/*****************************************************************************/
/* */
-/* coptstop.c */
+/* coptstop.c */
/* */
-/* Optimize operations that take operands via the stack */
+/* Optimize operations that take operands via the stack */
/* */
/* */
/* */
-/* (C) 2001-2009 Ullrich von Bassewitz */
-/* Roemerstrasse 52 */
-/* D-70794 Filderstadt */
-/* EMail: uz@cc65.org */
+/* (C) 2001-2019, Ullrich von Bassewitz */
+/* Roemerstrasse 52 */
+/* D-70794 Filderstadt */
+/* EMail: uz@cc65.org */
/* */
/* */
/* This software is provided 'as-is', without any expressed or implied */
#include "codeent.h"
#include "codeinfo.h"
#include "coptstop.h"
+#include "error.h"
LI_NONE = 0x00,
LI_DIRECT = 0x01, /* Direct op may be used */
LI_RELOAD_Y = 0x02, /* Reload index register Y */
- LI_TRANSFER = 0x04, /* Loaded value is transfered */
- LI_REMOVE = 0x08, /* Load may be removed */
+ LI_REMOVE = 0x04, /* Load may be removed */
+ LI_DONT_REMOVE = 0x08, /* Load may not be removed */
+ LI_DUP_LOAD = 0x10, /* Duplicate load */
} LI_FLAGS;
/* Structure that tells us how to load the lhs values */
CodeEntry* LoadEntry; /* The actual entry, 0 if invalid */
int XferIndex; /* Index of transfer insn */
CodeEntry* XferEntry; /* The actual transfer entry */
- unsigned char Offs; /* Stack offset if data is on stack */
+ int Offs; /* Stack offset if data is on stack */
};
/* Now combined for both registers */
const OptFuncDesc* OptFunc;
/* ZP register usage inside the sequence */
- unsigned UsedRegs;
+ unsigned ZPUsage;
/* Register load information for lhs and rhs */
LoadInfo Lhs;
static void ClearLoadRegInfo (LoadRegInfo* RI)
/* Clear a LoadRegInfo struct */
{
- RI->LoadIndex = -1;
- RI->XferIndex = -1;
RI->Flags = LI_NONE;
-}
-
-
-
-static void InvalidateLoadRegInfo (LoadRegInfo* RI)
-/* Invalidate a LoadRegInfo struct */
-{
RI->LoadIndex = -1;
RI->XferIndex = -1;
+ RI->Offs = 0;
}
static void AdjustLoadRegInfo (LoadRegInfo* RI, int Index, int Change)
/* Adjust a load register info struct after deleting or inserting an entry
- * with a given index
- */
+** with a given index
+*/
{
CHECK (abs (Change) == 1);
if (Change < 0) {
+static void HonourUseAndChg (LoadRegInfo* RI, unsigned Reg, const CodeEntry* E)
+/* Honour use and change flags for an instruction */
+{
+ if (E->Chg & Reg) {
+ ClearLoadRegInfo (RI);
+ } else if ((E->Use & Reg) && RI->LoadIndex >= 0) {
+ RI->Flags |= LI_DONT_REMOVE;
+ }
+}
+
+
+
static void TrackLoads (LoadInfo* LI, CodeEntry* E, int I)
/* Track loads for a code entry */
{
if (E->Info & OF_LOAD) {
+
+ LoadRegInfo* RI = 0;
+
+ /* Determine, which register was loaded */
if (E->Chg & REG_A) {
- LI->A.LoadIndex = I;
- LI->A.XferIndex = -1;
+ RI = &LI->A;
+ } else if (E->Chg & REG_X) {
+ RI = &LI->X;
+ } else if (E->Chg & REG_Y) {
+ RI = &LI->Y;
}
- if (E->Chg & REG_X) {
- LI->X.LoadIndex = I;
- LI->X.XferIndex = -1;
+ CHECK (RI != 0);
+
+ /* If we had a load or xfer op before, this is a duplicate load which
+ ** can cause problems if it encountered between the pushax and the op,
+ ** so remember it.
+ */
+ if (RI->LoadIndex >= 0 || RI->XferIndex >= 0) {
+ RI->Flags |= LI_DUP_LOAD;
}
- if (E->Chg & REG_Y) {
- LI->Y.LoadIndex = I;
- LI->Y.XferIndex = -1;
+
+ /* Remember the load */
+ RI->LoadIndex = I;
+ RI->XferIndex = -1;
+
+ /* Set load flags */
+ RI->Flags &= ~(LI_DIRECT | LI_RELOAD_Y);
+ if (E->AM == AM65_IMM || E->AM == AM65_ZP || E->AM == AM65_ABS) {
+ /* These insns are all ok and replaceable */
+ RI->Flags |= LI_DIRECT;
+ } else if (E->AM == AM65_ZP_INDY &&
+ RegValIsKnown (E->RI->In.RegY) &&
+ strcmp (E->Arg, "sp") == 0) {
+ /* A load from the stack with known offset is also ok, but in this
+ ** case we must reload the index register later. Please note that
+ ** a load indirect via other zero page locations is not ok, since
+ ** these locations may change between the push and the actual
+ ** operation.
