]> git.sur5r.net Git - cc65/blobdiff - src/cc65/coptind.c
New optimization step
[cc65] / src / cc65 / coptind.c
index f4bdecbb126f5cb817bcdf0a1a987bd572427426..80d6c772c65a56647da04e0bc68c03f65915e487 100644 (file)
@@ -6,8 +6,8 @@
 /*                                                                           */
 /*                                                                           */
 /*                                                                           */
-/* (C) 2001-2005, Ullrich von Bassewitz                                      */
-/*                Römerstraße 52                                             */
+/* (C) 2001-2009, Ullrich von Bassewitz                                      */
+/*                Roemerstrasse 52                                           */
 /*                D-70794 Filderstadt                                        */
 /* EMail:         uz@cc65.org                                                */
 /*                                                                           */
 
 /* cc65 */
 #include "codeent.h"
+#include "coptind.h"
 #include "codeinfo.h"
 #include "codeopt.h"
 #include "error.h"
-#include "coptind.h"
 
 
 
 
 
 
+static int MemAccess (CodeSeg* S, unsigned From, unsigned To, const CodeEntry* N)
+/* Checks a range of code entries if there are any memory accesses to N->Arg */
+{
+    /* Get the length of the argument */
+    unsigned NLen = strlen (N->Arg);
+
+    /* What to check for? */
+    enum {
+        None    = 0x00,
+        Base    = 0x01,         /* Check for location without "+1" */
+        Word    = 0x02,         /* Check for location with "+1" added */
+    } What = None;
+
+
+    /* If the argument of N is a zero page location that ends with "+1", we
+     * must also check for word accesses to the location without +1.
+     */
+    if (N->AM == AM65_ZP && NLen > 2 && strcmp (N->Arg + NLen - 2, "+1") == 0) {
+        What |= Base;
+    }
+
+    /* If the argument is zero page indirect, we must also check for accesses
+     * to "arg+1"
+     */
+    if (N->AM == AM65_ZP_INDY || N->AM == AM65_ZPX_IND || N->AM == AM65_ZP_IND) {
+        What |= Word;
+    }
+
+    /* Walk over all code entries */
+    while (From <= To) {
+
+       /* Get the next entry */
+       CodeEntry* E = CS_GetEntry (S, From);
+
+        /* Check if there is an argument and if this argument equals Arg in
+         * some variants.
+         */
+        if (E->Arg[0] != '\0') {
+
+            unsigned ELen;
+
+            if (strcmp (E->Arg, N->Arg) == 0) {
+                /* Found an access */
+                return 1;
+            }
+
+            ELen = strlen (E->Arg);
+            if ((What & Base) != 0) {
+                if (ELen == NLen - 2 && strncmp (E->Arg, N->Arg, NLen-2) == 0) {
+                    /* Found an access */
+                    return 1;
+                }
+            }
+
+            if ((What & Word) != 0) {
+                if (ELen == NLen + 2 && strncmp (E->Arg, N->Arg, NLen) == 0 &&
+                    E->Arg[NLen] == '+' && E->Arg[NLen+1] == '1') {
+                    /* Found an access */
+                    return 1;
+                }
+            }
+        }
+
+        /* Next entry */
+        ++From;
+    }
+
+    /* Nothing found */
+    return 0;
+}
+
+
+
 static int GetBranchDist (CodeSeg* S, unsigned From, CodeEntry* To)
 /* Get the branch distance between the two entries and return it. The distance
  * will be negative for backward jumps and positive for forward jumps.
@@ -406,17 +479,12 @@ unsigned OptJumpCascades (CodeSeg* S)
                /* Remember, we had changes */
                ++Changes;
 
-               /* Done */
-               continue;
-
-           }
-
            /* Check if both are conditional branches, and the condition of
             * the second is the inverse of that of the first. In this case,
             * the second branch will never be taken, and we may jump directly
             * to the instruction behind this one.
             */
-           if ((E->Info & OF_CBRA) != 0 && (N->Info & OF_CBRA) != 0) {
+           } else if ((E->Info & OF_CBRA) != 0 && (N->Info & OF_CBRA) != 0) {
 
                CodeEntry* X;   /* Instruction behind N */
                CodeLabel* LX;  /* Label attached to X */
@@ -447,10 +515,6 @@ unsigned OptJumpCascades (CodeSeg* S)
 
                /* Remember, we had changes */
                ++Changes;
-
-               /* Done */
-               continue;
-
            }
        }
 
@@ -520,7 +584,7 @@ unsigned OptRTS (CodeSeg* S)
 
 
 
-unsigned OptJumpTarget (CodeSeg* S)
+unsigned OptJumpTarget1 (CodeSeg* S)
 /* If the instruction preceeding an unconditional branch is the same as the
  * instruction preceeding the jump target, the jump target may be moved
  * one entry back. This is a size optimization, since the instruction before
@@ -540,12 +604,13 @@ unsigned OptJumpTarget (CodeSeg* S)
        /* Get next entry */
                E2 = CS_GetNextEntry (S, I);
 
