Improve rcp.h, remove `HW_REG` macro (#1425)

* Real rcp.h

* Correction to comment in initialize.c

* Try fix R4300.h

* Adjust rcp.h formatting, remove defines in other headers that are now in rcp.h

* Suggested changes, document a bug in the modified osAiSetNextBuffer

* More rcp.h formatting changes

46 files changed, 256 insertions, 229 deletions

diff --git a/src/code/graph.c b/src/code/graph.c
index 0deb81fb8..bbd87434c 100644
--- a/src/code/graph.c
+++ b/src/code/graph.c
@@ -175,8 +175,8 @@ void Graph_TaskSet00(GraphicsContext* gfxCtx) {
         osSyncPrintf("RCPが帰ってきませんでした。"); // "RCP did not return."
         osSyncPrintf(VT_RST);
 
-        LogUtils_LogHexDump((void*)&HW_REG(SP_MEM_ADDR_REG, u32), 0x20);
-        LogUtils_LogHexDump((void*)&DPC_START_REG, 0x20);
+        LogUtils_LogHexDump((void*)PHYS_TO_K1(SP_BASE_REG), 0x20);
+        LogUtils_LogHexDump((void*)PHYS_TO_K1(DPC_BASE_REG), 0x20);
         LogUtils_LogHexDump(gGfxSPTaskYieldBuffer, sizeof(gGfxSPTaskYieldBuffer));
 
         SREG(6) = -1;
@@ -321,8 +321,8 @@ void Graph_Update(GraphicsContext* gfxCtx, GameState* gameState) {
         }
 
         if (HREG(81) < 0) {
-            LogUtils_LogHexDump((void*)&HW_REG(SP_MEM_ADDR_REG, u32), 0x20);
-            LogUtils_LogHexDump((void*)&DPC_START_REG, 0x20);
+            LogUtils_LogHexDump((void*)PHYS_TO_K1(SP_BASE_REG), 0x20);
+            LogUtils_LogHexDump((void*)PHYS_TO_K1(DPC_BASE_REG), 0x20);
         }
 
         if (HREG(81) < 0) {
diff --git a/src/code/rcp_utils.c b/src/code/rcp_utils.c
index 5c14c39f9..e258ab02f 100644
--- a/src/code/rcp_utils.c
+++ b/src/code/rcp_utils.c
@@ -50,6 +50,6 @@ void RcpUtils_Reset(void) {
     // Flush the RDP pipeline and freeze clock counter
     osDpSetStatus(DPC_SET_FREEZE | DPC_SET_FLUSH);
     // Halt the RSP, disable interrupt on break and set "task done" signal
-    __osSpSetStatus(SP_SET_HALT | SP_SET_SIG2 | SP_CLR_INTR_BREAK);
+    __osSpSetStatus(SP_SET_HALT | SP_SET_TASKDONE | SP_CLR_INTR_BREAK);
     RcpUtils_PrintRegisterStatus();
 }
diff --git a/src/libultra/gu/sintable.inc.c b/src/libultra/gu/sintable.inc.c
index d430d65ae..2753d55a9 100644
--- a/src/libultra/gu/sintable.inc.c
+++ b/src/libultra/gu/sintable.inc.c
@@ -1,4 +1,4 @@
-#include "ultra64/types.h"
+#include "ultra64/ultratypes.h"
 
 static s16 sintable[0x400] = {
     0x0000, 0x0032, 0x0064, 0x0096, 0x00C9, 0x00FB, 0x012D, 0x0160, 0x0192, 0x01C4, 0x01F7, 0x0229, 0x025B, 0x028E,
diff --git a/src/libultra/io/aigetlen.c b/src/libultra/io/aigetlen.c
index 58d8db487..45be44251 100644
--- a/src/libultra/io/aigetlen.c
+++ b/src/libultra/io/aigetlen.c
@@ -1,5 +1,11 @@
 #include "global.h"
 
+/**
+ * Returns the number of bytes remaining in a currently ongoing audio DMA.
+ *
+ * Note that audio DMA is double-buffered, a DMA can be queued while another is in-progress. This only returns
+ * information about the currently in-progress DMA.
+ */
 u32 osAiGetLength(void) {
-    return HW_REG(AI_LEN_REG, u32);
+    return IO_READ(AI_LEN_REG);
 }
diff --git a/src/libultra/io/aisetfreq.c b/src/libultra/io/aisetfreq.c
index 8702941b9..9e1b8c44e 100644
--- a/src/libultra/io/aisetfreq.c
+++ b/src/libultra/io/aisetfreq.c
@@ -1,20 +1,35 @@
 #include "global.h"
 
+/**
+ * Programs the operating frequency of the Audio DAC.
+ *
+ * @param frequency Target Playback frequency.
+ * @return The actual playback frequency, or -1 if the supplied frequency cannot be used.
+ */
 s32 osAiSetFrequency(u32 frequency) {
-    u8 bitrate;
+    // Calculate the DAC sample period ("dperiod") (dperiod + 1 = vid_clock / frequency)
     f32 dacRateF = ((f32)osViClock / frequency) + 0.5f;
+    u8 bitrate;
     u32 dacRate = dacRateF;
 
-    if (dacRate < 132) {
+    // Upcoming division by 66. If dacRate is smaller than 2 * 66 = 132, bitrate will be 1 and AI_BITRATE_REG will be
+    // programmed with 0, which results in no audio output. Return -1 to indicate an unusable frequency.
+    if (dacRate < AI_MIN_DAC_RATE) {
         return -1;
     }
 
+    // Calculate the largest "bitrate" (ABUS clock half period, "aclockhp") supported for this dacrate. These two
+    // quantities must satisfy (dperiod + 1) >= 66 * (aclockhp + 1), here this is taken as equality.
     bitrate = (dacRate / 66);
-    if (bitrate > 16) {
-        bitrate = 16;
+    // Clamp to max value
+    if (bitrate > AI_MAX_BIT_RATE) {
+        bitrate = AI_MAX_BIT_RATE;
     }
 
-    HW_REG(AI_DACRATE_REG, u32) = dacRate - 1;
-    HW_REG(AI_BITRATE_REG, u32) = bitrate - 1;
+    IO_WRITE(AI_DACRATE_REG, dacRate - 1);
+    IO_WRITE(AI_BITRATE_REG, bitrate - 1);
+
+    // Return the true playback frequency (frequency = vid_clock / (dperiod + 1)), which may differ from the target
+    // frequency.
     return osViClock / (s32)dacRate;
 }
diff --git a/src/libultra/io/aisetnextbuf.c b/src/libultra/io/aisetnextbuf.c
index 56a7e4e6e..4cf5d1f7d 100644
--- a/src/libultra/io/aisetnextbuf.c
+++ b/src/libultra/io/aisetnextbuf.c
@@ -1,31 +1,44 @@
 #include "global.h"
 