+ */
+ RI->Offs = (unsigned char) E->RI->In.RegY;
+ RI->Flags |= (LI_DIRECT | LI_RELOAD_Y);
}
+
+
} else if (E->Info & OF_XFR) {
+
+ /* Determine source and target of the transfer and handle the TSX insn */
+ LoadRegInfo* Src;
+ LoadRegInfo* Tgt;
switch (E->OPC) {
- case OP65_TAX:
- LI->X.LoadIndex = LI->A.LoadIndex;
- LI->X.XferIndex = I;
- break;
- case OP65_TAY:
- LI->Y.LoadIndex = LI->A.LoadIndex;
- LI->Y.XferIndex = I;
- break;
- case OP65_TXA:
- LI->A.LoadIndex = LI->X.LoadIndex;
- LI->A.XferIndex = I;
- break;
- case OP65_TYA:
- LI->A.LoadIndex = LI->Y.LoadIndex;
- LI->A.XferIndex = I;
- break;
- default:
- break;
+ case OP65_TAX: Src = &LI->A; Tgt = &LI->X; break;
+ case OP65_TAY: Src = &LI->A; Tgt = &LI->Y; break;
+ case OP65_TXA: Src = &LI->X; Tgt = &LI->A; break;
+ case OP65_TYA: Src = &LI->Y; Tgt = &LI->A; break;
+ case OP65_TSX: ClearLoadRegInfo (&LI->X); return;
+ case OP65_TXS: return;
+ default: Internal ("Unknown XFR insn in TrackLoads");
+ }
+
+ /* If we had a load or xfer op before, this is a duplicate load which
+ ** can cause problems if it encountered between the pushax and the op,
+ ** so remember it.
+ */
+ if (Tgt->LoadIndex >= 0 || Tgt->XferIndex >= 0) {
+ Tgt->Flags |= LI_DUP_LOAD;
+ }
+
+ /* Transfer the data */
+ Tgt->LoadIndex = Src->LoadIndex;
+ Tgt->XferIndex = I;
+ Tgt->Offs = Src->Offs;
+ Tgt->Flags &= ~(LI_DIRECT | LI_RELOAD_Y);
+ Tgt->Flags |= Src->Flags & (LI_DIRECT | LI_RELOAD_Y);
+
+ } else if (CE_IsCallTo (E, "ldaxysp") && RegValIsKnown (E->RI->In.RegY)) {
+
+ /* If we had a load or xfer op before, this is a duplicate load which
+ ** can cause problems if it encountered between the pushax and the op,
+ ** so remember it for both registers involved.
+ */
+ if (LI->A.LoadIndex >= 0 || LI->A.XferIndex >= 0) {
+ LI->A.Flags |= LI_DUP_LOAD;
}
- } else if (CE_IsCallTo (E, "ldaxysp")) {
+ if (LI->X.LoadIndex >= 0 || LI->X.XferIndex >= 0) {
+ LI->X.Flags |= LI_DUP_LOAD;
+ }
+
/* Both registers set, Y changed */
LI->A.LoadIndex = I;
LI->A.XferIndex = -1;
+ LI->A.Flags |= (LI_DIRECT | LI_RELOAD_Y);
+ LI->A.Offs = (unsigned char) E->RI->In.RegY - 1;
+
LI->X.LoadIndex = I;
LI->X.XferIndex = -1;
- InvalidateLoadRegInfo (&LI->Y);
+ LI->X.Flags |= (LI_DIRECT | LI_RELOAD_Y);
+ LI->X.Offs = (unsigned char) E->RI->In.RegY;
+
+ ClearLoadRegInfo (&LI->Y);
} else {
- if (E->Chg & REG_A) {
- InvalidateLoadRegInfo (&LI->A);
- }
- if (E->Chg & REG_X) {
- InvalidateLoadRegInfo (&LI->X);
- }
- if (E->Chg & REG_Y) {
- InvalidateLoadRegInfo (&LI->Y);
- }
+ HonourUseAndChg (&LI->A, REG_A, E);
+ HonourUseAndChg (&LI->X, REG_X, E);
+ HonourUseAndChg (&LI->Y, REG_Y, E);
}
}
/*****************************************************************************/
-/* Helpers */
+/* Helpers */
/*****************************************************************************/
static void InsertEntry (StackOpData* D, CodeEntry* E, int Index)
/* Insert a new entry. Depending on Index, D->PushIndex and D->OpIndex will
- * be adjusted by this function.
- */
+** be adjusted by this function.
+*/
{
/* Insert the entry into the code segment */
CS_InsertEntry (D->Code, E, Index);
static void DelEntry (StackOpData* D, int Index)
/* Delete an entry. Depending on Index, D->PushIndex and D->OpIndex will be
- * adjusted by this function, and PushEntry/OpEntry may get invalidated.
- */
+** adjusted by this function, and PushEntry/OpEntry may get invalidated.
+*/
{
/* Delete the entry from the code segment */
CS_DelEntry (D->Code, Index);
static void AdjustStackOffset (StackOpData* D, unsigned Offs)
/* Adjust the offset for all stack accesses in the range PushIndex to OpIndex.
- * OpIndex is adjusted according to the insertions.
- */
+** OpIndex is adjusted according to the insertions.
+*/
{
/* Walk over all entries */
int I = D->PushIndex + 1;
while (I < D->OpIndex) {
- CodeEntry* E = CS_GetEntry (D->Code, I);
+ CodeEntry* E = CS_GetEntry (D->Code, I);
int NeedCorrection = 0;
- if ((E->Use & REG_SP) != 0) {
+ if ((E->Use & REG_SP) != 0) {
- /* Check for some things that should not happen */
- CHECK (E->AM == AM65_ZP_INDY || E->RI->In.RegY >= (short) Offs);
- CHECK (strcmp (E->Arg, "sp") == 0);
+ /* Check for some things that should not happen */
+ CHECK (E->AM == AM65_ZP_INDY || E->RI->In.RegY >= (short) Offs);
+ CHECK (strcmp (E->Arg, "sp") == 0);
/* We need to correct this one */
NeedCorrection = 1;
if (NeedCorrection) {
- /* Get the code entry before this one. If it's a LDY, adjust the
- * value.
- */
- CodeEntry* P = CS_GetPrevEntry (D->Code, I);
- if (P && P->OPC == OP65_LDY && CE_IsConstImm (P)) {
+ /* Get the code entry before this one. If it's a LDY, adjust the
+ ** value.
