Handwritten asm: exceptasm, get/setintmask (#1819)

* Handwritten asm: exceptasm, get/setintmask

* Fix maybe

* Update warnings

* Remove .global from jlabel

* Move os_internal_rsp inclusion to os.h

* Make HWINT_* defins consistent

17 files changed, 1952 insertions, 253 deletions

diff --git a/include/PR/os.h b/include/PR/os.h
index 226574e76..b9af86026 100644
--- a/include/PR/os.h
+++ b/include/PR/os.h
@@ -10,6 +10,7 @@
 #include "os_host.h"
 #include "os_internal_error.h"
 #include "os_internal_reg.h"
+#include "os_internal_rsp.h"
 #include "os_internal_si.h"
 #include "os_internal.h"
 #include "os_libc.h"
diff --git a/include/PR/os_exception.h b/include/PR/os_exception.h
index 7140c9663..5af2bd639 100644
--- a/include/PR/os_exception.h
+++ b/include/PR/os_exception.h
@@ -1,12 +1,6 @@
 #ifndef PR_OS_EXCEPTION_H
 #define PR_OS_EXCEPTION_H
 
-#include "ultratypes.h"
-
-
-typedef u32 OSIntMask;
-typedef u32 OSHWIntr;
-
 /* Flags for debugging purpose */
 
 #define OS_FLAG_CPU_BREAK   1   /* Break exception has occurred */
@@ -34,6 +28,15 @@ typedef u32 OSHWIntr;
 #define RCP_IMASK       0x003F0000
 #define RCP_IMASKSHIFT  16
 
+/* OSHWIntr values */
+#define OS_INTR_CART    1
+
+#ifdef _LANGUAGE_C
+
+#include "ultratypes.h"
+
+typedef u32 OSIntMask;
+typedef u32 OSHWIntr;
 
 OSIntMask osGetIntMask(void);
 OSIntMask osSetIntMask(OSIntMask im);
@@ -45,5 +48,6 @@ void __osGetHWIntrRoutine(OSHWIntr interrupt, s32 (**handler)(void), void** stac
 void __osSetGlobalIntMask(OSHWIntr mask);
 void __osResetGlobalIntMask(OSHWIntr mask);
 
+#endif
 
 #endif
diff --git a/include/PR/os_internal.h b/include/PR/os_internal.h
index 10ae5515e..d24be5505 100644
--- a/include/PR/os_internal.h
+++ b/include/PR/os_internal.h
@@ -1,14 +1,24 @@
 #ifndef PR_OS_INTERNAL_H
 #define PR_OS_INTERNAL_H
 
+#ifdef _LANGUAGE_C
+
 #include "ultratypes.h"
-#include "os_message.h"
-#include "os_pi.h"
-#include "os_internal_rsp.h"
 
 typedef struct __osHwInt {
     /* 0x0 */ s32 (*handler)(void);
     /* 0x4 */ void* stackEnd;
 } __osHwInt; // size = 0x8
 
+#else
+
+/* __osHwInt struct member offsets */
+#define HWINT_CALLBACK 0x0
+#define HWINT_SP       0x4
+
+/* __osHwInt struct size */
+#define HWINT_SIZE     0x8
+
+#endif
+
 #endif
diff --git a/include/PR/os_message.h b/include/PR/os_message.h
index bb577f57b..0a21272c1 100644
--- a/include/PR/os_message.h
+++ b/include/PR/os_message.h
@@ -1,20 +1,6 @@
 #ifndef PR_OS_MESSAGE_H
 #define PR_OS_MESSAGE_H
 
-#include "os_thread.h"
-
-typedef void* OSMesg;
-typedef u32 OSEvent;
-
-typedef struct OSMesgQueue {
-    /* 0x00 */ OSThread* mtQueue;
-    /* 0x04 */ OSThread* fullQueue;
-    /* 0x08 */ s32 validCount;
-    /* 0x0C */ s32 first;
-    /* 0x10 */ s32 msgCount;
-    /* 0x14 */ OSMesg* msg;
-} OSMesgQueue; // size = 0x18
-
 #define OS_NUM_EVENTS               15
 
 #define OS_EVENT_SW1                0     /* CPU SW1 interrupt */
@@ -45,6 +31,22 @@ typedef struct OSMesgQueue {
 #define OS_MESG_NOBLOCK  0
 #define OS_MESG_BLOCK    1
 
+#ifdef _LANGUAGE_C
+
+#include "os_thread.h"
+
+typedef void* OSMesg;
+typedef u32 OSEvent;
+
+typedef struct OSMesgQueue {
+    /* 0x00 */ OSThread* mtQueue;
+    /* 0x04 */ OSThread* fullQueue;
+    /* 0x08 */ s32 validCount;
+    /* 0x0C */ s32 first;
+    /* 0x10 */ s32 msgCount;
+    /* 0x14 */ OSMesg* msg;
+} OSMesgQueue; // size = 0x18
+
 #define MQ_GET_COUNT(mq)    ((mq)->validCount)
 
 #define MQ_IS_EMPTY(mq) (MQ_GET_COUNT(mq) == 0)
@@ -57,5 +59,17 @@ s32 osRecvMesg(OSMesgQueue* mq, OSMesg* msg, s32 flags);
 
 void osSetEventMesg(OSEvent e, OSMesgQueue* mq, OSMesg m);
 
+#else
+
+/* OSMesgQueue struct member offsets */
+
+#define MQ_MTQUEUE      0x00
+#define MQ_FULLQUEUE    0x04
+#define MQ_VALIDCOUNT   0x08
+#define MQ_FIRST        0x0C
+#define MQ_MSGCOUNT     0x10
+#define MQ_MSG          0x14
+
+#endif
 
 #endif
diff --git a/include/PR/os_thread.h b/include/PR/os_thread.h
index 4823377d0..95f01fb2b 100644
--- a/include/PR/os_thread.h
+++ b/include/PR/os_thread.h
@@ -1,11 +1,35 @@
 #ifndef PR_OS_THREAD_H
 #define PR_OS_THREAD_H
 
-#include "ultratypes.h"
+#define OS_STATE_STOPPED    (1 << 0)
+#define OS_STATE_RUNNABLE   (1 << 1)
+#define OS_STATE_RUNNING    (1 << 2)
+#define OS_STATE_WAITING    (1 << 3)
+
+#define OS_PRIORITY_IDLE          0
+#define OS_PRIORITY_MAIN         10
+#define OS_PRIORITY_GRAPH        11
+#define OS_PRIORITY_AUDIOMGR     12
+#define OS_PRIORITY_PADMGR       14
+#define OS_PRIORITY_SCHED        15
+#define OS_PRIORITY_DMAMGR       16
+#define OS_PRIORITY_IRQMGR       17
+#define OS_PRIORITY_PIMGR        150
+#define OS_PRIORITY_FAULTCLIENT  126
+#define OS_PRIORITY_FAULT        127
+#define OS_PRIORITY_APPMAX       127
+#define OS_PRIORITY_RMONSPIN     200
+#define OS_PRIORITY_RMON         250
+#define OS_PRIORITY_VIMGR        254
+#define OS_PRIORITY_MAX          255
 
 #define OS_FLAG_CPU_BREAK   1
 #define OS_FLAG_FAULT       2
 
+#ifdef _LANGUAGE_C
+
+#include "ultratypes.h"
+
 typedef s32 OSPri;
 typedef s32 OSId;
 
@@ -47,29 +71,6 @@ typedef struct OSThread {
     /* 0x20 */ __OSThreadContext context;
 } OSThread; // size = 0x1B0
 
-#define OS_STATE_STOPPED    (1 << 0)
-#define OS_STATE_RUNNABLE   (1 << 1)
-#define OS_STATE_RUNNING    (1 << 2)
-#define OS_STATE_WAITING    (1 << 3)
-
-
-#define OS_PRIORITY_IDLE          0
-#define OS_PRIORITY_MAIN         10
-#define OS_PRIORITY_GRAPH        11
-#define OS_PRIORITY_AUDIOMGR     12
-#define OS_PRIORITY_PADMGR       14
-#define OS_PRIORITY_SCHED        15
-#define OS_PRIORITY_DMAMGR       16
-#define OS_PRIORITY_IRQMGR       17
-#define OS_PRIORITY_PIMGR        150
-#define OS_PRIORITY_FAULTCLIENT  126
-#define OS_PRIORITY_FAULT        127
-#define OS_PRIORITY_APPMAX       127
-#define OS_PRIORITY_RMONSPIN     200
-#define OS_PRIORITY_RMON         250
-#define OS_PRIORITY_VIMGR        254
-#define OS_PRIORITY_MAX          255
-
 void osCreateThread(OSThread* thread, OSId id, void* entry, void* arg, void* sp, OSPri p);
 void osDestroyThread(OSThread* t);
 void osYieldThread(void);
@@ -82,4 +83,74 @@ OSPri osGetThreadPri(OSThread* t);
 // internal
 OSThread* __osGetActiveQueue(void);
 
+#else
+
+/* OSThread struct member offsets */
+
+#define THREAD_NEXT     0x00
+#define THREAD_PRI      0x04
+#define THREAD_QUEUE    0x08
+#define THREAD_TLNEXT   0x0C
+#define THREAD_STATE    0x10
+#define THREAD_FLAGS    0x12
+#define THREAD_ID       0x14
+#define THREAD_FP       0x18
+#define THREAD_PROFILE  0x1C
+#define THREAD_CONTEXT  0x20
+#define THREAD_AT       (THREAD_CONTEXT + 0x000)
+#define THREAD_V0       (THREAD_CONTEXT + 0x008)
+#define THREAD_V1       (THREAD_CONTEXT + 0x010)
+#define THREAD_A0       (THREAD_CONTEXT + 0x018)
+#define THREAD_A1       (THREAD_CONTEXT + 0x020)
+#define THREAD_A2       (THREAD_CONTEXT + 0x028)
+#define THREAD_A3       (THREAD_CONTEXT + 0x030)
+#define THREAD_T0       (THREAD_CONTEXT + 0x038)
+#define THREAD_T1       (THREAD_CONTEXT + 0x040)
+#define THREAD_T2       (THREAD_CONTEXT + 0x048)
+#define THREAD_T3       (THREAD_CONTEXT + 0x050)
+#define THREAD_T4       (THREAD_CONTEXT + 0x058)
+#define THREAD_T5       (THREAD_CONTEXT + 0x060)
+#define THREAD_T6       (THREAD_CONTEXT + 0x068)
+#define THREAD_T7       (THREAD_CONTEXT + 0x070)
+#define THREAD_S0       (THREAD_CONTEXT + 0x078)
+#define THREAD_S1       (THREAD_CONTEXT + 0x080)
+#define THREAD_S2       (THREAD_CONTEXT + 0x088)
+#define THREAD_S3       (THREAD_CONTEXT + 0x090)
+#define THREAD_S4       (THREAD_CONTEXT + 0x098)
+#define THREAD_S5       (THREAD_CONTEXT + 0x0A0)
+#define THREAD_S6       (THREAD_CONTEXT + 0x0A8)
+#define THREAD_S7       (THREAD_CONTEXT + 0x0B0)
+#define THREAD_T8       (THREAD_CONTEXT + 0x0B8)
+#define THREAD_T9       (THREAD_CONTEXT + 0x0C0)
+#define THREAD_GP       (THREAD_CONTEXT + 0x0C8)
+#define THREAD_SP       (THREAD_CONTEXT + 0x0D0)
+#define THREAD_S8       (THREAD_CONTEXT + 0x0D8)
+#define THREAD_RA       (THREAD_CONTEXT + 0x0E0)
+#define THREAD_LO       (THREAD_CONTEXT + 0x0E8)
+#define THREAD_HI       (THREAD_CONTEXT + 0x0F0)
+#define THREAD_SR       (THREAD_CONTEXT + 0x0F8)
+#define THREAD_PC       (THREAD_CONTEXT + 0x0FC)
+#define THREAD_CAUSE    (THREAD_CONTEXT + 0x100)
+#define THREAD_BADVADDR (THREAD_CONTEXT + 0x104)
+#define THREAD_RCP      (THREAD_CONTEXT + 0x108)
+#define THREAD_FPCSR    (THREAD_CONTEXT + 0x10C)
+#define THREAD_FP0      (THREAD_CONTEXT + 0x110)
+#define THREAD_FP2      (THREAD_CONTEXT + 0x118)
+#define THREAD_FP4      (THREAD_CONTEXT + 0x120)
+#define THREAD_FP6      (THREAD_CONTEXT + 0x128)
+#define THREAD_FP8      (THREAD_CONTEXT + 0x130)
+#define THREAD_FP10     (THREAD_CONTEXT + 0x138)
+#define THREAD_FP12     (THREAD_CONTEXT + 0x140)
+#define THREAD_FP14     (THREAD_CONTEXT + 0x148)
+#define THREAD_FP16     (THREAD_CONTEXT + 0x150)
+#define THREAD_FP18     (THREAD_CONTEXT + 0x158)
+#define THREAD_FP20     (THREAD_CONTEXT + 0x160)
+#define THREAD_FP22     (THREAD_CONTEXT + 0x168)
+#define THREAD_FP24     (THREAD_CONTEXT + 0x170)
+#define THREAD_FP26     (THREAD_CONTEXT + 0x178)
+#define THREAD_FP28     (THREAD_CONTEXT + 0x180)
+#define THREAD_FP30     (THREAD_CONTEXT + 0x188)
+
+#endif
+
 #endif
diff --git a/include/PR/rcp.h b/include/PR/rcp.h
index 1f2b57de5..5142de323 100644
--- a/include/PR/rcp.h
+++ b/include/PR/rcp.h
@@ -1,99 +1,108 @@
 #ifndef PR_RCP_H
 #define PR_RCP_H
 
