kernel: move VBR at the end of the user RAM area on fx-9860G

This leaves more space available for the heap.

README.md

@@ -21,7 +21,7 @@ Currently, this includes:
 * A stripped-down version of the [Grisu2b floating-point representation
   algorithm](https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf)
   with α=-59 and γ=-56, by Florian Loitsch. See `src/3rdparty/grisu2b_59_56/README`
-  for details [the original code here](https://drive.google.com/open?id=0BwvYOx00EwKmejFIMjRORTFLcTA).
+  for details, and [the original code here](https://drive.google.com/open?id=0BwvYOx00EwKmejFIMjRORTFLcTA).
 
 ## Programming interface
 

TODO

@@ -1,9 +1,8 @@
 Extensions on existing code:
+* kernel:    better restore to userspace before panic (ensure BL=0 IMASK=0)
 * kernel:    check if cpu_setVBR() really needs to be perma-mapped
-* stdio:     support %f in printf
 * project:   add license file
 * kernel:    group linker script symbols in a single header file
-* kernel:    be consistent about *tlb_mapped_memory() in hw_detect()
 * bopti:     try to display fullscreen images with TLB access + DMA on fxcg50
 * dma:       fx9860g support (need to switch it on and update the Makefile)
 * core:      try to leave add-in without reset in case of panic
@@ -12,7 +11,6 @@ Extensions on existing code:
 * core:      review forgotten globals and MPU addresses not in <gint/mpu/*.h>
 * core:      run destructors when a task-switch results in leaving the app
 * core rtc:  use qdiv10 to massively improve division performance
-* topti:     let the font specify letter and word spacing
 
 Future directions.
 * A complete file system abstraction

fx9860g.ld

@@ -16,12 +16,15 @@ MEMORY
 	/* Userspace mapping of the add-in (G1A header takes 0x200 bytes) */
 	rom   (rx):   o = 0x00300200, l = 500k
 
-	/* User RAM is mapped at 0x08100000 through MMU; 8k on SH3, 32k on SH4.
-	   Currently gint provides access to 8k, with three blocks:
+	/* User RAM is mapped at 0x08100000 through MMU; usually 8k on SH3, 32k
+	   on SH4. This script exposes only 6k to the user, reserving:
 	   * 0x200 bytes for text accessed without the TLB when SR.BL=1, linked
 	     into the rram region below, then loaded dynamically
-	   * 6k for user content
-	   * 0x600 bytes for the VBR space, also without MMU */
+	   * 0x600 bytes for the VBR space, also without MMU
+	   On SH3, the VBR space consumes these 0x600 bytes. On SH4, it spans
+	   0x1100 bytes near the end of the user RAM, which is larger; the 6k
+	   left for the user are honored in both cases. Unused memory from the
+	   exposed 6k and non-exposed memory is available through malloc(). */
 	ram   (rw):   o = 0x08100200, l = 6k
 
 	/* This region represents the first block of user RAM. Linker arranges
@@ -172,6 +175,9 @@ SECTIONS
 		*(.gint.bss .gint.bss.sh3)
 
 		. = ALIGN(16);
+
+		/* End of user RAM */
+		_euram = . ;
 	} > ram :NONE
 
 	_sgbss = SIZEOF(.gint.bss);

fxcg50.ld

@@ -179,6 +179,9 @@ SECTIONS
 	.gint.bss (NOLOAD) : {
 		*(.gint.bss)
 		. = ALIGN(16);
+
+		/* End of user RAM */
+		_euram = . ;
 	} > ram :NONE
 
 	_sgbss = SIZEOF(.gint.bss);

src/kernel/cpu.h

@@ -24,6 +24,9 @@
 extern uint32_t (*cpu_setVBR)(uint32_t vbr, void (*conf_intc)(int arg),
 	int arg);
 
+/* cpu_getVBR(): Query the current VBR address */
+uint32_t cpu_getVBR(void);
+
 /* cpu_setCPUOPM(): Change the CPU Operation Mode register
 
    Updates the CPU Operation Mode with the specified settings, then performs a

src/kernel/cpu.s

@@ -2,7 +2,8 @@
 **	gint:core:vbr - Assembler-level VBR management
 */
 
-.global	_cpu_setVBR
+.global _cpu_getVBR
+.global _cpu_setVBR
 .global _cpu_setCPUOPM
 .global _cpu_getCPUOPM
 .global _cpu_getSR
@@ -52,6 +53,12 @@ _cpu_setVBR:
 
 .text
 
+/* cpu_getVBR(): Query the current VBR address */
+_cpu_getVBR:
+	stc	vbr, r0
+	rts
+	nop
+
 /* cpu_setCPUOPM(): Change the CPU Operation Mode register */
 _cpu_setCPUOPM:
 	/* Set CPUOPM as requested */

src/kernel/kernel.c

@@ -127,9 +127,17 @@ static void kinit_cpu(void)
 void kinit(void)
 {
 	#ifdef FX9860G
-	/* VBR is loaded 0x600 bytes before end of the user RAM (0x100 bytes at
-	   the start of the VBR space are unused) */
-	gint_ctx.VBR = (uint32_t)mmu_uram() + 0x1a00 - 0x100;
+	/* VBR is loaded at the end of the user RAM. */
+	uint32_t uram_end = (uint32_t)mmu_uram() + mmu_uram_size();
+	/* On SH4, stack is at the end of the region, leave 8k */
+	if(isSH4()) uram_end -= 0x2000;
+	/* VBR size differs with models. On SH3, only 0x600 bytes are used due
+	   to the compact scheme. On SH4, 0x1100 bytes are needed to cover the
+	   expanded region. */
+	uram_end -= (isSH3() ? 0x600 : 0x1100);
+
+	/* There are 0x100 unused bytes at the start of the VBR area */
+	gint_ctx.VBR = uram_end - 0x100;
 	#endif
 
 	#ifdef FXCG50
@@ -191,6 +199,12 @@ void *gint_inthandler(int event_code, const void *handler, size_t size)
 	if(event_code < 0x400) return NULL;
 	event_code &= ~0x1f;
 
+	/* Prevent writing beyond the end of the VBR space on SH4. Using code
+	   0xfc0 into the interrupt handler space (which starts 0x540 bytes
+	   into VBR-reserved memory) would reach byte 0x540 + 0xfc0 - 0x400 =
+	   0x1100, which is out of gint's reserved VBR area.  */
+	if(isSH4() && event_code + size > 0xfc0) return NULL;
+
 	/* On SH3, make VBR compact. Use this offset specified in the VBR map
 	   above to avoid gaps */
 	if(isSH3())