-       /* Check if we have a jump or branch, and a matching label, which
-        * is not attached to the jump itself
+       /* Check if we have a jump or branch without a label attached, and
+         * a jump target, which is not attached to the jump itself
         */
-               if (E2 != 0                     &&
+               if (E2 != 0                     &&
            (E2->Info & OF_UBRA) != 0   &&
-           E2->JumpTo                  &&
+            !CE_HasLabel (E2)           &&
+           E2->JumpTo                  &&
            E2->JumpTo->Owner != E2) {
 
            /* Get the entry preceeding the branch target */
@@ -555,6 +620,13 @@ unsigned OptJumpTarget (CodeSeg* S)
                goto NextEntry;
            }
 
+            /* The entry preceeding the branch target may not be the branch
+             * insn.
+             */
+            if (T1 == E2) {
+                goto NextEntry;
+            }
+
            /* Get the entry preceeding the jump */
            E1 = CS_GetEntry (S, I);
 
@@ -599,13 +671,166 @@ NextEntry:
 
 
 
+unsigned OptJumpTarget2 (CodeSeg* S)
+/* If a bcs jumps to a sec insn or a bcc jumps to clc, skip this insn, since
+ * it's job is already done.
+ */
+{
+    unsigned Changes = 0;
+
+    /* Walk over the entries */
+    unsigned I = 0;
+    while (I < CS_GetEntryCount (S)) {
+
+        /* OP that may be skipped */
+        opc_t OPC;
+
+        /* Jump target insn, old and new */
+        CodeEntry* T;
+        CodeEntry* N;
+
+        /* New jump label */
+        CodeLabel* L;
+
+       /* Get next entry */
+               CodeEntry* E = CS_GetEntry (S, I);
+
+        /* Check if this is a bcc insn */
+        if (E->OPC == OP65_BCC || E->OPC == OP65_JCC) {
+            OPC = OP65_CLC;
+        } else if (E->OPC == OP65_BCS || E->OPC == OP65_JCS) {
+            OPC = OP65_SEC;
+        } else {
+            /* Not what we're looking for */
+            goto NextEntry;
+        }
+
+        /* Must have a jump target */
+        if (E->JumpTo == 0) {
+            goto NextEntry;
+        }
+
+        /* Get the owner insn of the jump target and check if it's the one, we
+         * will skip if present.
+         */
+        T = E->JumpTo->Owner;
+        if (T->OPC != OPC) {
+            goto NextEntry;
+        }
+
+        /* Get the entry following the branch target */
+        N = CS_GetNextEntry (S, CS_GetEntryIndex (S, T));
+        if (N == 0) {
+            /* There is no such entry */
+            goto NextEntry;
+        }
+
+        /* Get the label for the instruction following the jump target.
+         * This routine will create a new label if the instruction does
+         * not already have one.
+         */
+        L = CS_GenLabel (S, N);
+
+        /* Change the jump target to point to this new label */
+        CS_MoveLabelRef (S, E, L);
+
+        /* Remember that we had changes */
+        ++Changes;
+
+NextEntry:
+        /* Next entry */
+        ++I;
+    }
+
+    /* Return the number of changes made */
+    return Changes;
+}
+
+
+
+unsigned OptJumpTarget3 (CodeSeg* S)
+/* Jumps to load instructions of a register, that do already have the matching
+ * register contents may skip the load instruction, since it's job is already
+ * done.
+ */
+{
+    unsigned Changes = 0;
+    unsigned I;
+
+    /* Generate register info for this step */
+    CS_GenRegInfo (S);
+
+    /* Walk over the entries */
+    I = 0;
+    while (I < CS_GetEntryCount (S)) {
+
+        unsigned J, K;
+        CodeEntry* N;
+
+        /* New jump label */
+        CodeLabel* LN = 0;
+
+       /* Get next entry */
+               CodeEntry* E = CS_GetEntry (S, I);  
+
+        /* Check if this is a load insn with a label */
+        if ((E->Info & OF_LOAD) != 0            &&
+            CE_IsConstImm (E)                   &&
+            CE_HasLabel (E)                     &&
+            (N = CS_GetNextEntry (S, I)) != 0)    {
+
+            /* Walk over all insn that jump here */
+            for (J = 0; J < CE_GetLabelCount (E); ++J) {
+
+                /* Get the label */
+                CodeLabel* L = CE_GetLabel (E, J);
+                for (K = 0; K < CL_GetRefCount (L); ++K) {
+
+                    /* Get the entry that jumps here */
+                    CodeEntry* Jump = CL_GetRef (L, K);
+
+                    /* Get the register info from this insn */
+                    short Val = RegVal (E->Chg, &Jump->RI->Out2);
+
+                    /* Check if the outgoing value is the one thatr's loaded */
+                    if (Val == (unsigned char) E->Num) {
+
+                        /* Ok, skip the insn. First, generate a label */
+                        if (LN == 0) {
+                            LN = CS_GenLabel (S, N);
+                        }
+
+                        /* Change the jump target to point to this new label */
+                        CS_MoveLabelRef (S, Jump, LN);
+
+                        /* Remember that we had changes */
+                        ++Changes;
+                    }
+                }
+            }
+
+        }
+
+        /* Next entry */
+        ++I;
+    }
+
+    /* Free register info */
+    CS_FreeRegInfo (S);
+
+    /* Return the number of changes made */
+    return Changes;
+}
+
+
+
 /*****************************************************************************/
 /*                      Optimize conditional branches                       */
 /*****************************************************************************/
 