-//! Note that this is not the same as the original libultra
-//! osAiSetNextBuffer, see comments in the function
-
+/**
+ * Submits an audio buffer to be consumed by the Audio DAC. The audio interface can queue a second DMA while another
+ * is in progress and automatically begin the next one as soon as the current DMA completes. If there is already a
+ * second DMA queued (DMA is full), -1 is returned to indicate the buffer could not be submitted.
+ *
+ * Note that this is not the same as the original libultra osAiSetNextBuffer, see comments in the function.
+ *
+ * @param buf Next audio buffer. Must be an 8-byte aligned KSEG0 (0x80XXXXXX) address.
+ * @param size Length of next audio buffer in bytes, maximum size 0x40000 bytes / 256 KiB. Should be a multiple of 8.
+ * @return 0 if the DMA was enqueued successfully, -1 if the DMA could not yet be queued.
+ */
 s32 osAiSetNextBuffer(void* buf, u32 size) {
-    static u8 D_80130500 = false;
+    static u8 hdwrBugFlag = false;
     u32 bufAdjusted = (u32)buf;
     s32 status;
 
-    if (D_80130500) {
-        bufAdjusted = (u32)buf - 0x2000;
+    // Workaround for a hardware bug. If the end of the previous buffer was on an 0x2000 byte boundary, adjust the
+    // start of the next buffer.
+    if (hdwrBugFlag) {
+        bufAdjusted -= 0x2000;
     }
+    // Current buffer ends on an 0x2000 byte boundary, set flag to account for this in next buffer.
     if ((((u32)buf + size) & 0x1FFF) == 0) {
-        D_80130500 = true;
+        hdwrBugFlag = true;
     } else {
-        D_80130500 = false;
+        hdwrBugFlag = false;
     }
 
     // Originally a call to __osAiDeviceBusy
-    status = HW_REG(AI_STATUS_REG, s32);
-    if (status & AI_STATUS_AI_FULL) {
+    //! @bug The original __osAiDeviceBusy call was above the hardware bug workaround to ensure that it was only
+    //! performed when a transfer was guaranteed to start. If this condition passes and this function returns without
+    //! submitting a buffer for DMA, the code above will lose track of when to apply the workaround.
+    status = IO_READ(AI_STATUS_REG);
+    if (status & AI_STATUS_FIFO_FULL) {
         return -1;
     }
 
-    // OS_K0_TO_PHYSICAL replaces osVirtualToPhysical, this replacement
-    // assumes that only KSEG0 addresses are given
-    HW_REG(AI_DRAM_ADDR_REG, u32) = OS_K0_TO_PHYSICAL(bufAdjusted);
-    HW_REG(AI_LEN_REG, u32) = size;
+    // OS_K0_TO_PHYSICAL replaces osVirtualToPhysical, this replacement assumes that only KSEG0 addresses are given.
+    IO_WRITE(AI_DRAM_ADDR_REG, OS_K0_TO_PHYSICAL(bufAdjusted));
+    IO_WRITE(AI_LEN_REG, size);
     return 0;
 }
diff --git a/src/libultra/io/cartrominit.c b/src/libultra/io/cartrominit.c
index 231b00768..b2cf1d1e1 100644
--- a/src/libultra/io/cartrominit.c
+++ b/src/libultra/io/cartrominit.c
@@ -3,55 +3,54 @@
 OSPiHandle __CartRomHandle;
 
 OSPiHandle* osCartRomInit(void) {
-    register u32 a;
+    static u32 first = true;
+    register u32 value;
     register s32 status;
     register u32 prevInt;
-    register u32 lastLatency;
-    register u32 lastPageSize;
-    register u32 lastRelDuration;
-    register u32 lastPulse;
-
-    static u32 D_8000AF10 = 1;
+    register u32 latency;
+    register u32 pageSize;
+    register u32 relDuration;
+    register u32 pulse;
 
     __osPiGetAccess();
 
-    if (!D_8000AF10) {
+    if (!first) {
         __osPiRelAccess();
         return &__CartRomHandle;
     }
 
-    D_8000AF10 = 0;
+    first = false;
     __CartRomHandle.type = DEVICE_TYPE_CART;
-    __CartRomHandle.baseAddress = 0xB0000000;
+    __CartRomHandle.baseAddress = PHYS_TO_K1(PI_DOM1_ADDR2);
     __CartRomHandle.domain = PI_DOMAIN1;
     __CartRomHandle.speed = 0;
     bzero(&__CartRomHandle.transferInfo, sizeof(__OSTranxInfo));
 
-    status = HW_REG(PI_STATUS_REG, u32);
-    while (status & (PI_STATUS_BUSY | PI_STATUS_IOBUSY)) {
-        status = HW_REG(PI_STATUS_REG, u32);
+    status = IO_READ(PI_STATUS_REG);
+    while (status & (PI_STATUS_DMA_BUSY | PI_STATUS_IO_BUSY)) {
+        status = IO_READ(PI_STATUS_REG);
     }
 
-    lastLatency = HW_REG(PI_BSD_DOM1_LAT_REG, u32);
-    lastPageSize = HW_REG(PI_BSD_DOM1_PGS_REG, u32);
-    lastRelDuration = HW_REG(PI_BSD_DOM1_RLS_REG, u32);
-    lastPulse = HW_REG(PI_BSD_DOM1_PWD_REG, u32);
-
-    HW_REG(PI_BSD_DOM1_LAT_REG, u32) = 0xFF;
-    HW_REG(PI_BSD_DOM1_PGS_REG, u32) = 0;
-    HW_REG(PI_BSD_DOM1_RLS_REG, u32) = 3;
-    HW_REG(PI_BSD_DOM1_PWD_REG, u32) = 0xFF;
-
-    a = HW_REG(__CartRomHandle.baseAddress, u32);
-    __CartRomHandle.latency = a & 0xFF;
-    __CartRomHandle.pageSize = (a >> 0x10) & 0xF;
-    __CartRomHandle.relDuration = (a >> 0x14) & 0xF;
-    __CartRomHandle.pulse = (a >> 8) & 0xFF;
-
-    HW_REG(PI_BSD_DOM1_LAT_REG, u32) = lastLatency;
-    HW_REG(PI_BSD_DOM1_PGS_REG, u32) = lastPageSize;
-    HW_REG(PI_BSD_DOM1_RLS_REG, u32) = lastRelDuration;
-    HW_REG(PI_BSD_DOM1_PWD_REG, u32) = lastPulse;
+    latency = IO_READ(PI_BSD_DOM1_LAT_REG);
+    pageSize = IO_READ(PI_BSD_DOM1_PGS_REG);
+    relDuration = IO_READ(PI_BSD_DOM1_RLS_REG);
+    pulse = IO_READ(PI_BSD_DOM1_PWD_REG);
+
+    IO_WRITE(PI_BSD_DOM1_LAT_REG, 255);
+    IO_WRITE(PI_BSD_DOM1_PGS_REG, 0);
+    IO_WRITE(PI_BSD_DOM1_RLS_REG, 3);
+    IO_WRITE(PI_BSD_DOM1_PWD_REG, 255);
+
+    value = IO_READ(__CartRomHandle.baseAddress);
+    __CartRomHandle.latency = value & 0xFF;
+    __CartRomHandle.pageSize = (value >> 0x10) & 0xF;
+    __CartRomHandle.relDuration = (value >> 0x14) & 0xF;
+    __CartRomHandle.pulse = (value >> 8) & 0xFF;
+
+    IO_WRITE(PI_BSD_DOM1_LAT_REG, latency);
+    IO_WRITE(PI_BSD_DOM1_PGS_REG, pageSize);
+    IO_WRITE(PI_BSD_DOM1_RLS_REG, relDuration);
+    IO_WRITE(PI_BSD_DOM1_PWD_REG, pulse);
 
