[ntsc-1.2] Match N64 libultra (#2081)

* Match N64 libultra

* INITIALIZE_FUNC -> OSINITIALIZE_FUNC

* Document osGetIntMask

* Refer readers of osGetIntMask to osSetIntMask comment

* Whitelist new static functions in disasm unksyms check

Makefile

@@ -451,6 +451,9 @@ ifeq ($(PLATFORM),N64)
 $(BUILD_DIR)/src/code/z_rumble.o: CFLAGS += -DNO_SQRTF_INTRINSIC
 endif
 
+$(BUILD_DIR)/src/code/jpegutils.o: CC := $(CC_OLD)
+$(BUILD_DIR)/src/code/jpegdecoder.o: CC := $(CC_OLD)
+
 ifeq ($(DEBUG),1)
 $(BUILD_DIR)/src/libc/%.o: OPTFLAGS := -g
 else
@@ -470,28 +473,57 @@ endif
 $(BUILD_DIR)/src/audio/sfx.o: CFLAGS += -use_readwrite_const
 $(BUILD_DIR)/src/audio/sequence.o: CFLAGS += -use_readwrite_const
 
+$(BUILD_DIR)/src/libultra/%.o: CC := $(CC_OLD)
+
 $(BUILD_DIR)/src/libultra/libc/ll.o: OPTFLAGS := -O1
 $(BUILD_DIR)/src/libultra/libc/ll.o: MIPS_VERSION := -mips3 -32
 $(BUILD_DIR)/src/libultra/libc/llcvt.o: OPTFLAGS := -O1
 $(BUILD_DIR)/src/libultra/libc/llcvt.o: MIPS_VERSION := -mips3 -32
 
+ifeq ($(PLATFORM),N64)
+$(BUILD_DIR)/src/libultra/gu/%.o: OPTFLAGS := -O3
+$(BUILD_DIR)/src/libultra/io/%.o: OPTFLAGS := -O1
+$(BUILD_DIR)/src/libultra/libc/%.o: OPTFLAGS := -O3
 $(BUILD_DIR)/src/libultra/os/%.o: OPTFLAGS := -O1
+
+$(BUILD_DIR)/src/libultra/io/aisetfreq.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/cartrominit.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/contpfs.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/contramread.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/contramwrite.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/contreaddata.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/crc.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/devmgr.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/epiread.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/epiwrite.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/epirawdma.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/epirawread.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/epirawwrite.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/motor.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/pfsgetstatus.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/pfsselectbank.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/pimgr.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/pirawdma.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/sirawdma.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/sirawread.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/sirawwrite.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/sprawdma.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/vimgr.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/visetspecial.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/io/viswapcontext.o: OPTFLAGS := -O2
+
+$(BUILD_DIR)/src/libultra/gu/lookat.o: CFLAGS += -DNO_SQRTF_INTRINSIC
+$(BUILD_DIR)/src/libultra/gu/lookathil.o: CFLAGS += -DNO_SQRTF_INTRINSIC
+$(BUILD_DIR)/src/libultra/gu/normalize.o: CFLAGS += -DNO_SQRTF_INTRINSIC
+else
+$(BUILD_DIR)/src/libultra/gu/%.o: OPTFLAGS := -O2
 $(BUILD_DIR)/src/libultra/io/%.o: OPTFLAGS := -O2
 $(BUILD_DIR)/src/libultra/libc/%.o: OPTFLAGS := -O2
-$(BUILD_DIR)/src/libultra/rmon/%.o: OPTFLAGS := -O2
-$(BUILD_DIR)/src/libultra/gu/%.o: OPTFLAGS := -O2
+$(BUILD_DIR)/src/libultra/os/%.o: OPTFLAGS := -O1
+endif
 