 
 
-unsigned OptCondBranches (CodeSeg* S)
+unsigned OptCondBranches1 (CodeSeg* S)
 /* Performs several optimization steps:
  *
  *  - If an immidiate load of a register is followed by a conditional jump that
@@ -651,7 +876,7 @@ unsigned OptCondBranches (CodeSeg* S)
                CS_DelEntry (S, I+1);
 
                /* Remember, we had changes */
-               ++Changes;
+               ++Changes;
 
            } else if ((BC == BC_EQ && E->Num == 0)             ||
                       (BC == BC_NE && E->Num != 0)             ||
@@ -668,7 +893,7 @@ unsigned OptCondBranches (CodeSeg* S)
        }
 
        if ((E->Info & OF_CBRA) != 0              &&  /* It's a conditional branch */
-           (L = E->JumpTo) != 0                  &&  /* ..referencing a local label */
+           (L = E->JumpTo) != 0                  &&  /* ..referencing a local label */
                    (N = CS_GetNextEntry (S, I)) != 0     &&  /* There is a following entry */
            (N->Info & OF_UBRA) != 0              &&  /* ..which is an uncond branch, */
            !CE_HasLabel (N)                      &&  /* ..has no label attached */
@@ -698,6 +923,62 @@ unsigned OptCondBranches (CodeSeg* S)
 
 
 
+unsigned OptCondBranches2 (CodeSeg* S)
+/* If on entry to a "rol a" instruction the accu is zero, and a beq/bne follows,
+ * we can remove the rol and branch on the state of the carry flag.
+ */
+{
+    unsigned Changes = 0;
+    unsigned I;
+
+    /* Generate register info for this step */
+    CS_GenRegInfo (S);
+
+    /* Walk over the entries */
+    I = 0;
+    while (I < CS_GetEntryCount (S)) {
+
+               CodeEntry* N;
+
+               /* Get next entry */
+               CodeEntry* E = CS_GetEntry (S, I);
+
+       /* Check if it's a rol insn with A in accu and a branch follows */
+               if (E->OPC == OP65_ROL                  &&
+            E->AM == AM65_ACC                   &&
+            E->RI->In.RegA == 0                 &&
+            !CE_HasLabel (E)                    &&
+            (N = CS_GetNextEntry (S, I)) != 0   &&
+            (N->Info & OF_ZBRA) != 0            &&
+            !RegAUsed (S, I+1)) {
+
+           /* Replace the branch condition */
+           switch (GetBranchCond (N->OPC)) {
+                case BC_EQ:     CE_ReplaceOPC (N, OP65_JCC); break;
+                case BC_NE:     CE_ReplaceOPC (N, OP65_JCS); break;
+                default:        Internal ("Unknown branch condition in OptCondBranches2");
+            }
+
+            /* Delete the rol insn */
+            CS_DelEntry (S, I);
+
+            /* Remember, we had changes */
+            ++Changes;
+       }
+
+               /* Next entry */
+               ++I;
+    }
+
+    /* Free register info */
+    CS_FreeRegInfo (S);
+
+    /* Return the number of changes made */
+    return Changes;
+}
+
+
+
 /*****************************************************************************/
 /*                     Remove unused loads and stores                       */
 /*****************************************************************************/
@@ -726,17 +1007,17 @@ unsigned OptUnusedLoads (CodeSeg* S)
            /* Check which sort of load or transfer it is */
            unsigned R;
            switch (E->OPC) {
-               case OP65_DEA:
-               case OP65_INA:
-               case OP65_LDA:
+               case OP65_DEA:
+               case OP65_INA:
+               case OP65_LDA:
                case OP65_TXA:
                case OP65_TYA:  R = REG_A;      break;
-               case OP65_DEX:
-               case OP65_INX:
-               case OP65_LDX:
+               case OP65_DEX:
+               case OP65_INX:
+               case OP65_LDX:
                case OP65_TAX:  R = REG_X;      break;
-               case OP65_DEY:
-               case OP65_INY:
+               case OP65_DEY:
+               case OP65_INY:
                        case OP65_LDY:
                case OP65_TAY:  R = REG_Y;      break;
                default:        goto NextEntry;         /* OOPS */
@@ -746,11 +1027,11 @@ unsigned OptUnusedLoads (CodeSeg* S)
            if ((GetRegInfo (S, I+1, R) & R) == 0) {
 
                /* Register value is not used, remove the load */
-               CS_DelEntry (S, I);
+               CS_DelEntry (S, I);
 
-               /* Remember, we had changes. Account the deleted entry in I. */
-               ++Changes;
-               --I;
+               /* Remember, we had changes. Account the deleted entry in I. */
+               ++Changes;
+               --I;
 
            }
        }
@@ -798,6 +1079,8 @@ unsigned OptUnusedStores (CodeSeg* S)
                /* Remember, we had changes */
                ++Changes;
 
+                /* Continue with next insn */
+                continue;
            }
        }
 
