6502/cc65
1
0
Fork 0
mirror of https://github.com/cc65/cc65.git synced 2025-04-05 13:37:17 +00:00
Code Issues Projects Releases Wiki Activity

New/changed optimizations

Browse Source 
git-svn-id: svn://svn.cc65.org/cc65/trunk@1443 b7a2c559-68d2-44c3-8de9-860c34a00d81
This commit is contained in:
cuz 2002-10-05 20:32:31 +00:00
parent 3f1eba36f3
commit 7ce4196829
7 changed files with 461 additions and 380 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
src/cc65/codeopt.c
View File

@ -1013,7 +1013,7 @@ static unsigned OptDecouple (CodeSeg* S)
break;
case OP65_INY:
if (E->RI->In.RegY >= 0) {
if (E->RI->In.RegY >= 0) {
Arg = MakeHexArg ((E->RI->In.RegY + 1) & 0xFF);
X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, E->LI);
}
@ -1097,7 +1097,7 @@ static unsigned OptDecouple (CodeSeg* S)
break;
case OP65_TXA:
if (E->RI->In.RegX >= 0) {
if (E->RI->In.RegX >= 0) {
Arg = MakeHexArg (In->RegX);
X = NewCodeEntry (OP65_LDA, AM65_IMM, Arg, 0, E->LI);
}
@ -1425,6 +1425,7 @@ static OptFunc DOptPtrLoad6 = { OptPtrLoad6, "OptPtrLoad6", 100, 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 DOptPush1 = { OptPush1, "OptPush1", 65, 0, 0, 0, 0, 0 };
static OptFunc DOptPushPop = { OptPushPop, "OptPushPop", 0, 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 };*/
@ -1477,7 +1478,8 @@ static OptFunc* OptFuncs[] = {
&DOptPtrLoad6,
&DOptPtrStore1,
&DOptPtrStore2,
&DOptPush1,
&DOptPush1,
&DOptPushPop,
&DOptRTS,
&DOptRTSJumps1,
&DOptRTSJumps2,
@ -1805,6 +1807,7 @@ static unsigned RunOptGroup3 (CodeSeg* S)
C += RunOptFunc (S, &DOptDupLoads, 1);
C += RunOptFunc (S, &DOptStoreLoad, 1);
C += RunOptFunc (S, &DOptTransfers, 1);
C += RunOptFunc (S, &DOptPushPop, 1);
Changes += C;

2
src/cc65/codeseg.c
View File

@ -594,6 +594,8 @@ void CS_DelEntry (CodeSeg* S, unsigned Index)
/* Delete an entry from the code segment. This includes moving any associated
* labels, removing references to labels and even removing the referenced labels
* if the reference count drops to zero.
* Note: Labels are moved forward if possible, that is, they are moved to the
* next insn (not the preceeding one).
*/
{
/* Get the code entry for the given index */

4
src/cc65/codeseg.h
View File

@ -6,7 +6,7 @@
/* */
/* */
/* */
/* (C) 2001 Ullrich von Bassewitz */
/* (C) 2001-2002 Ullrich von Bassewitz */
/* Wacholderweg 14 */
/* D-70597 Stuttgart */
/* EMail: uz@cc65.org */
@ -112,6 +112,8 @@ void CS_DelEntry (CodeSeg* S, unsigned Index);
/* Delete an entry from the code segment. This includes moving any associated
* labels, removing references to labels and even removing the referenced labels
* if the reference count drops to zero.
* Note: Labels are moved forward if possible, that is, they are moved to the
* next insn (not the preceeding one).
*/
void CS_DelEntries (CodeSeg* S, unsigned Start, unsigned Count);