     prevInt = __osDisableInt();
     __CartRomHandle.next = __osPiTable;
diff --git a/src/libultra/io/devmgr.c b/src/libultra/io/devmgr.c
index e8581d821..5d1ce6e9a 100644
--- a/src/libultra/io/devmgr.c
+++ b/src/libultra/io/devmgr.c
@@ -43,7 +43,7 @@ void __osDevMgrMain(void* arg) {
                         __osEPiRawWriteIo(ioMesg->piHandle, 0x05000510, transfer->bmCtlShadow | 0x1000000);
                     }
                     block->errStatus = 4;
-                    HW_REG(PI_STATUS_REG, u32) = PI_STATUS_CLR_INTR;
+                    IO_WRITE(PI_STATUS_REG, PI_STATUS_CLR_INTR);
                     __osSetGlobalIntMask(OS_IM_CART | OS_IM_PI);
                 }
                 osSendMesg(ioMesg->hdr.retQueue, (OSMesg)ioMesg, OS_MESG_NOBLOCK);
diff --git a/src/libultra/io/dpgetstat.c b/src/libultra/io/dpgetstat.c
index b5267b060..1a944b2c7 100644
--- a/src/libultra/io/dpgetstat.c
+++ b/src/libultra/io/dpgetstat.c
@@ -1,5 +1,5 @@
 #include "global.h"
 
 u32 osDpGetStatus(void) {
-    return DPC_STATUS_REG;
+    return IO_READ(DPC_STATUS_REG);
 }
diff --git a/src/libultra/io/dpsetstat.c b/src/libultra/io/dpsetstat.c
index ecd3f77e0..4275e1fde 100644
--- a/src/libultra/io/dpsetstat.c
+++ b/src/libultra/io/dpsetstat.c
@@ -1,5 +1,5 @@
 #include "global.h"
 
 void osDpSetStatus(u32 status) {
-    DPC_STATUS_REG = status;
+    IO_WRITE(DPC_STATUS_REG, status);
 }
diff --git a/src/libultra/io/driverominit.c b/src/libultra/io/driverominit.c
index d6247298a..7d1441cdd 100644
--- a/src/libultra/io/driverominit.c
+++ b/src/libultra/io/driverominit.c
@@ -3,45 +3,45 @@
 OSPiHandle __DriveRomHandle;
 
 OSPiHandle* osDriveRomInit(void) {
+    static u32 first = true;
     register s32 status;
-    register u32 a;
+    register u32 value;
     register u32 prevInt;
-    static u32 D_8000AC70 = 1;
 
     __osPiGetAccess();
 
-    if (!D_8000AC70) {
+    if (!first) {
         __osPiRelAccess();
         return &__DriveRomHandle;
     }
 
-    D_8000AC70 = 0;
+    first = false;
     __DriveRomHandle.type = DEVICE_TYPE_BULK;
-    __DriveRomHandle.baseAddress = 0xA6000000;
+    __DriveRomHandle.baseAddress = PHYS_TO_K1(PI_DOM1_ADDR1);
     __DriveRomHandle.domain = PI_DOMAIN1;
     __DriveRomHandle.speed = 0;
     bzero(&__DriveRomHandle.transferInfo, sizeof(__OSTranxInfo));
 
-    status = HW_REG(PI_STATUS_REG, u32);
-    while (status & (PI_STATUS_BUSY | PI_STATUS_IOBUSY)) {
-        status = HW_REG(PI_STATUS_REG, u32);
+    status = IO_READ(PI_STATUS_REG);
+    while (status & (PI_STATUS_DMA_BUSY | PI_STATUS_IO_BUSY)) {
+        status = IO_READ(PI_STATUS_REG);
     }
 
-    HW_REG(PI_BSD_DOM1_LAT_REG, u32) = 0xFF;
-    HW_REG(PI_BSD_DOM1_PGS_REG, u32) = 0;
-    HW_REG(PI_BSD_DOM1_RLS_REG, u32) = 3;
-    HW_REG(PI_BSD_DOM1_PWD_REG, u32) = 0xFF;
-
-    a = HW_REG(__DriveRomHandle.baseAddress, u32);
-    __DriveRomHandle.latency = a & 0xFF;
-    __DriveRomHandle.pulse = (a >> 8) & 0xFF;
-    __DriveRomHandle.pageSize = (a >> 0x10) & 0xF;
-    __DriveRomHandle.relDuration = (a >> 0x14) & 0xF;
-
-    HW_REG(PI_BSD_DOM1_LAT_REG, u32) = (u8)a;
-    HW_REG(PI_BSD_DOM1_PGS_REG, u32) = __DriveRomHandle.pageSize;
-    HW_REG(PI_BSD_DOM1_RLS_REG, u32) = __DriveRomHandle.relDuration;
-    HW_REG(PI_BSD_DOM1_PWD_REG, u32) = __DriveRomHandle.pulse;
+    IO_WRITE(PI_BSD_DOM1_LAT_REG, 255);
+    IO_WRITE(PI_BSD_DOM1_PGS_REG, 0);
+    IO_WRITE(PI_BSD_DOM1_RLS_REG, 3);
+    IO_WRITE(PI_BSD_DOM1_PWD_REG, 255);
+
+    value = IO_READ(__DriveRomHandle.baseAddress);
+    __DriveRomHandle.latency = value & 0xFF;
+    __DriveRomHandle.pulse = (value >> 8) & 0xFF;
+    __DriveRomHandle.pageSize = (value >> 0x10) & 0xF;
+    __DriveRomHandle.relDuration = (value >> 0x14) & 0xF;
+
+    IO_WRITE(PI_BSD_DOM1_LAT_REG, __DriveRomHandle.latency);
+    IO_WRITE(PI_BSD_DOM1_PGS_REG, __DriveRomHandle.pageSize);
+    IO_WRITE(PI_BSD_DOM1_RLS_REG, __DriveRomHandle.relDuration);
+    IO_WRITE(PI_BSD_DOM1_PWD_REG, __DriveRomHandle.pulse);
 
     __osCurrentHandle[__DriveRomHandle.domain]->type = __DriveRomHandle.type;
     __osCurrentHandle[__DriveRomHandle.domain]->latency = __DriveRomHandle.latency;
diff --git a/src/libultra/io/epirawdma.c b/src/libultra/io/epirawdma.c
index ccc23f708..ff6bb9a65 100644
--- a/src/libultra/io/epirawdma.c
+++ b/src/libultra/io/epirawdma.c
@@ -4,9 +4,9 @@ s32 __osEPiRawStartDma(OSPiHandle* handle, s32 direction, u32 cartAddr, void* dr
     s32 status;
     OSPiHandle* curHandle;
 
