From 0fdf581d0651ed30f2e16e9c7771b542ad23e3e6 Mon Sep 17 00:00:00 2001 From: cuz Date: Thu, 4 Oct 2001 21:34:14 +0000 Subject: [PATCH] Added special multiplication routines git-svn-id: svn://svn.cc65.org/cc65/trunk@1012 b7a2c559-68d2-44c3-8de9-860c34a00d81 --- libsrc/runtime/Makefile | 3 ++ libsrc/runtime/mulax10.s | 32 +++++++++++++++ libsrc/runtime/mulax3.s | 27 +++++++++++++ libsrc/runtime/mulax5.s | 29 ++++++++++++++ src/cc65/codegen.c | 87 ++++++++++++++++++++++------------------ src/cc65/codeopt.c | 50 +++++++++++++++++++++++ 6 files changed, 190 insertions(+), 38 deletions(-) create mode 100644 libsrc/runtime/mulax10.s create mode 100644 libsrc/runtime/mulax3.s create mode 100644 libsrc/runtime/mulax5.s diff --git a/libsrc/runtime/Makefile b/libsrc/runtime/Makefile index cbb14f699..7b2f4732d 100644 --- a/libsrc/runtime/Makefile +++ b/libsrc/runtime/Makefile @@ -130,6 +130,9 @@ OBJS = add.o \ makebool.o \ mod.o \ mul.o \ + mulax3.o \ + mulax5.o \ + mulax10.o \ ne.o \ neg.o \ or.o \ diff --git a/libsrc/runtime/mulax10.s b/libsrc/runtime/mulax10.s new file mode 100644 index 000000000..18a17a450 --- /dev/null +++ b/libsrc/runtime/mulax10.s @@ -0,0 +1,32 @@ +; +; Ullrich von Bassewitz, 04.10.2001 +; +; CC65 runtime: Multiply the primary register +; + + .export mulax10 + .importzp ptr1 + +.proc mulax10 + + sta ptr1 + stx ptr1+1 + asl a + rol ptr1+1 + asl a + rol ptr1+1 + clc + adc ptr1 + pha + txa + adc ptr1+1 + sta ptr1+1 + pla + asl a + rol ptr1+1 + ldx ptr1+1 + rts + +.endproc + + diff --git a/libsrc/runtime/mulax3.s b/libsrc/runtime/mulax3.s new file mode 100644 index 000000000..07625f8cf --- /dev/null +++ b/libsrc/runtime/mulax3.s @@ -0,0 +1,27 @@ +; +; Ullrich von Bassewitz, 04.10.2001 +; +; CC65 runtime: Multiply the primary register +; + + .export mulax3 + .importzp ptr1 + +.proc mulax3 + + sta ptr1 + stx ptr1+1 + asl a + rol ptr1+1 + clc + adc ptr1 + pha + txa + adc ptr1+1 + tax + pla + rts + +.endproc + + diff --git a/libsrc/runtime/mulax5.s b/libsrc/runtime/mulax5.s new file mode 100644 index 000000000..dbfa9fc95 --- /dev/null +++ b/libsrc/runtime/mulax5.s @@ -0,0 +1,29 @@ +; +; Ullrich von Bassewitz, 04.10.2001 +; +; CC65 runtime: Multiply the primary register +; + + .export mulax5 + .importzp ptr1 + +.proc mulax5 + + sta ptr1 + stx ptr1+1 + asl a + rol ptr1+1 + asl a + rol ptr1+1 + clc + adc ptr1 + pha + txa + adc ptr1+1 + tax + pla + rts + +.endproc + + diff --git a/src/cc65/codegen.c b/src/cc65/codegen.c index 2e1ebc63b..b79b53304 100644 --- a/src/cc65/codegen.c +++ b/src/cc65/codegen.c @@ -2526,7 +2526,7 @@ void g_mul (unsigned flags, unsigned long val) if (flags & CF_CONST && (p2 = powerof2 (val)) >= 0) { /* Generate a shift instead */ g_asl (flags, p2); - return; + return; } /* If the right hand side is const, the lhs is not on stack but still @@ -2538,50 +2538,61 @@ void g_mul (unsigned flags, unsigned long val) case CF_CHAR: if (flags & CF_FORCECHAR) { - /* Handle some special cases */ - switch (val) { - - case 3: - AddCodeLine ("sta tmp1"); - AddCodeLine ("asl a"); - AddCodeLine ("clc"); - AddCodeLine ("adc tmp1"); - return; - - case 5: - AddCodeLine ("sta tmp1"); - AddCodeLine ("asl a"); - AddCodeLine ("asl a"); - AddCodeLine ("clc"); + /* Handle some special cases */ + switch (val) { + + case 3: + AddCodeLine ("sta tmp1"); + AddCodeLine ("asl a"); + AddCodeLine ("clc"); AddCodeLine ("adc tmp1"); - return; - - case 10: - AddCodeLine ("sta tmp1"); - AddCodeLine ("asl a"); - AddCodeLine ("asl a"); - AddCodeLine ("clc"); - AddCodeLine ("adc tmp1"); - AddCodeLine ("asl a"); - return; - } + return; + + case 5: + AddCodeLine ("sta tmp1"); + AddCodeLine ("asl a"); + AddCodeLine ("asl a"); + AddCodeLine ("clc"); + AddCodeLine ("adc tmp1"); + return; + + case 10: + AddCodeLine ("sta tmp1"); + AddCodeLine ("asl a"); + AddCodeLine ("asl a"); + AddCodeLine ("clc"); + AddCodeLine ("adc tmp1"); + AddCodeLine ("asl a"); + return; + } } /* FALLTHROUGH */ - case CF_INT: - break; + case CF_INT: + switch (val) { + case 3: + AddCodeLine ("jsr mulax3"); + return; + case 5: + AddCodeLine ("jsr mulax5"); + return; + case 10: + AddCodeLine ("jsr mulax10"); + return; + } + break; - case CF_LONG: - break; + case CF_LONG: + break; - default: - typeerror (flags); - } + default: + typeerror (flags); + } - /* If we go here, we didn't emit code. Push the lhs on stack and fall - * into the normal, non-optimized stuff. - */ - flags &= ~CF_FORCECHAR; /* Handle chars as ints */ + /* If we go here, we didn't emit code. Push the lhs on stack and fall + * into the normal, non-optimized stuff. + */ + flags &= ~CF_FORCECHAR; /* Handle chars as ints */ g_push (flags & ~CF_CONST, 0); } diff --git a/src/cc65/codeopt.c b/src/cc65/codeopt.c index 9ba4be02f..288a069b7 100644 --- a/src/cc65/codeopt.c +++ b/src/cc65/codeopt.c @@ -115,6 +115,53 @@ static unsigned OptShift1 (CodeSeg* S) +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; + + /* Walk over the entries */ + unsigned I = 0; + while (I < CS_GetEntryCount (S)) { + + /* Get next entry */ + CodeEntry* E = CS_GetEntry (S, I); + + /* 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); + } + + /* Delete the call to shlax */ + CS_DelEntry (S, I); + + /* Remember, we had changes */ + ++Changes; + + } + + /* Next entry */ + ++I; + + } + + /* Return the number of changes made */ + return Changes; +} + + + /*****************************************************************************/ /* Optimize stores through pointers */ /*****************************************************************************/ @@ -1299,6 +1346,7 @@ static OptFunc DOptPtrLoad6 = { OptPtrLoad6, "OptPtrLoad6", 0, 0, 0, 0 static OptFunc DOptPtrStore1 = { OptPtrStore1, "OptPtrStore1", 0, 0, 0, 0, 0 }; static OptFunc DOptPtrStore2 = { OptPtrStore2, "OptPtrStore2", 0, 0, 0, 0, 0 }; static OptFunc DOptShift1 = { OptShift1, "OptShift1", 0, 0, 0, 0, 0 }; +static OptFunc DOptShift2 = { OptShift2, "OptShift2", 0, 0, 0, 0, 0 }; static OptFunc DOptSize1 = { OptSize1, "OptSize1", 0, 0, 0, 0, 0 }; static OptFunc DOptSize2 = { OptSize2, "OptSize2", 0, 0, 0, 0, 0 }; static OptFunc DOptStackOps = { OptStackOps, "OptStackOps", 0, 0, 0, 0, 0 }; @@ -1349,6 +1397,7 @@ static OptFunc* OptFuncs[] = { &DOptRTS, &DOptRTSJumps, &DOptShift1, + &DOptShift2, &DOptSize1, &DOptSize2, &DOptStackOps, @@ -1596,6 +1645,7 @@ static void RunOptGroup1 (CodeSeg* S) RunOptFunc (S, &DOptAdd1, 1); RunOptFunc (S, &DOptAdd2, 1); RunOptFunc (S, &DOptShift1, 1); + RunOptFunc (S, &DOptShift2, 1); } -- 2.39.5