117
src/cc65/coptadd.c
View File

@ -49,20 +49,27 @@
unsigned OptAdd1 (CodeSeg* S)
/* Search for the sequence
*
* jsr pushax
* ldy xxx
* ldx #$00
* lda (sp),y
* ldy #xx
* jsr ldaxysp
* jsr pushax
* ldy #yy
* jsr ldaxysp
* jsr tosaddax
*
* and replace it by:
*
* ldy xxx-2
*
* ldy #xx-1
* lda (sp),y
* clc
* ldy #yy-3
* adc (sp),y
* bcc L
* inx
* L:
* pha
* ldy #xx
* lda (sp),y
* ldy #yy-2
* adc (sp),y
* tax
* pla
*/
{
unsigned Changes = 0;
@ -71,58 +78,78 @@ unsigned OptAdd1 (CodeSeg* S)
unsigned I = 0;
while (I < CS_GetEntryCount (S)) {
CodeEntry* L[5];
CodeEntry* L[6];
/* Get next entry */
CodeEntry* E = CS_GetEntry (S, I);
L[0] = CS_GetEntry (S, I);
/* Check for the sequence */
if (CE_IsCall (E, "pushax") &&
CS_GetEntries (S, L, I+1, 5) &&
L[0]->OPC == OP65_LDY &&
if (L[0]->OPC == OP65_LDY &&
CE_KnownImm (L[0]) &&
!CE_HasLabel (L[0]) &&
L[1]->OPC == OP65_LDX &&
CE_KnownImm (L[1]) &&
L[1]->Num == 0 &&
!CE_HasLabel (L[1]) &&
L[2]->OPC == OP65_LDA &&
!CE_HasLabel (L[2]) &&
CE_IsCall (L[3], "tosaddax") &&
!CE_HasLabel (L[3])) {
!CS_RangeHasLabel (S, I+1, 5) &&
CS_GetEntries (S, L+1, I+1, 5) &&
CE_IsCall (L[1], "ldaxysp") &&
CE_IsCall (L[2], "pushax") &&
L[3]->OPC == OP65_LDY &&
CE_KnownImm (L[3]) &&
CE_IsCall (L[4], "ldaxysp") &&
CE_IsCall (L[5], "tosaddax")) {
CodeEntry* X;
CodeLabel* Label;
const char* Arg;
/* Remove the call to pushax */
CS_DelEntry (S, I);
/* Correct the stack of the first Y register load */
CE_SetNumArg (L[0], L[0]->Num - 1);
/* Correct the stack offset (needed since pushax was removed) */
CE_SetNumArg (L[0], L[0]->Num - 2);
/* lda (sp),y */
X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, "sp", 0, L[1]->LI);
CS_InsertEntry (S, X, I+1);
/* Add the clc . */
X = NewCodeEntry (OP65_CLC, AM65_IMP, 0, 0, L[3]->LI);
CS_InsertEntry (S, X, I+1);
/* Remove the load */
CS_DelEntry (S, I+3); /* lda */
CS_DelEntry (S, I+2); /* ldx */
/* Add the adc */
X = NewCodeEntry (OP65_ADC, AM65_ZP_INDY, "sp", 0, L[3]->LI);
/* clc */
X = NewCodeEntry (OP65_CLC, AM65_IMP, 0, 0, L[5]->LI);
CS_InsertEntry (S, X, I+2);
/* Generate the branch label and the branch */
Label = CS_GenLabel (S, L[4]);
X = NewCodeEntry (OP65_BCC, AM65_BRA, Label->Name, Label, L[3]->LI);
/* ldy #yy-3 */
Arg = MakeHexArg (L[3]->Num - 3);
X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, L[4]->LI);
CS_InsertEntry (S, X, I+3);
/* Generate the increment of the high byte */
X = NewCodeEntry (OP65_INX, AM65_IMP, 0, 0, L[3]->LI);
/* adc (sp),y */
X = NewCodeEntry (OP65_ADC, AM65_ZP_INDY, "sp", 0, L[5]->LI);
CS_InsertEntry (S, X, I+4);
/* Delete the call to tosaddax */
CS_DelEntry (S, I+5);
/* pha */
X = NewCodeEntry (OP65_PHA, AM65_IMP, 0, 0, L[5]->LI);
CS_InsertEntry (S, X, I+5);
/* ldy #xx (beware: L[0] has changed) */
Arg = MakeHexArg (L[0]->Num + 1);
X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, L[1]->LI);
CS_InsertEntry (S, X, I+6);
/* lda (sp),y */
X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, "sp", 0, L[1]->LI);
CS_InsertEntry (S, X, I+7);
/* ldy #yy-2 */
Arg = MakeHexArg (L[3]->Num - 2);
X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, L[4]->LI);
CS_InsertEntry (S, X, I+8);
/* adc (sp),y */
X = NewCodeEntry (OP65_ADC, AM65_ZP_INDY, "sp", 0, L[5]->LI);
CS_InsertEntry (S, X, I+9);
/* tax */
X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, L[5]->LI);
CS_InsertEntry (S, X, I+10);
/* pla */
X = NewCodeEntry (OP65_PLA, AM65_IMP, 0, 0, L[5]->LI);
CS_InsertEntry (S, X, I+11);
/* Delete the old code */
CS_DelEntries (S, I+12, 5);
/* Remember, we had changes */
++Changes;

87
src/cc65/coptind.c
View File

@ -1054,9 +1054,9 @@ unsigned OptTransfers (CodeSeg* S)
(N->Info & OF_XFR) != 0) {
/* Check if it's a transfer and back */
if ((E->OPC == OP65_TAX && N->OPC == OP65_TXA && !RegXUsed (S, I+2)) ||
(E->OPC == OP65_TAY && N->OPC == OP65_TYA && !RegYUsed (S, I+2)) ||
(E->OPC == OP65_TXA && N->OPC == OP65_TAX && !RegAUsed (S, I+2)) ||
if ((E->OPC == OP65_TAX && N->OPC == OP65_TXA && !RegXUsed (S, I+1)) ||
(E->OPC == OP65_TAY && N->OPC == OP65_TYA && !RegYUsed (S, I+1)) ||
(E->OPC == OP65_TXA && N->OPC == OP65_TAX && !RegAUsed (S, I+1)) ||
(E->OPC == OP65_TYA && N->OPC == OP65_TAY && !RegAUsed (S, I+1))) {
/* If the next insn is a conditional branch, check if the insn
@ -1074,16 +1074,16 @@ unsigned OptTransfers (CodeSeg* S)
P = CS_GetEntry (S, I-1);
if ((P->Info & OF_SETF) == 0) {
/* Does not set the flags */
goto NextEntry;
}
}
goto NextEntry;
}
}
/* Remove both transfers */
CS_DelEntry (S, I+1);
CS_DelEntry (S, I);
/* Remove both transfers */
CS_DelEntry (S, I+1);
CS_DelEntry (S, I);
/* Remember, we had changes */
++Changes;
/* Remember, we had changes */
++Changes;
}
}
@ -1099,6 +1099,71 @@ NextEntry:
unsigned OptPushPop (CodeSeg* S)
/* Remove a PHA/PLA sequence were A is not used later */
{
unsigned Changes = 0;
/* Walk over the entries */
unsigned I = 0;
while (I < CS_GetEntryCount (S)) {
CodeEntry* N;
/* Get next entry */
CodeEntry* E = CS_GetEntry (S, I);
/* Check if it is a PLA instruction that does not have a label and
* where the register value is not used later and that is not followed
* by a conditional branch.
*/
if (E->OPC == OP65_PLA &&
!CE_HasLabel (E) &&
(N = CS_GetNextEntry (S, I)) != 0 &&
(N->Info & OF_CBRA) == 0 &&
!RegAUsed (S, I+1)) {
/* Search back until we find the matching PHA instruction. If we
* find a label or another PLA somewhere in between, bail out
* since this may have side effects.
*/
unsigned J = I;
while (J-- > 0) {
/* Get the previous entry */
CodeEntry* P = CS_GetEntry (S, J);
/* Check this entry */
if (P->OPC == OP65_PHA) {
/* Found the matching push, remove both */
CS_DelEntry (S, I);
CS_DelEntry (S, J);
/* Remember that we had changes and bail out */
++Changes;
break;
} else if (CE_HasLabel (P) || P->OPC == OP65_PLA) {
/* OOPS - too dangerous! */
break;
}
}
}
/* Next entry */
++I;
}
/* Return the number of changes made */
return Changes;
}
/*****************************************************************************/
/* Optimize branch types */
/*****************************************************************************/