-    status = HW_REG(PI_STATUS_REG, u32);
-    while (status & (PI_STATUS_BUSY | PI_STATUS_IOBUSY)) {
-        status = HW_REG(PI_STATUS_REG, u32);
+    status = IO_READ(PI_STATUS_REG);
+    while (status & (PI_STATUS_DMA_BUSY | PI_STATUS_IO_BUSY)) {
+        status = IO_READ(PI_STATUS_REG);
     }
 
     if (__osCurrentHandle[handle->domain]->type != handle->type) {
@@ -14,35 +14,35 @@ s32 __osEPiRawStartDma(OSPiHandle* handle, s32 direction, u32 cartAddr, void* dr
 
         if (handle->domain == 0) {
             if (curHandle->latency != handle->latency) {
-                HW_REG(PI_BSD_DOM1_LAT_REG, u32) = handle->latency;
+                IO_WRITE(PI_BSD_DOM1_LAT_REG, handle->latency);
             }
 
             if (curHandle->pageSize != handle->pageSize) {
-                HW_REG(PI_BSD_DOM1_PGS_REG, u32) = handle->pageSize;
+                IO_WRITE(PI_BSD_DOM1_PGS_REG, handle->pageSize);
             }
 
             if (curHandle->relDuration != handle->relDuration) {
-                HW_REG(PI_BSD_DOM1_RLS_REG, u32) = handle->relDuration;
+                IO_WRITE(PI_BSD_DOM1_RLS_REG, handle->relDuration);
             }
 
             if (curHandle->pulse != handle->pulse) {
-                HW_REG(PI_BSD_DOM1_PWD_REG, u32) = handle->pulse;
+                IO_WRITE(PI_BSD_DOM1_PWD_REG, handle->pulse);
             }
         } else {
             if (curHandle->latency != handle->latency) {
-                HW_REG(PI_BSD_DOM2_LAT_REG, u32) = handle->latency;
+                IO_WRITE(PI_BSD_DOM2_LAT_REG, handle->latency);
             }
 
             if (curHandle->pageSize != handle->pageSize) {
-                HW_REG(PI_BSD_DOM2_PGS_REG, u32) = handle->pageSize;
+                IO_WRITE(PI_BSD_DOM2_PGS_REG, handle->pageSize);
             }
 
             if (curHandle->relDuration != handle->relDuration) {
-                HW_REG(PI_BSD_DOM2_RLS_REG, u32) = handle->relDuration;
+                IO_WRITE(PI_BSD_DOM2_RLS_REG, handle->relDuration);
             }
 
             if (curHandle->pulse != handle->pulse) {
-                HW_REG(PI_BSD_DOM2_PWD_REG, u32) = handle->pulse;
+                IO_WRITE(PI_BSD_DOM2_PWD_REG, handle->pulse);
             }
         }
 
@@ -53,20 +53,18 @@ s32 __osEPiRawStartDma(OSPiHandle* handle, s32 direction, u32 cartAddr, void* dr
         curHandle->pulse = handle->pulse;
     }
 
-    HW_REG(PI_DRAM_ADDR_REG, void*) = (void*)osVirtualToPhysical(dramAddr);
-    HW_REG(PI_CART_ADDR_REG, void*) = (void*)((handle->baseAddress | cartAddr) & 0x1FFFFFFF);
+    IO_WRITE(PI_DRAM_ADDR_REG, osVirtualToPhysical(dramAddr));
+    IO_WRITE(PI_CART_ADDR_REG, K1_TO_PHYS(handle->baseAddress | cartAddr));
 
     switch (direction) {
         case OS_READ:
-            HW_REG(PI_WR_LEN_REG, u32) = size - 1;
+            IO_WRITE(PI_WR_LEN_REG, size - 1);
             break;
         case OS_WRITE:
-            HW_REG(PI_RD_LEN_REG, u32) = size - 1;
+            IO_WRITE(PI_RD_LEN_REG, size - 1);
             break;
         default:
             return -1;
-            break;
     }
-
     return 0;
 }
diff --git a/src/libultra/io/epirawread.c b/src/libultra/io/epirawread.c
index 5a3109364..41997df66 100644
--- a/src/libultra/io/epirawread.c
+++ b/src/libultra/io/epirawread.c
@@ -4,9 +4,9 @@ s32 __osEPiRawReadIo(OSPiHandle* handle, u32 devAddr, u32* data) {
     s32 status;
     OSPiHandle* curHandle;
 
-    status = HW_REG(PI_STATUS_REG, u32);
-    while (status & (PI_STATUS_BUSY | PI_STATUS_IOBUSY)) {
-        status = HW_REG(PI_STATUS_REG, u32);
+    status = IO_READ(PI_STATUS_REG);
+    while (status & (PI_STATUS_DMA_BUSY | PI_STATUS_IO_BUSY)) {
+        status = IO_READ(PI_STATUS_REG);
     }
 
     if (__osCurrentHandle[handle->domain]->type != handle->type) {
@@ -14,35 +14,35 @@ s32 __osEPiRawReadIo(OSPiHandle* handle, u32 devAddr, u32* data) {
 
         if (handle->domain == 0) {
             if (curHandle->latency != handle->latency) {
-                HW_REG(PI_BSD_DOM1_LAT_REG, u32) = handle->latency;
+                IO_WRITE(PI_BSD_DOM1_LAT_REG, handle->latency);
             }
 
             if (curHandle->pageSize != handle->pageSize) {
-                HW_REG(PI_BSD_DOM1_PGS_REG, u32) = handle->pageSize;
+                IO_WRITE(PI_BSD_DOM1_PGS_REG, handle->pageSize);
             }
 
             if (curHandle->relDuration != handle->relDuration) {
-                HW_REG(PI_BSD_DOM1_RLS_REG, u32) = handle->relDuration;
+                IO_WRITE(PI_BSD_DOM1_RLS_REG, handle->relDuration);
             }
 
             if (curHandle->pulse != handle->pulse) {
-                HW_REG(PI_BSD_DOM1_PWD_REG, u32) = handle->pulse;
+                IO_WRITE(PI_BSD_DOM1_PWD_REG, handle->pulse);
             }
         } else {
             if (curHandle->latency != handle->latency) {
-                HW_REG(PI_BSD_DOM2_LAT_REG, u32) = handle->latency;
+                IO_WRITE(PI_BSD_DOM2_LAT_REG, handle->latency);
             }
 
             if (curHandle->pageSize != handle->pageSize) {
-                HW_REG(PI_BSD_DOM2_PGS_REG, u32) = handle->pageSize;
+                IO_WRITE(PI_BSD_DOM2_PGS_REG, handle->pageSize);
             }
 
             if (curHandle->relDuration != handle->relDuration) {
-                HW_REG(PI_BSD_DOM2_RLS_REG, u32) = handle->relDuration;
+                IO_WRITE(PI_BSD_DOM2_RLS_REG, handle->relDuration);
             }
 
             if (curHandle->pulse != handle->pulse) {
-                HW_REG(PI_BSD_DOM2_PWD_REG, u32) = handle->pulse;
+                IO_WRITE(PI_BSD_DOM2_PWD_REG, handle->pulse);
             }
         }
 
@@ -53,6 +53,6 @@ s32 __osEPiRawReadIo(OSPiHandle* handle, u32 devAddr, u32* data) {
         curHandle->pulse = handle->pulse;
     }
 
-    *data = HW_REG(handle->baseAddress | devAddr | 0, u32);
+    *data = IO_READ(handle->baseAddress | devAddr);
     return 0;
 }
diff --git a/src/libultra/io/epirawwrite.c b/src/libultra/io/epirawwrite.c
index 2c24c5e25..4d704a192 100644
--- a/src/libultra/io/epirawwrite.c
+++ b/src/libultra/io/epirawwrite.c
@@ -4,9 +4,9 @@ s32 __osEPiRawWriteIo(OSPiHandle* handle, u32 devAddr, u32 data) {
     s32 status;
     OSPiHandle* curHandle;
 
