X-Git-Url: https://git.sur5r.net/?a=blobdiff_plain;f=src%2Fcc65%2Fcodeopt.c;h=4797e8b518668079942a3e13b94182a977a8a7de;hb=1366b6cbea1b374fa7502354cf2cb8a971f71c5f;hp=3f1fd1913dbf153eff1f05a1eaa8c6f72c4109e7;hpb=bbebbe3a1b86ae04b49f0fc37f3b9ecc86a955fb;p=cc65 diff --git a/src/cc65/codeopt.c b/src/cc65/codeopt.c index 3f1fd1913..4797e8b51 100644 --- a/src/cc65/codeopt.c +++ b/src/cc65/codeopt.c @@ -33,17 +33,29 @@ +#include #include /* common */ #include "abend.h" +#include "chartype.h" #include "print.h" +#include "xmalloc.h" +#include "xsprintf.h" /* cc65 */ #include "asmlabel.h" #include "codeent.h" #include "codeinfo.h" +#include "coptadd.h" +#include "coptc02.h" +#include "coptcmp.h" #include "coptind.h" +#include "coptneg.h" +#include "coptstop.h" +#include "coptsub.h" +#include "copttest.h" +#include "cpu.h" #include "error.h" #include "global.h" #include "codeopt.h" @@ -51,296 +63,62 @@ /*****************************************************************************/ -/* Data */ +/* Optimize shifts */ /*****************************************************************************/ -/* Defines for the conditions in a compare */ -typedef enum { - CMP_INV = -1, - CMP_EQ, - CMP_NE, - CMP_GT, - CMP_GE, - CMP_LT, - CMP_LE, - CMP_UGT, - CMP_UGE, - CMP_ULT, - CMP_ULE -} cmp_t; - -/* Table with the compare suffixes */ -static const char CmpSuffixTab [][4] = { - "eq", "ne", "gt", "ge", "lt", "le", "ugt", "uge", "ult", "ule" -}; - -/* Table used to invert a condition, indexed by condition */ -static const unsigned char CmpInvertTab [] = { - CMP_NE, CMP_EQ, - CMP_LE, CMP_LT, CMP_GE, CMP_GT, - CMP_ULE, CMP_ULT, CMP_UGE, CMP_UGT -}; - -/* Table to show which compares are signed (use the N flag) */ -static const char CmpSignedTab [] = { - 0, 0, 1, 1, 1, 1, 0, 0, 0, 0 -}; - - - -/*****************************************************************************/ -/* Helper functions */ -/*****************************************************************************/ - - - -static cmp_t FindCmpCond (const char* Code, unsigned CodeLen) -/* Search for a compare condition by the given code using the given length */ -{ - unsigned I; - - /* Linear search */ - for (I = 0; I < sizeof (CmpSuffixTab) / sizeof (CmpSuffixTab [0]); ++I) { - if (strncmp (Code, CmpSuffixTab [I], CodeLen) == 0) { - /* Found */ - return I; - } - } - - /* Not found */ - return CMP_INV; -} - - - -static cmp_t FindBoolCmpCond (const char* Name) -/* Map a condition suffix to a code. Return the code or CMP_INV on failure */ -{ - /* Check for the correct subroutine name */ - if (strncmp (Name, "bool", 4) == 0) { - /* Name is ok, search for the code in the table */ - return FindCmpCond (Name+4, strlen(Name)-4); - } else { - /* Not found */ - return CMP_INV; - } -} - - - -static cmp_t FindTosCmpCond (const char* Name) -/* Check if this is a call to one of the TOS compare functions (tosgtax). - * Return the condition code or CMP_INV on failure. +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 Len = strlen (Name); - - /* Check for the correct subroutine name */ - if (strncmp (Name, "tos", 3) == 0 && strcmp (Name+Len-2, "ax") == 0) { - /* Name is ok, search for the code in the table */ - return FindCmpCond (Name+3, Len-3-2); - } else { - /* Not found */ - return CMP_INV; - } -} - - - -static void ReplaceCmp (CodeSeg* S, unsigned I, cmp_t Cond) -/* Helper function for the replacement of routines that return a boolean - * followed by a conditional jump. Instead of the boolean value, the condition - * codes are evaluated directly. - * I is the index of the conditional branch, the sequence is already checked - * to be correct. - */ -{ - CodeEntry* N; - CodeLabel* L; - - /* Get the entry */ - CodeEntry* E = CS_GetEntry (S, I); - - /* Replace the conditional branch */ - switch (Cond) { + unsigned Changes = 0; - case CMP_EQ: - CE_ReplaceOPC (E, OP65_JEQ); - break; + /* Walk over the entries */ + unsigned I = 0; + while (I < CS_GetEntryCount (S)) { - case CMP_NE: - CE_ReplaceOPC (E, OP65_JNE); - break; + /* Get next entry */ + CodeEntry* E = CS_GetEntry (S, I); - case CMP_GT: - /* Replace by - * beq @L - * jpl Target - * @L: ... - */ - if ((N = CS_GetNextEntry (S, I)) == 0) { - /* No such entry */ - Internal ("Invalid program flow"); - } - L = CS_GenLabel (S, N); - N = NewCodeEntry (OP65_BEQ, AM65_BRA, L->Name, L, E->LI); - CS_InsertEntry (S, N, I); - CE_ReplaceOPC (E, OP65_JPL); - break; - - case CMP_GE: - CE_ReplaceOPC (E, OP65_JPL); - break; - - case CMP_LT: - CE_ReplaceOPC (E, OP65_JMI); - break; - - case CMP_LE: - /* Replace by - * jmi Target - * jeq Target - */ - CE_ReplaceOPC (E, OP65_JMI); - L = E->JumpTo; - N = NewCodeEntry (OP65_JEQ, AM65_BRA, L->Name, L, E->LI); - CS_InsertEntry (S, N, I+1); - break; - - case CMP_UGT: - /* Replace by - * beq @L - * jcs Target - * @L: ... - */ - if ((N = CS_GetNextEntry (S, I)) == 0) { - /* No such entry */ - Internal ("Invalid program flow"); + /* 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); } - L = CS_GenLabel (S, N); - N = NewCodeEntry (OP65_BEQ, AM65_BRA, L->Name, L, E->LI); - CS_InsertEntry (S, N, I); - CE_ReplaceOPC (E, OP65_JCS); - break; - - case CMP_UGE: - CE_ReplaceOPC (E, OP65_JCS); - break; - - case CMP_ULT: - CE_ReplaceOPC (E, OP65_JCC); - break; - - case CMP_ULE: - /* Replace by - * jcc Target - * jeq Target - */ - CE_ReplaceOPC (E, OP65_JCC); - L = E->JumpTo; - N = NewCodeEntry (OP65_JEQ, AM65_BRA, L->Name, L, E->LI); - CS_InsertEntry (S, N, I+1); - break; - - default: - Internal ("Unknown jump condition: %d", Cond); - - } - -} - - - -static int IsCmpToZero (const CodeEntry* E) -/* Check if the given instrcuction is a compare to zero instruction */ -{ - return (E->OPC == OP65_CMP && - E->AM == AM65_IMM && - (E->Flags & CEF_NUMARG) != 0 && - E->Num == 0); -} + /* Delete the call to shlax */ + CS_DelEntry (S, I); + /* Remember, we had changes */ + ++Changes; -static int IsSpLoad (const CodeEntry* E) -/* Return true if this is the load of A from the stack */ -{ - return E->OPC == OP65_LDA && E->AM == AM65_ZP_INDY && strcmp (E->Arg, "sp") == 0; -} + } + /* Next entry */ + ++I; + } -static int IsLocalLoad16 (CodeSeg* S, unsigned Index, - CodeEntry** L, unsigned Count) -/* Check if a 16 bit load of a local variable follows: - * - * ldy #$xx - * lda (sp),y - * tax - * dey - * lda (sp),y - * - * If so, read Count entries following the first ldy into L and return true - * if this is possible. Otherwise return false. - */ -{ - /* Be sure we read enough entries for the check */ - CHECK (Count >= 5); - - /* Read the first entry */ - L[0] = CS_GetEntry (S, Index); - - /* Check for the sequence */ - return (L[0]->OPC == OP65_LDY && - L[0]->AM == AM65_IMM && - (L[0]->Flags & CEF_NUMARG) != 0 && - CS_GetEntries (S, L+1, Index+1, Count-1) && - IsSpLoad (L[1]) && - !CE_HasLabel (L[1]) && - L[2]->OPC == OP65_TAX && - !CE_HasLabel (L[2]) && - L[3]->OPC == OP65_DEY && - !CE_HasLabel (L[3]) && - IsSpLoad (L[4]) && - !CE_HasLabel (L[4])); + /* Return the number of changes made */ + return Changes; } -static int IsImmCmp16 (CodeSeg* S, CodeEntry** L) -/* Check if the instructions at L are an immidiate compare of a/x: - * - * - */ -{ - return (L[0]->OPC == OP65_CPX && - L[0]->AM == AM65_IMM && - (L[0]->Flags & CEF_NUMARG) != 0 && - !CE_HasLabel (L[0]) && - (L[1]->OPC == OP65_JNE || L[1]->OPC == OP65_BNE) && - L[1]->JumpTo != 0 && - !CE_HasLabel (L[1]) && - L[2]->OPC == OP65_CMP && - L[2]->AM == AM65_IMM && - (L[2]->Flags & CEF_NUMARG) != 0 && - (L[3]->Info & OF_ZBRA) != 0 && - L[3]->JumpTo != 0 && - (L[1]->JumpTo->Owner == L[3] || L[1]->JumpTo == L[3]->JumpTo)); -} - - - -/*****************************************************************************/ -/* Remove calls to the bool transformer subroutines */ -/*****************************************************************************/ - - - -static unsigned OptBoolTransforms (CodeSeg* S) -/* Try to remove the call to boolean transformer routines where the call is - * not really needed. +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 not used later. */ { unsigned Changes = 0; @@ -349,33 +127,24 @@ static unsigned OptBoolTransforms (CodeSeg* S) unsigned I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* N; - cmp_t Cond; - /* Get next entry */ CodeEntry* E = CS_GetEntry (S, I); - /* Check for a boolean transformer */ - if (E->OPC == OP65_JSR && - (Cond = FindBoolCmpCond (E->Arg)) != CMP_INV && - (N = CS_GetNextEntry (S, I)) != 0 && - (N->Info & OF_ZBRA) != 0) { - - /* Make the boolean transformer unnecessary by changing the - * the conditional jump to evaluate the condition flags that - * are set after the compare directly. Note: jeq jumps if - * the condition is not met, jne jumps if the condition is met. - * Invert the code if we jump on condition not met. - */ - if (GetBranchCond (N->OPC) == BC_EQ) { - /* Jumps if condition false, invert condition */ - Cond = CmpInvertTab [Cond]; - } - - /* Check if we can replace the code by something better */ - ReplaceCmp (S, I+1, Cond); + /* Check for the sequence */ + if (E->OPC == OP65_JSR && + strncmp (E->Arg, "shrax", 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_LSR, AM65_ACC, "a", 0, E->LI); + CS_InsertEntry (S, X, I+1); + } - /* Remove the call to the bool transformer */ + /* Delete the call to shlax */ CS_DelEntry (S, I); /* Remember, we had changes */ @@ -395,79 +164,69 @@ static unsigned OptBoolTransforms (CodeSeg* S) /*****************************************************************************/ -/* Optimize subtractions */ +/* Optimize stores through pointers */ /*****************************************************************************/ -static unsigned OptSub1 (CodeSeg* S) -/* Search for the sequence - * - * sbc ... - * bcs L - * dex - * L: - * - * and remove the handling of the high byte if X is not used later. - */ +static unsigned OptPtrStore1Sub (CodeSeg* S, unsigned I, CodeEntry** const L) +/* Check if this is one of the allowed suboperation for OptPtrStore1 */ { - unsigned Changes = 0; - - /* Walk over the entries */ - unsigned I = 0; - while (I < CS_GetEntryCount (S)) { - - CodeEntry* L[3]; - - /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); - - /* Check for the sequence */ - if (E->OPC == OP65_SBC && - CS_GetEntries (S, L, I+1, 3) && - (L[0]->OPC == OP65_BCS || L[0]->OPC == OP65_JCS) && - L[0]->JumpTo != 0 && - !CE_HasLabel (L[0]) && - L[1]->OPC == OP65_DEX && - !CE_HasLabel (L[1]) && - L[0]->JumpTo->Owner == L[2] && - !RegXUsed (S, I+3)) { - - /* Remove the bcs/dex */ - CS_DelEntries (S, I+1, 2); - - /* Remember, we had changes */ - ++Changes; + /* 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; + } - /* Next entry */ - ++I; - } - /* Return the number of changes made */ - return Changes; + /* Not found */ + return 0; } -static unsigned OptSub2 (CodeSeg* S) -/* Search for the sequence +static unsigned OptPtrStore1 (CodeSeg* S) +/* Search for the sequence: * - * lda xx - * sec - * sta tmp1 - * lda yy - * sbc tmp1 - * sta yy + * jsr pushax + * ldy xxx + * jsr ldauidx + * subop + * ldy yyy + * jsr staspidx * - * and replace it by + * and replace it by: * - * sec - * lda yy - * sbc xx - * sta yy + * sta ptr1 + * stx ptr1+1 + * ldy xxx + * ldx #$00 + * lda (ptr1),y + * subop + * ldy yyy + * sta (ptr1),y */ { unsigned Changes = 0; @@ -476,49 +235,58 @@ static unsigned OptSub2 (CodeSeg* S) unsigned I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* L[5]; + unsigned K; + CodeEntry* L[10]; /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); + L[0] = CS_GetEntry (S, I); /* Check for the sequence */ - if (E->OPC == OP65_LDA && - CS_GetEntries (S, L, I+1, 5) && - L[0]->OPC == OP65_SEC && - !CE_HasLabel (L[0]) && - L[1]->OPC == OP65_STA && - strcmp (L[1]->Arg, "tmp1") == 0 && - !CE_HasLabel (L[1]) && - L[2]->OPC == OP65_LDA && - !CE_HasLabel (L[2]) && - L[3]->OPC == OP65_SBC && - strcmp (L[3]->Arg, "tmp1") == 0 && - !CE_HasLabel (L[3]) && - L[4]->OPC == OP65_STA && - strcmp (L[4]->Arg, L[2]->Arg) == 0 && - !CE_HasLabel (L[4])) { - - /* Remove the store to tmp1 */ - CS_DelEntry (S, I+2); + if (CE_IsCall (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_IsCall (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_IsCall (L[4+K], "staspidx") && + !CE_HasLabel (L[4+K])) { + + CodeEntry* X; + + /* Create and insert the stores */ + 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); - /* Remove the subtraction */ + /* Delete the call to pushax */ + CS_DelEntry (S, I); + + /* Delete the call to ldauidx */ CS_DelEntry (S, I+3); - /* Move the lda to the position of the subtraction and change the - * op to SBC. - */ - CS_MoveEntry (S, I, I+3); - CE_ReplaceOPC (E, OP65_SBC); + /* Insert the load from ptr1 */ + X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[3]->LI); + CS_InsertEntry (S, X, I+3); + X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, "ptr1", 0, L[2]->LI); + CS_InsertEntry (S, X, I+4); - /* If the sequence head had a label, move this label back to the - * head. - */ - if (CE_HasLabel (E)) { - CS_MoveLabels (S, E, L[0]); - } + /* Insert the store through ptr1 */ + X = NewCodeEntry (OP65_STA, AM65_ZP_INDY, "ptr1", 0, L[3]->LI); + CS_InsertEntry (S, X, I+6+K); + + /* Delete the call to staspidx */ + CS_DelEntry (S, I+7+K); /* Remember, we had changes */ - ++Changes; + ++Changes; } @@ -533,21 +301,21 @@ static unsigned OptSub2 (CodeSeg* S) -/*****************************************************************************/ -/* Optimize additions */ -/*****************************************************************************/ - - - -static unsigned OptAdd1 (CodeSeg* S) -/* Search for the sequence +static unsigned OptPtrStore2 (CodeSeg* S) +/* Search for the sequence: * - * adc ... - * bcc L - * inx - * L: + * jsr pushax + * lda xxx + * ldy yyy + * jsr staspidx + * + * and replace it by: * - * and remove the handling of the high byte if X is not used later. + * sta ptr1 + * stx ptr1+1 + * lda xxx + * ldy yyy + * sta (ptr1),y */ { unsigned Changes = 0; @@ -556,24 +324,39 @@ static unsigned OptAdd1 (CodeSeg* S) unsigned I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* L[3]; + CodeEntry* L[4]; /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); + L[0] = CS_GetEntry (S, I); /* Check for the sequence */ - if (E->OPC == OP65_ADC && - CS_GetEntries (S, L, I+1, 3) && - (L[0]->OPC == OP65_BCC || L[0]->OPC == OP65_JCC) && - L[0]->JumpTo != 0 && - !CE_HasLabel (L[0]) && - L[1]->OPC == OP65_INX && - !CE_HasLabel (L[1]) && - L[0]->JumpTo->Owner == L[2] && - !RegXUsed (S, I+3)) { - - /* Remove the bcs/dex */ - CS_DelEntries (S, I+1, 2); + if (CE_IsCall (L[0], "pushax") && + CS_GetEntries (S, L+1, I+1, 3) && + L[1]->OPC == OP65_LDA && + !CE_HasLabel (L[1]) && + L[2]->OPC == OP65_LDY && + !CE_HasLabel (L[2]) && + CE_IsCall (L[3], "staspidx") && + !CE_HasLabel (L[3])) { + + CodeEntry* X; + + /* Create and insert the stores */ + 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 the call to pushax */ + CS_DelEntry (S, I); + + /* Insert the store through ptr1 */ + X = NewCodeEntry (OP65_STA, AM65_ZP_INDY, "ptr1", 0, L[3]->LI); + CS_InsertEntry (S, X, I+4); + + /* Delete the call to staspidx */ + CS_DelEntry (S, I+5); /* Remember, we had changes */ ++Changes; @@ -592,22 +375,29 @@ static unsigned OptAdd1 (CodeSeg* S) /*****************************************************************************/ -/* Optimizations for compares */ +/* Optimize loads through pointers */ /*****************************************************************************/ -static unsigned OptCmp1 (CodeSeg* S) -/* Search for the sequence +static unsigned OptPtrLoad1 (CodeSeg* S) +/* Search for the sequence: * - * stx xx - * stx tmp1 - * ora tmp1 + * tax + * dey + * lda (sp),y # May be any destination + * ldy ... + * jsr ldauidx * - * and replace it by + * and replace it by: * - * stx xx - * ora xx + * sta ptr1+1 + * dey + * lda (sp),y + * sta ptr1 + * ldy ... + * ldx #$00 + * lda (ptr1),y */ { unsigned Changes = 0; @@ -616,26 +406,42 @@ static unsigned OptCmp1 (CodeSeg* S) unsigned I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* L[2]; + CodeEntry* L[5]; /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); + L[0] = CS_GetEntry (S, I); /* Check for the sequence */ - if (E->OPC == OP65_STX && - CS_GetEntries (S, L, I+1, 2) && - L[0]->OPC == OP65_STX && - strcmp (L[0]->Arg, "tmp1") == 0 && - !CE_HasLabel (L[0]) && - L[1]->OPC == OP65_ORA && - strcmp (L[1]->Arg, "tmp1") == 0 && - !CE_HasLabel (L[1])) { + if (L[0]->OPC == OP65_TAX && + CS_GetEntries (S, L+1, I+1, 4) && + L[1]->OPC == OP65_DEY && + !CE_HasLabel (L[1]) && + L[2]->OPC == OP65_LDA && + !CE_HasLabel (L[2]) && + L[3]->OPC == OP65_LDY && + !CE_HasLabel (L[3]) && + CE_IsCall (L[4], "ldauidx") && + !CE_HasLabel (L[4])) { + + CodeEntry* X; + + /* Store the high byte and remove the TAX instead */ + X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1+1", 0, L[0]->LI); + CS_InsertEntry (S, X, I+1); + CS_DelEntry (S, I); - /* Remove the remaining instructions */ - CS_DelEntries (S, I+1, 2); + /* Store the low byte */ + X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1", 0, L[2]->LI); + CS_InsertEntry (S, X, I+3); - /* Insert the ora instead */ - CS_InsertEntry (S, NewCodeEntry (OP65_ORA, E->AM, E->Arg, 0, E->LI), I+1); + /* Delete the call to ldauidx */ + CS_DelEntry (S, I+5); + + /* Load high and low byte */ + X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[3]->LI); + CS_InsertEntry (S, X, I+5); + X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, "ptr1", 0, L[3]->LI); + CS_InsertEntry (S, X, I+6); /* Remember, we had changes */ ++Changes; @@ -653,14 +459,30 @@ static unsigned OptCmp1 (CodeSeg* S) -static unsigned OptCmp2 (CodeSeg* S) -/* Search for +static unsigned OptPtrLoad2 (CodeSeg* S) +/* Search for the sequence: + * + * clc + * adc xxx + * tay + * txa + * adc yyy + * tax + * tya + * ldy + * jsr ldauidx * - * lda/and/ora/eor ... - * cmp #$00 - * jeq/jne + * and replace it by: * - * and remove the cmp. + * clc + * adc xxx + * sta ptr1 + * txa + * adc yyy + * sta ptr1+1 + * ldy + * ldx #$00 + * lda (ptr1),y */ { unsigned Changes = 0; @@ -669,31 +491,67 @@ static unsigned OptCmp2 (CodeSeg* S) unsigned I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* L[2]; + CodeEntry* L[9]; /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); + L[0] = CS_GetEntry (S, I); /* Check for the sequence */ - if ((E->OPC == OP65_ADC || - E->OPC == OP65_AND || - E->OPC == OP65_DEA || - E->OPC == OP65_EOR || - E->OPC == OP65_INA || - E->OPC == OP65_LDA || - E->OPC == OP65_ORA || - E->OPC == OP65_PLA || - E->OPC == OP65_SBC || - E->OPC == OP65_TXA || - E->OPC == OP65_TYA) && - CS_GetEntries (S, L, I+1, 2) && - IsCmpToZero (L[0]) && - !CE_HasLabel (L[0]) && - (L[1]->Info & OF_FBRA) != 0 && - !CE_HasLabel (L[1])) { - - /* Remove the compare */ - CS_DelEntry (S, I+1); + if (L[0]->OPC == OP65_CLC && + CS_GetEntries (S, L+1, I+1, 8) && + L[1]->OPC == OP65_ADC && + !CE_HasLabel (L[1]) && + L[2]->OPC == OP65_TAY && + !CE_HasLabel (L[2]) && + L[3]->OPC == OP65_TXA && + !CE_HasLabel (L[3]) && + L[4]->OPC == OP65_ADC && + !CE_HasLabel (L[4]) && + L[5]->OPC == OP65_TAX && + !CE_HasLabel (L[5]) && + L[6]->OPC == OP65_TYA && + !CE_HasLabel (L[6]) && + L[7]->OPC == OP65_LDY && + !CE_HasLabel (L[7]) && + CE_IsCall (L[8], "ldauidx") && + !CE_HasLabel (L[8])) { + + CodeEntry* X; + CodeEntry* P; + + /* Store the low byte and remove the TAY instead */ + X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1", 0, L[1]->LI); + CS_InsertEntry (S, X, I+2); + CS_DelEntry (S, I+3); + + /* Store the high byte */ + X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1+1", 0, L[4]->LI); + CS_InsertEntry (S, X, I+5); + + /* If the instruction before the adc is a ldx, replace the + * txa by and lda with the same location of the ldx. + */ + 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); + CS_InsertEntry (S, X, I+4); + CS_DelEntry (S, I+3); + } + + /* Delete more transfer insns */ + CS_DelEntry (S, I+7); + CS_DelEntry (S, I+6); + + /* Delete the call to ldauidx */ + CS_DelEntry (S, I+7); + + /* Load high and low byte */ + X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[7]->LI); + CS_InsertEntry (S, X, I+7); + X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, "ptr1", 0, L[7]->LI); + CS_InsertEntry (S, X, I+8); /* Remember, we had changes */ ++Changes; @@ -711,25 +569,30 @@ static unsigned OptCmp2 (CodeSeg* S) -static unsigned OptCmp3 (CodeSeg* S) -/* Search for - * - * lda x - * ldx y - * cpx #a - * bne L1 - * cmp #b - * jne/jeq L2 +static unsigned OptPtrLoad3 (CodeSeg* S) +/* Search for the sequence: * - * If a is zero, we may remove the compare. If a and b are both zero, we may - * replace it by the sequence + * adc xxx + * pha + * txa + * iny + * adc yyy + * tax + * pla + * ldy + * jsr ldauidx * - * lda x - * ora x+1 - * jne/jeq ... + * and replace it by: * - * L1 may be either the label at the branch instruction, or the target label - * of this instruction. + * adc xxx + * sta ptr1 + * txa + * iny + * adc yyy + * sta ptr1+1 + * ldy + * ldx #$00 + * lda (ptr1),y */ { unsigned Changes = 0; @@ -738,43 +601,58 @@ static unsigned OptCmp3 (CodeSeg* S) unsigned I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* L[5]; + CodeEntry* L[9]; /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); + L[0] = CS_GetEntry (S, I); /* Check for the sequence */ - if (E->OPC == OP65_LDA && - CS_GetEntries (S, L, I+1, 5) && - L[0]->OPC == OP65_LDX && - !CE_HasLabel (L[0]) && - IsImmCmp16 (S, L+1)) { - - if (L[1]->Num == 0 && L[3]->Num == 0) { - /* The value is zero, we may use the simple code version. */ - CE_ReplaceOPC (L[0], OP65_ORA); - CS_DelEntries (S, I+2, 3); - } else { - /* Move the lda instruction after the first branch. This will - * improve speed, since the load is delayed after the first - * test. - */ - CS_MoveEntry (S, I, I+4); - - /* We will replace the ldx/cpx by lda/cmp */ - CE_ReplaceOPC (L[0], OP65_LDA); - CE_ReplaceOPC (L[1], OP65_CMP); - - /* Beware: If the first LDA instruction had a label, we have - * to move this label to the top of the sequence again. - */ - if (CE_HasLabel (E)) { - CS_MoveLabels (S, E, L[0]); - } + if (L[0]->OPC == OP65_ADC && + CS_GetEntries (S, L+1, I+1, 8) && + L[1]->OPC == OP65_PHA && + !CE_HasLabel (L[1]) && + L[2]->OPC == OP65_TXA && + !CE_HasLabel (L[2]) && + L[3]->OPC == OP65_INY && + !CE_HasLabel (L[3]) && + L[4]->OPC == OP65_ADC && + !CE_HasLabel (L[4]) && + L[5]->OPC == OP65_TAX && + !CE_HasLabel (L[5]) && + L[6]->OPC == OP65_PLA && + !CE_HasLabel (L[6]) && + L[7]->OPC == OP65_LDY && + !CE_HasLabel (L[7]) && + CE_IsCall (L[8], "ldauidx") && + !CE_HasLabel (L[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); + CS_DelEntry (S, I+2); + + /* Store the high byte */ + X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1+1", 0, L[4]->LI); + CS_InsertEntry (S, X, I+5); + + /* Delete more transfer and PLA insns */ + CS_DelEntry (S, I+7); + CS_DelEntry (S, I+6); + + /* Delete the call to ldauidx */ + CS_DelEntry (S, I+7); + /* Load high and low byte */ + X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[6]->LI); + CS_InsertEntry (S, X, I+7); + X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, "ptr1", 0, L[6]->LI); + CS_InsertEntry (S, X, I+8); + + /* Remember, we had changes */ ++Changes; + } /* Next entry */ @@ -788,18 +666,23 @@ static unsigned OptCmp3 (CodeSeg* S) -static unsigned OptCmp4 (CodeSeg* S) -/* Optimize compares of local variables: +static unsigned OptPtrLoad4 (CodeSeg* S) +/* Search for the sequence: * - * ldy #o - * lda (sp),y - * tax - * dey - * lda (sp),y - * cpx #a - * bne L1 - * cmp #b - * jne/jeq L2 + * 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; @@ -808,46 +691,67 @@ static unsigned OptCmp4 (CodeSeg* S) unsigned I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* L[9]; + CodeEntry* L[8]; + unsigned Len; + + /* Get next entry */ + L[0] = CS_GetEntry (S, I); /* Check for the sequence */ - if (IsLocalLoad16 (S, I, L, 9) && IsImmCmp16 (S, L+5)) { - - if (L[5]->Num == 0 && L[7]->Num == 0) { - - /* The value is zero, we may use the simple code version: - * ldy #o - * lda (sp),y - * dey - * ora (sp),y - * jne/jeq ... - */ - CE_ReplaceOPC (L[4], OP65_ORA); - CS_DelEntries (S, I+5, 3); /* cpx/bne/cmp */ - CS_DelEntry (S, I+2); /* tax */ - - } else { - - /* Change the code to just use the A register. Move the load - * of the low byte after the first branch if possible: - * - * ldy #o - * lda (sp),y - * cmp #a - * bne L1 - * dey - * lda (sp),y - * cmp #b - * jne/jeq ... - */ - CS_DelEntry (S, I+2); /* tax */ - CE_ReplaceOPC (L[5], OP65_CMP); /* cpx -> cmp */ - CS_MoveEntry (S, I+4, I+2); /* cmp */ - CS_MoveEntry (S, I+5, I+3); /* bne */ + 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 && + !CE_HasLabel (L[1]) && + L[2]->OPC == OP65_CLC && + !CE_HasLabel (L[2]) && + L[3]->OPC == OP65_ADC && + (L[3]->AM == AM65_ABS || L[3]->AM == AM65_ZP) && + !CE_HasLabel (L[3]) && + (L[4]->OPC == OP65_BCC || L[4]->OPC == OP65_JCC) && + L[4]->JumpTo != 0 && + L[4]->JumpTo->Owner == L[6] && + !CE_HasLabel (L[4]) && + L[5]->OPC == OP65_INX && + !CE_HasLabel (L[5]) && + L[6]->OPC == OP65_LDY && + CE_KnownImm (L[6]) && + L[6]->Num == 0 && + CE_IsCall (L[7], "ldauidx") && + !