5
src/cc65/coptind.h
View File

@ -6,7 +6,7 @@
/* */
/* */
/* */
/* (C) 2001 Ullrich von Bassewitz */
/* (C) 2001-2002 Ullrich von Bassewitz */
/* Wacholderweg 14 */
/* D-70597 Stuttgart */
/* EMail: uz@cc65.org */
@ -105,6 +105,9 @@ unsigned OptStoreLoad (CodeSeg* S);
unsigned OptTransfers (CodeSeg* S);
/* Remove transfers from one register to another and back */
unsigned OptPushPop (CodeSeg* S);
/* Remove a PHA/PLA sequence were A is not used later */
unsigned OptBranchDist (CodeSeg* S);
/* Change branches for the distance needed. */

617
src/cc65/coptstop.c
View File

@ -50,6 +50,22 @@
/* Structure that holds the needed data */
typedef struct StackOpData StackOpData;
struct StackOpData {
CodeSeg* Code; /* Pointer to code segment */
unsigned Flags; /* Flags to remember things */
unsigned PushIndex; /* Index of call to pushax in codeseg */
unsigned OpIndex; /* Index of actual operation */
CodeEntry* PrevEntry; /* Entry before the call to pushax */
CodeEntry* PushEntry; /* Pointer to entry with call to pushax */
CodeEntry* OpEntry; /* Pointer to entry with op */
CodeEntry* NextEntry; /* Entry after the op */
const char* ZPLo; /* Lo byte of zero page loc to use */
const char* ZPHi; /* Hi byte of zero page loc to use */
unsigned IP; /* Insertion point used by some routines */
};
/* Flags returned by DirectOp */
#define OP_DIRECT 0x01 /* Direct op may be used */
#define OP_ONSTACK 0x02 /* Operand is on stack */
@ -57,13 +73,13 @@
/*****************************************************************************/
/* Helpers */
/* Helpers */
/*****************************************************************************/
static unsigned AdjustStackOffset (CodeSeg* S, unsigned Start, unsigned Stop,
unsigned Offs)
unsigned Offs)
/* Adjust the offset for all stack accesses in the range Start to Stop, both
* inclusive. The function returns the number of instructions that have been
* inserted.
@ -123,64 +139,173 @@ static unsigned AdjustStackOffset (CodeSeg* S, unsigned Start, unsigned Stop,
static unsigned DirectOp (CodeEntry* E)
static void InsertEntry (StackOpData* D, CodeEntry* E, unsigned Index)
/* Insert a new entry. Depending on Index, D->PushIndex and D->OpIndex will
* be adjusted by this function.
*/
{
/* Insert the entry into the code segment */
CS_InsertEntry (D->Code, E, Index);
/* Adjust the indices if necessary */
if (D->PushEntry && Index <= D->PushIndex) {
++D->PushIndex;
}
if (D->OpEntry && Index <= D->OpIndex) {
++D->OpIndex;
}
}
static void DelEntry (StackOpData* D, unsigned Index)
/* Delete an entry. Depending on Index, D->PushIndex and D->OpIndex will be
* adjusted by this function, and PushEntry/OpEntry may get invalidated.
*/
{
/* Delete the entry from the code segment */
CS_DelEntry (D->Code, Index);
/* Adjust the indices if necessary */
if (Index < D->PushIndex) {
--D->PushIndex;
} else if (Index == D->PushIndex) {
D->PushEntry = 0;
}
if (Index < D->OpIndex) {
--D->OpIndex;
} else if (Index == D->OpIndex) {
D->OpEntry = 0;
}
}
static void CheckDirectOp (StackOpData* D)
/* Check if the given entry is a lda instruction with an addressing mode
* that allows us to replace it by another operation (like ora). If so, we may
* use this location for the or and must not save the value in the zero
* page location.
*/
{
unsigned Flags = 0;
/* We need the entry before the push */
CodeEntry* E;
CHECK ((E = D->PrevEntry) != 0);
if (E->OPC == OP65_LDA) {
if (E->AM == AM65_IMM || E->AM == AM65_ZP || E->AM == AM65_ABS) {
/* These insns are all ok and replaceable */
Flags |= OP_DIRECT;
D->Flags |= OP_DIRECT;
} else if (E->AM == AM65_ZP_INDY &&
E->RI->In.RegY >= 0 &&
strcmp (E->Arg, "sp") == 0) {
(E->Use & REG_SP) != 0) {
/* Load from stack with known offset is also ok */
Flags |= (OP_DIRECT | OP_ONSTACK);
D->Flags |= (OP_DIRECT | OP_ONSTACK);
}
}
return Flags;
}
static void ReplacePushByStore (StackOpData* D)
/* Replace the call to the push subroutine by a store into the zero page
* location (actually, the push is not replaced, because we need it for
* later, but the name is still ok since the push will get removed at the
* end of each routine.
*/
{
CodeEntry* X;
/* Store the value into the zeropage instead of pushing it */
X = NewCodeEntry (OP65_STX, AM65_ZP, D->ZPHi, 0, D->PushEntry->LI);
InsertEntry (D, X, D->PushIndex+1);
if ((D->Flags & OP_DIRECT) == 0) {
X = NewCodeEntry (OP65_STA, AM65_ZP, D->ZPLo, 0, D->PushEntry->LI);
InsertEntry (D, X, D->PushIndex+1);
}
}
static void AddOpLow (StackOpData* D, opc_t OPC)