-    status = HW_REG(PI_STATUS_REG, u32);
-    while (status & (PI_STATUS_BUSY | PI_STATUS_IOBUSY)) {
-        status = HW_REG(PI_STATUS_REG, u32);
+    status = IO_READ(PI_STATUS_REG);
+    while (status & (PI_STATUS_DMA_BUSY | PI_STATUS_IO_BUSY)) {
+        status = IO_READ(PI_STATUS_REG);
     }
 
     if (__osCurrentHandle[handle->domain]->type != handle->type) {
@@ -14,35 +14,35 @@ s32 __osEPiRawWriteIo(OSPiHandle* handle, u32 devAddr, u32 data) {
 
         if (handle->domain == 0) {
             if (curHandle->latency != handle->latency) {
-                HW_REG(PI_BSD_DOM1_LAT_REG, u32) = handle->latency;
+                IO_WRITE(PI_BSD_DOM1_LAT_REG, handle->latency);
             }
 
             if (curHandle->pageSize != handle->pageSize) {
-                HW_REG(PI_BSD_DOM1_PGS_REG, u32) = handle->pageSize;
+                IO_WRITE(PI_BSD_DOM1_PGS_REG, handle->pageSize);
             }
 
             if (curHandle->relDuration != handle->relDuration) {
-                HW_REG(PI_BSD_DOM1_RLS_REG, u32) = handle->relDuration;
+                IO_WRITE(PI_BSD_DOM1_RLS_REG, handle->relDuration);
             }
 
             if (curHandle->pulse != handle->pulse) {
-                HW_REG(PI_BSD_DOM1_PWD_REG, u32) = handle->pulse;
+                IO_WRITE(PI_BSD_DOM1_PWD_REG, handle->pulse);
             }
         } else {
             if (curHandle->latency != handle->latency) {
-                HW_REG(PI_BSD_DOM2_LAT_REG, u32) = handle->latency;
+                IO_WRITE(PI_BSD_DOM2_LAT_REG, handle->latency);
             }
 
             if (curHandle->pageSize != handle->pageSize) {
-                HW_REG(PI_BSD_DOM2_PGS_REG, u32) = handle->pageSize;
+                IO_WRITE(PI_BSD_DOM2_PGS_REG, handle->pageSize);
             }
 
             if (curHandle->relDuration != handle->relDuration) {
-                HW_REG(PI_BSD_DOM2_RLS_REG, u32) = handle->relDuration;
+                IO_WRITE(PI_BSD_DOM2_RLS_REG, handle->relDuration);
             }
 
             if (curHandle->pulse != handle->pulse) {
-                HW_REG(PI_BSD_DOM2_PWD_REG, u32) = handle->pulse;
+                IO_WRITE(PI_BSD_DOM2_PWD_REG, handle->pulse);
             }
         }
 
@@ -53,6 +53,6 @@ s32 __osEPiRawWriteIo(OSPiHandle* handle, u32 devAddr, u32 data) {
         curHandle->pulse = handle->pulse;
     }
 
-    HW_REG(handle->baseAddress | devAddr, u32) = data;
+    IO_WRITE(handle->baseAddress | devAddr, data);
     return 0;
 }
diff --git a/src/libultra/io/pirawdma.c b/src/libultra/io/pirawdma.c
index d3dda984c..84f809c9c 100644
--- a/src/libultra/io/pirawdma.c
+++ b/src/libultra/io/pirawdma.c
@@ -3,21 +3,20 @@
 s32 __osPiRawStartDma(s32 dir, u32 cartAddr, void* dramAddr, size_t size) {
     s32 status;
 
-    status = HW_REG(PI_STATUS_REG, u32);
-    while (status & (PI_STATUS_BUSY | PI_STATUS_IOBUSY)) {
-        status = HW_REG(PI_STATUS_REG, u32);
+    status = IO_READ(PI_STATUS_REG);
+    while (status & (PI_STATUS_DMA_BUSY | PI_STATUS_IO_BUSY)) {
+        status = IO_READ(PI_STATUS_REG);
     }
 
-    HW_REG(PI_DRAM_ADDR_REG, void*) = (void*)osVirtualToPhysical(dramAddr);
-
-    HW_REG(PI_CART_ADDR_REG, void*) = (void*)((osRomBase | cartAddr) & 0x1FFFFFFF);
+    IO_WRITE(PI_DRAM_ADDR_REG, osVirtualToPhysical(dramAddr));
+    IO_WRITE(PI_CART_ADDR_REG, K1_TO_PHYS(osRomBase | cartAddr));
 
     switch (dir) {
         case OS_READ:
-            HW_REG(PI_WR_LEN_REG, u32) = size - 1;
+            IO_WRITE(PI_WR_LEN_REG, size - 1);
             break;
         case OS_WRITE:
-            HW_REG(PI_RD_LEN_REG, u32) = size - 1;
+            IO_WRITE(PI_RD_LEN_REG, size - 1);
             break;
         default:
             return -1;
diff --git a/src/libultra/io/si.c b/src/libultra/io/si.c
index 789815312..4afee0340 100644
--- a/src/libultra/io/si.c
+++ b/src/libultra/io/si.c
@@ -1,11 +1,10 @@
 #include "global.h"
 
 s32 __osSiDeviceBusy(void) {
-    register u32 status = HW_REG(SI_STATUS_REG, u32);
+    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;
     }
+    return false;
 }
diff --git a/src/libultra/io/sirawdma.c b/src/libultra/io/sirawdma.c
index 00b80b4b0..9bf499cc6 100644
--- a/src/libultra/io/sirawdma.c
+++ b/src/libultra/io/sirawdma.c
@@ -1,20 +1,22 @@
 #include "global.h"
 
+#define PIF_RAM_SIZE (PIF_RAM_END + 1 - PIF_RAM_START)
+
 s32 __osSiRawStartDma(s32 dir, void* addr) {
-    if (HW_REG(SI_STATUS_REG, u32) & (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;
     }
     if (dir == OS_WRITE) {
-        osWritebackDCache(addr, 0x40);
+        osWritebackDCache(addr, PIF_RAM_SIZE);
     }
-    HW_REG(SI_DRAM_ADDR_REG, void*) = (void*)osVirtualToPhysical(addr);
+    IO_WRITE(SI_DRAM_ADDR_REG, osVirtualToPhysical(addr));
     if (dir == OS_READ) {
-        HW_REG(SI_PIF_ADDR_RD64B_REG, void*) = (void*)PIF_RAM_START;
+        IO_WRITE(SI_PIF_ADDR_RD64B_REG, PIF_RAM_START);
     } else {
-        HW_REG(SI_PIF_ADDR_WR64B_REG, void*) = (void*)PIF_RAM_START;
+        IO_WRITE(SI_PIF_ADDR_WR64B_REG, PIF_RAM_START);
     }
     if (dir == OS_READ) {
-        osInvalDCache(addr, 0x40);
+        osInvalDCache(addr, PIF_RAM_SIZE);
     }
     return 0;
 }
diff --git a/src/libultra/io/sirawread.c b/src/libultra/io/sirawread.c
index 61b2d730e..8efa071ef 100644
--- a/src/libultra/io/sirawread.c
+++ b/src/libultra/io/sirawread.c
@@ -4,6 +4,6 @@ s32 __osSiRawReadIo(void* devAddr, u32* dst) {
     if (__osSiDeviceBusy()) {
         return -1;
     }
-    *dst = HW_REG((u32)devAddr, u32);
+    *dst = IO_READ(devAddr);
     return 0;
 }
diff --git a/src/libultra/io/sirawwrite.c b/src/libultra/io/sirawwrite.c
index a1935c2e2..f0e4514d7 100644
--- a/src/libultra/io/sirawwrite.c
+++ b/src/libultra/io/sirawwrite.c
@@ -4,6 +4,6 @@ s32 __osSiRawWriteIo(void* devAddr, u32 val) {
     if (__osSiDeviceBusy()) {
         return -1;
     }
-    HW_REG((u32)devAddr, u32) = val;
+    IO_WRITE(devAddr, val);
     return 0;
 }
diff --git a/src/libultra/io/sp.c b/src/libultra/io/sp.c
index 67ccd8939..1796dc8f6 100644
--- a/src/libultra/io/sp.c
+++ b/src/libultra/io/sp.c
@@ -1,7 +1,7 @@
 #include "global.h"
 
