--- trunk/src/cpus/cpu_arm_instr_loadstore.c	2007/10/08 16:18:51	14
+++ trunk/src/cpus/cpu_arm_instr_loadstore.c	2007/10/08 16:19:11	18
@@ -25,7 +25,7 @@
  *  SUCH DAMAGE.
  *
  *
- *  $Id: cpu_arm_instr_loadstore.c,v 1.8 2005/10/05 21:17:32 debug Exp $
+ *  $Id: cpu_arm_instr_loadstore.c,v 1.15 2005/10/27 14:01:13 debug Exp $
  *
  *
  *  TODO:  Many things...
@@ -35,6 +35,8 @@
  *	o)  Alignment checks!
  *
  *	o)  Native load/store if the endianness is the same as the host's
+ *	    (only implemented for little endian, so far, and it assumes that
+ *	    alignment is correct!)
  *
  *	o)  "Base Updated Abort Model", which updates the base register
  *	    even if the memory access failed.
@@ -51,6 +53,11 @@
  */
 
 
+#if defined(A__SIGNED) && defined(A__H) && !defined(A__L)
+#define A__STRD
+#endif
+
+
 /*
  *  General load/store, by using memory_rw(). If at all possible, memory_rw()
  *  then inserts the page into the translation array, so that the fast
@@ -63,25 +70,43 @@
 #else
 	const int memory_rw_flags = CACHE_DATA;
 #endif
+
+#ifdef A__REG
+	uint32_t (*reg_func)(struct cpu *, struct arm_instr_call *)
+	    = (void *)(size_t)ic->arg[1];
+#endif
+
+#ifdef A__STRD
+	unsigned char data[8];
+	const int datalen = 8;
+#else
 #ifdef A__B
 	unsigned char data[1];
+	const int datalen = 1;
 #else
 #ifdef A__H
 	unsigned char data[2];
+	const int datalen = 2;
+#else
+	const int datalen = 4;
+#ifdef HOST_LITTLE_ENDIAN
+	unsigned char *data = (unsigned char *) ic->arg[2];
 #else
 	unsigned char data[4];
 #endif
 #endif
+#endif
+#endif
+
 	uint32_t addr, low_pc, offset =
 #ifndef A__U
 	    -
 #endif
 #ifdef A__REG
-	    R(cpu, ic, ic->arg[1], 0)
+	    reg_func(cpu, ic);
 #else
-	    ic->arg[1]
+	    ic->arg[1];
 #endif
-	    ;
 
 	low_pc = ((size_t)ic - (size_t)cpu->cd.arm.
 	    cur_ic_page) / sizeof(struct arm_instr_call);
@@ -98,7 +123,7 @@
 
 #ifdef A__L
 	/*  Load:  */
-	if (!cpu->memory_rw(cpu, cpu->mem, addr, data, sizeof(data),
+	if (!cpu->memory_rw(cpu, cpu->mem, addr, data, datalen,
 	    MEM_READ, memory_rw_flags)) {
 		/*  load failed, an exception was generated  */
 		return;
@@ -117,26 +142,39 @@
 #endif
 	    (data[0] + (data[1] << 8));
 #else
+#ifdef HOST_LITTLE_ENDIAN
+	/*  Nothing.  */
+#else
 	reg(ic->arg[2]) = data[0] + (data[1] << 8) +
 	    (data[2] << 16) + (data[3] << 24);
 #endif
 #endif
+#endif
 #else
 	/*  Store:  */
-#ifdef A__B
-	data[0] = reg(ic->arg[2]);
+#if !defined(A__B) && !defined(A__H) && defined(HOST_LITTLE_ENDIAN)
+#ifdef A__STRD
+	*(uint32_t *)data = reg(ic->arg[2]);
+	*(uint32_t *)(data + 4) = reg(ic->arg[2] + 4);
+#endif
 #else
-#ifdef A__H
 	data[0] = reg(ic->arg[2]);
+#ifndef A__B
 	data[1] = reg(ic->arg[2]) >> 8;
-#else
-	data[0] = reg(ic->arg[2]);
+#if !defined(A__H) || defined(A__STRD)
 	data[1] = reg(ic->arg[2]) >> 8;
 	data[2] = reg(ic->arg[2]) >> 16;
 	data[3] = reg(ic->arg[2]) >> 24;
+#ifdef A__STRD
+	data[4] = reg(ic->arg[2] + 4);
+	data[5] = reg(ic->arg[2] + 4) >> 8;
+	data[6] = reg(ic->arg[2] + 4) >> 16;
+	data[7] = reg(ic->arg[2] + 4) >> 24;
+#endif
+#endif
 #endif
 #endif
-	if (!cpu->memory_rw(cpu, cpu->mem, addr, data, sizeof(data),
+	if (!cpu->memory_rw(cpu, cpu->mem, addr, data, datalen,
 	    MEM_WRITE, memory_rw_flags)) {
 		/*  store failed, an exception was generated  */
 		return;
@@ -158,20 +196,23 @@
  */
 void A__NAME(struct cpu *cpu, struct arm_instr_call *ic)
 {
-#if !defined(A__P) && defined(A__W)
-	/*  T-bit: userland access. Use the general routine for that.  */
+#ifdef A__STRD
+	/*  Chicken out, let's do this unoptimized for now:  */
 	A__NAME__general(cpu, ic);
 #else
+#ifdef A__REG
+	uint32_t (*reg_func)(struct cpu *, struct arm_instr_call *)
+	    = (void *)(size_t)ic->arg[1];
+#endif
 	uint32_t offset =
 #ifndef A__U
 	    -
 #endif
 #ifdef A__REG
-	    R(cpu, ic, ic->arg[1], 0)
+	    reg_func(cpu, ic);
 #else
-	    ic->arg[1]
+	    ic->arg[1];
 #endif
-	    ;
 	uint32_t addr = reg(ic->arg[0])
 #ifdef A__P
 	    + offset
@@ -185,8 +226,22 @@
 #endif
 	    [addr >> 12];
 