/* Add an op for the low byte of an operator. This function honours the
* OP_DIRECT and OP_ONSTACK flags and generates the necessary instructions.
* All code is inserted at the current insertion point.
*/
{
CodeEntry* X;
if ((D->Flags & OP_DIRECT) != 0) {
/* Op with a variable location. If the location is on the stack, we
* need to reload the Y register.
*/
if ((D->Flags & OP_ONSTACK) != 0) {
const char* Arg = MakeHexArg (D->PrevEntry->RI->In.RegY);
X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
}
X = NewCodeEntry (OPC, D->PrevEntry->AM, D->PrevEntry->Arg, 0, D->OpEntry->LI);
} else {
/* Op with temp storage */
X = NewCodeEntry (OPC, AM65_ZP, D->ZPLo, 0, D->OpEntry->LI);
}
InsertEntry (D, X, D->IP++);
}
static void AddOpHigh (StackOpData* D, opc_t OPC)
/* Add an op for the high byte of an operator. Special cases (constant values
* or similar have to be checked separately, the function covers only the
* generic case. Code is inserted at the insertion point.
*/
{
CodeEntry* X;
/* High byte is unknown */
X = NewCodeEntry (OP65_STA, AM65_ZP, D->ZPLo, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
X = NewCodeEntry (OP65_TXA, AM65_IMP, 0, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
X = NewCodeEntry (OPC, AM65_ZP, D->ZPHi, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
X = NewCodeEntry (OP65_LDA, AM65_ZP, D->ZPLo, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
}
static void RemovePushAndOp (StackOpData* D)
/* Remove the call to pushax and the call to the operator subroutine */
{
DelEntry (D, D->OpIndex);
DelEntry (D, D->PushIndex);
}
/*****************************************************************************/
/* Actual optimization functions */
/* Actual optimization functions */
/*****************************************************************************/
static unsigned Opt_staspidx (CodeSeg* S, unsigned Push, unsigned Store,
const char* ZPLo, const char* ZPHi)
static unsigned Opt_staspidx (StackOpData* D)
/* Optimize the staspidx sequence if possible */
{
CodeEntry* X;
CodeEntry* PushEntry;
CodeEntry* StoreEntry;
/* Get the push entry */
PushEntry = CS_GetEntry (S, Push);
/* Store the value into the zeropage instead of pushing it */
X = NewCodeEntry (OP65_STA, AM65_ZP, ZPLo, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+1);
X = NewCodeEntry (OP65_STX, AM65_ZP, ZPHi, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+2);
ReplacePushByStore (D);
/* Correct the index of the store and get a pointer to the entry */
Store += 2;
StoreEntry = CS_GetEntry (S, Store);
/* Inline the store */
X = NewCodeEntry (OP65_STA, AM65_ZP_INDY, ZPLo, 0, StoreEntry->LI);
CS_InsertEntry (S, X, Store+1);
/* Replace the store subroutine call by a direct op */
X = NewCodeEntry (OP65_STA, AM65_ZP_INDY, D->ZPLo, 0, D->OpEntry->LI);
InsertEntry (D, X, D->OpIndex+1);
/* Remove the push and the call to the staspidx function */
CS_DelEntry (S, Store);
CS_DelEntry (S, Push);
RemovePushAndOp (D);
/* We changed the sequence */
return 1;
@ -188,47 +313,33 @@ static unsigned Opt_staspidx (CodeSeg* S, unsigned Push, unsigned Store,
static unsigned Opt_staxspidx (CodeSeg* S, unsigned Push, unsigned Store,
const char* ZPLo, const char* ZPHi)
static unsigned Opt_staxspidx (StackOpData* D)
/* Optimize the staxspidx sequence if possible */
{
CodeEntry* X;
CodeEntry* PushEntry;
CodeEntry* StoreEntry;
/* Get the push entry */
PushEntry = CS_GetEntry (S, Push);
/* Store the value into the zeropage instead of pushing it */
X = NewCodeEntry (OP65_STA, AM65_ZP, ZPLo, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+1);
X = NewCodeEntry (OP65_STX, AM65_ZP, ZPHi, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+2);
/* Correct the index of the store and get a pointer to the entry */
Store += 2;
StoreEntry = CS_GetEntry (S, Store);
ReplacePushByStore (D);
/* Inline the store */
X = NewCodeEntry (OP65_STA, AM65_ZP_INDY, ZPLo, 0, StoreEntry->LI);
CS_InsertEntry (S, X, Store+1);
X = NewCodeEntry (OP65_INY, AM65_IMP, 0, 0, StoreEntry->LI);
CS_InsertEntry (S, X, Store+2);
if (StoreEntry->RI->In.RegX >= 0) {
X = NewCodeEntry (OP65_STA, AM65_ZP_INDY, D->ZPLo, 0, D->OpEntry->LI);
InsertEntry (D, X, D->OpIndex+1);
X = NewCodeEntry (OP65_INY, AM65_IMP, 0, 0, D->OpEntry->LI);
InsertEntry (D, X, D->OpIndex+2);
if (D->OpEntry->RI->In.RegX >= 0) {
/* Value of X is known */
const char* Arg = MakeHexArg (StoreEntry->RI->In.RegX);
X = NewCodeEntry (OP65_LDA, AM65_IMM, Arg, 0, StoreEntry->LI);
const char* Arg = MakeHexArg (D->OpEntry->RI->In.RegX);
X = NewCodeEntry (OP65_LDA, AM65_IMM, Arg, 0, D->OpEntry->LI);