+#include "R4300.h"
+#include "ultratypes.h"
+
 /**
- * PIF Physical memory map (total size = 2 KB)
+ * RCP memory map overview:
  *
- *            Size      Description     Mode
- *  1FC007FF    +-------+-----------------+-----+
- *              |  64 B | JoyChannel RAM  | R/W |
- *  1FC007C0    +-------+-----------------+-----+
- *              |1984 B |    Boot ROM     |  *  |  * = Reserved
- *  1FC00000    +-------+-----------------+-----+
+ *      0x0000_0000 .. 0x03EF_FFFF  RDRAM memory
+ *      0x03F0_0000 .. 0x03FF_FFFF  RDRAM registers
+ *
+ *      0x0400_0000 .. 0x0400_2000  SP memory
+ *      0x0404_0000 .. 0x040F_FFFF  SP registers
+ *      0x0410_0000 .. 0x041F_FFFF  DP command registers
+ *      0x0420_0000 .. 0x042F_FFFF  DP span registers
+ *      0x0430_0000 .. 0x043F_FFFF  MI registers
+ *      0x0440_0000 .. 0x044F_FFFF  VI registers
+ *      0x0450_0000 .. 0x045F_FFFF  AI registers
+ *      0x0460_0000 .. 0x046F_FFFF  PI registers
+ *      0x0470_0000 .. 0x047F_FFFF  RI registers
+ *      0x0480_0000 .. 0x048F_FFFF  SI registers
+ *      0x0490_0000 .. 0x04FF_FFFF  unused
+ *
+ *      0x0500_0000 .. 0x05FF_FFFF  cartridge domain 2
+ *      0x0600_0000 .. 0x07FF_FFFF  cartridge domain 1
+ *      0x0800_0000 .. 0x0FFF_FFFF  cartridge domain 2
+ *      0x1000_0000 .. 0x1FBF_FFFF  cartridge domain 1
+ *
+ *      0x1FC0_0000 .. 0x1FC0_07BF  PIF Boot Rom (1984 bytes)
+ *      0x1FC0_07C0 .. 0x1FC0_07FF  PIF (JoyChannel) RAM (64 bytes)
+ *      0x1FC0_0800 .. 0x1FCF_FFFF  Reserved
+ *      0x1FD0_0000 .. 0x7FFF_FFFF  cartridge domain 1
+ *      0x8000_0000 .. 0xFFFF_FFFF  external SysAD device
  */
-#define PIF_ROM_START   0x1FC00000
-#define PIF_ROM_END     0x1FC007BF
-#define PIF_RAM_START   0x1FC007C0
-#define PIF_RAM_END     0x1FC007FF
 
-/*
- * Patterns to interpret VI_CONTROL_REG
+
+/**
+ * RDRAM memory
  */
-#define VI_CTRL_TYPE_16             0x00002 /* [1:0] pixel size: 16 bit */
-#define VI_CTRL_TYPE_32             0x00003 /* [1:0] pixel size: 32 bit */
-#define VI_CTRL_GAMMA_DITHER_ON     0x00004 /* 2: default = on */
-#define VI_CTRL_GAMMA_ON            0x00008 /* 3: default = on */
-#define VI_CTRL_DIVOT_ON            0x00010 /* 4: default = on */
-#define VI_CTRL_SERRATE_ON          0x00040 /* 6: on if interlaced */
-#define VI_CTRL_ANTIALIAS_MASK      0x00300 /* [9:8] anti-alias mode */
-#define VI_CTRL_ANTIALIAS_MODE_1    0x00100 /* Bit [9:8] anti-alias mode */
-#define VI_CTRL_ANTIALIAS_MODE_2    0x00200 /* Bit [9:8] anti-alias mode */
-#define VI_CTRL_ANTIALIAS_MODE_3    0x00300 /* Bit [9:8] anti-alias mode */
-#define VI_CTRL_PIXEL_ADV_MASK      0x01000 /* [15:12] pixel advance mode? */
-#define VI_CTRL_PIXEL_ADV_1         0x01000 /* Bit [15:12] pixel advance mode? */
-#define VI_CTRL_PIXEL_ADV_2         0x02000 /* Bit [15:12] pixel advance mode? */
-#define VI_CTRL_PIXEL_ADV_3         0x03000 /* Bit [15:12] pixel advance mode? */
-#define VI_CTRL_DITHER_FILTER_ON    0x10000 /* 16: dither-filter mode */
 
+#define RDRAM_0_START   0x00000000
+#define RDRAM_0_END     0x001FFFFF
+#define RDRAM_1_START   0x00200000
+#define RDRAM_1_END     0x003FFFFF
+
+#define RDRAM_START     RDRAM_0_START
+#define RDRAM_END       RDRAM_1_END
 
-#define VI_NTSC_CLOCK       48681812        /* Hz = 48.681812 MHz */
-#define VI_PAL_CLOCK        49656530        /* Hz = 49.656530 MHz */
-#define VI_MPAL_CLOCK       48628316        /* Hz = 48.628316 MHz */
 
 /**
- * Audio Interface (AI) Registers
+ * RDRAM registers
  */
-#define AI_BASE_REG     0x04500000
+#define RDRAM_BASE_REG          0x03F00000
 
-/* AI DRAM address (W): [23:0] starting RDRAM address (8B-aligned) */
-#define AI_DRAM_ADDR_REG    (AI_BASE_REG + 0x00)    /* R0: DRAM address */
+#define RDRAM_CONFIG_REG        (RDRAM_BASE_REG + 0x00)
+#define RDRAM_DEVICE_TYPE_REG   (RDRAM_BASE_REG + 0x00)
+#define RDRAM_DEVICE_ID_REG     (RDRAM_BASE_REG + 0x04)
+#define RDRAM_DELAY_REG         (RDRAM_BASE_REG + 0x08)
+#define RDRAM_MODE_REG          (RDRAM_BASE_REG + 0x0C)
+#define RDRAM_REF_INTERVAL_REG  (RDRAM_BASE_REG + 0x10)
+#define RDRAM_REF_ROW_REG       (RDRAM_BASE_REG + 0x14)
+#define RDRAM_RAS_INTERVAL_REG  (RDRAM_BASE_REG + 0x18)
+#define RDRAM_MIN_INTERVAL_REG  (RDRAM_BASE_REG + 0x1C)
+#define RDRAM_ADDR_SELECT_REG   (RDRAM_BASE_REG + 0x20)
+#define RDRAM_DEVICE_MANUF_REG  (RDRAM_BASE_REG + 0x24)
 
-/* AI length (R/W): [14:0] transfer length (v1.0) - Bottom 3 bits are ignored */
-/*                  [17:0] transfer length (v2.0) - Bottom 3 bits are ignored */
-#define AI_LEN_REG          (AI_BASE_REG + 0x04)    /* R1: Length */
+#define RDRAM_0_DEVICE_ID   0
+#define RDRAM_1_DEVICE_ID   1
 
-/* AI control (W): [0] DMA enable - if LSB == 1, DMA is enabled */
-#define AI_CONTROL_REG      (AI_BASE_REG + 0x08)    /* R2: DMA Control */
+#define RDRAM_RESET_MODE    0
+#define RDRAM_ACTIVE_MODE   1
+#define RDRAM_STANDBY_MODE  2
 
-/* Value for control register */
-#define AI_CONTROL_DMA_ON   1   /* LSB = 1: DMA enable*/
-#define AI_CONTROL_DMA_OFF  0   /* LSB = 1: DMA enable*/
+#define RDRAM_LENGTH            (2 * 512 * 2048)
+#define RDRAM_0_BASE_ADDRESS    (RDRAM_0_DEVICE_ID * RDRAM_LENGTH)
+#define RDRAM_1_BASE_ADDRESS    (RDRAM_1_DEVICE_ID * RDRAM_LENGTH)
 
-/* 
- * AI status (R): [31]/[0] ai_full (addr & len buffer full), [30] ai_busy 
- *        Note that a 1->0 transition in ai_full will set interrupt
- *           (W): clear audio interrupt 
- */
-#define AI_STATUS_REG       (AI_BASE_REG + 0x0C)    /* R3: Status */
+#define RDRAM_0_CONFIG      0x00000
+#define RDRAM_1_CONFIG      0x00400
+#define RDRAM_GLOBAL_CONFIG 0x80000
 
-/* Value for status register */
-#define AI_STATUS_FIFO_FULL (1 << 31)
-#define AI_STATUS_DMA_BUSY  (1 << 30)
 
-/* 
- * AI DAC sample period register (W): [13:0] dac rate 
- *   - vid_clock/(dperiod + 1) is the DAC sample rate
- *   - (dperiod + 1) >= 66 * (aclockhp + 1) must be true
+/**
+ * PIF Physical memory map (total size = 2 KB)
+ *
+ *               Size    Description       Mode
+ *  1FC007FF    +-------+-----------------+-----+
+ *              |  64 B | JoyChannel RAM  | R/W |
+ *  1FC007C0    +-------+-----------------+-----+
+ *              |1984 B |    Boot ROM     |  *  |  * = Reserved
+ *  1FC00000    +-------+-----------------+-----+
  */
-#define AI_DACRATE_REG      (AI_BASE_REG + 0x10)    /* R4: DAC rate 14-lsb*/
+#define PIF_ROM_START   0x1FC00000
+#define PIF_ROM_END     0x1FC007BF
+#define PIF_RAM_START   0x1FC007C0
+#define PIF_RAM_END     0x1FC007FF
 
-/* DAC rate = video clock / audio frequency
- *   - DAC rate >= (66 * Bit rate) must be true
- */
-#define AI_MAX_DAC_RATE 16384   /* 14-bit+1 */
-#define AI_MIN_DAC_RATE 132
 
-/* 
- * AI bit rate (W): [3:0] bit rate (abus clock half period register - aclockhp)
- *   - vid_clock/(2 * (aclockhp + 1)) is the DAC clock rate
- *   - The abus clock stops if aclockhp is zero
+/**
+ * Controller channel
+ * Each game controller channel has 4 error bits that are defined in bit 6-7 of
+ * the Rx and Tx data size area bytes. Programmers need to clear these bits
+ * when setting the Tx/Rx size area values for a channel
  */
-#define AI_BITRATE_REG      (AI_BASE_REG + 0x14)    /* R5: Bit rate 4-lsb */
-
-/* Bit rate <= (DAC rate / 66) */
-#define AI_MAX_BIT_RATE 16      /* 4-bit+1 */
-#define AI_MIN_BIT_RATE 2
-
 #define CHNL_ERR_NORESP     0x80    /* Bit 7 (Rx): No response error */
 #define CHNL_ERR_OVERRUN    0x40    /* Bit 6 (Rx): Overrun error */
 #define CHNL_ERR_FRAME      0x80    /* Bit 7 (Tx): Frame error */
@@ -101,11 +110,18 @@
 
 #define CHNL_ERR_MASK       0xC0    /* Bit 6-7: channel errors */
 
+
+/**
+ * External device info
+ */
 #define DEVICE_TYPE_CART    0   /* ROM cartridge */
 #define DEVICE_TYPE_BULK    1   /* ROM bulk */
 #define DEVICE_TYPE_64DD    2   /* 64 Disk Drive */
 #define DEVICE_TYPE_SRAM    3   /* SRAM */
+/* 4-6 are reserved */
 #define DEVICE_TYPE_INIT    7   /* initial value */
+/* 8-14 are reserved */
+
 
 /**
  * Signal Processor (SP) Memory
@@ -115,43 +131,74 @@
 #define SP_IMEM_START   0x04001000
 #define SP_IMEM_END     0x04001FFF
 
-#define SP_MEM_ADDR_REG   0x04040000
-#define SP_DRAM_ADDR_REG  0x04040004
-#define SP_RD_LEN_REG     0x04040008
-#define SP_WR_LEN_REG     0x0404000C
-#define SP_STATUS_REG     0x04040010
-#define SP_DMA_FULL_REG   0x04040014
-#define SP_DMA_BUSY_REG   0x04040018
-#define SP_PC_REG         0x04080000
 
 /**
+ * Signal Processor (SP) CP0 Registers
+ */
+
+#define SP_BASE_REG         0x04040000
+
+/* SP memory address (R/W): [12] 0=DMEM,1=IMEM, [11:0] DMEM/IMEM address */
+#define SP_MEM_ADDR_REG     (SP_BASE_REG + 0x00)
+
+/* SP DRAM DMA address (R/W): [23:0] RDRAM address */
+#define SP_DRAM_ADDR_REG    (SP_BASE_REG + 0x04)
+
+/* SP read DMA length (R/W): [31:20] skip, [19:12] count, [11:0] length; RDRAM -> I/DMEM */
+#define SP_RD_LEN_REG       (SP_BASE_REG + 0x08)
+
+/* SP write DMA length (R/W): [31:20] skip, [19:12] count, [11:0] length; I/DMEM -> RDRAM */
+#define SP_WR_LEN_REG       (SP_BASE_REG + 0x0C)
+
+/* SP status (R/W): [14:0] valid bits; see below for write/read mode */
+#define SP_STATUS_REG       (SP_BASE_REG + 0x10)
+
+/* SP DMA full (R): [0] dma full */
+#define SP_DMA_FULL_REG     (SP_BASE_REG + 0x14)
+
+/* SP DMA busy (R): [0] dma busy */
+#define SP_DMA_BUSY_REG     (SP_BASE_REG + 0x18)
+
+/* SP semaphore (R/W): Read: [0] acquire semaphore; Write: [] release semaphore */
+#define SP_SEMAPHORE_REG    (SP_BASE_REG + 0x1C)
+
+/* SP PC (R/W): [11:0] program counter */
+#define SP_PC_REG       0x04080000
+
+/*
+ * SP_MEM_ADDR_REG: bit 12
+ */
+#define SP_DMA_DMEM     (0 << 12)
+#define SP_DMA_IMEM     (1 << 12)
+
+/*
  * SP_STATUS_REG: write bits
  */
-#define SP_CLR_HALT        (1 <<  0)
-#define SP_SET_HALT        (1 <<  1)
-#define SP_CLR_BROKE       (1 <<  2)
-#define SP_CLR_INTR        (1 <<  3)
-#define SP_SET_INTR        (1 <<  4)
-#define SP_CLR_SSTEP       (1 <<  5)
-#define SP_SET_SSTEP       (1 <<  6)
-#define SP_CLR_INTR_BREAK  (1 <<  7)
-#define SP_SET_INTR_BREAK  (1 <<  8)
-#define SP_CLR_SIG0        (1 <<  9)
-#define SP_SET_SIG0        (1 << 10)
-#define SP_CLR_SIG1        (1 << 11)
-#define SP_SET_SIG1        (1 << 12)
-#define SP_CLR_SIG2        (1 << 13)
-#define SP_SET_SIG2        (1 << 14)
-#define SP_CLR_SIG3        (1 << 15)
-#define SP_SET_SIG3        (1 << 16)
-#define SP_CLR_SIG4        (1 << 17)
-#define SP_SET_SIG4        (1 << 18)
-#define SP_CLR_SIG5        (1 << 19)
-#define SP_SET_SIG5        (1 << 20)
-#define SP_CLR_SIG6        (1 << 21)
-#define SP_SET_SIG6        (1 << 22)
-#define SP_CLR_SIG7        (1 << 23)
-#define SP_SET_SIG7        (1 << 24)
+#define SP_CLR_HALT         (1 <<  0)   /* clear halt */
+#define SP_SET_HALT         (1 <<  1)   /*   set halt */
+#define SP_CLR_BROKE        (1 <<  2)   /* clear broke */
+#define SP_CLR_INTR         (1 <<  3)   /* clear interrupt */
+#define SP_SET_INTR         (1 <<  4)   /*   set interrupt */
+#define SP_CLR_SSTEP        (1 <<  5)   /* clear sstep */
+#define SP_SET_SSTEP        (1 <<  6)   /*   set sstep */
+#define SP_CLR_INTR_BREAK   (1 <<  7)   /* clear interrupt on break */
+#define SP_SET_INTR_BREAK   (1 <<  8)   /*   set interrupt on break */
+#define SP_CLR_SIG0         (1 <<  9)   /* clear signal 0 */
+#define SP_SET_SIG0         (1 << 10)   /*   set signal 0 */
+#define SP_CLR_SIG1         (1 << 11)   /* clear signal 1 */
+#define SP_SET_SIG1         (1 << 12)   /*   set signal 1 */
+#define SP_CLR_SIG2         (1 << 13)   /* clear signal 2 */
+#define SP_SET_SIG2         (1 << 14)   /*   set signal 2 */
+#define SP_CLR_SIG3         (1 << 15)   /* clear signal 3 */
+#define SP_SET_SIG3         (1 << 16)   /*   set signal 3 */
+#define SP_CLR_SIG4         (1 << 17)   /* clear signal 4 */
+#define SP_SET_SIG4         (1 << 18)   /*   set signal 4 */
+#define SP_CLR_SIG5         (1 << 19)   /* clear signal 5 */
+#define SP_SET_SIG5         (1 << 20)   /*   set signal 5 */
+#define SP_CLR_SIG6         (1 << 21)   /* clear signal 6 */
+#define SP_SET_SIG6         (1 << 22)   /*   set signal 6 */
+#define SP_CLR_SIG7         (1 << 23)   /* clear signal 7 */
+#define SP_SET_SIG7         (1 << 24)   /*   set signal 7 */
 