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, "$00", 0, L[0]->LI); + CS_InsertEntry (S, X, I+9); + + Label = memcpy (xmalloc (Len-2), L[0]->Arg+2, Len-3); + Label[Len-3] = '\0'; + X = NewCodeEntry (OP65_LDA, AM65_ABSY, Label, 0, L[0]->LI); + CS_InsertEntry (S, X, I+10); + xfree (Label); + + /* Remove the old code */ + CS_DelEntries (S, I, 8); + + /* Remember, we had changes */ ++Changes; + } /* Next entry */ @@ -861,11 +765,26 @@ static unsigned OptCmp4 (CodeSeg* S) -static unsigned OptCmp5 (CodeSeg* S) -/* Search for calls to compare subroutines followed by a conditional branch - * and replace them by cheaper versions, since the branch means that the - * boolean value returned by these routines is not needed (we may also check - * that explicitly, but for the current code generator it is always true). +static unsigned OptPtrLoad5 (CodeSeg* S) +/* Search for the sequence: + * + * lda #<(label+0) + * ldx #>(label+0) + * ldy #$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; @@ -874,45 +793,78 @@ static unsigned OptCmp5 (CodeSeg* S) unsigned I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* N; - cmp_t Cond; + CodeEntry* L[9]; + unsigned Len; /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); + L[0] = CS_GetEntry (S, I); /* Check for the sequence */ - if (E->OPC == OP65_JSR && - (Cond = FindTosCmpCond (E->Arg)) != CMP_INV && - (N = CS_GetNextEntry (S, I)) != 0 && - (N->Info & OF_ZBRA) != 0 && - !CE_HasLabel (N)) { - - /* The tos... functions will return a boolean value in a/x and - * the Z flag says if this value is zero or not. We will call - * a cheaper subroutine instead, one that does not return a - * boolean value but only valid flags. Note: jeq jumps if - * the condition is not met, jne jumps if the condition is met. - * Invert the code if we jump on condition not met. - */ - if (GetBranchCond (N->OPC) == BC_EQ) { - /* Jumps if condition false, invert condition */ - Cond = CmpInvertTab [Cond]; - } - - /* Replace the subroutine call. */ - E = NewCodeEntry (OP65_JSR, AM65_ABS, "tosicmp", 0, E->LI); - CS_InsertEntry (S, E, I+1); - CS_DelEntry (S, I); + 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_IsCall (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); - /* Replace the conditional branch */ - ReplaceCmp (S, I+1, Cond); + /* 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 */ + /* Next entry */ ++I; } @@ -923,9 +875,21 @@ static unsigned OptCmp5 (CodeSeg* S) -static unsigned OptCmp6 (CodeSeg* S) -/* Search for a sequence ldx/txa/branch and remove the txa if A is not - * used later. +static unsigned OptPtrLoad6 (CodeSeg* S) +/* Search for the sequence + * + * ldy ... + * jsr ldauidx + * + * and replace it by: + * + * ldy ... + * stx ptr1+1 + * sta ptr1 + * ldx #$00 + * lda (ptr1),y + * + * This step must be execute *after* OptPtrLoad1! */ { unsigned Changes = 0; @@ -934,29 +898,42 @@ static unsigned OptCmp6 (CodeSeg* S) unsigned I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* L[2]; + CodeEntry* L[2]; - /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); + /* Get next entry */ + L[0] = CS_GetEntry (S, I); + + /* Check for the sequence */ + if (L[0]->OPC == OP65_LDY && + CS_GetEntries (S, L+1, I+1, 1) && + CE_IsCall (L[1], "ldauidx") && + !CE_HasLabel (L[1])) { + + CodeEntry* X; + + /* Store the high byte */ + X = NewCodeEntry (OP65_STA, AM65_ZP, "ptr1", 0, L[0]->LI); + CS_InsertEntry (S, X, I+1); - /* Check for the sequence */ - if ((E->OPC == OP65_LDX || E->OPC == OP65_TAX) && - CS_GetEntries (S, L, I+1, 2) && - L[0]->OPC == OP65_TXA && - !CE_HasLabel (L[0]) && - (L[1]->Info & OF_FBRA) != 0 && - !CE_HasLabel (L[1]) && - !RegAUsed (S, I+3)) { + /* Store the low byte */ + X = NewCodeEntry (OP65_STX, AM65_ZP, "ptr1+1", 0, L[0]->LI); + CS_InsertEntry (S, X, I+2); - /* Remove the txa */ - CS_DelEntry (S, I+1); + /* Delete the call to ldauidx */ + CS_DelEntry (S, I+3); + + /* Load the high and low byte */ + X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, L[0]->LI); + CS_InsertEntry (S, X, I+3); + X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, "ptr1", 0, L[0]->LI); + CS_InsertEntry (S, X, I+4); /* Remember, we had changes */ ++Changes; } - /* Next entry */ + /* Next entry */ ++I; } @@ -968,60 +945,152 @@ static unsigned OptCmp6 (CodeSeg* S) /*****************************************************************************/ -/* Optimize tests */ +/* Decouple operations */ /*****************************************************************************/ -static unsigned OptTest1 (CodeSeg* S) -/* On a sequence - * - * stx xxx - * ora xxx - * beq/bne ... - * - * if X is zero, the sequence may be changed +static unsigned OptDecouple (CodeSeg* S) +/* Decouple operations, that is, do the following replacements: * - * cmp #$00 - * beq/bne ... + * dex -> ldx #imm + * inx -> ldx #imm + * dey -> ldy #imm + * iny -> ldy #imm + * tax -> ldx #imm + * txa -> lda #imm + * tay -> ldy #imm + * tya -> lda #imm + * lda sreg -> lda #imm + * ldx sreg -> ldx #imm + * ldy sreg -> ldy #imm * - * which may be optimized further by another step. + * Provided that the register values are known of course. */ { unsigned Changes = 0; unsigned I; - /* Generate register info for this step */ + /* Generate register info for the following step */ CS_GenRegInfo (S); /* Walk over the entries */ I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* L[3]; + char Buf [16]; /* Get next entry */ - L[0] = CS_GetEntry (S, I); + CodeEntry* E = CS_GetEntry (S, I); - /* Check if it's the sequence we're searching for */ - if (L[0]->OPC == OP65_STX && - L[0]->RI->In.RegX == 0 && - CS_GetEntries (S, L+1, I+1, 2) && - !CE_HasLabel (L[1]) && - L[1]->OPC == OP65_ORA && - strcmp (L[0]->Arg, L[1]->Arg) == 0 && - !CE_HasLabel (L[2]) && - (L[2]->Info & OF_ZBRA) != 0) { - - /* Insert the compare */ - CodeEntry* N = NewCodeEntry (OP65_CMP, AM65_IMM, "$00", 0, L[0]->LI); - CS_InsertEntry (S, N, I); - - /* Remove the two other insns */ - CS_DelEntry (S, I+2); - CS_DelEntry (S, I+1); + /* Assume we have no replacement */ + CodeEntry* X = 0; + + /* Check the instruction */ + switch (E->OPC) { + + case OP65_DEX: + if (E->RI->In.RegX >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", (E->RI->In.RegX - 1) & 0xFF); + X = NewCodeEntry (OP65_LDX, AM65_IMM, Buf, 0, E->LI); + } + break; + + case OP65_DEY: + if (E->RI->In.RegY >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", (E->RI->In.RegY - 1) & 0xFF); + X = NewCodeEntry (OP65_LDY, AM65_IMM, Buf, 0, E->LI); + } + break; + + case OP65_INX: + if (E->RI->In.RegX >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", (E->RI->In.RegX + 1) & 0xFF); + X = NewCodeEntry (OP65_LDX, AM65_IMM, Buf, 0, E->LI); + } + break; + + case OP65_INY: + if (E->RI->In.RegY >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", (E->RI->In.RegY + 1) & 0xFF); + X = NewCodeEntry (OP65_LDY, AM65_IMM, Buf, 0, E->LI); + } + break; + + case OP65_LDA: + if (E->AM == AM65_ZP) { + if ((E->Use & REG_SREG_LO) != 0 && E->RI->In.SRegLo >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", E->RI->In.SRegLo); + X = NewCodeEntry (OP65_LDA, AM65_IMM, Buf, 0, E->LI); + } else if ((E->Use & REG_SREG_HI) != 0 && E->RI->In.SRegHi >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", E->RI->In.SRegHi); + X = NewCodeEntry (OP65_LDA, AM65_IMM, Buf, 0, E->LI); + } + } + break; + + case OP65_LDX: + if (E->AM == AM65_ZP) { + if ((E->Use & REG_SREG_LO) != 0 && E->RI->In.SRegLo >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", E->RI->In.SRegLo); + X = NewCodeEntry (OP65_LDX, AM65_IMM, Buf, 0, E->LI); + } else if ((E->Use & REG_SREG_HI) != 0 && E->RI->In.SRegHi >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", E->RI->In.SRegHi); + X = NewCodeEntry (OP65_LDX, AM65_IMM, Buf, 0, E->LI); + } + } + break; + + case OP65_LDY: + if (E->AM == AM65_ZP) { + if ((E->Use & REG_SREG_LO) != 0 && E->RI->In.SRegLo >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", E->RI->In.SRegLo); + X = NewCodeEntry (OP65_LDY, AM65_IMM, Buf, 0, E->LI); + } else if ((E->Use & REG_SREG_HI) != 0 && E->RI->In.SRegHi >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", E->RI->In.SRegHi); + X = NewCodeEntry (OP65_LDY, AM65_IMM, Buf, 0, E->LI); + } + } + break; + + case OP65_TAX: + if (E->RI->In.RegA >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", E->RI->In.RegA); + X = NewCodeEntry (OP65_LDX, AM65_IMM, Buf, 0, E->LI); + } + break; + + case OP65_TAY: + if (E->RI->In.RegA >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", E->RI->In.RegA); + X = NewCodeEntry (OP65_LDY, AM65_IMM, Buf, 0, E->LI); + } + break; + + case OP65_TXA: + if (E->RI->In.RegX >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", E->RI->In.RegX); + X = NewCodeEntry (OP65_LDA, AM65_IMM, Buf, 0, E->LI); + } + break; + + case OP65_TYA: + if (E->RI->In.RegY >= 0) { + xsprintf (Buf, sizeof (Buf), "$%02X", E->RI->In.RegY); + X = NewCodeEntry (OP65_LDA, AM65_IMM, Buf, 0, E->LI); + } + break; + + default: + /* Avoid gcc warnings */ + break; + + } - /* We had changes */ + /* Insert the replacement if we have one */ + if (X) { + CS_InsertEntry (S, X, I+1); + CS_DelEntry (S, I); ++Changes; } @@ -1039,115 +1108,203 @@ static unsigned OptTest1 (CodeSeg* S) - - - - /*****************************************************************************/ -/* nega optimizations */ +/* Size optimization */ /*****************************************************************************/ -static unsigned OptNegA1 (CodeSeg* S) -/* Check for - * - * ldx #$00 - * lda .. - * jsr bnega - * - * Remove the ldx if the lda does not use it. +#if 0 +static unsigned OptSize1 (CodeSeg* S) +/* Do size optimization by calling special subroutines that preload registers. + * This routine does not work standalone, it needs a following register load + * removal pass. */ { + static const char* Func = { + "stax0sp", /* staxysp, y = 0 */ + "addeq0sp", + "ldax0sp", /* ldaxysp, y = 1 */ + "ldeax0sp", /* ldeaxysp, y = 3 */ + "push0", /* pushax, a = 0, x = 0 */ + "pusha0", /* pushax, x = 0 */ + "pushaFF", /* pushax, x = ff */ + "pusha0sp", /* pushaysp, y = 0 */ + "tosadda0", /* tosaddax, x = 0 */ + "tosanda0", /* tosandax, x = 0 */ + "tosdiva0", /* tosdivax, x = 0 */ + "toseqa0", /* toseqax, x = 0 */ + "tosgea0", /* tosgeax, x = 0 */ + "tosgta0", /* tosgtax, x = 0 */ + "tosadd0ax", /* tosaddeax, sreg = 0 */ + "laddeqa", /* laddeq, sreg = 0, x = 0 */ + "laddeq1", /* laddeq, sreg = 0, x = 0, a = 1 */ + "laddeq0sp", /* laddeqysp, y = 0 */ + "tosand0ax", /* tosandeax, sreg = 0 */ + "ldaxi", /* ldaxidx, y = 1 */ + "ldeaxi", /* ldeaxidx, y = 3 */ + "ldeax0sp", /* ldeaxysp, y = 3 */ + "tosdiv0ax", /* tosdiveax, sreg = 0 */ + "toslea0", /* tosleax, x = 0 */ + "tosmod0ax", /* tosmodeax, sreg = 0 */ + "tosmul0ax", /* tosmuleax, sreg = 0 */ + "tosumul0ax", /* tosumuleax, sreg = 0 */ + "tosor0ax", /* tosoreax, sreg = 0 */ + "push0ax", /* pusheax, sreg = 0 */ + "tosrsub0ax", /* tosrsubeax, sreg = 0 */ + "tosshl0ax", /* tosshleax, sreg = 0 */ + "tosasl0ax", /* tosasleax, sreg = 0 */ + "tosshr0ax", /* tosshreax, sreg = 0 */ + "tosasr0ax", /* tosasreax, sreg = 0 */ + "tossub0ax", /* tossubeax, sreg = 0 */ + "lsubeqa", /* lsubeq, sreg = 0, x = 0 */ + "lsubeq1", /* lsubeq, sreg = 0, x = 0, a = 1 */ + "lsubeq0sp", /* lsubeqysp, y = 0 */ + "toslta0", /* tosltax, x = 0 */ + "tosudiv0ax", /* tosudiveax, sreg = 0 */ + "tosumod0ax", /* tosumodeax, sreg = 0 */ + "tosxor0ax", /* tosxoreax, sreg = 0 */ + "tosmoda0", /* tosmodax, x = 0 */ + "tosmula0", /* tosmulax, x = 0 */ + "tosumula0", /* tosumulax, x = 0 */ + "tosnea0", /* tosneax, x = 0 */ + "tosora0", /* tosorax, x = 0 */ + "push1", /* pushax, x = 0, a = 1 */ + "push2", /* pushax, x = 0, a = 2 */ + "push3", /* pushax, x = 0, a = 3 */ + "push4", /* pushax, x = 0, a = 4 */ + "push5", /* pushax, x = 0, a = 5 */ + "push6", /* pushax, x = 0, a = 6 */ + "push7", /* pushax, x = 0, a = 7 */ + "pushc0", /* pusha, a = 0 */ + "pushc1", /* pusha, a = 1 */ + "pushc2", /* pusha, a = 2 */ + "tosrsuba0", /* tosrsubax, x = 0 */ + "tosshla0", /* tosshlax, x = 0 */ + "tosasla0", /* tosaslax, x = 0 */ + "tosshra0", /* tosshrax, x = 0 */ + "tosasra0", /* tosasrax, x = 0 */ + "steax0sp", /* steaxsp, y = 0 */ + "tossuba0", /* tossubax, x = 0 */ + "subeq0sp", /* subeqysp, y = 0 */ + "tosudiva0", /* tosudivax, x = 0 */ + "tosugea0", /* tosugeax, x = 0 */ + "tosugta0", /* tosugtax, x = 0 */ + "tosulea0", /* tosuleax, x = 0 */ + "tosulta0", /* tosultax, x = 0 */ + "tosumoda0", /* tosumodax, x = 0 */ + "tosxora0", /* tosxorax, x = 0 */ + }; + unsigned Changes = 0; + unsigned I; + + /* Generate register info for the following step */ + CS_GenRegInfo (S); /* Walk over the entries */ - unsigned I = 0; + I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* L[2]; - /* Get next entry */ CodeEntry* E = CS_GetEntry (S, I); - /* Check for a ldx */ - if (E->OPC == OP65_LDX && - E->AM == AM65_IMM && - (E->Flags & CEF_NUMARG) != 0 && - E->Num == 0 && - CS_GetEntries (S, L, I+1, 2) && - L[0]->OPC == OP65_LDA && - (L[0]->Use & REG_X) == 0 && - L[1]->OPC == OP65_JSR && - strcmp (L[1]->Arg, "bnega") == 0) { - - /* Remove the ldx instruction */ - 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; } +#endif -static unsigned OptNegA2 (CodeSeg* S) -/* Check for - * - * lda .. - * jsr bnega - * jeq/jne .. - * - * Adjust the conditional branch and remove the call to the subroutine. +static unsigned OptSize2 (CodeSeg* S) +/* Do size optimization by using shorter code sequences, even if this + * introduces relations between instructions. This step must be one of the + * last steps, because it makes further work much more difficult. */ { unsigned Changes = 0; + unsigned I; + + /* Generate register info for the following step */ + CS_GenRegInfo (S); /* Walk over the entries */ - unsigned I = 0; + I = 0; while (I < CS_GetEntryCount (S)) { - CodeEntry* L[2]; /* Get next entry */ CodeEntry* E = CS_GetEntry (S, I); - /* Check for the sequence */ - if ((E->OPC == OP65_ADC || - E->OPC == OP65_AND || - E->OPC == OP65_DEA || - E->OPC == OP65_EOR || - E->OPC == OP65_INA || - E->OPC == OP65_LDA || - E->OPC == OP65_ORA || - E->OPC == OP65_PLA || - E->OPC == OP65_SBC || - E->OPC == OP65_TXA || - E->OPC == OP65_TYA) && - CS_GetEntries (S, L, I+1, 2) && - L[0]->OPC == OP65_JSR && - strcmp (L[0]->Arg, "bnega") == 0 && - !CE_HasLabel (L[0]) && - (L[1]->Info & OF_ZBRA) != 0) { - - /* Invert the branch */ - CE_ReplaceOPC (L[1], GetInverseBranch (L[1]->OPC)); - - /* Delete the subroutine call */ - CS_DelEntry (S, I+1); + /* Assume we have no replacement */ + CodeEntry* X = 0; + + /* Check the instruction */ + switch (E->OPC) { + + case OP65_LDA: + if (CE_KnownImm (E)) { + short Val = (short) E->Num; + if (Val == E->RI->In.RegX) { + X = NewCodeEntry (OP65_TXA, AM65_IMP, 0, 0, E->LI); + } else if (Val == E->RI->In.RegY) { + X = NewCodeEntry (OP65_TYA, AM65_IMP, 0, 0, E->LI); + } else if (E->RI->In.RegA >= 0 && CPU >= CPU_65C02) { + if (Val == ((E->RI->In.RegA - 1) & 0xFF)) { + X = NewCodeEntry (OP65_DEA, AM65_IMP, 0, 0, E->LI); + } else if (Val == ((E->RI->In.RegA + 1) & 0xFF)) { + X = NewCodeEntry (OP65_INA, AM65_IMP, 0, 0, E->LI); + } + } + } + break; + + case OP65_LDX: + if (CE_KnownImm (E)) { + short Val = (short) E->Num; + if (E->RI->In.RegX >= 0 && Val == ((E->RI->In.RegX - 1) & 0xFF)) { + X = NewCodeEntry (OP65_DEX, AM65_IMP, 0, 0, E->LI); + } else if (E->RI->In.RegX >= 0 && Val == ((E->RI->In.RegX + 1) & 0xFF)) { + X = NewCodeEntry (OP65_INX, AM65_IMP, 0, 0, E->LI); + } else if (Val == E->RI->In.RegA) { + X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, E->LI); + } + } + break; + + case OP65_LDY: + if (CE_KnownImm (E)) { + short Val = (short) E->Num; + if (E->RI->In.RegY >= 0 && Val == ((E->RI->In.RegY - 1) & 0xFF)) { + X = NewCodeEntry (OP65_DEY, AM65_IMP, 0, 0, E->LI); + } else if (E->RI->In.RegY >= 0 && Val == ((E->RI->In.RegY + 1) & 0xFF)) { + X = NewCodeEntry (OP65_INY, AM65_IMP, 0, 0, E->LI); + } else if (Val == E->RI->In.RegA) { + X = NewCodeEntry (OP65_TAY, AM65_IMP, 0, 0, E->LI); + } + } + break; - /* Remember, we had changes */ - ++Changes; + default: + /* Avoid gcc warnings */ + break; + + } + /* Insert the replacement if we have one */ + if (X) { + CS_InsertEntry (S, X, I+1); + CS_DelEntry (S, I); + ++Changes; } /* Next entry */ @@ -1155,6 +1312,9 @@ static unsigned OptNegA2 (CodeSeg* S) } + /* Free register info */ + CS_FreeRegInfo (S); + /* Return the number of changes made */ return Changes; } @@ -1162,378 +1322,479 @@ static unsigned OptNegA2 (CodeSeg* S) /*****************************************************************************/ -/* negax optimizations */ +/* struct OptFunc */ /*****************************************************************************/ -static unsigned OptNegAX1 (CodeSeg* S) -/* On a call to bnegax, if X is zero, the result depends only on the value in - * A, so change the call to a call to bnega. This will get further optimized - * later if possible. - */ -{ - unsigned Changes = 0; - unsigned I; +typedef struct OptFunc OptFunc; +struct OptFunc { + unsigned (*Func) (CodeSeg*); /* Optimizer function */ + const char* Name; /* Name of the function/group */ + unsigned CodeSizeFactor; /* Code size factor for this opt func */ + unsigned long TotalRuns; /* Total number of runs */ + unsigned long LastRuns; /* Last number of runs */ + unsigned long TotalChanges; /* Total number of changes */ + unsigned long LastChanges; /* Last number of changes */ + char Disabled; /* True if function disabled */ +}; - /* 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); +/*****************************************************************************/ +/* Code */ +/*****************************************************************************/ - /* Check if this is a call to bnegax, and if X is known and zero */ - if (E->OPC == OP65_JSR && - E->RI->In.RegX == 0 && - strcmp (E->Arg, "bnegax") == 0) { - /* We're cheating somewhat here ... */ - E->Arg[5] = '\0'; - E->Use &= ~REG_X; - /* We had changes */ - ++Changes; - } +/* Macro that builds a function description */ +#define OptFuncEntry(func) static OptFuncDesc D##func = { func, #func, 0 } + +/* A list of all the function descriptions */ +static OptFunc DOpt65C02Ind = { Opt65C02Ind, "Opt65C02Ind", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptAdd1 = { OptAdd1, "OptAdd1", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptAdd2 = { OptAdd2, "OptAdd2", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptAdd3 = { OptAdd3, "OptAdd3", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptBoolTrans = { OptBoolTrans, "OptBoolTrans", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptBranchDist = { OptBranchDist, "OptBranchDist", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptCmp1 = { OptCmp1, "OptCmp1", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptCmp2 = { OptCmp2, "OptCmp2", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptCmp3 = { OptCmp3, "OptCmp3", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptCmp4 = { OptCmp4, "OptCmp4", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptCmp5 = { OptCmp5, "OptCmp5", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptCmp6 = { OptCmp6, "OptCmp6", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptCmp7 = { OptCmp7, "OptCmp7", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptCondBranches = { OptCondBranches, "OptCondBranches", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptDeadCode = { OptDeadCode, "OptDeadCode", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptDeadJumps = { OptDeadJumps, "OptDeadJumps", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptDecouple = { OptDecouple, "OptDecouple", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptDupLoads = { OptDupLoads, "OptDupLoads", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptJumpCascades = { OptJumpCascades, "OptJumpCascades", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptJumpTarget = { OptJumpTarget, "OptJumpTarget", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptRTS = { OptRTS, "OptRTS", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptRTSJumps1 = { OptRTSJumps1, "OptRTSJumps1", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptRTSJumps2 = { OptRTSJumps2, "OptRTSJumps2", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptNegA1 = { OptNegA1, "OptNegA1", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptNegA2 = { OptNegA2, "OptNegA2", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptNegAX1 = { OptNegAX1, "OptNegAX1", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptNegAX2 = { OptNegAX2, "OptNegAX2", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptNegAX3 = { OptNegAX3, "OptNegAX3", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptNegAX4 = { OptNegAX4, "OptNegAX4", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptPtrLoad1 = { OptPtrLoad1, "OptPtrLoad1", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptPtrLoad2 = { OptPtrLoad2, "OptPtrLoad2", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptPtrLoad3 = { OptPtrLoad3, "OptPtrLoad3", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptPtrLoad4 = { OptPtrLoad4, "OptPtrLoad4", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptPtrLoad5 = { OptPtrLoad5, "OptPtrLoad5", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptPtrLoad6 = { OptPtrLoad6, "OptPtrLoad6", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptPtrStore1 = { OptPtrStore1, "OptPtrStore1", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptPtrStore2 = { OptPtrStore2, "OptPtrStore2", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptShift1 = { OptShift1, "OptShift1", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptShift2 = { OptShift2, "OptShift2", 100, 0, 0, 0, 0, 0 }; +/*static OptFunc DOptSize1 = { OptSize1, "OptSize1", 100, 0, 0, 0, 0, 0 };*/ +static OptFunc DOptSize2 = { OptSize2, "OptSize2", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptStackOps = { OptStackOps, "OptStackOps", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptStoreLoad = { OptStoreLoad, "OptStoreLoad", 100, 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 DOptTest1 = { OptTest1, "OptTest1", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptTransfers = { OptTransfers, "OptTransfers", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptUnusedLoads = { OptUnusedLoads, "OptUnusedLoads", 100, 0, 0, 0, 0, 0 }; +static OptFunc DOptUnusedStores = { OptUnusedStores, "OptUnusedStores", 100, 0, 0, 0, 0, 0 }; + + +/* Table containing all the steps in alphabetical order */ +static OptFunc* OptFuncs[] = { + &DOpt65C02Ind, + &DOptAdd1, + &DOptAdd2, + &DOptAdd3, + &DOptBoolTrans, + &DOptBranchDist, + &DOptCmp1, + &DOptCmp2, + &DOptCmp3, + &DOptCmp4, + &DOptCmp5, + &DOptCmp6, + &DOptCmp7, + &DOptCondBranches, + &DOptDeadCode, + &DOptDeadJumps, + &DOptDecouple, + &DOptDupLoads, + &DOptJumpCascades, + &DOptJumpTarget, + &DOptNegA1, + &DOptNegA2, + &DOptNegAX1, + &DOptNegAX2, + &DOptNegAX3, + &DOptNegAX4, + &DOptPtrLoad1, + &DOptPtrLoad2, + &DOptPtrLoad3, + &DOptPtrLoad4, + &DOptPtrLoad5, + &DOptPtrLoad6, + &DOptPtrStore1, + &DOptPtrStore2, + &DOptRTS, + &DOptRTSJumps1, + &DOptRTSJumps2, + &DOptShift1, + &DOptShift2, + /*&DOptSize1,*/ + &DOptSize2, + &DOptStackOps, + &DOptStoreLoad, + &DOptSub1, + &DOptSub2, + &DOptTest1, + &DOptTransfers, + &DOptUnusedLoads, + &DOptUnusedStores, +}; +#define OPTFUNC_COUNT (sizeof(OptFuncs) / sizeof(OptFuncs[0])) - /* Next entry */ - ++I; - } - /* Free register info */ - CS_FreeRegInfo (S); +static int