+ */
+ CodeEntry* P = CS_GetPrevEntry (D->Code, I);
+ if (P && P->OPC == OP65_LDY && CE_IsConstImm (P)) {
- /* The Y load is just before the stack access, adjust it */
- CE_SetNumArg (P, P->Num - Offs);
+ /* The Y load is just before the stack access, adjust it */
+ CE_SetNumArg (P, P->Num - Offs);
- } else {
+ } else {
- /* Insert a new load instruction before the stack access */
- const char* Arg = MakeHexArg (E->RI->In.RegY - Offs);
- CodeEntry* X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, E->LI);
- InsertEntry (D, X, I++);
+ /* Insert a new load instruction before the stack access */
+ const char* Arg = MakeHexArg (E->RI->In.RegY - Offs);
+ CodeEntry* X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, E->LI);
+ InsertEntry (D, X, I++);
- }
+ }
/* If we need the value of Y later, be sure to reload it */
if (RegYUsed (D->Code, I+1)) {
- const char* Arg = MakeHexArg (E->RI->In.RegY);
- CodeEntry* X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, E->LI);
- InsertEntry (D, X, I+1);
+ const char* Arg = MakeHexArg (E->RI->In.RegY);
+ CodeEntry* X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, E->LI);
+ InsertEntry (D, X, I+1);
- /* Skip this instruction in the next round */
- ++I;
+ /* Skip this instruction in the next round */
+ ++I;
}
- }
+ }
- /* Next entry */
- ++I;
+ /* Next entry */
+ ++I;
}
/* If we have rhs load insns that load from stack, we'll have to adjust
- * the offsets for these also.
- */
+ ** the offsets for these also.
+ */
if (D->Rhs.A.Flags & LI_RELOAD_Y) {
D->Rhs.A.Offs -= Offs;
}
-static void CheckOneDirectOp (LoadRegInfo* LI, unsigned char Offs)
-/* 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.
- */
-{
- /* Get the load entry */
- CodeEntry* E = LI->LoadEntry;
- if (E == 0) {
- /* No load insn */
- return;
- }
-
- /* Check the load entry */
- if (E) {
- /* Must check the call first since addressing mode is ABS, so second
- * "if" will catch otherwise.
- */
- if (CE_IsCallTo (E, "ldaxysp")) {
- /* Same as single loads from stack. Since we must distinguish
- * between A and X here, the necessary offset is passed to the
- * function as a parameter.
- */
- LI->Offs = (unsigned char) E->RI->In.RegY - Offs;
- LI->Flags |= (LI_DIRECT | LI_RELOAD_Y);
- } else if (E->AM == AM65_IMM || E->AM == AM65_ZP || E->AM == AM65_ABS) {
- /* These insns are all ok and replaceable */
- LI->Flags |= LI_DIRECT;
- } else if (E->AM == AM65_ZP_INDY &&
- RegValIsKnown (E->RI->In.RegY) &&
- strcmp (E->Arg, "sp") == 0) {
- /* A load from the stack with known offset is also ok, but in this
- * case we must reload the index register later. Please note that
- * a load indirect via other zero page locations is not ok, since
- * these locations may change between the push and the actual
- * operation.
- */
- LI->Offs = (unsigned char) E->RI->In.RegY;
- LI->Flags |= (LI_DIRECT | LI_RELOAD_Y);
- }
- }
-}
-
-
-
-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.
- */
-{
- /* Check flags for all load instructions */
- CheckOneDirectOp (&D->Lhs.A, 1);
- CheckOneDirectOp (&D->Lhs.X, 0);
- CheckOneDirectOp (&D->Rhs.A, 1);
- CheckOneDirectOp (&D->Rhs.X, 0);
-}
-
-
-
static void AddStoreA (StackOpData* D)
/* Add a store to zero page after the push insn */
{
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).
- */
+** 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).
+*/
{
/* Store the value into the zeropage instead of pushing it. Check high
- * byte first so that the store is later in A/X order.
- */
+ ** byte first so that the store is later in A/X order.
+ */
if ((D->Lhs.X.Flags & LI_DIRECT) == 0) {
AddStoreX (D);
}
static void AddOpLow (StackOpData* D, opc_t OPC, LoadInfo* LI)
/* Add an op for the low byte of an operator. This function honours the
- * OP_DIRECT and OP_RELOAD_Y flags and generates the necessary instructions.
- * All code is inserted at the current insertion point.
- */
+** OP_DIRECT and OP_RELOAD_Y flags and generates the necessary instructions.
+** All code is inserted at the current insertion point.
+*/
{
CodeEntry* X;
if ((LI->A.Flags & LI_DIRECT) != 0) {
- /* Op with a variable location. If the location is on the stack, we
- * need to reload the Y register.
- */
+ /* Op with a variable location. If the location is on the stack, we
+ ** need to reload the Y register.
+ */
if ((LI->A.Flags & LI_RELOAD_Y) == 0) {
/* opc ... */
} else {
- /* Op with temp storage */
- X = NewCodeEntry (OPC, AM65_ZP, D->ZPLo, 0, D->OpEntry->LI);
+ /* 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, LoadInfo* LI, int KeepResult)
/* 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.
- */
+** or similar) have to be checked separately, the function covers only the
+** generic case. Code is inserted at the insertion point.