 $(BUILD_DIR)/assets/misc/z_select_static/%.o: GBI_DEFINES := -DF3DEX_GBI
 
-$(BUILD_DIR)/src/libultra/gu/%.o: CC := $(CC_OLD)
-$(BUILD_DIR)/src/libultra/io/%.o: CC := $(CC_OLD)
-$(BUILD_DIR)/src/libultra/libc/%.o: CC := $(CC_OLD)
-$(BUILD_DIR)/src/libultra/os/%.o: CC := $(CC_OLD)
-$(BUILD_DIR)/src/libultra/rmon/%.o: CC := $(CC_OLD)
-
-$(BUILD_DIR)/src/code/jpegutils.o: CC := $(CC_OLD)
-$(BUILD_DIR)/src/code/jpegdecoder.o: CC := $(CC_OLD)
-
 # For using asm_processor on some files:
 #$(BUILD_DIR)/.../%.o: CC := $(PYTHON) tools/asm_processor/build.py $(CC) -- $(AS) $(ASFLAGS) --
 

include/functions.h

@@ -1677,7 +1677,7 @@ s32 __osCheckPackId(OSPfs* pfs, __OSPackId* check);
 s32 __osGetId(OSPfs* pfs);
 s32 __osCheckId(OSPfs* pfs);
 s32 __osPfsRWInode(OSPfs* pfs, __OSInode* inode, u8 flag, u8 bank);
-void guMtxL2F(MtxF* m1, Mtx* m2);
+void guMtxL2F(f32 mf[4][4], Mtx* m);
 s32 osPfsFindFile(OSPfs* pfs, u16 companyCode, u32 gameCode, u8* gameName, u8* extName, s32* fileNo);
 s32 osAfterPreNMI(void);
 s32 osContStartQuery(OSMesgQueue* mq);
@@ -1728,7 +1728,7 @@ u32 __osSpGetStatus(void);
 void __osSpSetStatus(u32 status);
 void osWritebackDCacheAll(void);
 OSThread* __osGetCurrFaultedThread(void);
-void guMtxF2L(MtxF* m1, Mtx* m2);
+void guMtxF2L(f32 mf[4][4], Mtx* m);
 // ? __d_to_ll(?);
 // ? __f_to_ll(?);
 // ? __d_to_ull(?);

include/ultra64/gu.h

@@ -5,6 +5,9 @@
 
 #define GU_PI 3.1415926
 
+#define FTOFIX32(x) (s32)((x) * (f32)0x00010000)
+#define FIX32TOF(x) ((f32)(x) * (1.0f / (f32)0x00010000))
+
 #define ROUND(x) (s32)(((x) >= 0.0) ? ((x) + 0.5) : ((x) - 0.5))
 