CmpOptStep (const void* Key, const void* Func) +/* Compare function for bsearch */ +{ + return strcmp (Key, (*(const OptFunc**)Func)->Name); +} - /* Return the number of changes made */ - return Changes; + + +static OptFunc* FindOptFunc (const char* Name) +/* Find an optimizer step by name in the table and return a pointer. Return + * NULL if no such step is found. + */ +{ + /* Search for the function in the list */ + OptFunc** O = bsearch (Name, OptFuncs, OPTFUNC_COUNT, sizeof (OptFuncs[0]), CmpOptStep); + return O? *O : 0; } -static unsigned OptNegAX2 (CodeSeg* S) -/* Search for the sequence: - * - * lda (xx),y - * tax - * dey - * lda (xx),y - * jsr bnegax - * jne/jeq ... - * - * and replace it by - * - * lda (xx),y - * dey - * ora (xx),y - * jeq/jne ... +static OptFunc* GetOptFunc (const char* Name) +/* Find an optimizer step by name in the table and return a pointer. Print an + * error and call AbEnd if not found. */ { - unsigned Changes = 0; + /* Search for the function in the list */ + OptFunc* F = FindOptFunc (Name); + if (F == 0) { + /* Not found */ + AbEnd ("Optimization step `%s' not found", Name); + } + return F; +} - /* Walk over the entries */ - unsigned I = 0; - while (I < CS_GetEntryCount (S)) { - CodeEntry* L[5]; - /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); +void DisableOpt (const char* Name) +/* Disable the optimization with the given name */ +{ + if (strcmp (Name, "any") == 0) { + unsigned I; + for (I = 0; I < OPTFUNC_COUNT; ++I) { + OptFuncs[I]->Disabled = 1; + } + } else { + GetOptFunc(Name)->Disabled = 1; + } +} - /* Check for the sequence */ - if (E->OPC == OP65_LDA && - E->AM == AM65_ZP_INDY && - CS_GetEntries (S, L, I+1, 5) && - L[0]->OPC == OP65_TAX && - L[1]->OPC == OP65_DEY && - L[2]->OPC == OP65_LDA && - L[2]->AM == AM65_ZP_INDY && - strcmp (L[2]->Arg, E->Arg) == 0 && - !CE_HasLabel (L[2]) && - L[3]->OPC == OP65_JSR && - strcmp (L[3]->Arg, "bnegax") == 0 && - !CE_HasLabel (L[3]) && - (L[4]->Info & OF_ZBRA) != 0) { - - /* lda --> ora */ - CE_ReplaceOPC (L[2], OP65_ORA); - - /* Invert the branch */ - CE_ReplaceOPC (L[4], GetInverseBranch (L[4]->OPC)); - - /* Delete the entries no longer needed. Beware: Deleting entries - * will change the indices. - */ - CS_DelEntry (S, I+4); /* jsr bnegax */ - CS_DelEntry (S, I+1); /* tax */ - /* Remember, we had changes */ - ++Changes; +void EnableOpt (const char* Name) +/* Enable the optimization with the given name */ +{ + if (strcmp (Name, "any") == 0) { + unsigned I; + for (I = 0; I < OPTFUNC_COUNT; ++I) { + OptFuncs[I]->Disabled = 0; } + } else { + GetOptFunc(Name)->Disabled = 0; + } +} - /* Next entry */ - ++I; - } - /* Return the number of changes made */ - return Changes; +void ListOptSteps (FILE* F) +/* List all optimization steps */ +{ + unsigned I; + for (I = 0; I < OPTFUNC_COUNT; ++I) { + fprintf (F, "%s\n", OptFuncs[I]->Name); + } } -static unsigned OptNegAX3 (CodeSeg* S) -/* Search for the sequence: - * - * lda xx - * ldx yy - * jsr bnegax - * jne/jeq ... - * - * and replace it by - * - * lda xx - * ora xx+1 - * jeq/jne ... - */ +static void ReadOptStats (const char* Name) +/* Read the optimizer statistics file */ { - unsigned Changes = 0; + char Buf [256]; + unsigned Lines; + + /* Try to open the file */ + FILE* F = fopen (Name, "r"); + if (F == 0) { + /* Ignore the error */ + return; + } - /* Walk over the entries */ - unsigned I = 0; - while (I < CS_GetEntryCount (S)) { + /* Read and parse the lines */ + Lines = 0; + while (fgets (Buf, sizeof (Buf), F) != 0) { - CodeEntry* L[3]; + char* B; + unsigned Len; + OptFunc* Func; - /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); + /* Fields */ + char Name[32]; + unsigned long TotalRuns; + unsigned long TotalChanges; - /* Check for the sequence */ - if (E->OPC == OP65_LDA && - CS_GetEntries (S, L, I+1, 3) && - L[0]->OPC == OP65_LDX && - !CE_HasLabel (L[0]) && - L[1]->OPC == OP65_JSR && - strcmp (L[1]->Arg, "bnegax") == 0 && - !CE_HasLabel (L[1]) && - (L[2]->Info & OF_ZBRA) != 0) { + /* Count lines */ + ++Lines; - /* ldx --> ora */ - CE_ReplaceOPC (L[0], OP65_ORA); + /* Remove trailing white space including the line terminator */ + B = Buf; + Len = strlen (B); + while (Len > 0 && IsSpace (B[Len-1])) { + --Len; + } + B[Len] = '\0'; - /* Invert the branch */ - CE_ReplaceOPC (L[2], GetInverseBranch (L[2]->OPC)); + /* Remove leading whitespace */ + while (IsSpace (*B)) { + ++B; + } - /* Delete the subroutine call */ - CS_DelEntry (S, I+2); + /* Check for empty and comment lines */ + if (*B == '\0' || *B == ';' || *B == '#') { + continue; + } - /* Remember, we had changes */ - ++Changes; + /* Parse the line */ + if (sscanf (B, "%31s %lu %*u %lu %*u", Name, &TotalRuns, &TotalChanges) != 3) { + /* Syntax error */ + continue; + } + /* Search for the optimizer step. */ + Func = FindOptFunc (Name); + if (Func == 0) { + /* Not found */ + continue; } - /* Next entry */ - ++I; + /* Found the step, set the fields */ + Func->TotalRuns = TotalRuns; + Func->TotalChanges = TotalChanges; } - /* Return the number of changes made */ - return Changes; + /* Close the file, ignore errors here. */ + fclose (F); } -static unsigned OptNegAX4 (CodeSeg* S) -/* Search for the sequence: - * - * jsr xxx - * jsr bnega(x) - * jeq/jne ... - * - * and replace it by: - * - * jsr xxx - * - * jne/jeq ... - */ +static void WriteOptStats (const char* Name) +/* Write the optimizer statistics file */ { - unsigned Changes = 0; - - /* Walk over the entries */ - unsigned I = 0; - while (I < CS_GetEntryCount (S)) { + unsigned I; - CodeEntry* L[2]; + /* Try to open the file */ + FILE* F = fopen (Name, "w"); + if (F == 0) { + /* Ignore the error */ + return; + } - /* Get next entry */ - CodeEntry* E = CS_GetEntry (S, I); + /* Write the file */ + for (I = 0; I < OPTFUNC_COUNT; ++I) { + const OptFunc* O = OptFuncs[I]; + fprintf (F, + "%-20s %6lu %6lu %6lu %6lu\n", + O->Name, + O->TotalRuns, + O->LastRuns, + O->TotalChanges, + O->LastChanges); + } - /* Check for the sequence */ - if (E->OPC == OP65_JSR && - CS_GetEntries (S, L, I+1, 2) && - L[0]->OPC == OP65_JSR && - strncmp (L[0]->Arg,"bnega",5) == 0 && - !CE_HasLabel (L[0]) && - (L[1]->Info & OF_ZBRA) != 0) { + /* Close the file, ignore errors here. */ + fclose (F); +} - CodeEntry* X; - /* Check if we're calling bnega or bnegax */ - int ByteSized = (strcmp (L[0]->Arg, "bnega") == 0); - - /* Insert apropriate test code */ - if (ByteSized) { - /* Test bytes */ - X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, L[0]->LI); - CS_InsertEntry (S, X, I+2); - } else { - /* Test words */ - X = NewCodeEntry (OP65_STX, AM65_ZP, "tmp1", 0, L[0]->LI); - CS_InsertEntry (S, X, I+2); - X = NewCodeEntry (OP65_ORA, AM65_ZP, "tmp1", 0, L[0]->LI); - CS_InsertEntry (S, X, I+3); - } - /* Delete the subroutine call */ - CS_DelEntry (S, I+1); +static unsigned RunOptFunc (CodeSeg* S, OptFunc* F, unsigned Max) +/* Run one optimizer function Max times or until there are no more changes */ +{ + unsigned Changes, C; - /* Invert the branch */ - CE_ReplaceOPC (L[1], GetInverseBranch (L[1]->OPC)); + /* Don't run the function if it is disabled or if it is prohibited by the + * code size factor + */ + if (F->Disabled || CodeSizeFactor < F->CodeSizeFactor) { + return 0; + } - /* Remember, we had changes */ - ++Changes; + /* Run this until there are no more changes */ + Changes = 0; + do { - } + /* Run the function */ + C = F->Func (S); + Changes += C; - /* Next entry */ - ++I; + /* Do statistics */ + ++F->TotalRuns; + ++F->LastRuns; + F->TotalChanges += C; + F->LastChanges += C; - } + } while (--Max && C > 0); - /* Return the number of changes made */ + /* Return the number of changes */ return Changes; } -/*****************************************************************************/ -/* Code */ -/*****************************************************************************/ - - +static void RunOptGroup1 (CodeSeg* S) +/* Run the first group of optimization steps. These steps translate known + * patterns emitted by the code generator into more optimal patterns. Order + * of the steps is important, because some of the steps done earlier cover + * the