+
+#if !defined(A__P) && defined(A__W)
+	/*
+	 *  T-bit: userland access: check the corresponding bit in the
+	 *  is_userpage array. If it is set, then we're ok. Otherwise: use the
+	 *  generic function.
+	 */
+	unsigned char x = cpu->cd.arm.is_userpage[addr >> 15];
+	if (!(x & (1 << ((addr >> 12) & 7))))
+		A__NAME__general(cpu, ic);
+	else
+#endif
+
+
 	if (page == NULL) {
-	        A__NAME__general(cpu, ic);
+		A__NAME__general(cpu, ic);
 	} else {
 #ifdef A__L
 #ifdef A__B
@@ -203,12 +258,16 @@
 #endif
 		    (page[addr & 0xfff] + (page[(addr & 0xfff) + 1] << 8));
 #else
+#ifdef HOST_LITTLE_ENDIAN
+		reg(ic->arg[2]) = *(uint32_t *)(page + (addr & 0xffc));
+#else
 		reg(ic->arg[2]) = page[addr & 0xfff] +
 		    (page[(addr & 0xfff) + 1] << 8) +
 		    (page[(addr & 0xfff) + 2] << 16) +
 		    (page[(addr & 0xfff) + 3] << 24);
 #endif
 #endif
+#endif
 #else
 #ifdef A__B
 		page[addr & 0xfff] = reg(ic->arg[2]);
@@ -217,6 +276,9 @@
 		page[addr & 0xfff] = reg(ic->arg[2]);
 		page[(addr & 0xfff)+1] = reg(ic->arg[2]) >> 8;
 #else
+#ifdef HOST_LITTLE_ENDIAN
+		*(uint32_t *)(page + (addr & 0xffc)) = reg(ic->arg[2]);
+#else
 		page[addr & 0xfff] = reg(ic->arg[2]);
 		page[(addr & 0xfff)+1] = reg(ic->arg[2]) >> 8;
 		page[(addr & 0xfff)+2] = reg(ic->arg[2]) >> 16;
@@ -224,6 +286,7 @@
 #endif
 #endif
 #endif
+#endif
 
 		/*  Index Write-back:  */
 #ifdef A__P
@@ -235,7 +298,7 @@
 		reg(ic->arg[0]) = addr + offset;
 #endif
 	}
-#endif	/*  not T-bit  */
+#endif	/*  not STRD  */
 }
 
 
@@ -272,7 +335,9 @@
 	A__NAME(cpu, ic);
 	if (ic->arg[2] == (size_t)(&cpu->cd.arm.r[ARM_PC])) {
 		cpu->pc = cpu->cd.arm.r[ARM_PC];
-		arm_pc_to_pointers(cpu);
+		if (cpu->machine->show_trace_tree)
+			cpu_functioncall_trace(cpu, cpu->pc);
+		quick_pc_to_pointers(cpu);
 	}
 #else
 	/*  Store:  */
@@ -369,3 +434,8 @@
 ((cpu->cd.arm.cpsr & ARM_FLAG_V)?1:0) ||
 (cpu->cd.arm.cpsr & ARM_FLAG_Z)) A__NAME_PC(cpu, ic); }
 #endif
+
+
+#ifdef A__STRD
+#undef A__STRD
+#endif