 /*
  * SP_STATUS_REG: read bits
@@ -172,7 +219,7 @@
 #define SP_STATUS_SIG6          (1 << 13)
 #define SP_STATUS_SIG7          (1 << 14)
 
-/**
+/*
  * SP_STATUS_REG: use of SIG bits
  */
 #define SP_CLR_YIELD            SP_CLR_SIG0
@@ -191,49 +238,465 @@
 #define SP_SET_CPUSIGNAL        SP_SET_SIG4
 #define SP_STATUS_CPUSIGNAL     SP_STATUS_SIG4
 
-#define VI_STATUS_REG          0x04400000
-#define VI_CONTROL_REG         0x04400000
-#define VI_ORIGIN_REG          0x04400004
-#define VI_DRAM_ADDR_REG       0x04400004
-#define VI_WIDTH_REG           0x04400008
-#define VI_H_WIDTH_REG         0x04400008
-#define VI_INTR_REG            0x0440000C
-#define VI_V_INTER_REG         0x0440000C
-#define VI_CURRENT_REG         0x04400010
-#define VI_V_CURRENT_LINE_REG  0x04400010
-#define VI_BURST_REG           0x04400014
-#define VI_TIMING_REG          0x04400014
-#define VI_V_SYNC_REG          0x04400018 // VI vertical sync
-#define VI_H_SYNC_REG          0x0440001C // VI horizontal sync
-#define VI_LEAP_REG            0x04400020 // VI horizontal sync leap
-#define VI_H_SYNC_LEAP_REG     0x04400020
-#define VI_H_START_REG         0x04400024 // VI horizontal video
-#define VI_H_VIDEO_REG         0x04400024
-#define VI_V_START_REG         0x04400028 // VI vertical video
-#define VI_V_VIDEO_REG         0x04400028
-#define VI_V_BURST_REG         0x0440002C // VI vertical burst
-#define VI_X_SCALE_REG         0x04400030 // VI x-scale
-#define VI_Y_SCALE_REG         0x04400034 // VI y-scale
-
-#define PI_DRAM_ADDR_REG     0x04600000 // PI DRAM address
-#define PI_CART_ADDR_REG     0x04600004 // PI pbus (cartridge) address
-#define PI_RD_LEN_REG        0x04600008 // PI read length
-#define PI_WR_LEN_REG        0x0460000C // PI write length
-#define PI_STATUS_REG        0x04600010 // PI status
-#define PI_BSD_DOM1_LAT_REG  0x04600014 // PI dom1 latency
-#define PI_DOMAIN1_REG       0x04600014
-#define PI_BSD_DOM1_PWD_REG  0x04600018 // PI dom1 pulse width
-#define PI_BSD_DOM1_PGS_REG  0x0460001C // PI dom1 page size
-#define PI_BSD_DOM1_RLS_REG  0x04600020 // PI dom1 release
-#define PI_BSD_DOM2_LAT_REG  0x04600024 // PI dom2 latency
-#define PI_DOMAIN2_REG       0x04600024
-#define PI_BSD_DOM2_PWD_REG  0x04600028 // PI dom2 pulse width
-#define PI_BSD_DOM2_PGS_REG  0x0460002C // PI dom2 page size
-#define PI_BSD_DOM2_RLS_REG  0x04600030 // PI dom2 release
-
-#define PI_STATUS_DMA_BUSY (1 << 0)
-#define PI_STATUS_IO_BUSY  (1 << 1)
-#define PI_STATUS_ERROR    (1 << 2)
+/* SP IMEM BIST REG (R/W): [6:0] BIST status bits; see below for detail */
+#define SP_IBIST_REG        0x04080004
+
+/*
+ * SP_IBIST_REG: write bits
+ */
+#define SP_IBIST_CHECK      (1 << 0)    /* BIST check */
+#define SP_IBIST_GO         (1 << 1)    /* BIST go */
+#define SP_IBIST_CLEAR      (1 << 2)    /* BIST clear */
+
+/*
+ * SP_BIST_REG: read bits
+ * First 2 bits are same as in write mode
+ */
+#define SP_IBIST_DONE       (1 << 2)
+#define SP_IBIST_FAILED     0x78        /* bits [6:3], BIST fail */
+
+
+/**
+ * Display Processor Command (DPC) Registers
+ */
+#define DPC_BASE_REG        0x04100000
+
+/* DP CMD DMA start (R/W): [23:0] DMEM/RDRAM start address */
+#define DPC_START_REG       (DPC_BASE_REG + 0x00)
+
+/* DP CMD DMA end (R/W): [23:0] DMEM/RDRAM end address */
+#define DPC_END_REG         (DPC_BASE_REG + 0x04)
+
+/* DP CMD DMA end (R): [23:0] DMEM/RDRAM current address */
+#define DPC_CURRENT_REG     (DPC_BASE_REG + 0x08)
+
+/* DP CMD status (R/W): [9:0] valid bits - see below for definitions */
+#define DPC_STATUS_REG      (DPC_BASE_REG + 0x0C)
+
+/* DP clock counter (R): [23:0] clock counter */
+#define DPC_CLOCK_REG       (DPC_BASE_REG + 0x10)
+
+/* DP buffer busy counter (R): [23:0] clock counter */
+#define DPC_BUFBUSY_REG     (DPC_BASE_REG + 0x14)
+
+/* DP pipe busy counter (R): [23:0] clock counter */
+#define DPC_PIPEBUSY_REG    (DPC_BASE_REG + 0x18)
+
+/* DP TMEM load counter (R): [23:0] clock counter */
+#define DPC_TMEM_REG        (DPC_BASE_REG + 0x1C)
+
+/*
+ * DPC_STATUS_REG: write bits
+ */
+#define DPC_CLR_XBUS_DMEM_DMA       (1 << 0)
+#define DPC_SET_XBUS_DMEM_DMA       (1 << 1)
+#define DPC_CLR_FREEZE              (1 << 2)
+#define DPC_SET_FREEZE              (1 << 3)
+#define DPC_CLR_FLUSH               (1 << 4)
+#define DPC_SET_FLUSH               (1 << 5)
+#define DPC_CLR_TMEM_CTR            (1 << 6)
+#define DPC_CLR_PIPE_CTR            (1 << 7)
+#define DPC_CLR_CMD_CTR             (1 << 8)
+#define DPC_CLR_CLOCK_CTR           (1 << 9)
+
+/*
+ * DPC_STATUS_REG: read bits
+ */
+#define DPC_STATUS_XBUS_DMEM_DMA    (1 <<  0)
+#define DPC_STATUS_FREEZE           (1 <<  1)
+#define DPC_STATUS_FLUSH            (1 <<  2)
+#define DPC_STATUS_START_GCLK       (1 <<  3)
+#define DPC_STATUS_TMEM_BUSY        (1 <<  4)
+#define DPC_STATUS_PIPE_BUSY        (1 <<  5)
+#define DPC_STATUS_CMD_BUSY         (1 <<  6)
+#define DPC_STATUS_CBUF_READY       (1 <<  7)
+#define DPC_STATUS_DMA_BUSY         (1 <<  8)
+#define DPC_STATUS_END_VALID        (1 <<  9)
+#define DPC_STATUS_START_VALID      (1 << 10)
+
+
+/**
+ * Display Processor Span (DPS) Registers
+ */
+#define DPS_BASE_REG            0x04200000
+
+/* DP tmem built-in self-test (R/W): [10:0] BIST status bits */
+#define DPS_TBIST_REG           (DPS_BASE_REG + 0x00)
+
+/* DP span test mode (R/W): [0] Span buffer test access enable */
+#define DPS_TEST_MODE_REG       (DPS_BASE_REG + 0x04)
+
+/* DP span buffer test address (R/W): [6:0] bits */
+#define DPS_BUFTEST_ADDR_REG    (DPS_BASE_REG + 0x08)
+
+/* DP span buffer test data (R/W): [31:0] span buffer data */
+#define DPS_BUFTEST_DATA_REG    (DPS_BASE_REG + 0x0C)
+
+/*
+ * DPS_TMEM_BIST_REG: write bits
+ */
+#define DPS_TBIST_CHECK     (1 << 0)
+#define DPS_TBIST_GO        (1 << 1)
+#define DPS_TBIST_CLEAR     (1 << 2)
+
+/*
+ * DPS_TMEM_BIST_REG: read bits
+ * First 2 bits are same as in write mode
+ */
+#define DPS_TBIST_DONE      (1 << 2)
+#define DPS_TBIST_FAILED    0x7F8       /* bits [10:3], BIST fail */
+
+
+/**
+ * MIPS Interface (MI) Registers
+ */
+#define MI_BASE_REG     0x04300000
+
+/* MI init mode (W): [11] clear DP interrupt, [9/10] clear/set ebus test mode */
+/*                   [8] set init mode, [7] clear init mode, [6:0] init length */
+/*              (R): [8] ebus test mode, [7] init mode, [6:0] init length */
+#define MI_INIT_MODE_REG    (MI_BASE_REG + 0x00)
+#define MI_MODE_REG         MI_INIT_MODE_REG
+
+/*
+ *  MI_MODE_REG: write bits
+ */
+#define MI_CLR_INIT     (1 <<  7)   /* clear init mode */
+#define MI_SET_INIT     (1 <<  8)   /*   set init mode */
+#define MI_CLR_EBUS     (1 <<  9)   /* clear ebus test */
+#define MI_SET_EBUS     (1 << 10)   /*   set ebus test mode */
+#define MI_CLR_DP_INTR  (1 << 11)   /* clear dp interrupt */
+#define MI_CLR_RDRAM    (1 << 12)   /* clear RDRAM reg */
+#define MI_SET_RDRAM    (1 << 13)   /*   set RDRAM reg mode */
+
+/*
+ *  MI_MODE_REG: read bits
+ */
+#define MI_MODE_INIT    (1 << 7)    /* init mode */
+#define MI_MODE_EBUS    (1 << 8)    /* ebus test mode */
+#define MI_MODE_RDRAM   (1 << 9)    /* RDRAM reg mode */
+
+/* MI version (R): [31:24] rsp, [23:16] rdp, [15:8] rac, [7:0] io */
+#define MI_VERSION_REG  (MI_BASE_REG + 0x04)
+#define MI_NOOP_REG     MI_VERSION_REG
+
+/* MI interrupt (R): [5:0] valid bits - see below for bit patterns */
+#define MI_INTR_REG     (MI_BASE_REG + 0x08)
+
+/* MI interrupt mask (R): [5:0] valid bits - see below for bit patterns */
+/*                   (W): [11:0] valid bits - see below for bit patterns */
+#define MI_INTR_MASK_REG    (MI_BASE_REG + 0x0C)
+
+/*
+ * MI_INTR_REG: read bits
+ */
+#define MI_INTR_SP  (1 << 0)    /* SP intr */
+#define MI_INTR_SI  (1 << 1)    /* SI intr */
+#define MI_INTR_AI  (1 << 2)    /* AI intr */
+#define MI_INTR_VI  (1 << 3)    /* VI intr */
+#define MI_INTR_PI  (1 << 4)    /* PI intr */
+#define MI_INTR_DP  (1 << 5)    /* DP intr */
+
+/*
+ * MI_INTR_MASK_REG: write bits
+ */
+#define MI_INTR_MASK_CLR_SP (1 <<  0)   /* clear SP mask */
+#define MI_INTR_MASK_SET_SP (1 <<  1)   /*   set SP mask */
+#define MI_INTR_MASK_CLR_SI (1 <<  2)   /* clear SI mask */
+#define MI_INTR_MASK_SET_SI (1 <<  3)   /*   set SI mask */
+#define MI_INTR_MASK_CLR_AI (1 <<  4)   /* clear AI mask */
+#define MI_INTR_MASK_SET_AI (1 <<  5)   /*   set AI mask */
+#define MI_INTR_MASK_CLR_VI (1 <<  6)   /* clear VI mask */
+#define MI_INTR_MASK_SET_VI (1 <<  7)   /*   set VI mask */
+#define MI_INTR_MASK_CLR_PI (1 <<  8)   /* clear PI mask */
+#define MI_INTR_MASK_SET_PI (1 <<  9)   /*   set PI mask */
+#define MI_INTR_MASK_CLR_DP (1 << 10)   /* clear DP mask */
+#define MI_INTR_MASK_SET_DP (1 << 11)   /*   set DP mask */
+
+/*
+ * MI_INTR_MASK_REG: read bits
+ */
+#define MI_INTR_MASK_SP (1 << 0)    /* SP intr mask */
+#define MI_INTR_MASK_SI (1 << 1)    /* SI intr mask */
+#define MI_INTR_MASK_AI (1 << 2)    /* AI intr mask */
+#define MI_INTR_MASK_VI (1 << 3)    /* VI intr mask */
+#define MI_INTR_MASK_PI (1 << 4)    /* PI intr mask */
+#define MI_INTR_MASK_DP (1 << 5)    /* DP intr mask */
+
+
+/**
+ *  Video Interface (VI) Registers
+ */
+#define VI_BASE_REG     0x04400000
+
+/*
+ * VI status/control (R/W): [15-0] valid bits:
+ *  [1:0]   = type[1:0] (pixel size)
+ *              0: blank (no data, no sync)
+ *              1: reserved
+ *              2: 5/5/5/3 ("16" bit)
+ *              3: 8/8/8/8 (32 bit)
+ *  [2]     = gamma_dither_enable (normally on, unless "special effect")
+ *  [3]     = gamma_enable (normally on, unless MPEG/JPEG)
+ *  [4]     = divot_enable (normally on if antialiased, unless decal lines)
+ *  [5]     = vbus_clock_enable - always off
+ *  [6]     = serrate (always on if interlaced, off if not)
+ *  [7]     = test_mode - diagnostics only
+ *  [9:8]   = anti-alias (aa) mode[1:0]
+ *              0: aa & resamp (always fetch extra lines)
+ *              1: aa & resamp (fetch extra lines if needed)
+ *              2: resamp only (treat as all fully covered)
+ *              3: neither (replicate pixels, no interpolate)
+ *  [11]    = kill_we - diagnostics only
+ *  [15:12] = pixel_advance
+ *  [16]    = dither_filter_enable
+ */
+#define VI_CONTROL_REG      (VI_BASE_REG + 0x00)
+#define VI_STATUS_REG       VI_CONTROL_REG
+
+/* VI origin (R/W): [23:0] frame buffer origin in bytes */
+#define VI_ORIGIN_REG       (VI_BASE_REG + 0x04)