 typedef union du {

spec

@@ -589,7 +589,14 @@ beginseg
 #if OOT_DEBUG
     include "$(BUILD_DIR)/src/libultra/io/pfsfreeblocks.o"
 #endif
+#if PLATFORM_N64
+    include "$(BUILD_DIR)/src/libultra/os/getintmask.o"
+#endif
+#if PLATFORM_N64
+    include "$(BUILD_DIR)/src/libultra/gu/scale.o"
+#else
     include "$(BUILD_DIR)/src/libultra/mgu/scale.o"
+#endif
     include "$(BUILD_DIR)/src/libultra/gu/sinf.o"
     include "$(BUILD_DIR)/src/libultra/gu/sins.o"
     include "$(BUILD_DIR)/src/libultra/io/sptask.o"
@@ -605,7 +612,12 @@ beginseg
     include "$(BUILD_DIR)/src/libultra/io/pfsreadwritefile.o"
     include "$(BUILD_DIR)/src/libultra/io/pfsgetstatus.o"
 #endif
+#if PLATFORM_N64
+    include "$(BUILD_DIR)/src/libultra/gu/mtxutil.o"
+#endif
+#if !PLATFORM_N64
     include "$(BUILD_DIR)/src/libultra/mgu/mtxidentf.o"
+#endif
     include "$(BUILD_DIR)/src/libultra/gu/lookat.o"
 #if OOT_DEBUG
     include "$(BUILD_DIR)/src/libultra/io/pfsallocatefile.o"
@@ -613,7 +625,9 @@ beginseg
     include "$(BUILD_DIR)/src/libultra/os/stoptimer.o"
 #if OOT_DEBUG
     include "$(BUILD_DIR)/src/libultra/io/contpfs.o"
+#if !PLATFORM_N64
     include "$(BUILD_DIR)/src/libultra/mgu/mtxl2f.o"
+#endif
     include "$(BUILD_DIR)/src/libultra/io/pfsfindfile.o"
 #endif
     include "$(BUILD_DIR)/src/libultra/gu/sqrtf.o"
@@ -627,13 +641,19 @@ beginseg
     include "$(BUILD_DIR)/src/libultra/libc/string.o"
 #endif
     include "$(BUILD_DIR)/src/libultra/io/sp.o"
+#if !PLATFORM_N64
     include "$(BUILD_DIR)/src/libultra/mgu/mtxident.o"
+#endif
     include "$(BUILD_DIR)/src/libultra/gu/position.o"
     include "$(BUILD_DIR)/src/libultra/io/sptaskyielded.o"
     include "$(BUILD_DIR)/src/libultra/gu/rotate.o"
     include "$(BUILD_DIR)/src/libultra/io/aisetfreq.o"
     include "$(BUILD_DIR)/src/libultra/os/getactivequeue.o"
+#if PLATFORM_N64
+    include "$(BUILD_DIR)/src/libultra/gu/normalize.o"
+#else
     include "$(BUILD_DIR)/src/libultra/mgu/normalize.o"
+#endif
     include "$(BUILD_DIR)/src/libultra/io/dpgetstat.o"
     include "$(BUILD_DIR)/src/libultra/io/dpsetstat.o"
 #if OOT_DEBUG
@@ -643,6 +663,9 @@ beginseg
     include "$(BUILD_DIR)/src/libultra/gu/cosf.o"
     include "$(BUILD_DIR)/src/libultra/gu/libm_vals.o"
     include "$(BUILD_DIR)/src/libultra/gu/coss.o"
+#if PLATFORM_N64
+    include "$(BUILD_DIR)/src/libultra/os/settime.o"
+#endif
     include "$(BUILD_DIR)/src/libultra/io/visetevent.o"
 #if OOT_DEBUG
     include "$(BUILD_DIR)/src/libultra/io/pfsisplug.o"
@@ -656,7 +679,11 @@ beginseg
     include "$(BUILD_DIR)/src/libultra/io/pfschecker.o"
 #endif
     include "$(BUILD_DIR)/src/libultra/io/aigetlen.o"
+#if PLATFORM_N64
+    include "$(BUILD_DIR)/src/libultra/gu/translate.o"
+#else
     include "$(BUILD_DIR)/src/libultra/mgu/translate.o"
+#endif
     include "$(BUILD_DIR)/src/libultra/io/contramwrite.o"
 #if !OOT_DEBUG
     include "$(BUILD_DIR)/src/libultra/io/vimodefpallan1.o"
@@ -678,7 +705,9 @@ beginseg
     include "$(BUILD_DIR)/src/libultra/io/spsetstat.o"
     include "$(BUILD_DIR)/src/libultra/os/writebackdcacheall.o"
     include "$(BUILD_DIR)/src/libultra/os/getcurrfaultedthread.o"
+#if !PLATFORM_N64
     include "$(BUILD_DIR)/src/libultra/mgu/mtxf2l.o"
+#endif
     include "$(BUILD_DIR)/src/libultra/libc/llcvt.o"
     include "$(BUILD_DIR)/src/libultra/io/vigetcurrframebuf.o"
     include "$(BUILD_DIR)/src/libultra/io/spsetpc.o"

src/code/mtxuty-cvt.c

@@ -20,5 +20,5 @@ void MtxConv_F2L(Mtx* m1, MtxF* m2) {
 void MtxConv_L2F(MtxF* m1, Mtx* m2) {
     LOG_UTILS_CHECK_NULL_POINTER("m1", m1, "../mtxuty-cvt.c", 55);
     LOG_UTILS_CHECK_NULL_POINTER("m2", m2, "../mtxuty-cvt.c", 56);
-    guMtxL2F(m1, m2);
+    guMtxL2F(m1->mf, m2);
 }

src/code/sys_matrix.c

@@ -1020,7 +1020,7 @@ void Matrix_SetTranslateUniformScaleMtx(Mtx* mtx, f32 scale, f32 translateX, f32
     MtxF mf;
 
     Matrix_SetTranslateUniformScaleMtxF(&mf, scale, translateX, translateY, translateZ);
-    guMtxF2L(&mf, mtx);
+    guMtxF2L(mf.mf, mtx);
 }
 
 void Matrix_SetTranslateUniformScaleMtx2(Mtx* mtx, f32 scale, f32 translateX, f32 translateY, f32 translateZ) {

src/libultra/gu/lookat.c

@@ -61,5 +61,5 @@ void guLookAt(Mtx* m, f32 xEye, f32 yEye, f32 zEye, f32 xAt, f32 yAt, f32 zAt, f
 
     guLookAtF(mf, xEye, yEye, zEye, xAt, yAt, zAt, xUp, yUp, zUp);
 