} else {
/* Value unknown */
X = NewCodeEntry (OP65_TXA, AM65_IMP, 0, 0, StoreEntry->LI);
X = NewCodeEntry (OP65_TXA, AM65_IMP, 0, 0, D->OpEntry->LI);
}
CS_InsertEntry (S, X, Store+3);
X = NewCodeEntry (OP65_STA, AM65_ZP_INDY, ZPLo, 0, StoreEntry->LI);
CS_InsertEntry (S, X, Store+4);
InsertEntry (D, X, D->OpIndex+3);
X = NewCodeEntry (OP65_STA, AM65_ZP_INDY, D->ZPLo, 0, D->OpEntry->LI);
InsertEntry (D, X, D->OpIndex+4);
/* Remove the push and the call to the staspidx function */
CS_DelEntry (S, Store);
CS_DelEntry (S, Push);
RemovePushAndOp (D);
/* We changed the sequence */
return 1;
@ -236,106 +347,65 @@ static unsigned Opt_staxspidx (CodeSeg* S, unsigned Push, unsigned Store,
static unsigned Opt_tosaddax (CodeSeg* S, unsigned Push, unsigned Add,
const char* ZPLo, const char* ZPHi)
static unsigned Opt_tosaddax (StackOpData* D)
/* Optimize the tosaddax sequence if possible */
{
CodeEntry* P;
CodeEntry* N;
CodeEntry* X;
CodeEntry* PushEntry;
CodeEntry* AddEntry;
unsigned Flags;
/* We need the entry behind the add */
CHECK ((N = CS_GetNextEntry (S, Add)) != 0);
/* And the entry before the push */
CHECK ((P = CS_GetPrevEntry (S, Push)) != 0);
/* Get the push entry */
PushEntry = CS_GetEntry (S, Push);
CHECK (D->NextEntry != 0);
/* Check the entry before the push. If it's a lda instruction with an
* addressing mode that allows us to replace it, we may use this
* location for the op and must not save the value in the zero page
* location.
*/
Flags = DirectOp (P);
CheckDirectOp (D);
/* Store the value into the zeropage instead of pushing it */
X = NewCodeEntry (OP65_STX, AM65_ZP, ZPHi, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+1);
++Add; /* Correct the index */
if ((Flags & OP_DIRECT) == 0) {
X = NewCodeEntry (OP65_STA, AM65_ZP, ZPLo, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+1);
++Add; /* Correct the index */
}
/* Get a pointer to the add entry */
AddEntry = CS_GetEntry (S, Add);
ReplacePushByStore (D);
/* Inline the add */
X = NewCodeEntry (OP65_CLC, AM65_IMP, 0, 0, AddEntry->LI);
CS_InsertEntry (S, X, Add+1);
if ((Flags & OP_DIRECT) != 0) {
/* Add a variable location. If the location is on the stack, we
* need to reload the Y register.
*/
if ((Flags & OP_ONSTACK) != 0) {
X = NewCodeEntry (OP65_LDY, AM65_IMM, MakeHexArg (P->RI->In.RegY), 0, AddEntry->LI);
CS_InsertEntry (S, X, Add);
++Add;
}
X = NewCodeEntry (OP65_ADC, P->AM, P->Arg, 0, AddEntry->LI);
} else {
/* Add from temp storage */
X = NewCodeEntry (OP65_ADC, AM65_ZP, ZPLo, 0, AddEntry->LI);
}
CS_InsertEntry (S, X, Add+2);
if (PushEntry->RI->In.RegX == 0) {
D->IP = D->OpIndex+1;
X = NewCodeEntry (OP65_CLC, AM65_IMP, 0, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
/* Low byte */
AddOpLow (D, OP65_ADC);
/* High byte */
if (D->PushEntry->RI->In.RegX == 0) {
/* The high byte is the value in X plus the carry */
CodeLabel* L = CS_GenLabel (S, N);
X = NewCodeEntry (OP65_BCC, AM65_BRA, L->Name, L, AddEntry->LI);
CS_InsertEntry (S, X, Add+3);
X = NewCodeEntry (OP65_INX, AM65_IMP, 0, 0, AddEntry->LI);
CS_InsertEntry (S, X, Add+4);
} else if (AddEntry->RI->In.RegX == 0) {
/* The high byte is that of the first operand plus carry */
CodeLabel* L = CS_GenLabel (D->Code, D->NextEntry);
X = NewCodeEntry (OP65_BCC, AM65_BRA, L->Name, L, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
X = NewCodeEntry (OP65_INX, AM65_IMP, 0, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
} else if (D->OpEntry->RI->In.RegX == 0) {
/* The high byte is that of the first operand plus carry */
CodeLabel* L;
if (PushEntry->RI->In.RegX >= 0) {
if (D->PushEntry->RI->In.RegX >= 0) {
/* Value of first op high byte is known */
const char* Arg = MakeHexArg (PushEntry->RI->In.RegX);
X = NewCodeEntry (OP65_LDX, AM65_IMM, Arg, 0, AddEntry->LI);
const char* Arg = MakeHexArg (D->PushEntry->RI->In.RegX);
X = NewCodeEntry (OP65_LDX, AM65_IMM, Arg, 0, D->OpEntry->LI);
} else {
/* Value of first op high byte is unknown */
X = NewCodeEntry (OP65_LDX, AM65_ZP, ZPHi, 0, AddEntry->LI);
X = NewCodeEntry (OP65_LDX, AM65_ZP, D->ZPHi, 0, D->OpEntry->LI);
}
CS_InsertEntry (S, X, Add+3);
L = CS_GenLabel (S, N);
X = NewCodeEntry (OP65_BCC, AM65_BRA, L->Name, L, AddEntry->LI);
CS_InsertEntry (S, X, Add+4);
X = NewCodeEntry (OP65_INX, AM65_IMP, 0, 0, AddEntry->LI);
CS_InsertEntry (S, X, Add+5);
InsertEntry (D, X, D->IP++);
L = CS_GenLabel (D->Code, D->NextEntry);
X = NewCodeEntry (OP65_BCC, AM65_BRA, L->Name, L, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