@@ -867,12 +1150,12 @@ unsigned OptDupLoads (CodeSeg* S)
                break;
 
            case OP65_STA:
-               /* If we store into a known zero page location, and this
+               /* If we store into a known zero page location, and this
                 * location does already contain the value to be stored,
                 * remove the store.
                 */
                if (RegValIsKnown (In->RegA)          && /* Value of A is known */
-                   E->AM == AM65_ZP                  && /* Store into zp */
+                           E->AM == AM65_ZP                  && /* Store into zp */
                    In->RegA == ZPRegVal (E->Chg, In)) { /* Value identical */
 
                    Delete = 1;
@@ -900,7 +1183,7 @@ unsigned OptDupLoads (CodeSeg* S)
                           In->RegX == In->RegA       &&
                           E->AM != AM65_ABSY         &&
                           E->AM != AM65_ZPY) {
-                   /* Use the A register instead */
+                   /* Use the A register instead */
                            CE_ReplaceOPC (E, OP65_STA);
                }
                break;
@@ -920,12 +1203,12 @@ unsigned OptDupLoads (CodeSeg* S)
                 * that in the A register, replace the store by a STA. The
                 * optimizer will then remove the load instruction for Y
                 * later. If replacement by A is not possible try a
-                * replacement by X, but check for invalid addressing modes
+                * replacement by X, but check for invalid addressing modes
                 * in this case.
                 */
                        } else if (RegValIsKnown (In->RegY)) {
                    if (In->RegY == In->RegA) {
-                       CE_ReplaceOPC (E, OP65_STA);
+                               CE_ReplaceOPC (E, OP65_STA);
                    } else if (In->RegY == In->RegX   &&
                               E->AM != AM65_ABSX     &&
                               E->AM != AM65_ZPX) {
@@ -953,7 +1236,7 @@ unsigned OptDupLoads (CodeSeg* S)
                            !CE_UseLoadFlags (N)) {
                    /* Value is identical and not followed by a branch */
                    Delete = 1;
-               }
+               }
                break;
 
            case OP65_TAY:
@@ -973,39 +1256,39 @@ unsigned OptDupLoads (CodeSeg* S)
                            !CE_UseLoadFlags (N)) {
                    /* Value is identical and not followed by a branch */
                    Delete = 1;
-               }
-               break;
+               }
+               break;
 
-           case OP65_TYA:
+           case OP65_TYA:
                 if (RegValIsKnown (In->RegY)            &&
-                   In->RegY == In->RegA                &&
-                   (N = CS_GetNextEntry (S, I)) != 0   &&
+                           In->RegY == In->RegA                &&
+                   (N = CS_GetNextEntry (S, I)) != 0   &&
                            !CE_UseLoadFlags (N)) {
-                   /* Value is identical and not followed by a branch */
-                   Delete = 1;
-               }
-               break;
+                   /* Value is identical and not followed by a branch */
+                   Delete = 1;
+               }
+               break;
 
-           default:
-               break;
+           default:
+               break;
 
-       }
+       }
 
-       /* Delete the entry if requested */
-       if (Delete) {
+       /* Delete the entry if requested */
+       if (Delete) {
 
-           /* Register value is not used, remove the load */
-           CS_DelEntry (S, I);
+           /* Register value is not used, remove the load */
+           CS_DelEntry (S, I);
 
-           /* Remember, we had changes */
-           ++Changes;
+           /* Remember, we had changes */
+           ++Changes;
 
-       } else {
+       } else {
 
-           /* Next entry */
-           ++I;
+           /* Next entry */
+           ++I;
 
-       }
+       }
 
     }
 
@@ -1027,10 +1310,10 @@ unsigned OptStoreLoad (CodeSeg* S)
     unsigned I = 0;
     while (I < CS_GetEntryCount (S)) {
 
-       CodeEntry* N;
-       CodeEntry* X;
+               CodeEntry* N;
+               CodeEntry* X;
 
-       /* Get next entry */
+               /* Get next entry */
                CodeEntry* E = CS_GetEntry (S, I);
 
        /* Check if it is a store instruction followed by a load from the
@@ -1105,7 +1388,7 @@ unsigned OptTransfers1 (CodeSeg* S)
                    if (I == 0) {
                        /* No preceeding entry */
                        goto NextEntry;
-                   }
+                   }
                    P = CS_GetEntry (S, I-1);
                    if ((P->Info & OF_SETF) == 0) {
                        /* Does not set the flags */
@@ -1210,12 +1493,355 @@ unsigned OptTransfers2 (CodeSeg* S)
 
 
 
