+
+ /* inc regbank+n */
+ X = NewCodeEntry (OP65_INC, AM65_ZP, L[0]->Arg, 0, L[5]->LI);
+ CS_InsertEntry (S, X, I+17);
+
+ /* bne ... */
+ X = NewCodeEntry (OP65_BNE, AM65_BRA, Label->Name, Label, L[6]->LI);
+ CS_InsertEntry (S, X, I+18);
+
+ /* inc regbank+n+1 */
+ X = NewCodeEntry (OP65_INC, AM65_ZP, L[1]->Arg, 0, L[7]->LI);
+ CS_InsertEntry (S, X, I+19);
+
+ /* Delete the old code */
+ CS_DelEntries (S, I, 14);
+
+ /* Remember, we had changes */
+ ++Changes;
+
+ }
+
+ /* Next entry */
+ ++I;