X = NewCodeEntry (OP65_INX, AM65_IMP, 0, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
} else {
/* High byte is unknown */
X = NewCodeEntry (OP65_STA, AM65_ZP, ZPLo, 0, AddEntry->LI);
CS_InsertEntry (S, X, Add+3);
X = NewCodeEntry (OP65_TXA, AM65_IMP, 0, 0, AddEntry->LI);
CS_InsertEntry (S, X, Add+4);
X = NewCodeEntry (OP65_ADC, AM65_ZP, ZPHi, 0, AddEntry->LI);
CS_InsertEntry (S, X, Add+5);
X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, AddEntry->LI);
CS_InsertEntry (S, X, Add+6);
X = NewCodeEntry (OP65_LDA, AM65_ZP, ZPLo, 0, AddEntry->LI);
CS_InsertEntry (S, X, Add+7);
AddOpHigh (D, OP65_ADC);
}
/* Remove the push and the call to the tosaddax function */
CS_DelEntry (S, Add);
CS_DelEntry (S, Push);
RemovePushAndOp (D);
/* We changed the sequence */
return 1;
@ -343,79 +413,37 @@ static unsigned Opt_tosaddax (CodeSeg* S, unsigned Push, unsigned Add,
static unsigned Opt_tosandax (CodeSeg* S, unsigned Push, unsigned And,
const char* ZPLo, const char* ZPHi)
static unsigned Opt_tosandax (StackOpData* D)
/* Optimize the tosandax sequence if possible */
{
CodeEntry* P;
CodeEntry* X;
CodeEntry* PushEntry;
CodeEntry* AndEntry;
unsigned Flags;
/* Get the entry before the push */
CHECK ((P = CS_GetPrevEntry (S, Push)) != 0);
/* Get the push entry */
PushEntry = CS_GetEntry (S, Push);
/* Check the entry before the push. If it's a lda instruction with an
* addressing mode that allows us to replace it, we may use this
* location for the op and must not save the value in the zero page
* location.
*/
Flags = DirectOp (P);
CheckDirectOp (D);
/* Store the value into the zeropage instead of pushing it */
X = NewCodeEntry (OP65_STX, AM65_ZP, ZPHi, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+1);
++And; /* Correct the index */
if ((Flags & OP_DIRECT) == 0) {
X = NewCodeEntry (OP65_STA, AM65_ZP, ZPLo, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+1);
++And; /* Correct the index */
}
ReplacePushByStore (D);
/* Get a pointer to the and entry */
AndEntry = CS_GetEntry (S, And);
/* Inline the and, low byte */
D->IP = D->OpIndex + 1;
AddOpLow (D, OP65_AND);
/* Inline the and */
if ((Flags & OP_DIRECT) != 0) {
/* And with variable location. If the location is on the stack, we
* need to reload the Y register.
*/
if ((Flags & OP_ONSTACK) != 0) {
X = NewCodeEntry (OP65_LDY, AM65_IMM, MakeHexArg (P->RI->In.RegY), 0, AndEntry->LI);
CS_InsertEntry (S, X, And);
++And;
}
X = NewCodeEntry (OP65_AND, P->AM, P->Arg, 0, AndEntry->LI);
} else {
/* And with temp storage */
X = NewCodeEntry (OP65_AND, AM65_ZP, ZPLo, 0, AndEntry->LI);
}
CS_InsertEntry (S, X, And+1);
if (PushEntry->RI->In.RegX == 0 || AndEntry->RI->In.RegX == 0) {
/* High byte */
if (D->PushEntry->RI->In.RegX == 0 || D->OpEntry->RI->In.RegX == 0) {
/* The high byte is zero */
X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, AndEntry->LI);
CS_InsertEntry (S, X, And+2);
X = NewCodeEntry (OP65_LDX, AM65_IMM, "$00", 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
} else {
/* High byte is unknown */
X = NewCodeEntry (OP65_STA, AM65_ZP, ZPLo, 0, AndEntry->LI);
CS_InsertEntry (S, X, And+2);
X = NewCodeEntry (OP65_TXA, AM65_IMP, 0, 0, AndEntry->LI);
CS_InsertEntry (S, X, And+3);
X = NewCodeEntry (OP65_AND, AM65_ZP, ZPHi, 0, AndEntry->LI);
CS_InsertEntry (S, X, And+4);
X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, AndEntry->LI);
CS_InsertEntry (S, X, And+5);
X = NewCodeEntry (OP65_LDA, AM65_ZP, ZPLo, 0, AndEntry->LI);
CS_InsertEntry (S, X, And+6);
AddOpHigh (D, OP65_AND);
}
/* Remove the push and the call to the tosandax function */
CS_DelEntry (S, And);
CS_DelEntry (S, Push);
RemovePushAndOp (D);
/* We changed the sequence */
return 1;
@ -423,80 +451,38 @@ static unsigned Opt_tosandax (CodeSeg* S, unsigned Push, unsigned And,
static unsigned Opt_tosorax (CodeSeg* S, unsigned Push, unsigned Or,
const char* ZPLo, const char* ZPHi)
static unsigned Opt_tosorax (StackOpData* D)
/* Optimize the tosorax sequence if possible */
{
CodeEntry* P;
CodeEntry* X;
CodeEntry* PushEntry;
CodeEntry* OrEntry;
unsigned Flags;
/* Get the entry before the push */
CHECK ((P = CS_GetPrevEntry (S, Push)) != 0);
/* Get the push entry */
PushEntry = CS_GetEntry (S, Push);
/* Check the entry before the push. If it's a lda instruction with an
* addressing mode that allows us to replace it, we may use this
* location for the op and must not save the value in the zero page
* location.
*/
Flags = DirectOp (P);
CheckDirectOp (D);
/* Store the value into the zeropage instead of pushing it */
