--- trunk/src/cpus/cpu_transputer_instr.c	2007/10/08 16:20:26	28
+++ trunk/src/cpus/cpu_transputer_instr.c	2007/10/08 16:20:40	30
@@ -25,7 +25,7 @@
  *  SUCH DAMAGE.
  *
  *
- *  $Id: cpu_transputer_instr.c,v 1.2 2006/07/21 05:46:45 debug Exp $
+ *  $Id: cpu_transputer_instr.c,v 1.6 2006/07/23 19:36:14 debug Exp $
  *
  *  INMOS transputer instructions.
  *
@@ -34,24 +34,18 @@
  *  call. If no instruction was executed, then it should be decreased. If, say,
  *  4 instructions were combined into one function and executed, then it should
  *  be increased by 3.)
+ *
+ *  NOTE: The PC does not need to be synched before e.g. memory_rw(), because
+ *        there is no MMU in the Transputer, and hence there can not be any
+ *        exceptions on memory accesses.
  */
 
 
-/*
- *  nop:  Do nothing.
- */
-X(nop)
-{
-}
-
-
 /*****************************************************************************/
 
 
 /*
  *  j:  Relative jump
- *
- *  arg[0] = relative distance from the _current_ instruction.
  */
 X(j)
 {
@@ -59,23 +53,320 @@
 	int low_pc = ((size_t)ic - (size_t)cpu->cd.transputer.cur_ic_page)
 	    / sizeof(struct transputer_instr_call);
 	cpu->pc &= ~(TRANSPUTER_IC_ENTRIES_PER_PAGE-1);
-	cpu->pc += low_pc + ic->arg[0];
+	cpu->pc += low_pc + ic->arg[0] + 1 + cpu->cd.transputer.oreg;
+	cpu->cd.transputer.oreg = 0;
 	quick_pc_to_pointers(cpu);
 }
 