 u32 __osSpDeviceBusy(void) {
-    register u32 status = HW_REG(SP_STATUS_REG, u32);
+    register u32 status = IO_READ(SP_STATUS_REG);
 
     if (status & (SP_STATUS_DMA_BUSY | SP_STATUS_DMA_FULL | SP_STATUS_IO_FULL)) {
         return 1;
diff --git a/src/libultra/io/spgetstat.c b/src/libultra/io/spgetstat.c
index 99a4ad0a1..e1545bdb7 100644
--- a/src/libultra/io/spgetstat.c
+++ b/src/libultra/io/spgetstat.c
@@ -1,5 +1,5 @@
 #include "global.h"
 
 u32 __osSpGetStatus(void) {
-    return HW_REG(SP_STATUS_REG, u32);
+    return IO_READ(SP_STATUS_REG);
 }
diff --git a/src/libultra/io/sprawdma.c b/src/libultra/io/sprawdma.c
index 6a1b01b2a..795282c4a 100644
--- a/src/libultra/io/sprawdma.c
+++ b/src/libultra/io/sprawdma.c
@@ -4,12 +4,12 @@ s32 __osSpRawStartDma(s32 direction, void* devAddr, void* dramAddr, u32 size) {
     if (__osSpDeviceBusy()) {
         return -1;
     }
-    HW_REG(SP_MEM_ADDR_REG, u32) = (u32)devAddr;
-    HW_REG(SP_DRAM_ADDR_REG, u32) = osVirtualToPhysical(dramAddr);
+    IO_WRITE(SP_MEM_ADDR_REG, devAddr);
+    IO_WRITE(SP_DRAM_ADDR_REG, osVirtualToPhysical(dramAddr));
     if (direction == OS_READ) {
-        HW_REG(SP_WR_LEN_REG, u32) = size - 1;
+        IO_WRITE(SP_WR_LEN_REG, size - 1);
     } else {
-        HW_REG(SP_RD_LEN_REG, u32) = size - 1;
+        IO_WRITE(SP_RD_LEN_REG, size - 1);
     }
     return 0;
 }
diff --git a/src/libultra/io/spsetpc.c b/src/libultra/io/spsetpc.c
index 2c3cc32ba..944fc9863 100644
--- a/src/libultra/io/spsetpc.c
+++ b/src/libultra/io/spsetpc.c
@@ -1,13 +1,12 @@
 #include "global.h"
 
 s32 __osSpSetPc(void* pc) {
-    register u32 spStatus = HW_REG(SP_STATUS_REG, u32);
+    register u32 spStatus = IO_READ(SP_STATUS_REG);
 
     if (!(spStatus & SP_STATUS_HALT)) {
         return -1;
-    } else {
-        HW_REG(SP_PC_REG, void*) = pc;
     }
 
+    IO_WRITE(SP_PC_REG, pc);
     return 0;
 }
diff --git a/src/libultra/io/spsetstat.c b/src/libultra/io/spsetstat.c
index b2f97a3ff..3fa7a4918 100644
--- a/src/libultra/io/spsetstat.c
+++ b/src/libultra/io/spsetstat.c
@@ -1,5 +1,5 @@
 #include "global.h"
 
 void __osSpSetStatus(u32 status) {
-    HW_REG(SP_STATUS_REG, u32) = status;
+    IO_WRITE(SP_STATUS_REG, status);
 }
diff --git a/src/libultra/io/sptask.c b/src/libultra/io/sptask.c
index bd9b5fdf2..8b3286a45 100644
--- a/src/libultra/io/sptask.c
+++ b/src/libultra/io/sptask.c
@@ -32,7 +32,7 @@ void osSpTaskLoad(OSTask* intp) {
         intp->t.flags &= ~OS_TASK_YIELDED;
 
         if (tp->t.flags & OS_TASK_LOADABLE) {
-            tp->t.ucode = (u64*)HW_REG((u32)intp->t.yield_data_ptr + OS_YIELD_DATA_SIZE - 4, u32);
+            tp->t.ucode = (u64*)IO_READ((u32)intp->t.yield_data_ptr + OS_YIELD_DATA_SIZE - 4);
         }
     }
     osWritebackDCache(tp, sizeof(OSTask));
diff --git a/src/libultra/io/vi.c b/src/libultra/io/vi.c
index 12905742e..6e4db3f8d 100644
--- a/src/libultra/io/vi.c
+++ b/src/libultra/io/vi.c
@@ -11,8 +11,8 @@ void __osViInit(void) {
 
     __osViNext->retraceCount = 1;
     __osViCurr->retraceCount = 1;
-    __osViNext->buffer = (void*)0x80000000;
-    __osViCurr->buffer = (void*)0x80000000;
+    __osViNext->buffer = (void*)K0BASE;
+    __osViCurr->buffer = (void*)K0BASE;
 
     if (osTvType == OS_TV_PAL) {
         __osViNext->modep = &osViModePalLan1;
@@ -25,10 +25,9 @@ void __osViInit(void) {
     __osViNext->state = 0x20;
     __osViNext->features = __osViNext->modep->comRegs.ctrl;
 