+*/
{
CodeEntry* X;
if (KeepResult) {
- /* pha */
- X = NewCodeEntry (OP65_PHA, AM65_IMP, 0, 0, D->OpEntry->LI);
- InsertEntry (D, X, D->IP++);
+ /* pha */
+ X = NewCodeEntry (OP65_PHA, AM65_IMP, 0, 0, D->OpEntry->LI);
+ InsertEntry (D, X, D->IP++);
}
/* txa */
/* opc xxx */
CodeEntry* LoadX = LI->X.LoadEntry;
- X = NewCodeEntry (OPC, LoadX->AM, LoadX->Arg, 0, D->OpEntry->LI);
+ X = NewCodeEntry (OPC, LoadX->AM, LoadX->Arg, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
} else {
}
if (KeepResult) {
- /* tax */
- X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, D->OpEntry->LI);
- InsertEntry (D, X, D->IP++);
+ /* tax */
+ X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, D->OpEntry->LI);
+ InsertEntry (D, X, D->IP++);
- /* pla */
- X = NewCodeEntry (OP65_PLA, AM65_IMP, 0, 0, D->OpEntry->LI);
- InsertEntry (D, X, D->IP++);
+ /* pla */
+ X = NewCodeEntry (OP65_PLA, AM65_IMP, 0, 0, D->OpEntry->LI);
+ InsertEntry (D, X, D->IP++);
}
}
/* Remove register load insns */
{
/* Both registers may be loaded with one insn, but DelEntry will in this
- * case clear the other one.
- */
- if (LI->A.Flags & LI_REMOVE) {
+ ** case clear the other one.
+ */
+ if ((LI->A.Flags & (LI_REMOVE | LI_DONT_REMOVE)) == LI_REMOVE) {
if (LI->A.LoadIndex >= 0) {
DelEntry (D, LI->A.LoadIndex);
}
DelEntry (D, LI->A.XferIndex);
}
}
- if (LI->X.Flags & LI_REMOVE) {
+ if ((LI->X.Flags & (LI_REMOVE | LI_DONT_REMOVE)) == LI_REMOVE) {
if (LI->X.LoadIndex >= 0) {
DelEntry (D, LI->X.LoadIndex);
}
static int IsRegVar (StackOpData* D)
/* If the value pushed is that of a zeropage variable, replace ZPLo and ZPHi
- * in the given StackOpData struct by the variable and return true. Otherwise
- * leave D untouched and return false.
- */
+** in the given StackOpData struct by the variable and return true. Otherwise
+** leave D untouched and return false.
+*/
{
CodeEntry* LoadA = D->Lhs.A.LoadEntry;
CodeEntry* LoadX = D->Lhs.X.LoadEntry;
/*****************************************************************************/
-/* Actual optimization functions */
+/* Actual optimization functions */
/*****************************************************************************/
CodeLabel* L;
/* Create a call to the boolean transformer function and a label for this
- * insn. This is needed for all variants. Other insns are inserted *before*
- * the call.
- */
+ ** insn. This is needed for all variants. Other insns are inserted *before*
+ ** the call.
+ */
X = NewCodeEntry (OP65_JSR, AM65_ABS, BoolTransformer, 0, D->OpEntry->LI);
InsertEntry (D, X, D->OpIndex + 1);
L = CS_GenLabel (D->Code, X);
/* If the lhs is direct (but not stack relative), encode compares with lhs
- * effectively reverting the order (which doesn't matter for ==).
- */
+ ** effectively reverting the order (which doesn't matter for ==).
+ */
if ((D->Lhs.A.Flags & (LI_DIRECT | LI_RELOAD_Y)) == LI_DIRECT &&
(D->Lhs.X.Flags & (LI_DIRECT | LI_RELOAD_Y)) == LI_DIRECT) {
D->IP = D->OpIndex+1;
- /* Add operand for low byte */
- AddOpLow (D, OP65_CMP, &D->Rhs);
+ /* Add operand for low byte */
+ AddOpLow (D, OP65_CMP, &D->Rhs);
/* bne L */
X = NewCodeEntry (OP65_BNE, AM65_BRA, L->Name, L, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
- /* Add operand for high byte */
- AddOpHigh (D, OP65_CMP, &D->Rhs, 0);
+ /* Add operand for high byte */
+ AddOpHigh (D, OP65_CMP, &D->Rhs, 0);
} else {
+static unsigned Opt_tosshift (StackOpData* D, const char* Name)
+/* Optimize shift sequences. */
+{
+ CodeEntry* X;
+
+ /* Store the value into the zeropage instead of pushing it */
+ ReplacePushByStore (D);
+
+ /* If the lhs is direct (but not stack relative), we can just reload the
+ ** data later.
+ */
+ if ((D->Lhs.A.Flags & (LI_DIRECT | LI_RELOAD_Y)) == LI_DIRECT &&
+ (D->Lhs.X.Flags & (LI_DIRECT | LI_RELOAD_Y)) == LI_DIRECT) {
+
+ CodeEntry* LoadX = D->Lhs.X.LoadEntry;
+ CodeEntry* LoadA = D->Lhs.A.LoadEntry;
+
+ /* Inline the shift */
+ D->IP = D->OpIndex+1;
+
+ /* tay */
+ X = NewCodeEntry (OP65_TAY, AM65_IMP, 0, 0, D->OpEntry->LI);
+ InsertEntry (D, X, D->IP++);
+
+ /* lda */
+ X = NewCodeEntry (OP65_LDA, LoadA->AM, LoadA->Arg, 0, D->OpEntry->LI);
+ InsertEntry (D, X, D->IP++);
+
+ /* ldx */
+ X = NewCodeEntry (OP65_LDX, LoadX->AM, LoadX->Arg, 0, D->OpEntry->LI);
+ InsertEntry (D, X, D->IP++);
+
+ /* Lhs load entries can be removed */
+ D->Lhs.X.Flags |= LI_REMOVE;
+ D->Lhs.A.Flags |= LI_REMOVE;
+
+ } else {
+
+ /* Save lhs into zeropage and reload later */
+ AddStoreX (D);
+ AddStoreA (D);
+
+ /* Be sure to setup IP after adding the stores, otherwise it will get
+ ** messed up.