 
 /*
- *  ldc: Load constant
+ *  ldlp: Load local pointer
+ */
+X(ldlp)
+{
+	cpu->cd.transputer.c = cpu->cd.transputer.b;
+	cpu->cd.transputer.b = cpu->cd.transputer.a;
+	cpu->cd.transputer.oreg |= ic->arg[0];
+	cpu->cd.transputer.a = cpu->cd.transputer.oreg * sizeof(uint32_t) +
+	    cpu->cd.transputer.wptr;
+	cpu->cd.transputer.oreg = 0;
+}
+
+
+/*
+ *  pfix: Prefix
  *
- *  arg[0] = constant
+ *  arg[0] = nibble
+ */
+X(pfix)
+{
+	cpu->cd.transputer.oreg |= ic->arg[0];
+	cpu->cd.transputer.oreg <<= 4;
+}
+
+
+/*
+ *  ldnl: Load non-local
+ */
+X(ldnl)
+{
+	uint32_t addr, w32;
+	uint8_t word[sizeof(uint32_t)];
+	unsigned char *page;
+
+	cpu->cd.transputer.oreg |= ic->arg[0];
+	addr = cpu->cd.transputer.oreg * sizeof(uint32_t) +
+	    cpu->cd.transputer.a;
+
+	page = cpu->cd.transputer.host_load[TRANSPUTER_ADDR_TO_PAGENR(addr)];
+	if (page == NULL) {
+		cpu->memory_rw(cpu, cpu->mem, addr, word, sizeof(word),
+		    MEM_READ, CACHE_DATA);
+		w32 = *(uint32_t *) &word[0];
+	} else {
+		w32 = *(uint32_t *) &page[TRANSPUTER_PC_TO_IC_ENTRY(addr)];
+	}
+
+	cpu->cd.transputer.a = LE32_TO_HOST(w32);
+	cpu->cd.transputer.oreg = 0;
+}
+
+
+/*
+ *  ldc: Load constant
  */
 X(ldc)
 {
-	/*  TODO: Is oreg really cleared like this?  */
-	cpu->cd.transputer.oreg = ic->arg[0];
 	cpu->cd.transputer.c = cpu->cd.transputer.b;
 	cpu->cd.transputer.b = cpu->cd.transputer.a;
-	cpu->cd.transputer.a = ic->arg[0];
+	cpu->cd.transputer.a = cpu->cd.transputer.oreg | ic->arg[0];
+	cpu->cd.transputer.oreg = 0;
+}
+
+
+/*
+ *  ldnlp: Load non-local pointer
+ */
+X(ldnlp)
+{
+	cpu->cd.transputer.a += (cpu->cd.transputer.oreg | ic->arg[0])
+	    * sizeof(uint32_t);
+	cpu->cd.transputer.oreg = 0;
+}
+
+
+/*
+ *  nfix: Negative prefix
+ */
+X(nfix)
+{
+	cpu->cd.transputer.oreg |= ic->arg[0];
+	cpu->cd.transputer.oreg = (~cpu->cd.transputer.oreg) << 4;
+}
+
+
+/*
+ *  ldl: Load local
+ */
+X(ldl)
+{
+	uint32_t addr;
+	uint8_t word[sizeof(uint32_t)];
+	uint32_t w32;
+	unsigned char *page;
+
+	cpu->cd.transputer.c = cpu->cd.transputer.b;
+	cpu->cd.transputer.b = cpu->cd.transputer.a;
+	cpu->cd.transputer.oreg |= ic->arg[0];
+	addr = cpu->cd.transputer.oreg * sizeof(uint32_t) +
+	    cpu->cd.transputer.wptr;
+
+	page = cpu->cd.transputer.host_load[TRANSPUTER_ADDR_TO_PAGENR(addr)];
+	if (page == NULL) {
+		cpu->memory_rw(cpu, cpu->mem, addr, word, sizeof(word),
+		    MEM_READ, CACHE_DATA);
+		w32 = *(uint32_t *) &word[0];
+printf("A w32 = %08"PRIx32"\n", w32);
+	} else {
+		w32 = *(uint32_t *) &page[TRANSPUTER_PC_TO_IC_ENTRY(addr)];
+printf("B w32 = %08"PRIx32"\n", w32);
+	}
+
+	cpu->cd.transputer.a = LE32_TO_HOST(w32);
+	cpu->cd.transputer.oreg = 0;
+}
+
+
+/*
+ *  ajw: Adjust workspace
+ */
+X(ajw)
+{
+	cpu->cd.transputer.oreg |= ic->arg[0];
+	cpu->cd.transputer.wptr += (cpu->cd.transputer.oreg * sizeof(uint32_t));
+	cpu->cd.transputer.oreg = 0;
+}
+
+
+/*
+ *  eqc: Equal to constant
+ */
+X(eqc)
+{
+	cpu->cd.transputer.a = (cpu->cd.transputer.a ==
+	    (cpu->cd.transputer.oreg | ic->arg[0]));
+	cpu->cd.transputer.oreg = 0;
+}
+
+
+/*
+ *  stl: Store local
+ */
+X(stl)
+{
+	uint32_t addr, w32;
+	unsigned char *page;
+
+	w32 = LE32_TO_HOST(cpu->cd.transputer.a);
+	cpu->cd.transputer.oreg |= ic->arg[0];
+	addr = cpu->cd.transputer.oreg * sizeof(uint32_t) +
+	    cpu->cd.transputer.wptr;
+
+	page = cpu->cd.transputer.host_store[TRANSPUTER_ADDR_TO_PAGENR(addr)];
+	if (page == NULL) {
+		cpu->memory_rw(cpu, cpu->mem, addr, (void *)&w32, sizeof(w32),
+		    MEM_WRITE, CACHE_DATA);
+	} else {
+		*((uint32_t *) &page[TRANSPUTER_PC_TO_IC_ENTRY(addr)]) = w32;
+	}
+
+	cpu->cd.transputer.a = cpu->cd.transputer.b;
+	cpu->cd.transputer.b = cpu->cd.transputer.c;
+	cpu->cd.transputer.oreg = 0;
+}
+
+
+/*
+ *  stnl: Store non-local
