From d604ed5e3faa4763cec25146c6b4ddda93816efc Mon Sep 17 00:00:00 2001 From: cuz Date: Wed, 25 Jul 2001 21:51:40 +0000 Subject: [PATCH] More optimizations git-svn-id: svn://svn.cc65.org/cc65/trunk@822 b7a2c559-68d2-44c3-8de9-860c34a00d81 --- src/cc65/codegen.c | 54 +++++++++---- src/cc65/codeinfo.c | 4 +- src/cc65/codeopt.c | 188 ++++++++++++++++++++++++++++++++++++++------ 3 files changed, 202 insertions(+), 44 deletions(-) diff --git a/src/cc65/codegen.c b/src/cc65/codegen.c index 92b954693..9ca4eb657 100644 --- a/src/cc65/codegen.c +++ b/src/cc65/codegen.c @@ -1339,9 +1339,9 @@ void g_scale (unsigned flags, long val) } else if (val > 0) { /* Scale up */ - if ((p2 = powerof2 (val)) > 0 && p2 <= 3) { + if ((p2 = powerof2 (val)) > 0 && p2 <= 4) { - /* Factor is 2, 4 or 8, use special function */ + /* Factor is 2, 4, 8 and 16, use special function */ switch (flags & CF_TYPE) { case CF_CHAR: @@ -1354,11 +1354,11 @@ void g_scale (unsigned flags, long val) /* FALLTHROUGH */ case CF_INT: - if (CodeSizeFactor >= (p2+1)*130U) { + if (CodeSizeFactor >= (p2+1)*130U) { AddCodeLine ("stx tmp1"); while (p2--) { AddCodeLine ("asl a"); - AddCodeLine ("rol tmp1"); + AddCodeLine ("rol tmp1"); } AddCodeLine ("ldx tmp1"); } else { @@ -1394,9 +1394,9 @@ void g_scale (unsigned flags, long val) /* Scale down */ val = -val; - if ((p2 = powerof2 (val)) > 0 && p2 <= 3) { + if ((p2 = powerof2 (val)) > 0 && p2 <= 4) { - /* Factor is 2, 4 or 8, use special function */ + /* Factor is 2, 4, 8 and 16 use special function */ switch (flags & CF_TYPE) { case CF_CHAR: @@ -2959,7 +2959,7 @@ void g_asr (unsigned flags, unsigned long val) case CF_CHAR: case CF_INT: - if (val >= 1 && val <= 3) { + if (val >= 1 && val <= 4) { if (flags & CF_UNSIGNED) { AddCodeLine ("jsr shrax%ld", val); } else { @@ -2974,7 +2974,7 @@ void g_asr (unsigned flags, unsigned long val) break; case CF_LONG: - if (val >= 1 && val <= 3) { + if (val >= 1 && val <= 4) { if (flags & CF_UNSIGNED) { AddCodeLine ("jsr shreax%ld", val); } else { @@ -3042,7 +3042,7 @@ void g_asl (unsigned flags, unsigned long val) case CF_CHAR: case CF_INT: - if (val >= 1 && val <= 3) { + if (val >= 1 && val <= 4) { if (flags & CF_UNSIGNED) { AddCodeLine ("jsr shlax%ld", val); } else { @@ -3057,7 +3057,7 @@ void g_asl (unsigned flags, unsigned long val) break; case CF_LONG: - if (val >= 1 && val <= 3) { + if (val >= 1 && val <= 4) { if (flags & CF_UNSIGNED) { AddCodeLine ("jsr shleax%ld", val); } else { @@ -3559,16 +3559,31 @@ void g_le (unsigned flags, unsigned long val) */ if (flags & CF_CONST) { + /* <= is not very effective on the 6502, so try to convert + * it into < if the value is in a valid range. + */ + /* Look at the type */ switch (flags & CF_TYPE) { case CF_CHAR: if (flags & CF_FORCECHAR) { - AddCodeLine ("cmp #$%02X", (unsigned char)val); if (flags & CF_UNSIGNED) { - AddCodeLine ("jsr boolule"); + if (val < 255) { + AddCodeLine ("cmp #$%02X", (unsigned char)val+1); + AddCodeLine ("jsr boolult"); + } else { + AddCodeLine ("cmp #$%02X", (unsigned char)val); + AddCodeLine ("jsr boolule"); + } } else { - AddCodeLine ("jsr boolle"); + if (val < 127) { + AddCodeLine ("cmp #$%02X", (unsigned char)val+1); + AddCodeLine ("jsr boollt"); + } else { + AddCodeLine ("cmp #$%02X", (unsigned char)val); + AddCodeLine ("jsr boolle"); + } } return; } @@ -3577,11 +3592,16 @@ void g_le (unsigned flags, unsigned long val) case CF_INT: if (flags & CF_UNSIGNED) { unsigned L = GetLocalLabel(); + const char* Name = "boolule"; + if (val < 65535) { + ++val; + Name = "boolult"; + } AddCodeLine ("cpx #$%02X", (unsigned char)(val >> 8)); - AddCodeLine ("bne %s", LocalLabelName (L)); - AddCodeLine ("cmp #$%02X", (unsigned char)val); - g_defcodelabel (L); - AddCodeLine ("jsr boolule"); + AddCodeLine ("bne %s", LocalLabelName (L)); + AddCodeLine ("cmp #$%02X", (unsigned char)val); + g_defcodelabel (L); + AddCodeLine ("jsr %s", Name); return; } break; diff --git a/src/cc65/codeinfo.c b/src/cc65/codeinfo.c index d3dee341d..32c7041a7 100644 --- a/src/cc65/codeinfo.c +++ b/src/cc65/codeinfo.c @@ -125,9 +125,11 @@ static const FuncInfo FuncInfoTable[] = { { "shrax1", REG_AX, REG_AX }, { "shrax2", REG_AX, REG_AX }, { "shrax3", REG_AX, REG_AX }, + { "shrax4", REG_AX, REG_AX }, { "shreax1", REG_AX, REG_AX }, { "shreax2", REG_AX, REG_AX }, { "shreax3", REG_AX, REG_AX }, + { "shreax4", REG_AX, REG_AX }, { "staspidx", REG_A | REG_Y, REG_Y }, { "tosicmp", REG_AX, REG_AXY }, { "tosdiva0", REG_AX, REG_AXY }, @@ -140,7 +142,7 @@ static const FuncInfo FuncInfoTable[] = { { "tosumula0", REG_AX, REG_AXY }, { "tosumulax", REG_AX, REG_AXY }, { "tosumuleax", REG_AX, REG_AXY }, -}; +}; #define FuncInfoCount (sizeof(FuncInfoTable) / sizeof(FuncInfoTable[0])) /* Table with names of zero page locations used by the compiler */ diff --git a/src/cc65/codeopt.c b/src/cc65/codeopt.c index c596068be..b9eb3d118 100644 --- a/src/cc65/codeopt.c +++ b/src/cc65/codeopt.c @@ -181,8 +181,8 @@ static void ReplaceCmp (CodeSeg* S, unsigned I, cmp_t Cond) * @L: ... */ if ((N = CS_GetNextEntry (S, I)) == 0) { - /* No such entry */ - Internal ("Invalid program flow"); + /* No such entry */ + Internal ("Invalid program flow"); } L = CS_GenLabel (S, N); N = NewCodeEntry (OP65_BEQ, AM65_BRA, L->Name, L, E->LI); @@ -253,14 +253,6 @@ static void ReplaceCmp (CodeSeg* S, unsigned I, cmp_t Cond) -static int IsBitOp (const CodeEntry* E) -/* Check if E is one of the bit operations (and, or, eor) */ -{ - return (E->OPC == OP65_AND || E->OPC == OP65_ORA || E->OPC == OP65_EOR); -} - - - static int IsCmpToZero (const CodeEntry* E) /* Check if the given instrcuction is a compare to zero instruction */ { @@ -523,7 +515,7 @@ static unsigned OptSub2 (CodeSeg* S) */ if (CE_HasLabel (E)) { CS_MoveLabels (S, E, L[0]); - } + } /* Remember, we had changes */ ++Changes; @@ -683,7 +675,17 @@ static unsigned OptCmp2 (CodeSeg* S) CodeEntry* E = CS_GetEntry (S, I); /* Check for the sequence */ - if ((E->OPC == OP65_LDA || IsBitOp (E)) && + 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]) && @@ -760,7 +762,7 @@ static unsigned OptCmp3 (CodeSeg* S) CS_MoveEntry (S, I, I+4); /* We will replace the ldx/cpx by lda/cmp */ - CE_ReplaceOPC (L[0], OP65_LDA); + CE_ReplaceOPC (L[0], OP65_LDA); CE_ReplaceOPC (L[1], OP65_CMP); /* Beware: If the first LDA instruction had a label, we have @@ -796,7 +798,7 @@ static unsigned OptCmp4 (CodeSeg* S) * lda (sp),y * cpx #a * bne L1 - * cmp #b + * cmp #b * jne/jeq L2 */ { @@ -817,7 +819,7 @@ static unsigned OptCmp4 (CodeSeg* S) * ldy #o * lda (sp),y * dey - * ora (sp),y + * ora (sp),y * jne/jeq ... */ CE_ReplaceOPC (L[4], OP65_ORA); @@ -910,17 +912,93 @@ static unsigned OptCmp5 (CodeSeg* S) } + /* Next entry */ + ++I; + + } + + /* Return the number of changes made */ + return Changes; +} + + + +/*****************************************************************************/ +/* Optimize tests */ +/*****************************************************************************/ + + + +static unsigned OptTest1 (CodeSeg* S) +/* On a sequence + * + * stx xxx + * ora xxx + * beq/bne ... + * + * if X is zero, the sequence may be changed + * + * cmp #$00 + * beq/bne ... + * + * which may be optimized further by another step. + */ +{ + unsigned Changes = 0; + unsigned I; + + /* Generate register info for this step */ + CS_GenRegInfo (S); + + /* Walk over the entries */ + I = 0; + while (I < CS_GetEntryCount (S)) { + + CodeEntry* L[3]; + + /* Get next entry */ + L[0] = 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); + + /* We had changes */ + ++Changes; + } + /* Next entry */ ++I; } + /* Free register info */ + CS_FreeRegInfo (S); + /* Return the number of changes made */ return Changes; } + + + + /*****************************************************************************/ /* nega optimizations */ /*****************************************************************************/ @@ -930,7 +1008,7 @@ static unsigned OptCmp5 (CodeSeg* S) static unsigned OptNegA1 (CodeSeg* S) /* Check for * - * ldx #$00 + * ldx #$00 * lda .. * jsr bnega * @@ -953,7 +1031,7 @@ static unsigned OptNegA1 (CodeSeg* S) E->AM == AM65_IMM && (E->Flags & CEF_NUMARG) != 0 && E->Num == 0 && - CS_GetEntries (S, L, I+1, 2) && + CS_GetEntries (S, L, I+1, 2) && L[0]->OPC == OP65_LDA && (L[0]->Use & REG_X) == 0 && L[1]->OPC == OP65_JSR && @@ -1000,11 +1078,21 @@ static unsigned OptNegA2 (CodeSeg* S) CodeEntry* E = CS_GetEntry (S, I); /* Check for the sequence */ - if (E->OPC == OP65_LDA && + 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]) && + !CE_HasLabel (L[0]) && (L[1]->Info & OF_ZBRA) != 0) { /* Invert the branch */ @@ -1036,6 +1124,52 @@ static unsigned OptNegA2 (CodeSeg* S) 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; + + /* Generate register info for this step */ + CS_GenRegInfo (S); + + /* Walk over the entries */ + I = 0; + while (I < CS_GetEntryCount (S)) { + + /* Get next entry */ + CodeEntry* E = CS_GetEntry (S, I); + + /* Check if 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; + } + + /* Next entry */ + ++I; + + } + + /* Free register info */ + CS_FreeRegInfo (S); + + /* Return the number of changes made */ + return Changes; +} + + + +static unsigned OptNegAX2 (CodeSeg* S) /* Search for the sequence: * * lda (xx),y @@ -1082,7 +1216,7 @@ static unsigned OptNegAX1 (CodeSeg* S) /* lda --> ora */ CE_ReplaceOPC (L[2], OP65_ORA); - /* Invert the branch */ + /* Invert the branch */ CE_ReplaceOPC (L[4], GetInverseBranch (L[4]->OPC)); /* Delete the entries no longer needed. Beware: Deleting entries @@ -1107,7 +1241,7 @@ static unsigned OptNegAX1 (CodeSeg* S) -static unsigned OptNegAX2 (CodeSeg* S) +static unsigned OptNegAX3 (CodeSeg* S) /* Search for the sequence: * * lda xx @@ -1168,16 +1302,16 @@ static unsigned OptNegAX2 (CodeSeg* S) -static unsigned OptNegAX3 (CodeSeg* S) +static unsigned OptNegAX4 (CodeSeg* S) /* Search for the sequence: * - * jsr _xxx + * jsr xxx * jsr bnega(x) * jeq/jne ... * * and replace it by: * - * jsr _xxx + * jsr xxx * * jne/jeq ... */ @@ -1195,7 +1329,6 @@ static unsigned OptNegAX3 (CodeSeg* S) /* Check for the sequence */ if (E->OPC == OP65_JSR && - E->Arg[0] == '_' && CS_GetEntries (S, L, I+1, 2) && L[0]->OPC == OP65_JSR && strncmp (L[0]->Arg,"bnega",5) == 0 && @@ -1211,7 +1344,7 @@ static unsigned OptNegAX3 (CodeSeg* S) if (ByteSized) { /* Test bytes */ X = NewCodeEntry (OP65_TAX, AM65_IMP, 0, 0, L[0]->LI); - CS_InsertEntry (S, X, I+2); + CS_InsertEntry (S, X, I+2); } else { /* Test words */ X = NewCodeEntry (OP65_STX, AM65_ZP, "tmp1", 0, L[0]->LI); @@ -1288,12 +1421,15 @@ static OptFunc OptFuncs [] = { { 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 }, + /* Optimize tests */ + { OptTest1, "OptTest1", 0 }, /* Remove unused loads */ { OptUnusedLoads, "OptUnusedLoads", 0 }, { OptDuplicateLoads, "OptDuplicateLoads", 0 }, -- 2.39.5