+#define VI_DRAM_ADDR_REG    VI_ORIGIN_REG
+
+/* VI width (R/W): [11:0] frame buffer line width in pixels */
+#define VI_WIDTH_REG        (VI_BASE_REG + 0x08)
+#define VI_H_WIDTH_REG      VI_WIDTH_REG
+
+/* VI vertical intr (R/W): [9:0] interrupt when current half-line = V_INTR */
+#define VI_INTR_REG         (VI_BASE_REG + 0x0C)
+#define VI_V_INTR_REG       VI_INTR_REG
+
+/* VI current vertical line (R/W): [9:0] current half line, sampled once per */
+/*  line (the lsb of V_CURRENT is constant within a field, and in interlaced */
+/*  modes gives the field number - which is constant for non-interlaced modes) */
+/*  - Any write to this register will clear interrupt line */
+#define VI_CURRENT_REG          (VI_BASE_REG + 0x10)
+#define VI_V_CURRENT_LINE_REG   VI_CURRENT_REG
+
+/* VI video timing (R/W): [29:20] start of color burst in pixels from h-sync */
+/*                        [19:16] vertical sync width in half lines, */
+/*                        [15: 8] color burst width in pixels, */
+/*                        [ 7: 0] horizontal sync width in pixels, */
+#define VI_BURST_REG        (VI_BASE_REG + 0x14)
+#define VI_TIMING_REG       VI_BURST_REG
+
+/* VI vertical sync (R/W): [9:0] number of half-lines per field */
+#define VI_V_SYNC_REG       (VI_BASE_REG + 0x18)
+
+/* VI horizontal sync (R/W): [20:16] a 5-bit leap pattern used for PAL only (h_sync_period) */
+/*                           [11: 0] total duration of a line in 1/4 pixel */
+#define VI_H_SYNC_REG       (VI_BASE_REG + 0x1C)
+
+/* VI horizontal sync leap (R/W): [27:16] identical to h_sync_period */
+/*                                [11: 0] identical to h_sync_period */
+#define VI_LEAP_REG         (VI_BASE_REG + 0x20)
+#define VI_H_SYNC_LEAP_REG  VI_LEAP_REG
+
+/* VI horizontal video (R/W): [25:16] start of active video in screen pixels */
+/*                            [ 9: 0] end of active video in screen pixels */
+#define VI_H_START_REG      (VI_BASE_REG + 0x24)
+#define VI_H_VIDEO_REG      VI_H_START_REG
+
+/* VI vertical video (R/W): [25:16] start of active video in screen half-lines */
+/*                          [ 9: 0] end of active video in screen half-lines */
+#define VI_V_START_REG      (VI_BASE_REG + 0x28)
+#define VI_V_VIDEO_REG      VI_V_START_REG
+
+/* VI vertical burst (R/W): [25:16] start of color burst enable in half-lines */
+/*                          [ 9: 0] end of color burst enable in half-lines */
+#define VI_V_BURST_REG      (VI_BASE_REG + 0x2C)
+
+/* VI x-scale (R/W): [27:16] horizontal subpixel offset (2.10 format) */
+/*                   [11: 0] 1/horizontal scale up factor (2.10 format) */
+#define VI_X_SCALE_REG      (VI_BASE_REG + 0x30)
+
+/* VI y-scale (R/W): [27:16] vertical subpixel offset (2.10 format) */
+/*                   [11: 0] 1/vertical scale up factor (2.10 format) */
+#define VI_Y_SCALE_REG      (VI_BASE_REG + 0x34)
+
+/*
+ * VI_CONTROL_REG: read bits
+ */
+#define VI_CTRL_TYPE_16             0x00002 /* [1:0] pixel size: 16 bit */
+#define VI_CTRL_TYPE_32             0x00003 /* [1:0] pixel size: 32 bit */
+#define VI_CTRL_GAMMA_DITHER_ON     0x00004 /* 2: default = on */
+#define VI_CTRL_GAMMA_ON            0x00008 /* 3: default = on */
+#define VI_CTRL_DIVOT_ON            0x00010 /* 4: default = on */
+#define VI_CTRL_SERRATE_ON          0x00040 /* 6: on if interlaced */
+#define VI_CTRL_ANTIALIAS_MASK      0x00300 /* [9:8] anti-alias mode */
+#define VI_CTRL_ANTIALIAS_MODE_0    0x00000 /* Bit [9:8] anti-alias mode: AA enabled, resampling enabled, always fetch extra lines */
+#define VI_CTRL_ANTIALIAS_MODE_1    0x00100 /* Bit [9:8] anti-alias mode: AA enabled, resampling enabled, fetch extra lines as-needed */
+#define VI_CTRL_ANTIALIAS_MODE_2    0x00200 /* Bit [9:8] anti-alias mode: AA disabled, resampling enabled, operate as if everything is covered */
+#define VI_CTRL_ANTIALIAS_MODE_3    0x00300 /* Bit [9:8] anti-alias mode: AA disabled, resampling disabled, replicate pixels */
+#define VI_CTRL_PIXEL_ADV_MASK      0x0F000 /* [15:12] pixel advance mode */
+#define VI_CTRL_PIXEL_ADV(n)        (((n) << 12) & VI_CTRL_PIXEL_ADV_MASK) /* Bit [15:12] pixel advance mode: Always 3 on N64 */
+#define VI_CTRL_DITHER_FILTER_ON    0x10000 /* 16: dither-filter mode */
+
+#define VI_CTRL_PIXEL_ADV_3         VI_CTRL_PIXEL_ADV(3)
+
+/*
+ * Possible video clocks (NTSC or PAL)
+ */
+#define VI_NTSC_CLOCK   48681812    /* Hz = 48.681812 MHz */
+#define VI_PAL_CLOCK    49656530    /* Hz = 49.656530 MHz */
+#define VI_MPAL_CLOCK   48628316    /* Hz = 48.628316 MHz */
+
+
+/**
+ * Audio Interface (AI) Registers
+ *
+ * The address and length registers are double buffered; that is, they
+ * can be written twice before becoming full.
+ * The address must be written before the length.
+ */
+#define AI_BASE_REG     0x04500000
+
+/* AI DRAM address (W): [23:0] starting RDRAM address (8B-aligned) */
+#define AI_DRAM_ADDR_REG    (AI_BASE_REG + 0x00)
+
+/* AI length (R/W): [14:0] transfer length (v1.0) - Bottom 3 bits are ignored */
+/*                  [17:0] transfer length (v2.0) - Bottom 3 bits are ignored */
+#define AI_LEN_REG          (AI_BASE_REG + 0x04)
+
+/* AI control (W): [0] DMA enable - if LSB == 1, DMA is enabled */
+#define AI_CONTROL_REG      (AI_BASE_REG + 0x08)
+
+/*
+ * AI_CONTROL_REG: write bits
+ */
+#define AI_CONTROL_DMA_ON   1   /* LSB = 1: DMA enable */
+#define AI_CONTROL_DMA_OFF  0   /* LSB = 1: DMA enable */
+
+/* AI status (R): [31]/[0] ai_full (addr & len buffer full), [30] ai_busy */
+/*                Note that a 1->0 transition in ai_full will set interrupt */
+/*           (W): clear audio interrupt */
+#define AI_STATUS_REG       (AI_BASE_REG + 0x0C)
+
+/*
+ * AI_STATUS_REG: read bits
+ */
+#define AI_STATUS_FIFO_FULL (1 << 31)
+#define AI_STATUS_DMA_BUSY  (1 << 30)
+
+/* AI DAC sample period register (W): [13:0] dac rate */
+/*   - vid_clock/(dperiod + 1) is the DAC sample rate */
+/*   - (dperiod + 1) >= 66 * (aclockhp + 1) must be true */
+#define AI_DACRATE_REG      (AI_BASE_REG + 0x10)
+
+/* DAC rate = video clock / audio frequency */
+/*   - DAC rate >= (66 * Bit rate) must be true */
+#define AI_MAX_DAC_RATE 16384   /* 14-bit+1 */
+#define AI_MIN_DAC_RATE 132
+
+/* AI bit rate (W): [3:0] bit rate (abus clock half period register - aclockhp) */
+/*   - vid_clock/(2 * (aclockhp + 1)) is the DAC clock rate */
+/*   - The abus clock stops if aclockhp is zero */
+#define AI_BITRATE_REG      (AI_BASE_REG + 0x14)
+
+/* Bit rate <= (DAC rate / 66) */
+#define AI_MAX_BIT_RATE 16      /* 4-bit+1 */
+#define AI_MIN_BIT_RATE 2
+
+/*
+ * Maximum and minimum values for audio frequency based on video clocks
+ *   max frequency = (video clock / min dac rate)
+ *   min frequency = (video clock / max dac rate)
+ */
+#define AI_NTSC_MAX_FREQ    368000      /* 368 KHz */
+#define AI_NTSC_MIN_FREQ    3000        /*   3 KHz ~ 2971 Hz */
+
+#define AI_PAL_MAX_FREQ     376000      /* 376 KHz */
+#define AI_PAL_MIN_FREQ     3050        /*   3 KHz ~ 3031 Hz */
+
+#define AI_MPAL_MAX_FREQ    368000      /* 368 KHz */
+#define AI_MPAL_MIN_FREQ    3000        /*   3 KHz ~ 2968 Hz */
+
+
+/**
+ * Peripheral Interface (PI) Registers
+ */
+#define PI_BASE_REG         0x04600000
+
+/* PI DRAM address (R/W): [23:0] starting RDRAM address */
+#define PI_DRAM_ADDR_REG    (PI_BASE_REG + 0x00)
+
+/* PI pbus (cartridge) address (R/W): [31:0] starting AD16 address */
+#define PI_CART_ADDR_REG    (PI_BASE_REG + 0x04)
+
+/* PI read length (R/W): [23:0] read data length */
+#define PI_RD_LEN_REG       (PI_BASE_REG + 0x08)
+
+/* PI write length (R/W): [23:0] write data length */
+#define PI_WR_LEN_REG       (PI_BASE_REG + 0x0C)
+
+/* PI status (R): [3] interrupt flag, [2] error, [1] IO busy, [0] DMA busy */
+/*           (W): [1] clear intr, [0] reset controller (and abort current op) */
+#define PI_STATUS_REG       (PI_BASE_REG + 0x10)
+
+/* PI dom1 latency (R/W): [7:0] domain 1 device latency */
+#define PI_BSD_DOM1_LAT_REG (PI_BASE_REG + 0x14)
+
+/* PI dom1 pulse width (R/W): [7:0] domain 1 device R/W strobe pulse width */
+#define PI_BSD_DOM1_PWD_REG (PI_BASE_REG + 0x18)
+
+/* PI dom1 page size (R/W): [3:0] domain 1 device page size */
+#define PI_BSD_DOM1_PGS_REG (PI_BASE_REG + 0x1C)
+
+/* PI dom1 release (R/W): [1:0] domain 1 device R/W release duration */
+#define PI_BSD_DOM1_RLS_REG (PI_BASE_REG + 0x20)
+
+/* PI dom2 latency (R/W): [7:0] domain 2 device latency */
+#define PI_BSD_DOM2_LAT_REG (PI_BASE_REG + 0x24)
+
+/* PI dom2 pulse width (R/W): [7:0] domain 2 device R/W strobe pulse width */
+#define PI_BSD_DOM2_PWD_REG (PI_BASE_REG + 0x28)
+
+/* PI dom2 page size (R/W): [3:0] domain 2 device page size */
+#define PI_BSD_DOM2_PGS_REG (PI_BASE_REG + 0x2C)
+
+/* PI dom2 release (R/W): [1:0] domain 2 device R/W release duration */
+#define PI_BSD_DOM2_RLS_REG (PI_BASE_REG + 0x30)
+
+#define PI_DOMAIN1_REG      PI_BSD_DOM1_LAT_REG
+#define PI_DOMAIN2_REG      PI_BSD_DOM2_LAT_REG
+
+#define PI_DOM_LAT_OFS      0x00
+#define PI_DOM_PWD_OFS      0x04
+#define PI_DOM_PGS_OFS      0x08
+#define PI_DOM_RLS_OFS      0x0C
+
+/*
+ * PI_STATUS_REG: read bits
+ *  Bit 0: DMA busy - set when DMA is in progress
+ *  Bit 1: IO busy  - set when IO is in progress
+ *  Bit 2: Error    - set when CPU issues IO request while DMA is busy
+ */
+#define PI_STATUS_DMA_BUSY  (1 << 0)
+#define PI_STATUS_IO_BUSY   (1 << 1)
+#define PI_STATUS_ERROR     (1 << 2)
+
+/*
+ * PI status register has 2 bits active when written to:
+ *  Bit 0: When set, reset PIC
+ *  Bit 1: When set, clear interrupt flag
+ * The values of the two bits can be ORed together to both reset PIC and
+ * clear interrupt at the same time.
+ *
+ * Note:
+ *  - The PIC does generate an interrupt at the end of each DMA. CPU
+ *  needs to clear the interrupt flag explicitly (from an interrupt
+ *  handler) by writing into the STATUS register with bit 1 set.
+ *
+ *  - When a DMA completes, the interrupt flag is set.  CPU can issue
+ *  another request even while the interrupt flag is set (as long as
+ *  PIC is idle).  However, it is the CPU's responsibility for
+ *  maintaining accurate correspondence between DMA completions and
+ *  interrupts.
+ *
+ *  - When PIC is reset, if PIC happens to be busy, an interrupt will
+ *  be generated as PIC returns to idle.  Otherwise, no interrupt will
+ *  be generated and PIC remains idle.
+ */
 