X = NewCodeEntry (OP65_STX, AM65_ZP, ZPHi, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+1);
++Or; /* Correct the index */
if ((Flags & OP_DIRECT) == 0) {
X = NewCodeEntry (OP65_STA, AM65_ZP, ZPLo, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+1);
++Or; /* Correct the index */
}
ReplacePushByStore (D);
/* Get a pointer to the or entry */
OrEntry = CS_GetEntry (S, Or);
/* Inline the or, low byte */
D->IP = D->OpIndex + 1;
AddOpLow (D, OP65_ORA);
/* Inline the or */
if ((Flags & OP_DIRECT) != 0) {
/* Or with variable location. If the location is on the stack, we
* need to reload the Y register.
*/
if ((Flags & OP_ONSTACK) != 0) {
X = NewCodeEntry (OP65_LDY, AM65_IMM, MakeHexArg (P->RI->In.RegY), 0, OrEntry->LI);
CS_InsertEntry (S, X, Or);
++Or;
}
X = NewCodeEntry (OP65_ORA, P->AM, P->Arg, 0, OrEntry->LI);
} else {
/* Or with temp storage */
X = NewCodeEntry (OP65_ORA, AM65_ZP, ZPLo, 0, OrEntry->LI);
}
CS_InsertEntry (S, X, Or+1);
if (PushEntry->RI->In.RegX >= 0 && OrEntry->RI->In.RegX >= 0) {
/* High byte */
if (D->PushEntry->RI->In.RegX >= 0 && D->OpEntry->RI->In.RegX >= 0) {
/* Both values known, precalculate the result */
const char* Arg = MakeHexArg (PushEntry->RI->In.RegX | OrEntry->RI->In.RegX);
X = NewCodeEntry (OP65_LDX, AM65_IMM, Arg, 0, OrEntry->LI);
CS_InsertEntry (S, X, Or+2);
} else if (PushEntry->RI->In.RegX != 0) {
const char* Arg = MakeHexArg (D->PushEntry->RI->In.RegX | D->OpEntry->RI->In.RegX);
X = NewCodeEntry (OP65_LDX, AM65_IMM, Arg, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
} else if (D->PushEntry->RI->In.RegX != 0) {
/* High byte is unknown */
X = NewCodeEntry (OP65_STA, AM65_ZP, ZPLo, 0, OrEntry->LI);
CS_InsertEntry (S, X, Or+2);
X = NewCodeEntry (OP65_TXA, AM65_IMP, 0, 0, OrEntry->LI);
CS_InsertEntry (S, X, Or+3);
X = NewCodeEntry (OP65_ORA, AM65_ZP, ZPHi, 0, OrEntry->LI);
CS_InsertEntry (S, X, Or+4);
X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, OrEntry->LI);
CS_InsertEntry (S, X, Or+5);
X = NewCodeEntry (OP65_LDA, AM65_ZP, ZPLo, 0, OrEntry->LI);
CS_InsertEntry (S, X, Or+6);
AddOpHigh (D, OP65_ORA);
}
/* Remove the push and the call to the tosorax function */
CS_DelEntry (S, Or);
CS_DelEntry (S, Push);
RemovePushAndOp (D);
/* We changed the sequence */
return 1;
@ -504,80 +490,38 @@ static unsigned Opt_tosorax (CodeSeg* S, unsigned Push, unsigned Or,
static unsigned Opt_tosxorax (CodeSeg* S, unsigned Push, unsigned Xor,
const char* ZPLo, const char* ZPHi)
static unsigned Opt_tosxorax (StackOpData* D)
/* Optimize the tosxorax sequence if possible */
{
CodeEntry* P;
CodeEntry* X;
CodeEntry* PushEntry;
CodeEntry* XorEntry;
unsigned Flags;
/* Get the entry before the push */
CHECK ((P = CS_GetPrevEntry (S, Push)) != 0);
/* Get the push entry */
PushEntry = CS_GetEntry (S, Push);
/* Check the entry before the push. If it's a lda instruction with an
* addressing mode that allows us to replace it, we may use this
* location for the op and must not save the value in the zero page
* location.
*/
Flags = DirectOp (P);
CheckDirectOp (D);
/* Store the value into the zeropage instead of pushing it */
X = NewCodeEntry (OP65_STX, AM65_ZP, ZPHi, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+1);
++Xor; /* Correct the index */
if ((Flags & OP_DIRECT) != 0) {
X = NewCodeEntry (OP65_STA, AM65_ZP, ZPLo, 0, PushEntry->LI);
CS_InsertEntry (S, X, Push+1);
++Xor; /* Correct the index */
}
ReplacePushByStore (D);
/* Get a pointer to the entry */
XorEntry = CS_GetEntry (S, Xor);
/* Inline the xor, low byte */
D->IP = D->OpIndex + 1;
AddOpLow (D, OP65_EOR);
/* Inline the xor */
if ((Flags & OP_DIRECT) != 0) {
/* Xor with a variable location. If the location is on the stack, we
* need to reload the Y register.
*/
if ((Flags & OP_ONSTACK) != 0) {
X = NewCodeEntry (OP65_LDY, AM65_IMM, MakeHexArg (P->RI->In.RegY), 0, XorEntry->LI);
CS_InsertEntry (S, X, Xor);
++Xor;
}
X = NewCodeEntry (OP65_EOR, P->AM, P->Arg, 0, XorEntry->LI);
} else {
/* Xor with temp storage */
X = NewCodeEntry (OP65_EOR, AM65_ZP, ZPLo, 0, XorEntry->LI);
}
CS_InsertEntry (S, X, Xor+1);
if (PushEntry->RI->In.RegX >= 0 && XorEntry->RI->In.RegX >= 0) {
/* High byte */
if (D->PushEntry->RI->In.RegX >= 0 && D->OpEntry->RI->In.RegX >= 0) {
/* Both values known, precalculate the result */
const char* Arg = MakeHexArg (PushEntry->RI->In.RegX ^ XorEntry->RI->In.RegX);
X = NewCodeEntry (OP65_LDX, AM65_IMM, Arg, 0, XorEntry->LI);
CS_InsertEntry (S, X, Xor+2);
} else if (PushEntry->RI->In.RegX != 0) {
const char* Arg = MakeHexArg (D->PushEntry->RI->In.RegX ^ D->OpEntry->RI->In.RegX);