+unsigned OptTransfers3 (CodeSeg* S)
+/* Replace a register transfer followed by a store of the second register by a
+ * store of the first register if this is possible.
+ */
+{
+    unsigned Changes      = 0;
+    unsigned UsedRegs     = REG_NONE;   /* Track used registers */
+    unsigned Xfer         = 0;          /* Index of transfer insn */
+    unsigned Store        = 0;          /* Index of store insn */
+    CodeEntry* XferEntry  = 0;          /* Pointer to xfer insn */
+    CodeEntry* StoreEntry = 0;          /* Pointer to store insn */
+
+    enum {
+        Initialize,
+        Search,
+        FoundXfer,
+        FoundStore
+    } State = Initialize;
+
+    /* Walk over the entries. Look for a xfer instruction that is followed by
+     * a store later, where the value of the register is not used later.
+     */
+    unsigned I = 0;
+    while (I < CS_GetEntryCount (S)) {
+
+       /* Get next entry */
+               CodeEntry* E = CS_GetEntry (S, I);
+
+        switch (State) {
+
+            case Initialize:
+                /* Clear the list of used registers */
+                UsedRegs = REG_NONE;
+                /* FALLTHROUGH */
+
+            case Search:
+                if (E->Info & OF_XFR) {
+                    /* Found start of sequence */
+                    Xfer = I;
+                    XferEntry = E;
+                    State = FoundXfer;
+                }
+                break;
+
+            case FoundXfer:
+                /* If we find a conditional jump, abort the sequence, since
+                 * handling them makes things really complicated.
+                 */
+                if (E->Info & OF_CBRA) {
+
+                    /* Switch back to searching */
+                    I = Xfer;
+                    State = Initialize;
+
+                /* Does this insn use the target register of the transfer? */
+                } else if ((E->Use & XferEntry->Chg) != 0) {
+
+                    /* It it's a store instruction, and the block is a basic
+                     * block, proceed. Otherwise restart
+                     */
+                    if ((E->Info & OF_STORE) != 0       &&
+                        CS_IsBasicBlock (S, Xfer, I)) {
+                        Store = I;
+                        StoreEntry = E;
+                        State = FoundStore;
+                    } else {
+                        I = Xfer;
+                        State = Initialize;
+                    }
+
+                /* Does this insn change the target register of the transfer? */
+                } else if (E->Chg & XferEntry->Chg) {
+
+                    /* We *may* add code here to remove the transfer, but I'm
+                     * currently not sure about the consequences, so I won't
+                     * do that and bail out instead.
+                     */
+                    I = Xfer;
+                    State = Initialize;
+
+                /* Does this insn have a label? */
+                } else if (CE_HasLabel (E)) {
+
+                    /* Too complex to handle - bail out */
+                    I = Xfer;
+                    State = Initialize;
+
+                } else {
+                    /* Track used registers */
+                    UsedRegs |= E->Use;
+                }
+                break;
+
+            case FoundStore:
+                /* We are at the instruction behind the store. If the register
+                 * isn't used later, and we have an address mode match, we can
+                 * replace the transfer by a store and remove the store here.
+                 */
+                if ((GetRegInfo (S, I, XferEntry->Chg) & XferEntry->Chg) == 0   &&
+                    (StoreEntry->AM == AM65_ABS         ||
+                     StoreEntry->AM == AM65_ZP)                                 &&
+                    (StoreEntry->AM != AM65_ZP ||
+                     (StoreEntry->Chg & UsedRegs) == 0)                         &&
+                    !MemAccess (S, Xfer+1, Store-1, StoreEntry)) {
+
+                    /* Generate the replacement store insn */
+                    CodeEntry* X = 0;
+                    switch (XferEntry->OPC) {
+
+                        case OP65_TXA:
+                            X = NewCodeEntry (OP65_STX,
+                                              StoreEntry->AM,
+                                              StoreEntry->Arg,
+                                              0,
+                                              StoreEntry->LI);
+                            break;
+
+                        case OP65_TAX:
+                            X = NewCodeEntry (OP65_STA,
+                                              StoreEntry->AM,
+                                              StoreEntry->Arg,
+                                              0,
+                                              StoreEntry->LI);
+                            break;
+
+                        case OP65_TYA:
+                            X = NewCodeEntry (OP65_STY,
+                                              StoreEntry->AM,
+                                              StoreEntry->Arg,
+                                              0,
+                                              StoreEntry->LI);
+                            break;
+
+                        case OP65_TAY:
+                            X = NewCodeEntry (OP65_STA,
+                                              StoreEntry->AM,
+                                              StoreEntry->Arg,
+                                              0,
+                                              StoreEntry->LI);
+                            break;
+
+                        default:
+                            break;
+                    }
+
+                    /* If we have a replacement store, change the code */
+                    if (X) {
+                        /* Insert after the xfer insn */
+                        CS_InsertEntry (S, X, Xfer+1);
+
+                        /* Remove the xfer instead */
+                        CS_DelEntry (S, Xfer);
+
+                        /* Remove the final store */
+                        CS_DelEntry (S, Store);
+
+                        /* Correct I so we continue with the next insn */
+                        I -= 2;
+
+                        /* Remember we had changes */
+                        ++Changes;
+                    } else {
+                        /* Restart after last xfer insn */
+                        I = Xfer;