-    while (HW_REG(VI_CURRENT_REG, u32) > 10) {
+    while (IO_READ(VI_CURRENT_REG) > 10) {
         ;
     }
-
-    HW_REG(VI_CONTROL_REG, u32) = 0;
+    IO_WRITE(VI_CONTROL_REG, 0);
     __osViSwapContext();
 }
diff --git a/src/libultra/io/viswapcontext.c b/src/libultra/io/viswapcontext.c
index d82d5cce6..7a46fb7bd 100644
--- a/src/libultra/io/viswapcontext.c
+++ b/src/libultra/io/viswapcontext.c
@@ -13,7 +13,7 @@ void __osViSwapContext(void) {
     field = 0;
     viNext = __osViNext;
     viMode = viNext->modep;
-    field = HW_REG(VI_V_CURRENT_LINE_REG, u32) & 1;
+    field = IO_READ(VI_V_CURRENT_LINE_REG) & 1;
     s2 = osVirtualToPhysical(viNext->buffer);
     origin = (viMode->fldRegs[field].origin) + s2;
     if (viNext->state & 2) {
@@ -43,19 +43,19 @@ void __osViSwapContext(void) {
         viNext->y.scale = (viNext->y.offset << 0x10) & 0x3FF0000;
         origin = osVirtualToPhysical(viNext->buffer);
     }
-    HW_REG(VI_ORIGIN_REG, u32) = origin;
-    HW_REG(VI_WIDTH_REG, u32) = viMode->comRegs.width;
-    HW_REG(VI_BURST_REG, u32) = viMode->comRegs.burst;
-    HW_REG(VI_V_SYNC_REG, u32) = viMode->comRegs.vSync;
-    HW_REG(VI_H_SYNC_REG, u32) = viMode->comRegs.hSync;
-    HW_REG(VI_LEAP_REG, u32) = viMode->comRegs.leap;
-    HW_REG(VI_H_START_REG, u32) = hStart;
-    HW_REG(VI_V_START_REG, u32) = vstart;
-    HW_REG(VI_V_BURST_REG, u32) = viMode->fldRegs[field].vBurst;
-    HW_REG(VI_INTR_REG, u32) = viMode->fldRegs[field].vIntr;
-    HW_REG(VI_X_SCALE_REG, u32) = viNext->x.scale;
-    HW_REG(VI_Y_SCALE_REG, u32) = viNext->y.scale;
-    HW_REG(VI_CONTROL_REG, u32) = viNext->features;
+    IO_WRITE(VI_ORIGIN_REG, origin);
+    IO_WRITE(VI_WIDTH_REG, viMode->comRegs.width);
+    IO_WRITE(VI_BURST_REG, viMode->comRegs.burst);
+    IO_WRITE(VI_V_SYNC_REG, viMode->comRegs.vSync);
+    IO_WRITE(VI_H_SYNC_REG, viMode->comRegs.hSync);
+    IO_WRITE(VI_LEAP_REG, viMode->comRegs.leap);
+    IO_WRITE(VI_H_START_REG, hStart);
+    IO_WRITE(VI_V_START_REG, vstart);
+    IO_WRITE(VI_V_BURST_REG, viMode->fldRegs[field].vBurst);
+    IO_WRITE(VI_INTR_REG, viMode->fldRegs[field].vIntr);
+    IO_WRITE(VI_X_SCALE_REG, viNext->x.scale);
+    IO_WRITE(VI_Y_SCALE_REG, viNext->y.scale);
+    IO_WRITE(VI_CONTROL_REG, viNext->features);
     __osViNext = __osViCurr;
     __osViCurr = viNext;
     *__osViNext = *__osViCurr;
diff --git a/src/libultra/os/exceptasm.s b/src/libultra/os/exceptasm.s
index d6bd3082c..16ff4f579 100644
--- a/src/libultra/os/exceptasm.s
+++ b/src/libultra/os/exceptasm.s
@@ -1,7 +1,6 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 #include "ultra64/rcp.h"
-#include "ultra64/rsp.h"
 #include "ultra64/message.h"
 #include "ultra64/thread.h"
 #include "ultra64/exception.h"
diff --git a/src/libultra/os/getcause.s b/src/libultra/os/getcause.s
index 48fbf3236..a621e9736 100644
--- a/src/libultra/os/getcause.s
+++ b/src/libultra/os/getcause.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noreorder
 
diff --git a/src/libultra/os/getcount.s b/src/libultra/os/getcount.s
index bad47d0bc..242c9da69 100644
--- a/src/libultra/os/getcount.s
+++ b/src/libultra/os/getcount.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noreorder
 
diff --git a/src/libultra/os/getfpccsr.s b/src/libultra/os/getfpccsr.s
index d86734ad0..93d4e51b8 100644
--- a/src/libultra/os/getfpccsr.s
+++ b/src/libultra/os/getfpccsr.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noreorder
 
diff --git a/src/libultra/os/getsr.s b/src/libultra/os/getsr.s
index aed66d40a..dc6901b98 100644
--- a/src/libultra/os/getsr.s
+++ b/src/libultra/os/getsr.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noreorder
 
diff --git a/src/libultra/os/initialize.c b/src/libultra/os/initialize.c
index d3998c0bb..3266d1e4d 100644
--- a/src/libultra/os/initialize.c
+++ b/src/libultra/os/initialize.c
@@ -1,63 +1,62 @@
 #include "global.h"
 
 typedef struct {
-    u32 ins_00; // lui     k0, 0x8000
-    u32 ins_04; // addiu   k0, k0, 0x39E0
-    u32 ins_08; // jr      k0 ; __osException
-    u32 ins_0C; // nop
-} struct_exceptionPreamble;
+    u32 inst1; // lui     $k0, %hi(__osException)
+    u32 inst2; // addiu   $k0, $k0, %lo(__osException)
+    u32 inst3; // jr      $k0
+    u32 inst4; //  nop
+} __osExceptionVector;
 
-void __osExceptionPreamble(void);
+extern __osExceptionVector __osExceptionPreamble;
 
 u64 osClockRate = OS_CLOCK_RATE;
 s32 osViClock = VI_NTSC_CLOCK;
-u32 __osShutdown = 0;
+u32 __osShutdown = false;
 OSHWIntr __OSGlobalIntMask = OS_IM_ALL;
 
-u32 D_800145C0;
+u32 __osFinalrom;
 
 void __createSpeedParam(void) {
     __Dom1SpeedParam.type = DEVICE_TYPE_INIT;
-    __Dom1SpeedParam.latency = HW_REG(PI_BSD_DOM1_LAT_REG, u32);
-    __Dom1SpeedParam.pulse = HW_REG(PI_BSD_DOM1_PWD_REG, u32);
-    __Dom1SpeedParam.pageSize = HW_REG(PI_BSD_DOM1_PGS_REG, u32);
-    __Dom1SpeedParam.relDuration = HW_REG(PI_BSD_DOM1_RLS_REG, u32);
+    __Dom1SpeedParam.latency = IO_READ(PI_BSD_DOM1_LAT_REG);
+    __Dom1SpeedParam.pulse = IO_READ(PI_BSD_DOM1_PWD_REG);
+    __Dom1SpeedParam.pageSize = IO_READ(PI_BSD_DOM1_PGS_REG);
+    __Dom1SpeedParam.relDuration = IO_READ(PI_BSD_DOM1_RLS_REG);
 
     __Dom2SpeedParam.type = DEVICE_TYPE_INIT;
-    __Dom2SpeedParam.latency = HW_REG(PI_BSD_DOM2_LAT_REG, u32);
-    __Dom2SpeedParam.pulse = HW_REG(PI_BSD_DOM2_PWD_REG, u32);
-    __Dom2SpeedParam.pageSize = HW_REG(PI_BSD_DOM2_PGS_REG, u32);
-    __Dom2SpeedParam.relDuration = HW_REG(PI_BSD_DOM2_RLS_REG, u32);
+    __Dom2SpeedParam.latency = IO_READ(PI_BSD_DOM2_LAT_REG);
+    __Dom2SpeedParam.pulse = IO_READ(PI_BSD_DOM2_PWD_REG);
+    __Dom2SpeedParam.pageSize = IO_READ(PI_BSD_DOM2_PGS_REG);
+    __Dom2SpeedParam.relDuration = IO_READ(PI_BSD_DOM2_RLS_REG);
 }
 
 void __osInitialize_common(void) {
-    u32 sp2C;
+    u32 pifdata;
 