 /*
  * PI_STATUS_REG: write bits
@@ -252,22 +715,99 @@
 #define PI_DOM2_ADDR1   0x05000000  /* to 0x05FFFFFF */
 #define PI_DOM2_ADDR2   0x08000000  /* to 0x0FFFFFFF */
 
+
+/**
+ * RDRAM Interface (RI) Registers
+ */
+#define RI_BASE_REG         0x04700000
+
+/* RI mode (R/W): [3] stop R active, [2] stop T active, [1:0] operating mode */
+#define RI_MODE_REG         (RI_BASE_REG + 0x00)
+
+/* RI config (R/W): [6] current control enable, [5:0] current control input */
+#define RI_CONFIG_REG       (RI_BASE_REG + 0x04)
+
+/* RI current load (W): [] any write updates current control register */
+#define RI_CURRENT_LOAD_REG (RI_BASE_REG + 0x08)
+
+/* RI select (R/W): [3:2] receive select, [1:0] transmit select */
+#define RI_SELECT_REG       (RI_BASE_REG + 0x0C)
+
+/* RI refresh (R/W): [16] refresh bank, [17] refresh enable, [18] refresh optimize */
+/*                   [7:0] clean refresh delay, [15:8] dirty refresh dela */
+#define RI_REFRESH_REG      (RI_BASE_REG + 0x10)
+#define RI_COUNT_REG        RI_REFRESH_REG
+
+/* RI latency (R/W): [3:0] DMA latency/overlap */
+#define RI_LATENCY_REG      (RI_BASE_REG + 0x14)
+
+/* RI error (R): [1] ack error, [0] nack error */
+#define RI_RERROR_REG       (RI_BASE_REG + 0x18)
+
+/* RI error (W): [] any write clears all error bits */
+#define RI_WERROR_REG       (RI_BASE_REG + 0x1C)
+
+
 /**
  * Serial Interface (SI) Registers
  */
 #define SI_BASE_REG             0x04800000
 
+/* SI DRAM address (R/W): [23:0] starting RDRAM address */
 #define SI_DRAM_ADDR_REG        (SI_BASE_REG + 0x00)
+
+/* SI address read 64B (W): [] write begins a 64B DMA write PIF RAM -> RDRAM */
 #define SI_PIF_ADDR_RD64B_REG   (SI_BASE_REG + 0x04)
+
+/* Address SI_BASE_REG + (0x08, 0x0C, 0x14) are reserved */
+
+/* SI address write 64B (W): [] write begins a 64B DMA read RDRAM -> PIF RAM */
 #define SI_PIF_ADDR_WR64B_REG   (SI_BASE_REG + 0x10)
+
+/* SI status (R/W): [] any write clears interrupt */
 #define SI_STATUS_REG           (SI_BASE_REG + 0x18)
 
-#define SI_STATUS_DMA_BUSY      (1 << 0)
-#define SI_STATUS_IO_READ_BUSY  (1 << 1)
-#define SI_STATUS_DMA_ERROR     (1 << 3)
-#define SI_STATUS_INTERRUPT     (1 << 12)
+/*
+ * SI_STATUS_REG: read bits
+ */
+#define SI_STATUS_DMA_BUSY  (1 <<  0) /* DMA in progress */
+#define SI_STATUS_RD_BUSY   (1 <<  1) /* IO access in progress */
+#define SI_STATUS_DMA_ERROR (1 <<  3) /* Overlapping DMA requests */
+#define SI_STATUS_INTERRUPT (1 << 12) /* Interrupt is set */
+
+
+/**
+ * Development Board GIO Control Registers
+ */
+
+#define GIO_BASE_REG        0x18000000
+
+/* Game to Host Interrupt */
+#define GIO_GIO_INTR_REG    (GIO_BASE_REG+0x000)
+
+/* Game to Host SYNC */
+#define GIO_GIO_SYNC_REG    (GIO_BASE_REG+0x400)
+
+/* Host to Game Interrupt */
+#define GIO_CART_INTR_REG   (GIO_BASE_REG+0x800)
+
+
+/**
+ * Common macros
+ */
+#if defined(_LANGUAGE_C) || defined(_LANGUAGE_C_PLUS_PLUS)
 
 #define IO_READ(addr)       (*(vu32*)PHYS_TO_K1(addr))
 #define IO_WRITE(addr,data) (*(vu32*)PHYS_TO_K1(addr)=(u32)(data))
 
+#define RCP_STAT_PRINT                                              \
+    rmonPrintf("current=%x start=%x end=%x dpstat=%x spstat=%x\n",  \
+        IO_READ(DPC_CURRENT_REG),                                   \
+        IO_READ(DPC_START_REG),                                     \
+        IO_READ(DPC_END_REG),                                       \
+        IO_READ(DPC_STATUS_REG),                                    \
+        IO_READ(SP_STATUS_REG))
+
+#endif
+
 #endif
diff --git a/include/PR/rdp.h b/include/PR/rdp.h
index 4bd6bc99d..445359f68 100644
--- a/include/PR/rdp.h
+++ b/include/PR/rdp.h
@@ -1,50 +1,7 @@
 #ifndef PR_RDP_H
 #define PR_RDP_H
 
-/* DP Command Registers */
-#define DPC_START_REG    0x04100000
-#define DPC_END_REG      0x04100004
-#define DPC_CURRENT_REG  0x04100008
-#define DPC_STATUS_REG   0x0410000C
-#define DPC_CLOCK_REG    0x04100010
-#define DPC_BUFBUSY_REG  0x04100014
-#define DPC_PIPEBUSY_REG 0x04100018
-#define DPC_TMEM_REG     0x0410001C
-
-/* DP Span Registers */
-#define DPS_TBIST_REG        0x04200000
-#define DPS_TEST_MODE_REG    0x04200004
-#define DPS_BUFTEST_ADDR_REG 0x04200008
-#define DPS_BUFTEST_DATA_REG 0x0420000C
-
-/* DP Status Read Flags */
-#define DPC_STATUS_XBUS_DMEM_DMA    (1 <<  0)
-#define DPC_STATUS_FREEZE           (1 <<  1)
-#define DPC_STATUS_FLUSH            (1 <<  2)
-#define DPC_STATUS_START_GCLK       (1 <<  3)
-#define DPC_STATUS_TMEM_BUSY        (1 <<  4)
-#define DPC_STATUS_PIPE_BUSY        (1 <<  5)
-#define DPC_STATUS_CMD_BUSY         (1 <<  6)
-#define DPC_STATUS_CBUF_READY       (1 <<  7)
-#define DPC_STATUS_DMA_BUSY         (1 <<  8)
-#define DPC_STATUS_END_VALID        (1 <<  9)
-#define DPC_STATUS_START_VALID      (1 << 10)
-
-/* DP Status Write Flags */
-#define DPC_CLR_XBUS_DMEM_DMA   (1 << 0)
-#define DPC_SET_XBUS_DMEM_DMA   (1 << 1)
-#define DPC_CLR_FREEZE          (1 << 2)
-#define DPC_SET_FREEZE          (1 << 3)
-#define DPC_CLR_FLUSH           (1 << 4)
-#define DPC_SET_FLUSH           (1 << 5)
-#define DPC_CLR_TMEM_CTR        (1 << 6)
-#define DPC_CLR_PIPE_CTR        (1 << 7)
-#define DPC_CLR_CMD_CTR         (1 << 8)
-#define DPC_CLR_CLOCK_CTR       (1 << 9)
-
-
 u32 osDpGetStatus(void);
 void osDpSetStatus(u32 data);
 
-
 #endif
diff --git a/include/PR/ultratypes.h b/include/PR/ultratypes.h
index 0e48d50dc..76f5e74ea 100644
--- a/include/PR/ultratypes.h
+++ b/include/PR/ultratypes.h
@@ -1,6 +1,8 @@
 #ifndef PR_ULTRATYPES_H
 #define PR_ULTRATYPES_H
 
+#ifdef _LANGUAGE_C
+
 typedef signed char            s8;
 typedef unsigned char          u8;
 typedef signed short int       s16;
@@ -39,3 +41,5 @@ typedef unsigned int  size_t;
 typedef void* TexturePtr;
 
 #endif
+
+#endif
diff --git a/include/PR/version.h b/include/PR/version.h
new file mode 100644
index 000000000..5aa8c3c9c
--- /dev/null
+++ b/include/PR/version.h
@@ -0,0 +1,14 @@
+#ifndef PR_VERSION_H
+#define PR_VERSION_H
+
+#define LIBULTRA_VERSION_D  'D'
+#define LIBULTRA_VERSION_E  'E'
+#define LIBULTRA_VERSION_F  'F'
+#define LIBULTRA_VERSION_G  'G'
+#define LIBULTRA_VERSION_H  'H'
+#define LIBULTRA_VERSION_I  'I'
+#define LIBULTRA_VERSION_J  'J'
+#define LIBULTRA_VERSION_K  'K'
+#define LIBULTRA_VERSION_L  'L'
+
+#endif
diff --git a/include/macro.inc b/include/macro.inc
index 4840002b9..9700d05aa 100644
--- a/include/macro.inc
+++ b/include/macro.inc
@@ -21,7 +21,6 @@
 .endm
 
 .macro jlabel label
-    .global \label
     \label:
 .endm
 
diff --git a/spec/spec b/spec/spec
index bd0f82c0b..03a2911fd 100644
--- a/spec/spec
+++ b/spec/spec
@@ -55,9 +55,8 @@ beginseg
     include "$(BUILD_DIR)/src/libultra/io/viextendvstart.o"
     include "$(BUILD_DIR)/src/libultra/os/stopthread.o"
     include "$(BUILD_DIR)/src/libultra/os/recvmesg.o"
-    include "$(BUILD_DIR)/asm/boot/setintmask.text.o"
-    include "$(BUILD_DIR)/data/boot/setintmask.rodata.o"
-    include "$(BUILD_DIR)/asm/boot/getintmask.text.o"
+    include "$(BUILD_DIR)/src/libultra/os/setintmask.o"
+    include "$(BUILD_DIR)/src/libultra/os/getintmask.o"
     include "$(BUILD_DIR)/src/libultra/voice/voicesetword.o"
     include "$(BUILD_DIR)/src/libultra/vimodes/vimodentschpf1.o"
     include "$(BUILD_DIR)/src/libultra/vimodes/vimodepallan1.o"
@@ -66,9 +65,7 @@ beginseg
     include "$(BUILD_DIR)/src/libultra/gu/sins.o"
     include "$(BUILD_DIR)/src/libultra/io/sptask.o"
     include "$(BUILD_DIR)/src/libultra/libc/ll.o"
-    include "$(BUILD_DIR)/asm/boot/exceptasm.text.o"
-    include "$(BUILD_DIR)/data/boot/exceptasm.data.o"
-    include "$(BUILD_DIR)/data/boot/exceptasm.rodata.o"
+    include "$(BUILD_DIR)/src/libultra/os/exceptasm.o"
     include "$(BUILD_DIR)/src/libultra/os/thread.o"
     include "$(BUILD_DIR)/src/libultra/os/destroythread.o"
     include "$(BUILD_DIR)/src/libultra/voice/voicecheckresult.o"
diff --git a/src/libultra/io/si.c b/src/libultra/io/si.c
index 4f9dff154..688c5e700 100644
--- a/src/libultra/io/si.c
+++ b/src/libultra/io/si.c
@@ -4,7 +4,7 @@
 s32 __osSiDeviceBusy() {
     register u32 status = IO_READ(SI_STATUS_REG);
 