+                    }
+                } else {
+                    /* Restart after last xfer insn */
+                    I = Xfer;
+                }
+                State = Initialize;
+                break;
+
+        }
+
+       /* Next entry */
+       ++I;
+    }
+
+    /* Return the number of changes made */
+    return Changes;
+}
+
+
+
+unsigned OptTransfers4 (CodeSeg* S)
+/* Replace a load of a register followed by a transfer insn of the same register
+ * by a load of the second register if possible.
+ */
+{
+    unsigned Changes      = 0;
+    unsigned Load         = 0;  /* Index of load insn */
+    unsigned Xfer         = 0;  /* Index of transfer insn */
+    CodeEntry* LoadEntry  = 0;  /* Pointer to load insn */
+    CodeEntry* XferEntry  = 0;  /* Pointer to xfer insn */
+
+    enum {
+        Search,
+        FoundLoad,
+        FoundXfer
+    } State = Search;
+
+    /* Walk over the entries. Look for a load instruction that is followed by
+     * a load later.
+     */
+    unsigned I = 0;
+    while (I < CS_GetEntryCount (S)) {
+
+       /* Get next entry */
+               CodeEntry* E = CS_GetEntry (S, I);
+
+        switch (State) {
+
+            case Search:
+                if (E->Info & OF_LOAD) {
+                    /* Found start of sequence */
+                    Load = I;
+                    LoadEntry = E;
+                    State = FoundLoad;
+                }
+                break;
+
+            case FoundLoad:
+                /* If we find a conditional jump, abort the sequence, since
+                 * handling them makes things really complicated.
+                 */
+                if (E->Info & OF_CBRA) {
+
+                    /* Switch back to searching */
+                    I = Load;
+                    State = Search;
+
+                /* Does this insn use the target register of the load? */
+                } else if ((E->Use & LoadEntry->Chg) != 0) {
+
+                    /* It it's a xfer instruction, and the block is a basic
+                     * block, proceed. Otherwise restart
+                     */
+                    if ((E->Info & OF_XFR) != 0       &&
+                        CS_IsBasicBlock (S, Load, I)) {
+                        Xfer = I;
+                        XferEntry = E;
+                        State = FoundXfer;
+                    } else {
+                        I = Load;
+                        State = Search;
+                    }
+
+                /* Does this insn change the target register of the load? */
+                } else if (E->Chg & LoadEntry->Chg) {
+
+                    /* We *may* add code here to remove the load, but I'm
+                     * currently not sure about the consequences, so I won't
+                     * do that and bail out instead.
+                     */
+                    I = Load;
+                    State = Search;
+                }
+                break;
+
+            case FoundXfer:
+                /* We are at the instruction behind the xfer. If the register
+                 * isn't used later, and we have an address mode match, we can
+                 * replace the transfer by a load and remove the initial load.
+                 */
+                if ((GetRegInfo (S, I, LoadEntry->Chg) & LoadEntry->Chg) == 0   &&
+                    (LoadEntry->AM == AM65_ABS          ||
+                     LoadEntry->AM == AM65_ZP           ||
+                     LoadEntry->AM == AM65_IMM)                                 &&
+                    !MemAccess (S, Load+1, Xfer-1, LoadEntry)) {
+
+                    /* Generate the replacement load insn */
+                    CodeEntry* X = 0;
+                    switch (XferEntry->OPC) {
+
+                        case OP65_TXA:
+                        case OP65_TYA:
+                            X = NewCodeEntry (OP65_LDA,
+                                              LoadEntry->AM,
+                                              LoadEntry->Arg,
+                                              0,
+                                              LoadEntry->LI);
+                            break;
+
+                        case OP65_TAX:
+                            X = NewCodeEntry (OP65_LDX,
+                                              LoadEntry->AM,
+                                              LoadEntry->Arg,
+                                              0,
+                                              LoadEntry->LI);
+                            break;
+
+                        case OP65_TAY:
+                            X = NewCodeEntry (OP65_LDY,
+                                              LoadEntry->AM,
+                                              LoadEntry->Arg,
+                                              0,
+                                              LoadEntry->LI);
+                            break;
+
+                        default:
+                            break;
+                    }
+
+                    /* If we have a replacement load, change the code */
+                    if (X) {
+                        /* Insert after the xfer insn */
+                        CS_InsertEntry (S, X, Xfer+1);
+
+                        /* Remove the xfer instead */
+                        CS_DelEntry (S, Xfer);
+
+                        /* Remove the initial load */
+                        CS_DelEntry (S, Load);
+
+                        /* Correct I so we continue with the next insn */
+                        I -= 2;
+
+                        /* Remember we had changes */
+                        ++Changes;
+                    } else {
+                        /* Restart after last xfer insn */
+                        I = Xfer;
+                    }
+                } else {
+                    /* Restart after last xfer insn */
+                    I = Xfer;
+                }
+                State = Search;
+                break;
+
+        }
+
+       /* Next entry */
+       ++I;
+    }
+
+    /* Return the number of changes made */
+    return Changes;
+}
+
+
+
 unsigned OptPushPop (CodeSeg* S)
 /* Remove a PHA/PLA sequence were A is not used later */
 {
     unsigned Changes = 0;
     unsigned Push    = 0;       /* Index of push insn */
     unsigned Pop     = 0;       /* Index of pop insn */
+    unsigned ChgA    = 0;       /* Flag for A changed */
     enum {
         Searching,
         FoundPush,
@@ -1234,6 +1860,8 @@ unsigned OptPushPop (CodeSeg* S)
     unsigned I = 0;
     while (I < CS_GetEntryCount (S)) {
 