+ */
+ D->IP = D->OpIndex+1;
+
+ /* tay */
+ X = NewCodeEntry (OP65_TAY, AM65_IMP, 0, 0, D->OpEntry->LI);
+ InsertEntry (D, X, D->IP++);
+
+ /* lda zp */
+ X = NewCodeEntry (OP65_LDA, AM65_ZP, D->ZPLo, 0, D->OpEntry->LI);
+ InsertEntry (D, X, D->IP++);
+
+ /* ldx zp+1 */
+ X = NewCodeEntry (OP65_LDX, AM65_ZP, D->ZPHi, 0, D->OpEntry->LI);
+ InsertEntry (D, X, D->IP++);
+
+ }
+
+ /* jsr shlaxy/aslaxy/whatever */
+ X = NewCodeEntry (OP65_JSR, AM65_ABS, Name, 0, D->OpEntry->LI);
+ InsertEntry (D, X, D->IP++);
+
+ /* Remove the push and the call to the shift function */
+ RemoveRemainders (D);
+
+ /* We changed the sequence */
+ return 1;
+}
+
+
+
static unsigned Opt___bzero (StackOpData* D)
/* Optimize the __bzero sequence */
{
}
/* If the return value of __bzero is used, we have to add code to reload
- * a/x from the pointer variable.
- */
+ ** a/x from the pointer variable.
+ */
if (RegAXUsed (D->Code, D->OpIndex+1)) {
X = NewCodeEntry (OP65_LDA, AM65_ZP, D->ZPLo, 0, D->OpEntry->LI);
InsertEntry (D, X, D->OpIndex+1);
}
/* X is always zero, A contains the size of the data area to zero.
- * Note: A may be zero, in which case the operation is null op.
- */
+ ** Note: A may be zero, in which case the operation is null op.
+ */
if (D->OpEntry->RI->In.RegA != 0) {
/* lda #$00 */
/* Loop using the sign bit */
/* ldy #count-1 */
- Arg = MakeHexArg (D->OpEntry->RI->In.RegA - 1);
+ Arg = MakeHexArg (D->OpEntry->RI->In.RegA - 1);
X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, D->OpEntry->LI);
InsertEntry (D, X, D->OpIndex+2);
InsertEntry (D, X, D->OpIndex+4);
/* cpy #count */
- Arg = MakeHexArg (D->OpEntry->RI->In.RegA);
+ Arg = MakeHexArg (D->OpEntry->RI->In.RegA);
X = NewCodeEntry (OP65_CPY, AM65_IMM, Arg, 0, D->OpEntry->LI);
InsertEntry (D, X, D->OpIndex+5);
if (RegValIsKnown (D->OpEntry->RI->In.RegY)) {
/* Value of Y is known */
- const char* Arg = MakeHexArg (D->OpEntry->RI->In.RegY + 1);
- X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, D->OpEntry->LI);
+ const char* Arg = MakeHexArg (D->OpEntry->RI->In.RegY + 1);
+ X = NewCodeEntry (OP65_LDY, AM65_IMM, Arg, 0, D->OpEntry->LI);
} else {
X = NewCodeEntry (OP65_INY, AM65_IMP, 0, 0, D->OpEntry->LI);
}
InsertEntry (D, X, D->OpIndex+2);
if (RegValIsKnown (D->OpEntry->RI->In.RegX)) {
- /* Value of X is known */
- const char* Arg = MakeHexArg (D->OpEntry->RI->In.RegX);
- X = NewCodeEntry (OP65_LDA, AM65_IMM, Arg, 0, D->OpEntry->LI);
+ /* Value of X is known */
+ 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, D->OpEntry->LI);
+ /* Value unknown */
+ X = NewCodeEntry (OP65_TXA, AM65_IMP, 0, 0, D->OpEntry->LI);
}
InsertEntry (D, X, D->OpIndex+3);
InsertEntry (D, X, D->OpIndex+4);
/* If we remove staxspidx, we must restore the Y register to what the
- * function would return.
- */
+ ** function would return.
+ */
X = NewCodeEntry (OP65_LDY, AM65_IMM, "$00", 0, D->OpEntry->LI);
InsertEntry (D, X, D->OpIndex+5);
CHECK (D->NextEntry != 0);
/* Check if the X register is known and zero when the add is done, and
- * if the add is followed by
- *
- * ldy #$00
- * jsr ldauidx ; or ldaidx
- *
- * If this is true, the addition does actually add an offset to a pointer
- * before it is dereferenced. Since both subroutines take an offset in Y,
- * we can pass the offset (instead of #$00) and remove the addition
- * alltogether.
- */
+ ** if the add is followed by
+ **
+ ** ldy #$00
+ ** jsr ldauidx ; or ldaidx
+ **
+ ** If this is true, the addition does actually add an offset to a pointer
+ ** before it is dereferenced. Since both subroutines take an offset in Y,
+ ** we can pass the offset (instead of #$00) and remove the addition
+ ** alltogether.
+ */
if (D->OpEntry->RI->In.RegX == 0 &&
D->NextEntry->OPC == OP65_LDY &&
CE_IsKnownImm (D->NextEntry, 0) &&
AddStoreA (D);
/* Replace the ldy by a tay. Be sure to create the new entry before
- * deleting the ldy, since we will reference the line info from this
- * insn.
- */
+ ** deleting the ldy, since we will reference the line info from this
+ ** insn.
+ */
X = NewCodeEntry (OP65_TAY, AM65_IMP, 0, 0, D->NextEntry->LI);
DelEntry (D, D->OpIndex + 1);
InsertEntry (D, X, D->OpIndex + 1);
/* Replace the call to ldaidx/ldauidx. Since X is already zero, and
- * the ptr is in the zero page location, we just need to load from
- * the pointer, and fix X in case of ldaidx.
- */
+ ** the ptr is in the zero page location, we just need to load from
+ ** the pointer, and fix X in case of ldaidx.