-    if (status & (SI_STATUS_DMA_BUSY | SI_STATUS_IO_READ_BUSY)) {
+    if (status & (SI_STATUS_DMA_BUSY | SI_STATUS_RD_BUSY)) {
         return true;
     } else {
         return false;
diff --git a/src/libultra/io/sirawdma.c b/src/libultra/io/sirawdma.c
index 4c5dc8cc9..eb6611c1c 100644
--- a/src/libultra/io/sirawdma.c
+++ b/src/libultra/io/sirawdma.c
@@ -2,7 +2,7 @@
 #include "alignment.h"
 
 s32 __osSiRawStartDma(s32 direction, void* dramAddr) {
-    if (IO_READ(SI_STATUS_REG) & (SI_STATUS_DMA_BUSY | SI_STATUS_IO_READ_BUSY)) {
+    if (IO_READ(SI_STATUS_REG) & (SI_STATUS_DMA_BUSY | SI_STATUS_RD_BUSY)) {
         return -1;
     }
 
diff --git a/src/libultra/os/exceptasm.s b/src/libultra/os/exceptasm.s
new file mode 100644
index 000000000..d6f4e1a03
--- /dev/null
+++ b/src/libultra/os/exceptasm.s
@@ -0,0 +1,879 @@
+#ifdef __GNUC__
+.set gp=64
+#endif
+#include "PR/asm.h"
+#include "PR/regdef.h"
+#include "PR/R4300.h"
+#include "PR/rcp.h"
+#include "PR/os_message.h"
+#include "PR/os_thread.h"
+#include "PR/os_exception.h"
+#include "PR/os_internal.h"
+#include "PR/version.h"
+
+#define MESG(x) ((x) << 3)
+
+.data
+.align 2
+
+DATA(__osHwIntTable)
+    .word 0, 0
+    .word 0, 0 /* cart */
+    .word 0, 0
+    .word 0, 0
+    .word 0, 0
+ENDDATA(__osHwIntTable)
+
+DATA(__osPiIntTable)
+    .word 0, 0
+ENDDATA(__osPiIntTable)
+
+.rdata
+.align 2
+
+__osIntOffTable:
+    .byte 0x00 /* redispatch */
+    .byte 0x14 /* prenmi */
+    .byte 0x18 /* IP6_Hdlr */
+    .byte 0x18 /* IP6_Hdlr */
+    .byte 0x1C /* IP7_Hdlr */
+    .byte 0x1C /* IP7_Hdlr */
+    .byte 0x1C /* IP7_Hdlr */
+    .byte 0x1C /* IP7_Hdlr */
+    .byte 0x20 /* counter */
+    .byte 0x20 /* counter */
+    .byte 0x20 /* counter */
+    .byte 0x20 /* counter */
+    .byte 0x20 /* counter */
+    .byte 0x20 /* counter */
+    .byte 0x20 /* counter */
+    .byte 0x20 /* counter */
+    .byte 0x00 /* redispatch */
+    .byte 0x04 /* sw1 */
+    .byte 0x08 /* sw2 */
+    .byte 0x08 /* sw2 */
+    .byte 0x0C /* rcp */
+    .byte 0x0C /* rcp */
+    .byte 0x0C /* rcp */
+    .byte 0x0C /* rcp */
+    .byte 0x10 /* cart */
+    .byte 0x10 /* cart */
+    .byte 0x10 /* cart */
+    .byte 0x10 /* cart */
+    .byte 0x10 /* cart */
+    .byte 0x10 /* cart */
+    .byte 0x10 /* cart */
+    .byte 0x10 /* cart */
+
+__osIntTable:
+    .word redispatch
+    .word sw1
+    .word sw2
+    .word rcp
+    .word cart
+    .word prenmi
+    .word IP6_Hdlr
+    .word IP7_Hdlr
+    .word counter
+
+.text
+
+/**
+ * The 16-byte exception preamble is copied to the exception vectors at
+ * UT_VEC, XUT_VEC, ECC_VEC, E_VEC, to direct execution to __osException
+ */
+LEAF(__osExceptionPreamble)
+    la      k0, __osException
+    jr      k0
+END(__osExceptionPreamble)
+
+LEAF(__osException)
+.set noat
+    /* Load scratch space for thread saving */
+    la      k0, __osThreadSave
+    /* Save at */
+    sd      AT, THREAD_AT(k0)
+.set at
+    /* Save sr */
+    MFC0(   k1, C0_SR)
+    sw      k1, THREAD_SR(k0)
+    /* Disable interrupts */
+    and     k1, k1, ~(SR_IE | SR_EXL)
+    MTC0(   k1, C0_SR)
+    /* Save some temp registers for use in the following */
+    sd      t0, THREAD_T0(k0)
+    sd      t1, THREAD_T1(k0)
+    sd      t2, THREAD_T2(k0)
+    /* Mark FPU as unused */
+    sw      zero, THREAD_FP(k0)
+    /* Left over from misplaced ifdef, immediately overwritten on next instruction */
+    MFC0(   t0, C0_CAUSE)
+savecontext:
+.set noreorder
+    /* Save the previously running thread's context to be restored when it resumes */
+    move    t0, k0
+    lw      k0, __osRunningThread
+    ld      t1, THREAD_AT(t0)
+    sd      t1, THREAD_AT(k0)
+    ld      t1, THREAD_SR(t0)
+    sd      t1, THREAD_SR(k0)
+    ld      t1, THREAD_T0(t0)
+    sd      t1, THREAD_T0(k0)
+    ld      t1, THREAD_T1(t0)
+    sd      t1, THREAD_T1(k0)
+    ld      t1, THREAD_T2(t0)
+    sd      t1, THREAD_T2(k0)
+.set reorder
+    sd      $2, THREAD_V0(k0)
+    sd      $3, THREAD_V1(k0)
+    sd      $4, THREAD_A0(k0)
+    sd      $5, THREAD_A1(k0)
+    sd      $6, THREAD_A2(k0)
+    sd      $7, THREAD_A3(k0)
+    sd      $11, THREAD_T3(k0)
+    sd      $12, THREAD_T4(k0)
+    sd      $13, THREAD_T5(k0)
+    sd      $14, THREAD_T6(k0)
+    sd      $15, THREAD_T7(k0)
+    sd      $16, THREAD_S0(k0)
+    sd      $17, THREAD_S1(k0)
+    sd      $18, THREAD_S2(k0)
+    sd      $19, THREAD_S3(k0)
+    sd      $20, THREAD_S4(k0)
+    sd      $21, THREAD_S5(k0)
+    sd      $22, THREAD_S6(k0)
+    sd      $23, THREAD_S7(k0)
+    sd      $24, THREAD_T8(k0)
+    sd      $25, THREAD_T9(k0)
+    sd      $28, THREAD_GP(k0)
+    sd      $29, THREAD_SP(k0)
+    sd      $30, THREAD_S8(k0)
+    sd      $31, THREAD_RA(k0)
+    mflo    t0
+    sd      t0, THREAD_LO(k0)
+    mfhi    t0
+    sd      t0, THREAD_HI(k0)
+    lw      k1, THREAD_SR(k0)
+    andi    t1, k1, SR_IMASK
+    beqz    t1, savercp
+    /* If any CPU interrupts are enabled in the previous thread's SR, bitwise-OR in the */
+    /* disabled CPU interrupts from the global interrupt mask. */
+    /* This is an attempt at reverting the effect of masking the thread's SR with the */
+    /* global interrupt mask. This is however broken, see comments for osSetIntMask. */
+    la      t0, __OSGlobalIntMask
+    lw      t0, (t0)
+    xor     t2, t0, 0xFFFFFFFF
+    andi    t2, t2, SR_IMASK
+    or      t4, t1, t2
+    and     t3, k1, ~SR_IMASK
+    or      t3, t3, t4
+    sw      t3, THREAD_SR(k0)
+    andi    t0, t0, SR_IMASK
+    and     t1, t1, t0
+    and     k1, k1, ~SR_IMASK
+    or      k1, k1, t1
+savercp:
+    /* Save the currently masked RCP interrupts. */
+    lw      t1, PHYS_TO_K1(MI_INTR_MASK_REG)
+    beqz    t1, endrcp
+    /* Similar to the above comment, but for RCP interrupt enable bits rather than CPU. */
+    /* This suffers from the same problem as above. */
+    la      t0, __OSGlobalIntMask
+    lw      t0, (t0)
+    srl     t0, t0, RCP_IMASKSHIFT
+    xor     t0, t0, 0xFFFFFFFF
+    andi    t0, t0, (RCP_IMASK >> RCP_IMASKSHIFT)
+    lw      t4, THREAD_RCP(k0)
+    and     t0, t0, t4
+    or      t1, t1, t0
+endrcp:
+    sw      t1, THREAD_RCP(k0)
+    MFC0(   t0, C0_EPC)
+    sw      t0, THREAD_PC(k0)
+    lw      t0, THREAD_FP(k0)
+    beqz    t0, handle_interrupt
+    /* Save FP Registers if FPU was used by the thread */
+.set noreorder 
+    cfc1    t0, C1_FPCSR
+    nop
+    sw      t0, THREAD_FPCSR(k0)
+.set reorder
+    sdc1    $f0, THREAD_FP0(k0)
+    sdc1    $f2, THREAD_FP2(k0)
+    sdc1    $f4, THREAD_FP4(k0)
+    sdc1    $f6, THREAD_FP6(k0)
+    sdc1    $f8, THREAD_FP8(k0)
+    sdc1    $f10, THREAD_FP10(k0)
+    sdc1    $f12, THREAD_FP12(k0)
+    sdc1    $f14, THREAD_FP14(k0)
+    sdc1    $f16, THREAD_FP16(k0)
+    sdc1    $f18, THREAD_FP18(k0)
+    sdc1    $f20, THREAD_FP20(k0)
+    sdc1    $f22, THREAD_FP22(k0)
+    sdc1    $f24, THREAD_FP24(k0)
+    sdc1    $f26, THREAD_FP26(k0)
+    sdc1    $f28, THREAD_FP28(k0)
+    sdc1    $f30, THREAD_FP30(k0)
+
+handle_interrupt:
+    /* Determine the cause of the exception or interrupt and */
+    /* enter appropriate handling routine */
+    MFC0(   t0, C0_CAUSE)
+    sw      t0, THREAD_CAUSE(k0)
+label:
+    li      t1, OS_STATE_RUNNABLE
+    sh      t1, THREAD_STATE(k0)
+    andi    t1, t0, CAUSE_EXCMASK
+    /* Test for break exception */
+    li      t2, EXC_BREAK
+    beq     t1, t2, handle_break
+    /* Test for CpU (coprocessor unusable) exception */
+    li      t2, EXC_CPU
+    beq     t1, t2, handle_CpU
+    /* Test for interrupt, if it's not an interrupt, panic */
+    li      t2, EXC_INT
+    bne     t1, t2, panic
+
+    and     s0, k1, t0
+
+next_interrupt:
+    /* Handle external interrupt causes, using a jump table */
+    /* to enter into the appropriate handler */
+    andi    t1, s0, CAUSE_IPMASK
+    srl     t2, t1, CAUSE_IPSHIFT + 4
+    bnez    t2, 1f
+
+    srl     t2, t1, CAUSE_IPSHIFT
+    addi    t2, t2, 0x10
+1:
+    lbu     t2, __osIntOffTable(t2)
+    lw      t2, __osIntTable(t2)
+    jr      t2
+
+/**
+ *  IP6 Interrupt
+ *  Only signalled by development hardware
+ */
+IP6_Hdlr:
+    /* Mask out interrupt and continue */
+    and     s0, s0, ~CAUSE_IP6
+    b       next_interrupt
+
+/**
+ *  IP7 Interrupt
+ *  Only signalled by development hardware
+ */
+IP7_Hdlr:
+    /* Mask out interrupt and continue */
+    and     s0, s0, ~CAUSE_IP7
+    b       next_interrupt
+
+/**
+ *  IP8/Counter Interrupt
+ *  Once the cop0 count register reaches the value of the
+ *  cop0 compare register, this interrupt is triggered
+ */
+counter:
+    MFC0(   t1, C0_COMPARE)
+    MTC0(   t1, C0_COMPARE)
+    /* Post counter message */
+    li      a0, MESG(OS_EVENT_COUNTER)
+    jal     send_mesg
+    /* Mask out interrupt and continue */
+    and     s0, s0, ~CAUSE_IP8
+    b       next_interrupt
+
+/**
+ *  IP4/Cartridge Interrupt
+ *  Signalled by the N64 Disk Drive
+ */
+cart:
+    /* Mask out interrupt */
+    and     s0, s0, ~CAUSE_IP4
+    /* Load cart callback set by __osSetHWIntrRoutine */
+    la      t1, __osHwIntTable
+    addi    t1, t1, (OS_INTR_CART * HWINT_SIZE)
+    lw      t2, HWINT_CALLBACK(t1)
+    /* If the callback is NULL, handling is done */
+    beqz    t2, 1f
+    /* Set up a stack and run the callback */
+    lw      sp, HWINT_SP(t1)
+    jalr    t2
+    beqz    v0, 1f
+    /* Redispatch immediately if the callback returned nonzero */
+    b       redispatch
+1:
+    /* Post a cart event message */
+    li      a0, MESG(OS_EVENT_CART)
+    jal     send_mesg
+    /* Continue */
+    b       next_interrupt
+
+/**
+ *  IP3/RCP Interrupt
+ *  Signalled by the RCP for various reasons, described below
+ */
+rcp:
+    /* Load the MI interrupts and mask with the RCP bits in the global interrupt mask */
+    /*! @bug this clobbers the t0 register which is expected to hold the value of the */
+    /*! C0_CAUSE register in the sw1 and sw2 handlers. If the sw1 or sw2 handler runs */
+    /*! after this, the interrupt will not be cleared properly. */
+    lw      s1, PHYS_TO_K1(MI_INTR_REG)
+    la      t0, __OSGlobalIntMask
+    lw      t0, (t0)
+    srl     t0, t0, RCP_IMASKSHIFT
+    and     s1, s1, t0
+
+/**
+ *  Signal Processor (SP) Interrupt
+ */
+    /* Test for sp interrupt */
+    andi    t1, s1, MI_INTR_SP
+    beqz    t1, vi
+    /* Mask out SP interrupt */
+    andi    s1, s1, (MI_INTR_SI | MI_INTR_AI | MI_INTR_VI | MI_INTR_PI | MI_INTR_DP)
+    lw      t4, PHYS_TO_K1(SP_STATUS_REG)
+    /* Clear interrupt and signal 3 */
+    li      t1, (SP_CLR_INTR | SP_CLR_SIG3)
+    sw      t1, PHYS_TO_K1(SP_STATUS_REG)
+    /* Test for yielded or done signals in particular */
+    andi    t4, t4, (SP_STATUS_YIELDED | SP_STATUS_TASKDONE)
+    beqz    t4, sp_other_break
+    /* Post an SP event message */
+    li      a0, MESG(OS_EVENT_SP)
+    jal     send_mesg
+    beqz    s1, NoMoreRcpInts
+    /* Step over sp_other_break handler */
+    b       vi
+
+sp_other_break:
+    /* An sp signal that is not due to yielding or task completion, such as */
+    /* an sp breakpoint. Post a different event message */
+    li      a0, MESG(OS_EVENT_SP_BREAK)
+    jal     send_mesg
+    beqz    s1, NoMoreRcpInts
+
+/**
+ *  Video Interface (VI) Interrupt