+        CodeEntry* X;
+
        /* Get next entry */
                CodeEntry* E = CS_GetEntry (S, I);
 
@@ -1243,6 +1871,7 @@ unsigned OptPushPop (CodeSeg* S)
                 if (E->OPC == OP65_PHA) {
                     /* Found start of sequence */
                     Push  = I;
+                    ChgA  = 0;
                     State = FoundPush;
                 }
                 break;
@@ -1251,29 +1880,67 @@ unsigned OptPushPop (CodeSeg* S)
                 if (E->OPC == OP65_PHA) {
                     /* Inner push/pop, restart */
                     Push = I;
+                    ChgA = 0;
                 } else if (E->OPC == OP65_PLA) {
                     /* Found a matching pop */
                     Pop = I;
-                    State = FoundPop;
+                    /* Check that the block between Push and Pop is a basic
+                     * block (one entry, one exit). Otherwise ignore it.
+                     */
+                    if (CS_IsBasicBlock (S, Push, Pop)) {
+                        State = FoundPop;
+                    } else {
+                        /* Go into searching mode again */
+                        State = Searching;
+                    }
+                } else if (E->Chg & REG_A) {
+                    ChgA = 1;
                 }
                 break;
 
             case FoundPop:
-                /* Next insn, just check if it is no conditional branch and
-                 * that A is not used later. Check also that the range we have
-                 * found now is a basic block, which means that the PHA is the
-                 * only entrance and the PLA the only exit.
+                /* We're at the instruction after the PLA.
+                 * Check for the following conditions:
+                 *   - If this instruction is a store of A, and A is not used
+                 *     later, we may replace the PHA by the store and remove
+                 *     pla if several other conditions are met.
+                 *   - If this instruction is not a conditional branch, and A
+                 *     is either unused later, or not changed by the code
+                 *     between push and pop, we may remove PHA and PLA.
                  */
-                if ((E->Info & OF_CBRA) == 0    &&
-                    !RegAUsed (S, I)            &&
-                    CS_IsBasicBlock (S, Push, Pop)) {
+                if (E->OPC == OP65_STA                  &&
+                    !RegAUsed (S, I+1)                  &&
+                    !MemAccess (S, Push+1, Pop-1, E)) {
+
+                    /* Insert a STA after the PHA */
+                    X = NewCodeEntry (E->OPC, E->AM, E->Arg, E->JumpTo, E->LI);
+                    CS_InsertEntry (S, X, Push+1);
+
+                    /* Remove the PHA instead */
+                    CS_DelEntry (S, Push);
+
+                    /* Remove the PLA/STA sequence */
+                    CS_DelEntries (S, Pop, 2);
+
+                    /* Correct I so we continue with the next insn */
+                    I -= 2;
+
+                    /* Remember we had changes */
+                    ++Changes;
+
+                } else if ((E->Info & OF_CBRA) == 0     &&
+                           (!RegAUsed (S, I) || !ChgA)) {
+
                     /* We can remove the PHA and PLA instructions */
                     CS_DelEntry (S, Pop);
                     CS_DelEntry (S, Push);
+
                     /* Correct I so we continue with the next insn */
                     I -= 2;
+
                     /* Remember we had changes */
                     ++Changes;
+
                 }
                 /* Go into search mode again */
                 State = Searching;
@@ -1309,11 +1976,13 @@ unsigned OptPrecalc (CodeSeg* S)
        /* Get next entry */
                CodeEntry* E = CS_GetEntry (S, I);
 
-               /* Get a pointer to the output registers of the insn */
+               /* Get pointers to the input and output registers of the insn */
                const RegContents* Out = &E->RI->Out;
+        const RegContents* In  = &E->RI->In;
 
-        /* Argument for LDA and flag */
+        /* Argument for LDn and flag */
         const char* Arg = 0;
+        opc_t OPC = OP65_LDA;
 
         /* Handle the different instructions */
         switch (E->OPC) {
@@ -1327,37 +1996,57 @@ unsigned OptPrecalc (CodeSeg* S)
 
             case OP65_LDX:
                 if (E->AM != AM65_IMM && RegValIsKnown (Out->RegX)) {
-                    /* Result of load is known */
+                    /* Result of load is known but register is X */
                     Arg = MakeHexArg (Out->RegX);
+                    OPC = OP65_LDX;
                 }
                 break;
 
             case OP65_LDY:
                 if (E->AM != AM65_IMM && RegValIsKnown (Out->RegY)) {
-                    /* Result of load is known */
+                    /* Result of load is known but register is Y */
                     Arg = MakeHexArg (Out->RegY);
+                    OPC = OP65_LDY;
                 }
                 break;
 
-            case OP65_ADC:
-            case OP65_ASL:
             case OP65_EOR:
-            case OP65_LSR:
-            case OP65_SBC:
                 if (RegValIsKnown (Out->RegA)) {
                     /* Accu op zp with known contents */
                     Arg = MakeHexArg (Out->RegA);
                 }
                 break;
 