same patterns as later steps as subpatterns. + */ +{ + RunOptFunc (S, &DOptPtrStore1, 1); + RunOptFunc (S, &DOptPtrStore2, 1); + RunOptFunc (S, &DOptPtrLoad1, 1); + RunOptFunc (S, &DOptPtrLoad2, 1); + RunOptFunc (S, &DOptPtrLoad3, 1); + RunOptFunc (S, &DOptPtrLoad4, 1); + RunOptFunc (S, &DOptPtrLoad5, 1); + RunOptFunc (S, &DOptNegAX1, 1); + RunOptFunc (S, &DOptNegAX2, 1); + RunOptFunc (S, &DOptNegAX3, 1); + RunOptFunc (S, &DOptNegAX4, 1); + RunOptFunc (S, &DOptAdd1, 1); + RunOptFunc (S, &DOptAdd2, 1); + RunOptFunc (S, &DOptShift1, 1); + RunOptFunc (S, &DOptShift2, 1); +} -/* Table with all the optimization functions */ -typedef struct OptFunc OptFunc; -struct OptFunc { - unsigned (*Func) (CodeSeg*);/* Optimizer function */ - const char* Name; /* Name of optimizer step */ - char Disabled; /* True if pass disabled */ -}; +static void RunOptGroup2 (CodeSeg* S) +/* Run one group of optimization steps. This step involves just decoupling + * instructions by replacing them by instructions that do not depend on + * previous instructions. This makes it easier to find instructions that + * aren't used. + */ +{ + RunOptFunc (S, &DOptDecouple, 1); +} -/* Table with optimizer steps - are called in this order */ -static OptFunc OptFuncs [] = { - /* Optimize subtractions */ - { OptSub1, "OptSub1", 0 }, - { OptSub2, "OptSub2", 0 }, - /* Optimize additions */ - { OptAdd1, "OptAdd1", 0 }, - /* Optimize jump cascades */ - { OptJumpCascades, "OptJumpCascades", 0 }, - /* Remove dead jumps */ - { OptDeadJumps, "OptDeadJumps", 0 }, - /* Change jsr/rts to jmp */ - { OptRTS, "OptRTS", 0 }, - /* Remove dead code */ - { OptDeadCode, "OptDeadCode", 0 }, - /* Optimize jump targets */ - { OptJumpTarget, "OptJumpTarget", 0 }, - /* Optimize conditional branches */ - { OptCondBranches, "OptCondBranches", 0 }, - /* Replace jumps to RTS by RTS */ - { OptRTSJumps, "OptRTSJumps", 0 }, - /* Remove calls to the bool transformer subroutines */ - { OptBoolTransforms, "OptBoolTransforms", 0 }, - /* Optimize calls to nega */ - { OptNegA1, "OptNegA1", 0 }, - { OptNegA2, "OptNegA2", 0 }, - /* Optimize calls to negax */ - { OptNegAX1, "OptNegAX1", 0 }, - { OptNegAX2, "OptNegAX2", 0 }, - { OptNegAX3, "OptNegAX3", 0 }, - { OptNegAX4, "OptNegAX4", 0 }, - /* Optimize compares */ - { OptCmp1, "OptCmp1", 0 }, - { OptCmp2, "OptCmp2", 0 }, - { OptCmp3, "OptCmp3", 0 }, - { OptCmp4, "OptCmp4", 0 }, - { OptCmp5, "OptCmp5", 0 }, - { OptCmp6, "OptCmp6", 0 }, - /* Optimize tests */ - { OptTest1, "OptTest1", 0 }, - /* Remove unused loads */ - { OptUnusedLoads, "OptUnusedLoads", 0 }, - { OptDuplicateLoads, "OptDuplicateLoads", 0 }, - { OptStoreLoad, "OptStoreLoad", 0 }, - /* Optimize branch distance */ - { OptBranchDist, "OptBranchDist", 0 }, -}; -static OptFunc* FindOptStep (const char* Name) -/* Find an optimizer step by name in the table and return a pointer. Print an - * error and call AbEnd if not found. +static void RunOptGroup3 (CodeSeg* S) +/* Run one group of optimization steps. These steps depend on each other, + * that means that one step may allow another step to do additional work, + * so we will repeat the steps as long as we see any changes. */ { - unsigned I; - - /* Run all optimization steps */ - for (I = 0; I < sizeof(OptFuncs)/sizeof(OptFuncs[0]); ++I) { - if (strcmp (OptFuncs[I].Name, Name) == 0) { - /* Found */ - return OptFuncs+I; - } - } + unsigned Changes; - /* Not found */ - AbEnd ("Optimization step `%s' not found", Name); - return 0; + do { + Changes = 0; + + Changes += RunOptFunc (S, &DOptPtrLoad6, 1); + Changes += RunOptFunc (S, &DOptNegA1, 1); + Changes += RunOptFunc (S, &DOptNegA2, 1); + Changes += RunOptFunc (S, &DOptSub1, 1); + Changes += RunOptFunc (S, &DOptSub2, 1); + Changes += RunOptFunc (S, &DOptAdd3, 1); + Changes += RunOptFunc (S, &DOptStackOps, 1); + Changes += RunOptFunc (S, &DOptJumpCascades, 1); + Changes += RunOptFunc (S, &DOptDeadJumps, 1); + Changes += RunOptFunc (S, &DOptRTS, 1); + Changes += RunOptFunc (S, &DOptDeadCode, 1); + Changes += RunOptFunc (S, &DOptJumpTarget, 1); + Changes += RunOptFunc (S, &DOptCondBranches, 1); + Changes += RunOptFunc (S, &DOptRTSJumps1, 1); + Changes += RunOptFunc (S, &DOptBoolTrans, 1); + Changes += RunOptFunc (S, &DOptCmp1, 1); + Changes += RunOptFunc (S, &DOptCmp2, 1); + Changes += RunOptFunc (S, &DOptCmp3, 1); + Changes += RunOptFunc (S, &DOptCmp4, 1); + Changes += RunOptFunc (S, &DOptCmp5, 1); + Changes += RunOptFunc (S, &DOptCmp6, 1); + Changes += RunOptFunc (S, &DOptCmp7, 1); + Changes += RunOptFunc (S, &DOptTest1, 1); + Changes += RunOptFunc (S, &DOptUnusedLoads, 1); + Changes += RunOptFunc (S, &DOptUnusedStores, 1); + Changes += RunOptFunc (S, &DOptDupLoads, 1); + Changes += RunOptFunc (S, &DOptStoreLoad, 1); + Changes += RunOptFunc (S, &DOptTransfers, 1); + + } while (Changes); } -void DisableOpt (const char* Name) -/* Disable the optimization with the given name */ +static void RunOptGroup4 (CodeSeg* S) +/* 65C02 specific optimizations. */ { - if (strcmp (Name, "any") == 0) { - unsigned I; - for (I = 0; I < sizeof(OptFuncs)/sizeof(OptFuncs[0]); ++I) { - OptFuncs[I].Disabled = 1; - } - } else { - OptFunc* F = FindOptStep (Name); - F->Disabled = 1; + if (CPU < CPU_65C02) { + return; + } + + /* Replace (zp),y by (zp) if Y is zero. If we have changes, run register + * load optimization again, since loads of Y may have become unnecessary. + */ + if (RunOptFunc (S, &DOpt65C02Ind, 1) > 0) { + RunOptFunc (S, &DOptUnusedLoads, 1); } } -void EnableOpt (const char* Name) -/* Enable the optimization with the given name */ +static void RunOptGroup5 (CodeSeg* S) +/* The last group of optimization steps. Adjust branches, do size optimizations. + */ { - if (strcmp (Name, "any") == 0) { - unsigned I; - for (I = 0; I < sizeof(OptFuncs)/sizeof(OptFuncs[0]); ++I) { - OptFuncs[I].Disabled = 0; - } - } else { - OptFunc* F = FindOptStep (Name); - F->Disabled = 0; + /* Optimize for size, that is replace operations by shorter ones, even + * if this does hinder further optimizations (no problem since we're + * done soon). + */ + RunOptFunc (S, &DOptSize2, 1); + + /* Run the jump target optimization again, since the size optimization + * above may have opened new oportunities. + */ + RunOptFunc (S, &DOptJumpTarget, 5); + + /* Adjust branch distances */ + RunOptFunc (S, &DOptBranchDist, 3); + + /* Replace conditional branches to RTS. If we had changes, we must run dead + * code elimination again, since the change may have introduced dead code. + */ + if (RunOptFunc (S, &DOptRTSJumps2, 1)) { + RunOptFunc (S, &DOptDeadCode, 1); } } @@ -1542,13 +1803,17 @@ void EnableOpt (const char* Name) void RunOpt (CodeSeg* S) /* Run the optimizer */ { - unsigned I; - unsigned Pass = 0; - unsigned OptChanges; + const char* StatFileName; /* If we shouldn't run the optimizer, bail out */ if (!Optimize) { - return; + return; + } + + /* Check if we are requested to write optimizer statistics */ + StatFileName = getenv ("CC65_OPTSTATS"); + if (StatFileName) { + ReadOptStats (StatFileName); } /* Print the name of the function we are working on */ @@ -1558,32 +1823,17 @@ void RunOpt (CodeSeg* S) Print (stdout, 1, "Running optimizer for global code segment\n"); } - /* Repeat all steps until there are no more changes */ - do { - /* Reset the number of changes */ - OptChanges = 0; - - /* Keep the user hapy */ - Print (stdout, 1, " Optimizer pass %u:\n", ++Pass); - - /* Run all optimization steps */ - for (I = 0; I < sizeof(OptFuncs)/sizeof(OptFuncs[0]); ++I) { - - /* Print the name of the following optimizer step */ - Print (stdout, 1, " %s:%*s", OptFuncs[I].Name, - (int) (30-strlen(OptFuncs[I].Name)), ""); - - /* Check if the step is disabled */ - if (OptFuncs[I].Disabled) { - Print (stdout, 1, "Disabled\n"); - } else { - unsigned Changes = OptFuncs[I].Func (S); - OptChanges += Changes; - Print (stdout, 1, "%u Changes\n", Changes); - } - } + /* Run groups of optimizations */ + RunOptGroup1 (S); + RunOptGroup2 (S); + RunOptGroup3 (S); + RunOptGroup4 (S); + RunOptGroup5 (S); - } while (OptChanges > 0); + /* Write statistics */ + if (StatFileName) { + WriteOptStats (StatFileName); + } }