-    D_800145C0 = 1;
+    __osFinalrom = true;
     __osSetSR(__osGetSR() | SR_CU1);
     __osSetFpcCsr(FPCSR_FS | FPCSR_EV);
-    __osSetWatchLo(0x4900000);
+    __osSetWatchLo(0x04900000);
 
-    while (__osSiRawReadIo((void*)(PIF_RAM_START + 0x3C), &sp2C)) {
+    while (__osSiRawReadIo((void*)(PIF_RAM_END - 3), &pifdata)) {
         ;
     }
-
-    while (__osSiRawWriteIo((void*)(PIF_RAM_START + 0x3C), sp2C | 8)) {
+    while (__osSiRawWriteIo((void*)(PIF_RAM_END - 3), pifdata | 8)) {
         ;
     }
 
-    *(struct_exceptionPreamble*)UT_VEC = *(struct_exceptionPreamble*)__osExceptionPreamble;  // TLB miss
-    *(struct_exceptionPreamble*)XUT_VEC = *(struct_exceptionPreamble*)__osExceptionPreamble; // XTLB miss
-    *(struct_exceptionPreamble*)ECC_VEC = *(struct_exceptionPreamble*)__osExceptionPreamble; // cache errors
-    *(struct_exceptionPreamble*)E_VEC = *(struct_exceptionPreamble*)__osExceptionPreamble;   // general exceptions
+    *(__osExceptionVector*)UT_VEC = __osExceptionPreamble;  // TLB miss
+    *(__osExceptionVector*)XUT_VEC = __osExceptionPreamble; // XTLB miss
+    *(__osExceptionVector*)ECC_VEC = __osExceptionPreamble; // cache errors
+    *(__osExceptionVector*)E_VEC = __osExceptionPreamble;   // general exceptions
 
-    osWritebackDCache((void*)K0BASE, E_VEC - K0BASE + sizeof(struct_exceptionPreamble));
-    osInvalICache((void*)K0BASE, E_VEC - K0BASE + sizeof(struct_exceptionPreamble));
+    osWritebackDCache((void*)K0BASE, E_VEC - K0BASE + sizeof(__osExceptionVector));
+    osInvalICache((void*)K0BASE, E_VEC - K0BASE + sizeof(__osExceptionVector));
     __createSpeedParam();
     osUnmapTLBAll();
     osMapTLBRdb();
 
-    osClockRate = (u64)((osClockRate * 3ll) / 4ull);
+    osClockRate = osClockRate * 3 / 4;
 
     if (!osResetType) {
         bzero(osAppNMIBuffer, sizeof(osAppNMIBuffer));
@@ -71,16 +70,16 @@ void __osInitialize_common(void) {
         osViClock = VI_NTSC_CLOCK;
     }
 
-    // Wait until there are no RCP interrupts
+    // If PreNMI is pending, loop until reset
     if (__osGetCause() & CAUSE_IP5) {
         while (true) {
             ;
         }
     }
 
-    HW_REG(AI_CONTROL_REG, u32) = 1;
-    HW_REG(AI_DACRATE_REG, u32) = 0x3FFF;
-    HW_REG(AI_BITRATE_REG, u32) = 0xF;
+    IO_WRITE(AI_CONTROL_REG, AI_CONTROL_DMA_ON);
+    IO_WRITE(AI_DACRATE_REG, AI_MAX_DAC_RATE - 1);
+    IO_WRITE(AI_BITRATE_REG, AI_MAX_BIT_RATE - 1);
 }
 
 void __osInitialize_autodetect(void) {
diff --git a/src/libultra/os/interrupt.s b/src/libultra/os/interrupt.s
index 82090742b..829377bb0 100644
--- a/src/libultra/os/interrupt.s
+++ b/src/libultra/os/interrupt.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 #include "ultra64/thread.h"
 
 .set noat
diff --git a/src/libultra/os/invaldcache.s b/src/libultra/os/invaldcache.s
index 41799d318..9464a24da 100644
--- a/src/libultra/os/invaldcache.s
+++ b/src/libultra/os/invaldcache.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noat
 .set noreorder
diff --git a/src/libultra/os/invalicache.s b/src/libultra/os/invalicache.s
index 79d576fba..55f831238 100644
--- a/src/libultra/os/invalicache.s
+++ b/src/libultra/os/invalicache.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noat
 .set noreorder
diff --git a/src/libultra/os/maptlbrdb.s b/src/libultra/os/maptlbrdb.s
index ea8fe4aa4..c3ac20948 100644
--- a/src/libultra/os/maptlbrdb.s
+++ b/src/libultra/os/maptlbrdb.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 #include "ultra64/rdb.h"
 
 .set noreorder
diff --git a/src/libultra/os/probetlb.s b/src/libultra/os/probetlb.s
index ce655faca..7bc7856e1 100644
--- a/src/libultra/os/probetlb.s
+++ b/src/libultra/os/probetlb.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noat
 .set noreorder
diff --git a/src/libultra/os/setcompare.s b/src/libultra/os/setcompare.s
index 4b69b2ea9..bd1653377 100644
--- a/src/libultra/os/setcompare.s
+++ b/src/libultra/os/setcompare.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noreorder
 
diff --git a/src/libultra/os/setfpccsr.s b/src/libultra/os/setfpccsr.s
index 9b6b368f3..1aa0e7d29 100644
--- a/src/libultra/os/setfpccsr.s
+++ b/src/libultra/os/setfpccsr.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noreorder
 
diff --git a/src/libultra/os/setintmask.s b/src/libultra/os/setintmask.s
index 3afadb73a..3d2bc5b9c 100644
--- a/src/libultra/os/setintmask.s
+++ b/src/libultra/os/setintmask.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 #include "ultra64/rcp.h"
 #include "ultra64/exception.h"
 
diff --git a/src/libultra/os/setsr.s b/src/libultra/os/setsr.s
index 382f2eada..b754359ea 100644
--- a/src/libultra/os/setsr.s
+++ b/src/libultra/os/setsr.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noreorder
 
diff --git a/src/libultra/os/setwatchlo.s b/src/libultra/os/setwatchlo.s
index 749b1dbd8..aee3ad4d6 100644
--- a/src/libultra/os/setwatchlo.s
+++ b/src/libultra/os/setwatchlo.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noreorder
 
diff --git a/src/libultra/os/unmaptlball.s b/src/libultra/os/unmaptlball.s
index d7fe7cf53..e1a03b1a2 100644
--- a/src/libultra/os/unmaptlball.s
+++ b/src/libultra/os/unmaptlball.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noreorder
 
diff --git a/src/libultra/os/writebackdcache.s b/src/libultra/os/writebackdcache.s
index 4eea0d83a..829f6f4a0 100644
--- a/src/libultra/os/writebackdcache.s
+++ b/src/libultra/os/writebackdcache.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noat
 .set noreorder
diff --git a/src/libultra/os/writebackdcacheall.s b/src/libultra/os/writebackdcacheall.s
index 837495c76..54a58dff8 100644
--- a/src/libultra/os/writebackdcacheall.s
+++ b/src/libultra/os/writebackdcacheall.s
@@ -1,5 +1,5 @@
 #include "ultra64/asm.h"
-#include "ultra64/r4300.h"
+#include "ultra64/R4300.h"
 
 .set noat
 .set noreorder