+            case OP65_ADC:
+            case OP65_SBC:
+                /* If this is an operation with an immediate operand of zero,
+                 * and the register is zero, the operation won't give us any
+                 * results we don't already have (including the flags), so
+                 * remove it. Something like this is generated as a result of
+                 * a compare where parts of the values are known to be zero.
+                 */
+                if (In->RegA == 0 && CE_IsKnownImm (E, 0x00)) {
+                    /* 0-0 or 0+0 -> remove */
+                    CS_DelEntry (S, I);
+                    ++Changes;
+                }
+                break;
+
             case OP65_AND:
                 if (CE_IsKnownImm (E, 0xFF)) {
                     /* AND with 0xFF, remove */
                     CS_DelEntry (S, I);
                     ++Changes;
+                } else if (CE_IsKnownImm (E, 0x00)) {
+                    /* AND with 0x00, replace by lda #$00 */
+                    Arg = MakeHexArg (0x00);
                 } else if (RegValIsKnown (Out->RegA)) {
                     /* Accu AND zp with known contents */
                     Arg = MakeHexArg (Out->RegA);
+                } else if (In->RegA == 0xFF) {
+                    /* AND but A contains 0xFF - replace by lda */
+                    CE_ReplaceOPC (E, OP65_LDA);
+                    ++Changes;
                 }
                 break;
 
@@ -1366,9 +2055,16 @@ unsigned OptPrecalc (CodeSeg* S)
                     /* ORA with zero, remove */
                     CS_DelEntry (S, I);
                     ++Changes;
+                } else if (CE_IsKnownImm (E, 0xFF)) {
+                    /* ORA with 0xFF, replace by lda #$ff */
+                    Arg = MakeHexArg (0xFF);
                 } else if (RegValIsKnown (Out->RegA)) {
                     /* Accu AND zp with known contents */
                     Arg = MakeHexArg (Out->RegA);
+                } else if (In->RegA == 0) {
+                    /* ORA but A contains 0x00 - replace by lda */
+                    CE_ReplaceOPC (E, OP65_LDA);
+                    ++Changes;
                 }
                 break;
 
@@ -1379,7 +2075,7 @@ unsigned OptPrecalc (CodeSeg* S)
 
         /* Check if we have to replace the insn by LDA */
         if (Arg) {
-            CodeEntry* X = NewCodeEntry (OP65_LDA, AM65_IMM, Arg, 0, E->LI);
+            CodeEntry* X = NewCodeEntry (OPC, AM65_IMM, Arg, 0, E->LI);
             CS_InsertEntry (S, X, I+1);
             CS_DelEntry (S, I);
             ++Changes;
@@ -1465,3 +2161,117 @@ unsigned OptBranchDist (CodeSeg* S)
 
 
 
+/*****************************************************************************/
+/*                          Optimize indirect loads                          */
+/*****************************************************************************/
+
+
+
+unsigned OptIndLoads1 (CodeSeg* S)
+/* Change
+ *
+ *     lda      (zp),y
+ *
+ * into
+ *
+ *     lda      (zp,x)
+ *
+ * provided that x and y are both zero.
+ */
+{
+    unsigned Changes = 0;
+    unsigned I;
+
+    /* Generate register info for this step */
+    CS_GenRegInfo (S);
+
+    /* Walk over the entries */
+    I = 0;
+    while (I < CS_GetEntryCount (S)) {
+
+       /* Get next entry */
+               CodeEntry* E = CS_GetEntry (S, I);
+
+       /* Check if it's what we're looking for */
+               if (E->OPC == OP65_LDA          &&
+            E->AM == AM65_ZP_INDY       &&
+            E->RI->In.RegY == 0         &&
+            E->RI->In.RegX == 0) {
+
+           /* Replace by the same insn with other addressing mode */
+            CodeEntry* X = NewCodeEntry (E->OPC, AM65_ZPX_IND, E->Arg, 0, E->LI);
+            CS_InsertEntry (S, X, I+1);
+
+            /* Remove the old insn */
+            CS_DelEntry (S, I);
+            ++Changes;
+       }
+
+       /* Next entry */
+       ++I;
+
+    }
+
+    /* Free register info */
+    CS_FreeRegInfo (S);
+
+    /* Return the number of changes made */
+    return Changes;
+}
+
+
+
+unsigned OptIndLoads2 (CodeSeg* S)
+/* Change
+ *
+ *     lda      (zp,x)
+ *
+ * into
+ *
+ *     lda      (zp),y
+ *
+ * provided that x and y are both zero.
+ */
+{
+    unsigned Changes = 0;
+    unsigned I;
+
+    /* Generate register info for this step */
+    CS_GenRegInfo (S);
+
+    /* Walk over the entries */
+    I = 0;
+    while (I < CS_GetEntryCount (S)) {
+
+       /* Get next entry */
+               CodeEntry* E = CS_GetEntry (S, I);
+
+       /* Check if it's what we're looking for */
+               if (E->OPC == OP65_LDA          &&
+            E->AM == AM65_ZPX_IND       &&
+            E->RI->In.RegY == 0         &&
+            E->RI->In.RegX == 0) {
+
+           /* Replace by the same insn with other addressing mode */
+            CodeEntry* X = NewCodeEntry (E->OPC, AM65_ZP_INDY, E->Arg, 0, E->LI);
+            CS_InsertEntry (S, X, I+1);
+
+            /* Remove the old insn */
+            CS_DelEntry (S, I);
+            ++Changes;
+       }
+
+       /* Next entry */
+       ++I;
+
+    }
+
+    /* Free register info */
+    CS_FreeRegInfo (S);
+
+    /* Return the number of changes made */
+    return Changes;
+}
+
+
+