+ */
+vi:
+    /* Test for vi interrupt */
+    andi    t1, s1, MI_INTR_VI
+    beqz    t1, ai
+    /* Mask out vi interrupt */
+    andi    s1, s1, (MI_INTR_SP | MI_INTR_SI | MI_INTR_AI | MI_INTR_PI | MI_INTR_DP)
+    /* Clear interrupt */
+    sw      zero, PHYS_TO_K1(VI_CURRENT_REG)
+    /* Post vi event message */
+    li      a0, MESG(OS_EVENT_VI)
+    jal     send_mesg
+    beqz    s1, NoMoreRcpInts
+
+/**
+ *  Audio Interface (AI) Interrupt
+ */
+ai:
+    /* Test for ai interrupt */
+    andi    t1, s1, MI_INTR_AI
+    beqz    t1, si
+
+    /* Mask out ai interrupt */
+    andi    s1, s1, (MI_INTR_SP | MI_INTR_SI | MI_INTR_VI | MI_INTR_PI | MI_INTR_DP)
+    /* Clear interrupt */
+    li      t1, 1
+    sw      t1, PHYS_TO_K1(AI_STATUS_REG)
+    /* Post ai event message */
+    li      a0, MESG(OS_EVENT_AI)
+    jal     send_mesg
+    beqz    s1, NoMoreRcpInts
+
+/**
+ *  Serial Interface (SI) Interrupt
+ */
+si:
+    /* Test for si interrupt */
+    andi    t1, s1, MI_INTR_SI
+    beqz    t1, pi
+
+    /* Mask out si interrupt */
+    andi    s1, s1, (MI_INTR_SP | MI_INTR_AI | MI_INTR_VI | MI_INTR_PI | MI_INTR_DP)
+    /* Clear interrupt */
+    sw      zero, PHYS_TO_K1(SI_STATUS_REG)
+    /* Post si event message */
+    li      a0, MESG(OS_EVENT_SI)
+    jal     send_mesg
+    beqz    s1, NoMoreRcpInts
+
+/**
+ *  Parallel Interface (PI) Interrupt
+ */
+pi:
+    /* Test for pi interrupt */
+    andi    t1, s1, MI_INTR_PI
+    beqz    t1, dp
+
+    /* Mask out pi interrupt */
+    andi    s1, s1, (MI_INTR_SP | MI_INTR_SI | MI_INTR_AI | MI_INTR_VI | MI_INTR_DP)
+    /* Clear the interrupt */
+    li      t1, PI_STATUS_CLR_INTR
+    sw      t1, PHYS_TO_K1(PI_STATUS_REG)
+    /* Load pi callback */
+    la      t1, __osPiIntTable
+    lw      t2, HWINT_CALLBACK(t1)
+    /* Skip callback if NULL */
+    beqz    t2, no_pi_callback
+    /* Set up a stack and run the callback */
+    lw      sp, HWINT_SP(t1)
+    move    a0, v0
+    jalr    t2
+    /* If the callback returns non-zero, don't post a pi event message */
+    bnez    v0, skip_pi_mesg
+no_pi_callback:
+    /* Post pi event message */
+    li      a0, MESG(OS_EVENT_PI)
+    jal     send_mesg
+skip_pi_mesg:
+    beqz    s1, NoMoreRcpInts
+
+/**
+ *  Display Processor (DP) Interrupt
+ */
+dp:
+    /* Test for dp interrupt */
+    andi    t1, s1, MI_INTR_DP
+    beqz    t1, NoMoreRcpInts
+
+    /* Mask out dp interrupt */
+    andi    s1, s1, (MI_INTR_SP | MI_INTR_SI | MI_INTR_AI | MI_INTR_VI | MI_INTR_PI)
+    /* Clear dp interrupt */
+    li      t1, MI_CLR_DP_INTR
+    sw      t1, PHYS_TO_K1(MI_INIT_MODE_REG)
+    /* Post dp event message */
+    li      a0, MESG(OS_EVENT_DP)
+    jal     send_mesg
+
+NoMoreRcpInts:
+    /* Mask out interrupt and continue */
+    and     s0, s0, ~CAUSE_IP3
+    b       next_interrupt
+
+/**
+ *  IP5/PreNMI Interrupt
+ *  Reset button has been pressed
+ */
+prenmi:
+    /* Disable IP5/PreNMI interrupt for the previously running thread */
+    lw      k1, THREAD_SR(k0)
+    and     k1, k1, ~SR_IBIT5
+    sw      k1, THREAD_SR(k0)
+    /* Test __osShutdown for first PreNMI event */
+    la      t1, __osShutdown
+    lw      t2, (t1)
+    beqz    t2, firstnmi
+    /* Mask out interrupt and redispatch immediately */
+    and     s0, s0, ~CAUSE_IP5
+    b       redispatch
+
+firstnmi:
+    /* Set __osShutdown */
+    li      t2, 1
+    sw      t2, (t1)
+    /* Post a PreNMI event message */
+    li      a0, MESG(OS_EVENT_PRENMI)
+    jal     send_mesg
+    /* Mask out and disable IP5/PreNMI interrupt for the highest priority thread */
+    and     s0, s0, ~SR_IBIT5
+    lw      t2, __osRunQueue
+    lw      k1, THREAD_SR(t2)
+    and     k1, k1, ~SR_IBIT5
+    sw      k1, THREAD_SR(t2)
+    /* Redispatch immediately */
+    b       redispatch
+
+sw2:
+    /* Mask out interrupt */
+    and     t0, t0, ~CAUSE_SW2
+    MTC0(   t0, C0_CAUSE)
+    /* Post sw2 event message */
+    li      a0, MESG(OS_EVENT_SW2)
+    jal     send_mesg
+    /* Mask out interrupt and continue */
+    and     s0, s0, ~CAUSE_SW2
+    b       next_interrupt
+
+sw1:
+    /* Mask out interrupt */
+    and     t0, t0, ~CAUSE_SW1
+    MTC0(   t0, C0_CAUSE)
+    /* Post sw1 event message */
+    li      a0, MESG(OS_EVENT_SW1)
+    jal     send_mesg
+    /* Mask out interrupt and continue */
+    and     s0, s0, ~CAUSE_SW1
+    b       next_interrupt
+
+handle_break:
+    /* Set last thread as having hit a break exception */
+    li      t1, OS_FLAG_CPU_BREAK
+    sh      t1, THREAD_FLAGS(k0)
+    /* Post a cpu break event message */
+    li      a0, MESG(OS_EVENT_CPU_BREAK)
+    jal     send_mesg
+    /* Redispatch */
+    b       redispatch
+
+redispatch:
+    /* Get priority of previously running thread */
+    lw      t1, THREAD_PRI(k0)
+    lw      t2, __osRunQueue
+    /* Get highest priority from waiting threads */
+    lw      t3, THREAD_PRI(t2)
+    bge     t1, t3, enqueueRunning
+    /* The previously running thread is no longer the highest priority, */
+    /*  enqueue it to the run queue to wait its turn again */
+    move    a1, k0
+    la      a0, __osRunQueue
+    jal     __osEnqueueThread
+
+    j       __osDispatchThread
+
+/**
+ *  Resume the previously running thread by placing it at the top of
+ *  the run queue and dispatching it
+ */
+enqueueRunning:
+    la      t1, __osRunQueue
+    lw      t2, (t1)
+    sw      t2, THREAD_NEXT(k0)
+    sw      k0, (t1)
+    j       __osDispatchThread
+
+/**
+ *  Unhandled exceptions & interrupts end up here,
+ *  trap to software by posting a fault message
+ */
+panic:
+    /* Mark the thread as having faulted */
+    sw      k0, __osFaultedThread
+    li      t1, OS_STATE_STOPPED
+    sh      t1, THREAD_STATE(k0)
+    li      t1, OS_FLAG_FAULT
+    sh      t1, THREAD_FLAGS(k0)
+    /* Save C0_BADVADDR */
+    MFC0(   t2, C0_BADVADDR)
+    sw      t2, THREAD_BADVADDR(k0)
+    /* Post the fault message */
+    li      a0, MESG(OS_EVENT_FAULT)
+    jal     send_mesg
+    /* Dispatch next thread */
+    j       __osDispatchThread
+
+/**
+ *  Handles posting event messages to the listening message queue, if there is one
+ */
+send_mesg:
+    /* Save return address */
+    move    s2, ra
+    /* Load pointer to listening message queue */
+    la      t2, __osEventStateTab
+    addu    t2, t2, a0
+    lw      t1, (t2)
+    /* If there is no listening message queue, done */
+    beqz    t1, send_done
+
+    /* Test if the message queue is full, if so don't post the message */
+    lw      t3, MQ_VALIDCOUNT(t1)
+    lw      t4, MQ_MSGCOUNT(t1)
+    bge     t3, t4, send_done
+
+    /* Add validcount to first and modulo with msgcount */
+    lw      t5, MQ_FIRST(t1)
+    addu    t5, t5, t3
+    rem     t5, t5, t4
+    lw      t4, MQ_MSG(t1)
+    mul     t5, t5, 4
+    addu    t4, t4, t5
+    /* Post the message to the message queue */
+    lw      t5, 4(t2)
+    sw      t5, (t4)
+    /* Increment the validCount */
+    addiu   t2, t3, 1
+    sw      t2, MQ_VALIDCOUNT(t1)
+    /* If there was a thread blocked on this message queue, */
+    /*  wake it up */
+    lw      t2, MQ_MTQUEUE(t1)
+    lw      t3, (t2)
+    beqz    t3, send_done
+    move    a0, t1
+    jal     __osPopThread
+    move    t2, v0
+    move    a1, t2
+    la      a0, __osRunQueue
+    jal     __osEnqueueThread
+send_done:
+    jr      s2
+
+/**
+ *  Handle coprocessor unusable exception
+ */
+handle_CpU:
+    and     t1, t0, CAUSE_CEMASK
+    srl     t1, t1, CAUSE_CESHIFT
+    li      t2, 1 /* if not coprocessor 1, panic */
+    bne     t1, t2, panic
+    /* Mark cop1 as usable for previous thread */
+    li      t1, 1
+    sw      t1, THREAD_FP(k0)
+    lw      k1, THREAD_SR(k0)
+    or      k1, k1, SR_CU1
+    sw      k1, THREAD_SR(k0)
+    b       enqueueRunning
+END(__osException)
+
+/**
+ *  void __osEnqueueAndYield(OSThread** threadQueue);
+ *
+ *  Voluntary thread yielding.
+ *  Enqueues the currently running thread to the top of the
+ *   thread queue `threadQueue` and yields to the highest priority
+ *   unblocked runnable thread.
+ */
+LEAF(__osEnqueueAndYield)
+    lw      a1, __osRunningThread
+    /* Save SR */
+    MFC0(   t0, C0_SR)
+    ori     t0, t0, SR_EXL
+    sw      t0, THREAD_SR(a1)
+    /* Save callee-saved registers */
+    sd      s0, THREAD_S0(a1)
+    sd      s1, THREAD_S1(a1)
+    sd      s2, THREAD_S2(a1)
+    sd      s3, THREAD_S3(a1)
+    sd      s4, THREAD_S4(a1)
+    sd      s5, THREAD_S5(a1)
+    sd      s6, THREAD_S6(a1)
+    sd      s7, THREAD_S7(a1)
+    sd      gp, THREAD_GP(a1)
+    sd      sp, THREAD_SP(a1)
+    sd      fp, THREAD_S8(a1)
+    sd      ra, THREAD_RA(a1)
+    sw      ra, THREAD_PC(a1)
+    /* Save FPU callee-saved registers if the current thread has used the FPU */
+    lw      k1, THREAD_FP(a1)
+    beqz    k1, 1f
+    cfc1    k1, C1_FPCSR
+    sw      k1, THREAD_FPCSR(a1)
+    sdc1    $f20, THREAD_FP20(a1)
+    sdc1    $f22, THREAD_FP22(a1)
+    sdc1    $f24, THREAD_FP24(a1)
+    sdc1    $f26, THREAD_FP26(a1)
+    sdc1    $f28, THREAD_FP28(a1)
+    sdc1    $f30, THREAD_FP30(a1)
+1:
+    lw      k1, THREAD_SR(a1)
+    andi    t1, k1, SR_IMASK
+    beqz    t1, 2f
+    /* This code does the same thing as the block just above the `savercp` label. */
+    /* See the comment there for more about this. */
+    la      t0, __OSGlobalIntMask
+    lw      t0, (t0)
+    xor     t0, t0, 0xFFFFFFFF
+    andi    t0, t0, SR_IMASK
+    or      t1, t1, t0
+    and     k1, k1, ~SR_IMASK
+    or      k1, k1, t1
+    sw      k1, THREAD_SR(a1)
+2:
+    lw      k1, PHYS_TO_K1(MI_INTR_MASK_REG)
+    beqz    k1, 3f
+    /* This code does the same thing as the block just below the `savercp` label. */
+    /* See the comment there for more about this. */
+    la      k0, __OSGlobalIntMask
+    lw      k0, (k0)
+    srl     k0, k0, RCP_IMASKSHIFT
+    xor     k0, k0, 0xFFFFFFFF
+    andi    k0, k0, (RCP_IMASK >> RCP_IMASKSHIFT)
+    lw      t0, THREAD_RCP(a1)
+    and     k0, k0, t0
+    or      k1, k1, k0
+3:
+    /* If the specified thread queue is null, skip */
+    /*  straight to dispatching */
+    sw      k1, THREAD_RCP(a1)
+    beqz    a0, no_enqueue
+    jal     __osEnqueueThread
+no_enqueue:
+    j       __osDispatchThread
+END(__osEnqueueAndYield)
+
+/**
+ *  void __osEnqueueThread(OSThread** threadQueue, OSThread* thread);
+ *
+ *  Enqueues `thread` to the thread queue `threadQueue`, inserted by priority
+ */
+LEAF(__osEnqueueThread)
+    move    t9, a0
+    lw      t8, (a0)
+    lw      t7, THREAD_PRI(a1)
+    lw      t6, THREAD_PRI(t8)
+    /* If the current highest priority thread is a lower priority than */
+    /*  the new thread, skip searching the queue */
+    blt     t6, t7, 2f
+1:
+    /* Search the queue for the position to insert the thread to maintain */
+    /*  ordering by priority */
+    move    t9, t8
+    lw      t8, THREAD_NEXT(t8)
+    lw      t6, THREAD_PRI(t8)
+    bge     t6, t7, 1b
+2:
+    /* Insert the thread into the queue */
+    lw      t8, (t9)
+    sw      t8, THREAD_NEXT(a1)
+    sw      a1, (t9)
+    sw      a0, THREAD_QUEUE(a1)
+    jr      ra
+END(__osEnqueueThread)
+
+/**
+ *  OSThread* __osPopThread(OSThread** threadQueue);
+ *