+ */
X = NewCodeEntry (OP65_LDA, AM65_ZP_INDY, D->ZPLo, 0, N->LI);
DelEntry (D, D->OpIndex + 2);
InsertEntry (D, X, D->OpIndex + 2);
X = NewCodeEntry (OP65_INX, AM65_IMP, 0, 0, D->OpEntry->LI);
InsertEntry (D, X, D->IP++);
- } else if (D->OpEntry->RI->In.RegX == 0 &&
- (RegValIsKnown (D->PushEntry->RI->In.RegX) ||
- (D->Lhs.X.Flags & LI_RELOAD_Y) == 0)) {
+ } else if (D->OpEntry->RI->In.RegX == 0 &&
+ (RegValIsKnown (D->PushEntry->RI->In.RegX) ||
+ (D->Lhs.X.Flags & LI_RELOAD_Y) == 0)) {
/* The high byte is that of the first operand plus carry */
CodeLabel* L;
X = NewCodeEntry (OP65_LDX, AM65_IMM, Arg, 0, D->OpEntry->LI);
} else {
/* Value of first op high byte is unknown. Load from ZP or
- * original storage.
- */
- if (D->Lhs.X.Flags & LI_DIRECT) {
- CodeEntry* LoadX = D->Lhs.X.LoadEntry;
- X = NewCodeEntry (OP65_LDX, LoadX->AM, LoadX->Arg, 0, D->OpEntry->LI);
- } else {
+ ** original storage.
+ */
+ if (D->Lhs.X.Flags & LI_DIRECT) {
+ CodeEntry* LoadX = D->Lhs.X.LoadEntry;
+ X = NewCodeEntry (OP65_LDX, LoadX->AM, LoadX->Arg, 0, D->OpEntry->LI);
+ } else {
X = NewCodeEntry (OP65_LDX, AM65_ZP, D->ZPHi, 0, D->OpEntry->LI);
- }
+ }
}
InsertEntry (D, X, D->IP++);
+static unsigned Opt_tosaslax (StackOpData* D)
+/* Optimize the tosaslax sequence */
+{
+ return Opt_tosshift (D, "aslaxy");
+}
+
+
+
+static unsigned Opt_tosasrax (StackOpData* D)
+/* Optimize the tosasrax sequence */
+{
+ return Opt_tosshift (D, "asraxy");
+}
+
+
+
static unsigned Opt_toseqax (StackOpData* D)
/* Optimize the toseqax sequence */
{
CodeLabel* L;
- /* Inline the sbc */
+ /* Inline the compare */
D->IP = D->OpIndex+1;
/* Must be true because of OP_RHS_LOAD */
+static unsigned Opt_tosshlax (StackOpData* D)
+/* Optimize the tosshlax sequence */
+{
+ return Opt_tosshift (D, "shlaxy");
+}
+
+
+
+static unsigned Opt_tosshrax (StackOpData* D)
+/* Optimize the tosshrax sequence */
+{
+ return Opt_tosshift (D, "shraxy");
+}
+
+
+
static unsigned Opt_tossubax (StackOpData* D)
/* Optimize the tossubax sequence. Note: subtraction is not commutative! */
{
/* High byte */
if (RegValIsKnown (D->PushEntry->RI->In.RegX) &&
RegValIsKnown (D->OpEntry->RI->In.RegX)) {
- /* Both values known, precalculate the result */
- 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++);
+ /* Both values known, precalculate the result */
+ 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 */
+ /* High byte is unknown */
AddOpHigh (D, OP65_EOR, &D->Lhs, 1);
}
/*****************************************************************************/
-/* Code */
+/* Code */
/*****************************************************************************/
{ "staxspidx", Opt_staxspidx, REG_AX, OP_NONE },
{ "tosaddax", Opt_tosaddax, REG_NONE, OP_NONE },
{ "tosandax", Opt_tosandax, REG_NONE, OP_NONE },
+ { "tosaslax", Opt_tosaslax, REG_NONE, OP_NONE },
+ { "tosasrax", Opt_tosasrax, REG_NONE, OP_NONE },
{ "toseqax", Opt_toseqax, REG_NONE, OP_NONE },
{ "tosgeax", Opt_tosgeax, REG_NONE, OP_RHS_LOAD_DIRECT },
{ "tosltax", Opt_tosltax, REG_NONE, OP_RHS_LOAD_DIRECT },
{ "tosneax", Opt_tosneax, REG_NONE, OP_NONE },
{ "tosorax", Opt_tosorax, REG_NONE, OP_NONE },
+ { "tosshlax", Opt_tosshlax, REG_NONE, OP_NONE },
+ { "tosshrax", Opt_tosshrax, REG_NONE, OP_NONE },
{ "tossubax", Opt_tossubax, REG_NONE, OP_RHS_LOAD_DIRECT },
{ "tosugeax", Opt_tosugeax, REG_NONE, OP_RHS_LOAD_DIRECT },
{ "tosugtax", Opt_tosugtax, REG_NONE, OP_RHS_LOAD_DIRECT },
static int CmpFunc (const void* Key, const void* Func)
/* Compare function for bsearch */
{
- return strcmp (Key, ((const OptFuncDesc*) Func)->Name);
+ return strcmp (Key, ((const OptFuncDesc*) Func)->Name);
}
static const OptFuncDesc* FindFunc (const char* Name)
/* Find the function with the given name. Return a pointer to the table entry
- * or NULL if the function was not found.
- */
+** or NULL if the function was not found.
+*/
{
return bsearch (Name, FuncTable, FUNC_COUNT, sizeof(OptFuncDesc), CmpFunc);
}
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.
- */
+** the pushax/op sequence when encountered.