+ */
+X(stnl)
+{
+	uint32_t addr, w32;
+	unsigned char *page;
+
+	w32 = LE32_TO_HOST(cpu->cd.transputer.b);
+	cpu->cd.transputer.oreg |= ic->arg[0];
+	addr = cpu->cd.transputer.oreg * sizeof(uint32_t) +
+	    cpu->cd.transputer.a;
+
+	page = cpu->cd.transputer.host_store[TRANSPUTER_ADDR_TO_PAGENR(addr)];
+	if (page == NULL) {
+		cpu->memory_rw(cpu, cpu->mem, addr, (void *)&w32, sizeof(w32),
+		    MEM_WRITE, CACHE_DATA);
+	} else {
+		*((uint32_t *) &page[TRANSPUTER_PC_TO_IC_ENTRY(addr)]) = w32;
+	}
+
+	cpu->cd.transputer.a = cpu->cd.transputer.c;
+	cpu->cd.transputer.oreg = 0;
+}
+
+
+/*
+ *  opr: Operate
+ *
+ *  TODO/NOTE: This doesn't work too well with the way dyntrans is designed
+ *             to work. Maybe it should be rewritten some day. But how?
+ *             Right now it works almost 100% like an old-style interpretation
+ *             function.
+ */
+X(opr)
+{
+	cpu->cd.transputer.oreg |= ic->arg[0];
+
+	switch (cpu->cd.transputer.oreg) {
+
+	case T_OPC_F_REV:
+		{
+			uint32_t tmp = cpu->cd.transputer.b;
+			cpu->cd.transputer.b = cpu->cd.transputer.a;
+			cpu->cd.transputer.a = tmp;
+		}
+		break;
+
+	case T_OPC_F_ADD:
+		{
+			uint32_t old_a = cpu->cd.transputer.a;
+			cpu->cd.transputer.a = cpu->cd.transputer.b + old_a;
+			if ((cpu->cd.transputer.a & 0x80000000) !=
+			    (cpu->cd.transputer.b & old_a & 0x80000000))
+				cpu->cd.transputer.error = 1;
+			cpu->cd.transputer.b = cpu->cd.transputer.c;
+		}
+		break;
+
+	case T_OPC_F_SUB:
+		{
+			uint32_t old_a = cpu->cd.transputer.a;
+			cpu->cd.transputer.a = cpu->cd.transputer.b - old_a;
+			if ((cpu->cd.transputer.a & 0x80000000) !=
+			    (cpu->cd.transputer.b & old_a & 0x80000000))
+				cpu->cd.transputer.error = 1;
+			cpu->cd.transputer.b = cpu->cd.transputer.c;
+		}
+		break;
+
+	case T_OPC_F_LDPI:
+		/*  Load pointer to (next) instruction  */
+		{
+			int low_pc = ((size_t)ic -
+			    (size_t)cpu->cd.transputer.cur_ic_page)
+			    / sizeof(struct transputer_instr_call);
+			cpu->pc &= ~(TRANSPUTER_IC_ENTRIES_PER_PAGE-1);
+			cpu->pc += low_pc;
+			cpu->cd.transputer.a += cpu->pc + 1;
+		}
+		break;
+
+	case T_OPC_F_STHF:
+		cpu->cd.transputer.fptrreg0 = cpu->cd.transputer.a;
+		cpu->cd.transputer.a = cpu->cd.transputer.b;
+		cpu->cd.transputer.b = cpu->cd.transputer.c;
+		break;
+
+	case T_OPC_F_STLF:
+		cpu->cd.transputer.fptrreg1 = cpu->cd.transputer.a;
+		cpu->cd.transputer.a = cpu->cd.transputer.b;
+		cpu->cd.transputer.b = cpu->cd.transputer.c;
+		break;
+
+	case T_OPC_F_BCNT:
+		cpu->cd.transputer.a <<= 2;
+		break;
+
+	case T_OPC_F_GAJW:
+		{
+			uint32_t old_wptr = cpu->cd.transputer.wptr;
+			cpu->cd.transputer.wptr = cpu->cd.transputer.a & ~3;
+			cpu->cd.transputer.a = old_wptr;
+		}
+		break;
+
+	case T_OPC_F_WCNT:
+		cpu->cd.transputer.c = cpu->cd.transputer.b;
+		cpu->cd.transputer.b = cpu->cd.transputer.a & 3;
+		cpu->cd.transputer.a >>= 2;
+		break;
+
+	case T_OPC_F_MINT:
+		cpu->cd.transputer.c = cpu->cd.transputer.b;
+		cpu->cd.transputer.b = cpu->cd.transputer.a;
+		cpu->cd.transputer.a = 0x80000000;
+		break;
+
+	case T_OPC_F_MOVE:
+		/*  TODO: This can be optimized a lot by using the host's
+		    memmove(). The only important thing to consider is
+		    if src or dst crosses a host memblock boundary.  */
+		{
+			uint32_t i, src = cpu->cd.transputer.c,
+			    dst = cpu->cd.transputer.b;
+			uint8_t byte;
+			for (i=1; i<=cpu->cd.transputer.a; i++) {
+				cpu->memory_rw(cpu, cpu->mem, src ++, &byte,
+				    1, MEM_READ, CACHE_DATA);
+				cpu->memory_rw(cpu, cpu->mem, dst ++, &byte,
+				    1, MEM_WRITE, CACHE_DATA);
+			}
+		}
+		break;
+
+	default:fatal("UNIMPLEMENTED opr oreg 0x%"PRIx32"\n",
+		    cpu->cd.transputer.oreg);
+		exit(1);
+	}
+
+	cpu->cd.transputer.oreg = 0;
 }
 