+ *  Pops the highest priority thread from the top of the
+ *   thread queue `threadQueue` and returns it
+ */
+LEAF(__osPopThread)
+    lw      v0, (a0)
+    lw      t9, THREAD_NEXT(v0)
+    sw      t9, (a0)
+    jr      ra
+END(__osPopThread)
+
+LEAF(__osNop)
+    jr      ra
+END(__osNop)
+
+/**
+ *  void __osDispatchThread(void);
+ *
+ *  Dispatches the next thread to run after restoring the context
+ */
+LEAF(__osDispatchThread)
+    /* Obtain highest priority thread from the active run queue */
+    la      a0, __osRunQueue
+    jal     __osPopThread
+    /* Set thread as running */
+    sw      v0, __osRunningThread
+    li      t0, OS_STATE_RUNNING
+    sh      t0, THREAD_STATE(v0)
+    /* Restore SR, masking out any interrupts that are not also */
+    /* enabled in the global interrupt mask */
+    move    k0, v0
+    lw      k1, THREAD_SR(k0)
+    la      t0, __OSGlobalIntMask
+    lw      t0, (t0)
+    andi    t0, t0, SR_IMASK
+    andi    t1, k1, SR_IMASK
+    and     t1, t1, t0
+    and     k1, k1, ~SR_IMASK
+    or      k1, k1, t1
+    MTC0(   k1, C0_SR)
+    /* Restore GPRs */
+.set noat
+    ld      AT, THREAD_AT(k0)
+    ld      v0, THREAD_V0(k0)
+    ld      v1, THREAD_V1(k0)
+    ld      a0, THREAD_A0(k0)
+    ld      a1, THREAD_A1(k0)
+    ld      a2, THREAD_A2(k0)
+    ld      a3, THREAD_A3(k0)
+    ld      t0, THREAD_T0(k0)
+    ld      t1, THREAD_T1(k0)
+    ld      t2, THREAD_T2(k0)
+    ld      t3, THREAD_T3(k0)
+    ld      t4, THREAD_T4(k0)
+    ld      t5, THREAD_T5(k0)
+    ld      t6, THREAD_T6(k0)
+    ld      t7, THREAD_T7(k0)
+    ld      s0, THREAD_S0(k0)
+    ld      s1, THREAD_S1(k0)
+    ld      s2, THREAD_S2(k0)
+    ld      s3, THREAD_S3(k0)
+    ld      s4, THREAD_S4(k0)
+    ld      s5, THREAD_S5(k0)
+    ld      s6, THREAD_S6(k0)
+    ld      s7, THREAD_S7(k0)
+    ld      t8, THREAD_T8(k0)
+    ld      t9, THREAD_T9(k0)
+    ld      gp, THREAD_GP(k0)
+    ld      sp, THREAD_SP(k0)
+    ld      fp, THREAD_S8(k0)
+    ld      ra, THREAD_RA(k0)
+    ld      k1, THREAD_LO(k0)
+    mtlo    k1
+    ld      k1, THREAD_HI(k0)
+    mthi    k1
+    /* Move thread pc to EPC so that eret will return execution to where the thread left off */
+    lw      k1, THREAD_PC(k0)
+    MTC0(   k1, C0_EPC)
+    /* Check if the FPU was used by this thread and if so also restore the FPU registers */
+    lw      k1, THREAD_FP(k0)
+    beqz    k1, 1f
+
+.set noreorder 
+    lw      k1, THREAD_FPCSR(k0)
+    ctc1    k1, C1_FPCSR
+.set reorder
+    ldc1    $f0, THREAD_FP0(k0)
+    ldc1    $f2, THREAD_FP2(k0)
+    ldc1    $f4, THREAD_FP4(k0)
+    ldc1    $f6, THREAD_FP6(k0)
+    ldc1    $f8, THREAD_FP8(k0)
+    ldc1    $f10, THREAD_FP10(k0)
+    ldc1    $f12, THREAD_FP12(k0)
+    ldc1    $f14, THREAD_FP14(k0)
+    ldc1    $f16, THREAD_FP16(k0)
+    ldc1    $f18, THREAD_FP18(k0)
+    ldc1    $f20, THREAD_FP20(k0)
+    ldc1    $f22, THREAD_FP22(k0)
+    ldc1    $f24, THREAD_FP24(k0)
+    ldc1    $f26, THREAD_FP26(k0)
+    ldc1    $f28, THREAD_FP28(k0)
+    ldc1    $f30, THREAD_FP30(k0)
+1:
+    /* Restore RCP interrupt mask, masking out any RCP interrupts that */
+    /* are not also enabled in the global interrupt mask */
+.set noreorder
+    lw      k1, THREAD_RCP(k0)
+    la      k0, __OSGlobalIntMask
+    lw      k0, (k0)
+    srl     k0, k0, RCP_IMASKSHIFT
+    and     k1, k1, k0
+    sll     k1, k1, 1
+    la      k0, __osRcpImTable
+    addu    k1, k1, k0
+    lhu     k1, (k1)
+    la      k0, PHYS_TO_K1(MI_INTR_MASK_REG)
+    sw      k1, (k0)
+    /* Empty pipeline */
+    nop
+    nop
+    nop
+    nop
+    /* Resume thread execution */
+    eret
+.set reorder
+.set at
+END(__osDispatchThread)
+
+/**
+ *  void __osCleanupThread(void);
+ *
+ *  When a thread entrypoint function returns, it returns to this function.
+ *  This function is responsible for cleaning up the thread, signalling for the
+ *  current thread to be destroyed.
+ */
+LEAF(__osCleanupThread)
+    move    a0, zero
+    jal     osDestroyThread
+    /* Despite being a jal, this function does not return as the thread will have been destroyed */
+END(__osCleanupThread)
diff --git a/src/libultra/os/getintmask.s b/src/libultra/os/getintmask.s
new file mode 100644
index 000000000..f5e7cf770
--- /dev/null
+++ b/src/libultra/os/getintmask.s
@@ -0,0 +1,53 @@
+#include "PR/asm.h"
+#include "PR/regdef.h"
+#include "PR/R4300.h"
+#include "PR/rcp.h"
+#include "PR/os_exception.h"
+
+.text
+
+/**
+ *  OSIntMask osGetIntMask(void);
+ *
+ *  Gets the interrupt enable mask for the current thread.
+ *  Interrupts that are not enabled in the global interrupt mask __OSGlobalIntMask
+ *   are not returned here. The global interrupt mask is OS-internal and is not
+ *   expected to change during runtime.
+ *
+ * @bug Some usage of the global interrupt mask is broken both in the
+ *       get/set interrupt mask routines and in the exception handler routines.
+ *       See the comment for osSetIntMask for more details.
+ */
+LEAF(osGetIntMask)
+.set noreorder
+    /* Extract interrupt enable bits from current SR */
+    mfc0    v0, C0_SR
+    andi    v0, v0, (SR_IMASK | SR_IE)
+    /* Get value of __OSGlobalIntMask */
+    la      t0, __OSGlobalIntMask
+    lw      t1, (t0)
+    /* Bitwise-OR in the disabled CPU bits of __OSGlobalIntMask */
+    xor     t0, t1, ~0
+    andi    t0, t0, SR_IMASK
+    or      v0, v0, t0
+    /* Fetch MI_INTR_MASK_REG */
+    lw      t1, PHYS_TO_K1(MI_INTR_MASK_REG)
+    /* If there are RCP interrupts masked */
+    beqz    t1, 1f
+    /* Get value of __OSGlobalIntMask */
+     la     t0, __OSGlobalIntMask /* Note: macro expansion in delay slot */
+    lw      t0, (t0)
+    /* Bitwise-OR in the disabled RCP bits of __OSGlobalIntMask */
+    srl     t0, t0, RCP_IMASKSHIFT
+    xor     t0, t0, ~0
+    andi    t0, t0, (RCP_IMASK >> RCP_IMASKSHIFT)
+    or      t1, t1, t0
+1:
+    /* Shift the RCP bits to not conflict with the CPU bits */
+    sll     t2, t1, RCP_IMASKSHIFT
+    /* OR the CPU and RCP bits together */
+    or      v0, v0, t2
+    jr      ra
+     nop
+.set reorder
+END(osGetIntMask)
diff --git a/src/libultra/os/setintmask.s b/src/libultra/os/setintmask.s
new file mode 100644
index 000000000..68a96bc8b
--- /dev/null
+++ b/src/libultra/os/setintmask.s
@@ -0,0 +1,154 @@
+#include "PR/asm.h"
+#include "PR/regdef.h"
+#include "PR/R4300.h"
+#include "PR/rcp.h"
+#include "PR/os_exception.h"
+
+.rdata
+.align 2
+
+/**
+ *  LUT to convert between an interrupt mask value and a value for MI_INTR_MASK_REG.
+ *  The interrupt mask value is a single bit 0 = disabled, 1 = enabled, while writes to MI_INTR_MASK_REG has two distinct non-zero
+ *   values for set and clear, hence the need for a conversion step.
+ */
+DATA(__osRcpImTable)
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_CLR_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_CLR_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_CLR_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_CLR_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_CLR_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_CLR_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+    .half MI_INTR_MASK_SET_SP | MI_INTR_MASK_SET_SI | MI_INTR_MASK_SET_AI | MI_INTR_MASK_SET_VI | MI_INTR_MASK_SET_PI | MI_INTR_MASK_SET_DP
+ENDDATA(__osRcpImTable)
+
+.text
+
+/**
+ *  OSIntMask osSetIntMask(OSIntMask);
+ *
+ *  Sets the interrupt enable mask for the current thread. External interrupts
+ *   originating either in the CPU or the RCP may be "masked out" so that they
+ *   are not handled. This is sometimes important for critical code sections
+ *   that must not be interrupted.
+ *  Interrupts that are not enabled in the global interrupt mask __OSGlobalIntMask
+ *   cannot be set here. The global interrupt mask is OS-internal and is not
+ *   expected to change during runtime.
+ *  The returned value is the previous interrupt enable mask so that it can be
+ *   restored later.
+ *
+ * @bug Some usage of the global interrupt mask is broken both in the
+ *       get/set interrupt mask routines and in the exception handler routines.
+ *      While a thread is running, the C0_SR interrupt enable bits contain the
+ *       interrupt enable bits for the current thread masked by the global
+ *       interrupt mask. There is an attempt to recover only the original interrupt
+ *       enable bits belonging to the thread itself using the operation
+ *          (SR | ~__OSGlobalIntMask).
+ *      However, this does not work as intended and can cause interrupts to end
+ *       up enabled when not intended to be. The same issue is present for the
+ *       RCP interrupt enable bits in MI_INTR_MASK_REG.
+ *      This does not cause issues in practice as __OSGlobalIntMask is almost always
+ *       OS_IM_ALL, so the operation is usually simply (SR | 0).
+ */
+LEAF(osSetIntMask)
+     /* Extract interrupt enable bits from current SR */
+    MFC0(   t4, C0_SR)
+.set noreorder
+    and     v0, t4, (SR_IMASK | SR_IE)
+     /* Get value of __OSGlobalIntMask */
+    la      t0, __OSGlobalIntMask
+    lw      t3, (t0)
+     /* Bitwise-OR in the disabled CPU bits of __OSGlobalIntMask */
+    xor     t0, t3, 0xFFFFFFFF
+    and     t0, t0, SR_IMASK
+    or      v0, v0, t0
+     /* Fetch MI_INTR_MASK_REG */
+    lw      t2, PHYS_TO_K1(MI_INTR_MASK_REG)
+     /* If there are RCP interrupts masked */
+    beqz    t2, 1f
+     srl    t1, t3, RCP_IMASKSHIFT
+    /* Bitwise-OR in the disabled RCP bits of __OSGlobalIntMask */
+    xor     t1, t1, 0xFFFFFFFF
+    and     t1, t1, (RCP_IMASK >> RCP_IMASKSHIFT)
+    or      t2, t2, t1
+1:
+     /* Shift the RCP bits to not conflict with the CPU bits */
+    sll     t2, t2, RCP_IMASKSHIFT
+     /* OR the CPU and RCP bits together */
+    or      v0, v0, t2
+     /* Extract RCP interrupt enable bits from requested mask and mask with __OSGlobalIntMask */
+    and     t0, a0, RCP_IMASK
+    and     t0, t0, t3
+     /* Convert to a value for MI_INTR_MASK_REG and set it */
+    srl     t0, t0, (RCP_IMASKSHIFT-1)
+    lhu     t2, __osRcpImTable(t0)
+    sw      t2, PHYS_TO_K1(MI_INTR_MASK_REG)
+     /* Extract CPU interrupt enable bits from requested mask and mask with __OSGlobalIntMask */
+    and     t0, a0, OS_IM_CPU
+    and     t1, t3, SR_IMASK
+    and     t0, t0, t1
+    and     t4, t4, ~SR_IMASK
+     /* Bitwise OR in the remaining bits of SR and set new SR */
+    or      t4, t4, t0
+    MTC0(   t4, C0_SR)
+    NOP
+    NOP
+    jr      ra
+END(osSetIntMask)
diff --git a/tools/warnings_count/warnings_build_current.txt b/tools/warnings_count/warnings_build_current.txt
index fe092b13c..96fff2b09 100644
--- a/tools/warnings_count/warnings_build_current.txt
+++ b/tools/warnings_count/warnings_build_current.txt
@@ -1,3 +1,5 @@
-Warning: Invalid pointer indirect: 0 for samplebank SampleBank_2
+Warning: In Soundfont Soundfont_40: Invalid pointer indirect 0 for samplebank SampleBank_2
+cc: Warning: -mips3 should not be used for ucode 32-bit compiles
+as1: Warning: src/libultra/os/getintmask.s, line 38: Macro instruction used in branch delay slot
 cc: Warning: -mips3 should not be used for ucode 32-bit compiles
 cc: Warning: -mips3 should not be used for ucode 32-bit compiles