+*/
{
- static const char* Tab[] = {
+ static const char* const Tab[] = {
"aslax1",
"aslax2",
"aslax3",
"aslax4",
+ "aslaxy",
"asrax1",
"asrax2",
"asrax3",
"asrax4",
+ "asraxy",
"bnegax",
+ "complax",
+ "decax1",
+ "decax2",
+ "decax3",
+ "decax4",
+ "decax5",
+ "decax6",
+ "decax7",
+ "decax8",
+ "decaxy",
+ "incax1",
+ "incax2",
+ "incax3",
+ "incax4",
+ "incax5",
+ "incax6",
+ "incax7",
+ "incax8",
+ "incaxy",
"ldaxidx",
"ldaxysp",
"negax",
"shlax2",
"shlax3",
"shlax4",
+ "shlaxy",
"shrax1",
"shrax2",
"shrax3",
"shrax4",
+ "shraxy",
};
void* R = bsearch (Name,
/* Reset the given data structure */
{
Data->OptFunc = 0;
- Data->UsedRegs = REG_NONE;
+ Data->ZPUsage = REG_NONE;
ClearLoadInfo (&Data->Lhs);
ClearLoadInfo (&Data->Rhs);
static int PreCondOk (StackOpData* D)
/* Check if the preconditions for a call to the optimizer subfunction are
- * satisfied. As a side effect, this function will also choose the zero page
- * register to use.
- */
+** satisfied. As a side effect, this function will also choose the zero page
+** register to use.
+*/
{
/* Check the flags */
unsigned UnusedRegs = D->OptFunc->UnusedRegs;
}
}
}
+ if ((D->Rhs.A.Flags | D->Rhs.X.Flags) & LI_DUP_LOAD) {
+ /* Cannot optimize */
+ return 0;
+ }
- /* Determine the zero page locations to use */
- if ((D->UsedRegs & REG_PTR1) == REG_NONE) {
+ /* Determine the zero page locations to use. We've tracked the used
+ ** ZP locations, so try to find some for us that are unused.
+ */
+ if ((D->ZPUsage & REG_PTR1) == REG_NONE) {
D->ZPLo = "ptr1";
D->ZPHi = "ptr1+1";
- } else if ((D->UsedRegs & REG_SREG) == REG_NONE) {
+ } else if ((D->ZPUsage & REG_SREG) == REG_NONE) {
D->ZPLo = "sreg";
D->ZPHi = "sreg+1";
- } else if ((D->UsedRegs & REG_PTR2) == REG_NONE) {
+ } else if ((D->ZPUsage & REG_PTR2) == REG_NONE) {
D->ZPLo = "ptr2";
D->ZPHi = "ptr2+1";
} else {
/*****************************************************************************/
-/* Code */
+/* Code */
/*****************************************************************************/
{
unsigned Changes = 0; /* Number of changes in one run */
StackOpData Data;
- unsigned I;
+ int I;
+ int OldEntryCount; /* Old number of entries */
+ unsigned UsedRegs = 0; /* Registers used */
+ unsigned ChangedRegs = 0;/* Registers changed */
+
enum {
Initialize,
} State = Initialize;
- /* Generate register info */
- CS_GenRegInfo (S);
-
/* Remember the code segment in the info struct */
Data.Code = S;
/* Look for a call to pushax followed by a call to some other function
- * that takes it's first argument on the stack, and the second argument
- * in the primary register.
- * It depends on the code between the two if we can handle/transform the
- * sequence, so check this code for the following list of things:
- *
- * - the range must be a basic block (one entry, one exit)
- * - there may not be accesses to local variables with unknown
- * offsets (because we have to adjust these offsets).
- * - no subroutine calls
- * - no jump labels
- *
- * Since we need a zero page register later, do also check the
- * intermediate code for zero page use.
- */
+ ** that takes it's first argument on the stack, and the second argument
+ ** in the primary register.
+ ** It depends on the code between the two if we can handle/transform the
+ ** sequence, so check this code for the following list of things:
+ **
+ ** - the range must be a basic block (one entry, one exit)
+ ** - there may not be accesses to local variables with unknown
+ ** offsets (because we have to adjust these offsets).
+ ** - no subroutine calls
+ ** - no jump labels
+ **
+ ** Since we need a zero page register later, do also check the
+ ** intermediate code for zero page use.
+ */
I = 0;
- while (I < CS_GetEntryCount (S)) {
+ while (I < (int)CS_GetEntryCount (S)) {
- /* Get the next entry */
- CodeEntry* E = CS_GetEntry (S, I);
+ /* Get the next entry */
+ CodeEntry* E = CS_GetEntry (S, I);
/* Actions depend on state */
switch (State) {
case Initialize:
ResetStackOpData (&Data);
+ UsedRegs = ChangedRegs = REG_NONE;
State = Search;
/* FALLTHROUGH */
case Search:
/* While searching, track register load insns, so we can tell
- * what is in a register once pushax is encountered.
- */
+ ** what is in a register once pushax is encountered.
+ */
+ if (CE_HasLabel (E)) {
+ /* Currently we don't track across branches */
+ ClearLoadInfo (&Data.Lhs);
+ }
if (CE_IsCallTo (E, "pushax")) {
Data.PushIndex = I;
State = FoundPush;
case FoundPush:
/* We' found a pushax before. Search for a stack op that may
- * follow and in the meantime, track zeropage usage and check
- * for code that will disable us from translating the sequence.
- */
+ ** follow and in the meantime, track zeropage usage and check
+ ** for code that will disable us from translating the sequence.
+ */
+ if (CE_HasLabel (E)) {
+ /* Currently we don't track across branches */
+ ClearLoadInfo (&Data.Rhs);
+ }
if (E->OPC == OP65_JSR) {
/* Subroutine call: Check if this is one of the functions,
- * we're going to replace.
- */
+ ** we're going to replace.
+ */
Data.OptFunc = FindFunc (E->Arg);
if (Data.OptFunc) {
/* Remember the op index and go on */
break;
} else if (!HarmlessCall (E->Arg)) {
/* A call to an unkown subroutine: We need to start
- * over after the last pushax. Note: This will also
- * happen if we encounter a call to pushax!