 
@@ -157,9 +448,9 @@
 
 	switch (ib[0] >> 4) {
 
-	case 0:	/*  j, relative jump  */
+	case T_OPC_J:
+		/*  relative jump  */
 		ic->f = instr(j);
-		ic->arg[0] = (ib[0] & 0xf) + 1;
 		/*  TODO: Samepage jump!  */
 
 		if (cpu->cd.transputer.cpu_type.features & T_DEBUG
@@ -174,10 +465,66 @@
 		}
 		break;
 
-	case 4:	/*  ldc, load constant  */
+	case T_OPC_LDLP:
+		/*  load local pointer  */
+		ic->f = instr(ldlp);
+		break;
+
+	case T_OPC_PFIX:
+		/*  prefix  */
+		ic->f = instr(pfix);
+		break;
+
+	case T_OPC_LDNL:
+		/*  load non-local  */
+		ic->f = instr(ldnl);
+		break;
+
+	case T_OPC_LDC:
+		/*  load constant  */
 		ic->f = instr(ldc);
 		break;
 
+	case T_OPC_LDNLP:
+		/*  load non-local pointer  */
+		ic->f = instr(ldnlp);
+		break;
+
+	case T_OPC_NFIX:
+		/*  negative prefix  */
+		ic->f = instr(nfix);
+		break;
+
+	case T_OPC_LDL:
+		/*  load local  */
+		ic->f = instr(ldl);
+		break;
+
+	case T_OPC_AJW:
+		/*  adjust workspace  */
+		ic->f = instr(ajw);
+		break;
+
+	case T_OPC_EQC:
+		/*  equal to constant  */
+		ic->f = instr(eqc);
+		break;
+
+	case T_OPC_STL:
+		/*  store local  */
+		ic->f = instr(stl);
+		break;
+
+	case T_OPC_STNL:
+		/*  store non-local  */
+		ic->f = instr(stnl);
+		break;
+
+	case T_OPC_OPR:
+		/*  operate  */
+		ic->f = instr(opr);
+		break;
+
 	default:fatal("UNIMPLEMENTED opcode 0x%02x\n", ib[0]);
 		goto bad;
 	}