X = NewCodeEntry (OP65_LDX, AM65_IMM, Arg, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
} else if (D->PushEntry->RI->In.RegX != 0) {
/* High byte is unknown */
X = NewCodeEntry (OP65_STA, AM65_ZP, ZPLo, 0, XorEntry->LI);
CS_InsertEntry (S, X, Xor+2);
X = NewCodeEntry (OP65_TXA, AM65_IMP, 0, 0, XorEntry->LI);
CS_InsertEntry (S, X, Xor+3);
X = NewCodeEntry (OP65_EOR, AM65_ZP, ZPHi, 0, XorEntry->LI);
CS_InsertEntry (S, X, Xor+4);
X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, XorEntry->LI);
CS_InsertEntry (S, X, Xor+5);
X = NewCodeEntry (OP65_LDA, AM65_ZP, ZPLo, 0, XorEntry->LI);
CS_InsertEntry (S, X, Xor+6);
AddOpHigh (D, OP65_EOR);
}
/* Remove the push and the call to the tosandax function */
CS_DelEntry (S, Xor);
CS_DelEntry (S, Push);
RemovePushAndOp (D);
/* We changed the sequence */
return 1;
@ -593,13 +537,12 @@ static unsigned Opt_tosxorax (CodeSeg* S, unsigned Push, unsigned Xor,
/* Flags for the functions */
typedef enum {
STOP_NONE, /* Nothing special */
STOP_NONE, /* Nothing special */
STOP_A_UNUSED /* Call only if a unused later */
} STOP_FLAGS;
typedef unsigned (*OptFunc) (CodeSeg* S, unsigned Push, unsigned Store,
const char* ZPLo, const char* ZPHi);
typedef unsigned (*OptFunc) (StackOpData* D);
typedef struct OptFuncDesc OptFuncDesc;
struct OptFuncDesc {
const char* Name; /* Name of the replaced runtime function */
@ -637,6 +580,33 @@ static const OptFuncDesc* FindFunc (const char* Name)
static int CmpHarmless (const void* Key, const void* Entry)
/* Compare function for bsearch */
{
return strcmp (Key, *(const char**)Entry);
}
static int HarmlessCall (const char* Name)
/* Check if this is a call to a harmless subroutine that will not interrupt
* the pushax/op sequence when encountered.
*/
{
static const char* Tab[] = {
"ldaxysp",
};
void* R = bsearch (Name,
Tab,
sizeof (Tab) / sizeof (Tab[0]),
sizeof (Tab[0]),
CmpHarmless);
return (R != 0);
}
/*****************************************************************************/
/* Code */
/*****************************************************************************/
@ -646,11 +616,12 @@ static const OptFuncDesc* FindFunc (const char* Name)
unsigned OptStackOps (CodeSeg* S)
/* Optimize operations that take operands via the stack */
{
unsigned Changes = 0; /* Number of changes in one run */
int InSeq = 0; /* Inside a sequence */
unsigned Push = 0; /* Index of pushax */
unsigned UsedRegs = 0; /* Zeropage registers used in sequence */
unsigned I;
unsigned Changes = 0; /* Number of changes in one run */
int InSeq = 0; /* Inside a sequence */
unsigned Push = 0; /* Index of pushax */
unsigned UsedRegs = 0; /* Zeropage registers used in sequence */
unsigned I;
/* Generate register info */
CS_GenRegInfo (S);
@ -694,9 +665,8 @@ unsigned OptStackOps (CodeSeg* S)
const OptFuncDesc* F = FindFunc (E->Arg);
if (F) {
const char* ZPLo = 0;
const char* ZPHi = 0;
int PreCondOk = 1;
StackOpData Data;
int PreCondOk = 1;
/* Check the flags */
if (F->Flags & STOP_A_UNUSED) {
@ -712,14 +682,14 @@ unsigned OptStackOps (CodeSeg* S)
UsedRegs |= GetRegInfo (S, I+1, REG_SREG | REG_PTR1 | REG_PTR2);
if ((UsedRegs & REG_SREG) == REG_NONE) {
/* SREG is available */
ZPLo = "sreg";
ZPHi = "sreg+1";
Data.ZPLo = "sreg";
Data.ZPHi = "sreg+1";
} else if ((UsedRegs & REG_PTR1) == REG_NONE) {
ZPLo = "ptr1";
ZPHi = "ptr1+1";
Data.ZPLo = "ptr1";
Data.ZPHi = "ptr1+1";
} else if ((UsedRegs & REG_PTR2) == REG_NONE) {
ZPLo = "ptr2";
ZPHi = "ptr2+1";
Data.ZPLo = "ptr2";
Data.ZPHi = "ptr2+1";
} else {
/* No registers available */
PreCondOk = 0;
@ -730,13 +700,22 @@ unsigned OptStackOps (CodeSeg* S)
if (PreCondOk) {
/* Adjust stack offsets */
unsigned Op = I + AdjustStackOffset (S, Push, I, 2);
Data.OpIndex = I + AdjustStackOffset (S, Push, I, 2);
/* Call the optimizer function */
Changes += F->Func (S, Push, Op, ZPLo, ZPHi);
/* Prepare the remainder of the data structure */
Data.Code = S;
Data.Flags = 0;
Data.PushIndex = Push;
Data.PrevEntry = CS_GetPrevEntry (S, Data.PushIndex);
Data.PushEntry = CS_GetEntry (S, Data.PushIndex);
Data.OpEntry = E;
Data.NextEntry = CS_GetNextEntry (S, Data.OpIndex);
/* Regenerate register info */
CS_GenRegInfo (S);
/* Call the optimizer function */
Changes += F->Func (&Data);
/* Regenerate register info */
CS_GenRegInfo (S);
}
/* End of sequence */
@ -746,7 +725,7 @@ unsigned OptStackOps (CodeSeg* S)
/* Restart the sequence */
Push = I;
UsedRegs = REG_NONE;
} else {
} else if (!HarmlessCall (E->Arg)) {
/* A call to an unkown subroutine ends the sequence */
InSeq = 0;
}