- */
+ ** over after the last pushax. Note: This will also
+ ** happen if we encounter a call to pushax!
+ */
I = Data.PushIndex;
State = Initialize;
break;
} else {
/* Track register usage */
- Data.UsedRegs |= (E->Use | E->Chg);
+ Data.ZPUsage |= (E->Use | E->Chg);
TrackLoads (&Data.Rhs, E, I);
}
} else if (E->Info & OF_STORE && (E->Chg & REG_ZP) == 0) {
/* Too dangerous - there may be a change of a variable
- * within the sequence.
- */
+ ** within the sequence.
+ */
I = Data.PushIndex;
State = Initialize;
break;
E->RI->In.RegY < 2)) {
/* If we are using the stack, and we don't have "indirect Y"
- * addressing mode, or the value of Y is unknown, or less
- * than two, we cannot cope with this piece of code. Having
- * an unknown value of Y means that we cannot correct the
- * stack offset, while having an offset less than two means
- * that the code works with the value on stack which is to
- * be removed.
- */
+ ** addressing mode, or the value of Y is unknown, or less
+ ** than two, we cannot cope with this piece of code. Having
+ ** an unknown value of Y means that we cannot correct the
+ ** stack offset, while having an offset less than two means
+ ** that the code works with the value on stack which is to
+ ** be removed.
+ */
I = Data.PushIndex;
State = Initialize;
break;
} else {
/* Other stuff: Track register usage */
- Data.UsedRegs |= (E->Use | E->Chg);
+ Data.ZPUsage |= (E->Use | E->Chg);
TrackLoads (&Data.Rhs, E, I);
}
+ /* If the registers from the push (A/X) are used before they're
+ ** changed, we cannot change the sequence, because this would
+ ** with a high probability change the register contents.
+ */
+ UsedRegs |= E->Use;
+ if ((UsedRegs & ~ChangedRegs) & REG_AX) {
+ I = Data.PushIndex;
+ State = Initialize;
+ break;
+ }
+ ChangedRegs |= E->Chg;
break;
case FoundOp:
/* Track zero page location usage beyond this point */
- Data.UsedRegs |= GetRegInfo (S, I, REG_SREG | REG_PTR1 | REG_PTR2);
+ Data.ZPUsage |= GetRegInfo (S, I, REG_SREG | REG_PTR1 | REG_PTR2);
/* Finalize the load info */
FinalizeLoadInfo (&Data.Lhs, S);
FinalizeLoadInfo (&Data.Rhs, S);
- /* Set flags for direct operations */
- CheckDirectOp (&Data);
-
- /* If the Lhs loads do load from zeropage, we have to include
- * them into UsedRegs registers used. The Rhs loads have already
- * been tracked.
- */
+ /* Check if the lhs loads from zeropage. If this is true, these
+ ** zero page locations have to be added to ZPUsage, because
+ ** they cannot be used for intermediate storage. In addition,
+ ** if one of these zero page locations is destroyed between
+ ** pushing the lhs and the actual operation, we cannot use the
+ ** original zero page locations for the final op, but must
+ ** use another ZP location to save them.
+ */
+ ChangedRegs &= REG_ZP;
if (Data.Lhs.A.LoadEntry && Data.Lhs.A.LoadEntry->AM == AM65_ZP) {
- Data.UsedRegs |= Data.Lhs.A.LoadEntry->Use;
+ Data.ZPUsage |= Data.Lhs.A.LoadEntry->Use;
+ if ((Data.Lhs.A.LoadEntry->Use & ChangedRegs) != 0) {
+ Data.Lhs.A.Flags &= ~(LI_DIRECT | LI_RELOAD_Y);
+ }
}
if (Data.Lhs.X.LoadEntry && Data.Lhs.X.LoadEntry->AM == AM65_ZP) {
- Data.UsedRegs |= Data.Lhs.X.LoadEntry->Use;
+ Data.ZPUsage |= Data.Lhs.X.LoadEntry->Use;
+ if ((Data.Lhs.X.LoadEntry->Use & ChangedRegs) != 0) {
+ Data.Lhs.X.Flags &= ~(LI_DIRECT | LI_RELOAD_Y);
+ }
}
/* Check the preconditions. If they aren't ok, reset the insn
- * pointer to the pushax and start over. We will loose part of
- * load tracking but at least a/x has probably lost between
- * pushax and here and will be tracked again when restarting.
- */
+ ** pointer to the pushax and start over. We will loose part of
+ ** load tracking but at least a/x has probably lost between
+ ** pushax and here and will be tracked again when restarting.
+ */
if (!PreCondOk (&Data)) {
I = Data.PushIndex;
State = Initialize;
Data.OpEntry = CS_GetEntry (S, Data.OpIndex);
Data.NextEntry = CS_GetNextEntry (S, Data.OpIndex);
+ /* Remember the current number of code lines */
+ OldEntryCount = CS_GetEntryCount (S);
+
/* Adjust stack offsets to account for the upcoming removal */
AdjustStackOffset (&Data, 2);
/* Regenerate register info, since AdjustStackOffset changed
- * the code
- */
+ ** the code
+ */
CS_GenRegInfo (S);
/* Call the optimizer function */
Changes += Data.OptFunc->Func (&Data);
+ /* Since the function may have added or deleted entries,
+ ** correct the index.
+ */
+ I += CS_GetEntryCount (S) - OldEntryCount;
+
/* Regenerate register info */
CS_GenRegInfo (S);
/* Done */
State = Initialize;
- break;
+ continue;
- }
+ }
- /* Next entry */
- ++I;
+ /* Next entry */
+ ++I;
}
- /* Free the register info */
- CS_FreeRegInfo (S);
-
/* Return the number of changes made */
return Changes;
}
-
-
-