-    guMtxF2L((MtxF*)mf, m);
+    guMtxF2L(mf, m);
 }

src/libultra/gu/lookathil.c

@@ -157,5 +157,5 @@ void guLookAtHilite(Mtx* m, LookAt* l, Hilite* h, f32 xEye, f32 yEye, f32 zEye,
     guLookAtHiliteF(mf, l, h, xEye, yEye, zEye, xAt, yAt, zAt, xUp, yUp, zUp, xl1, yl1, zl1, xl2, yl2, zl2, hiliteWidth,
                     hiliteHeight);
 
-    guMtxF2L((MtxF*)mf, m);
+    guMtxF2L(mf, m);
 }

src/libultra/gu/mtxutil.c

@@ -1 +1,63 @@
+#include "global.h"
 
+void guMtxF2L(f32 mf[4][4], Mtx* m) {
+    s32 i, j;
+    s32 e1, e2;
+    s32 *ai, *af;
+
+    ai = (s32*)&m->m[0][0];
+    af = (s32*)&m->m[2][0];
+
+    for (i = 0; i < 4; i++) {
+        for (j = 0; j < 2; j++) {
+            e1 = FTOFIX32(mf[i][j * 2]);
+            e2 = FTOFIX32(mf[i][j * 2 + 1]);
+            *(ai++) = (e1 & 0xFFFF0000) | ((e2 >> 16) & 0xFFFF);
+            *(af++) = ((e1 << 16) & 0xFFFF0000) | (e2 & 0xFFFF);
+        }
+    }
+}
+
+void guMtxL2F(f32 mf[4][4], Mtx* m) {
+    s32 i, j;
+    u32 e1, e2;
+    u32 *ai, *af;
+    s32 q1, q2;
+
+    ai = (u32*)&m->m[0][0];
+    af = (u32*)&m->m[2][0];
+
+    for (i = 0; i < 4; i++) {
+        for (j = 0; j < 2; j++) {
+            e1 = (*ai & 0xFFFF0000) | ((*af >> 16) & 0xFFFF);
+            e2 = ((*(ai++) << 16) & 0xFFFF0000) | (*(af++) & 0xFFFF);
+            q1 = *((s32*)&e1);
+            q2 = *((s32*)&e2);
+
+            mf[i][j * 2] = FIX32TOF(q1);
+            mf[i][j * 2 + 1] = FIX32TOF(q2);
+        }
+    }
+}
+
+void guMtxIdentF(f32 mf[4][4]) {
+    s32 i, j;
+
+    for (i = 0; i < 4; i++) {
+        for (j = 0; j < 4; j++) {
+            if (i == j) {
+                mf[i][j] = 1.0;
+            } else {
+                mf[i][j] = 0.0;
+            }
+        }
+    }
+}
+
+void guMtxIdent(Mtx* m) {
+    f32 mf[4][4];
+
+    guMtxIdentF(mf);
+
+    guMtxF2L(mf, m);
+}

src/libultra/gu/normalize.c

@@ -0,0 +1,9 @@
+#include "global.h"
+
+void guNormalize(f32* x, f32* y, f32* z) {
+    f32 m = 1 / sqrtf(SQ(*x) + SQ(*y) + SQ(*z));
+
+    *x *= m;
+    *y *= m;
+    *z *= m;
+}

src/libultra/gu/ortho.c

@@ -25,5 +25,5 @@ void guOrtho(Mtx* mtx, f32 left, f32 right, f32 bottom, f32 top, f32 near, f32 f
 
     guOrthoF(mf, left, right, bottom, top, near, far, scale);
 
-    guMtxF2L((MtxF*)mf, mtx);
+    guMtxF2L(mf, mtx);
 }

src/libultra/gu/perspective.c

@@ -38,5 +38,6 @@ void guPerspective(Mtx* m, u16* perspNorm, f32 fovy, f32 aspect, f32 near, f32 f
     f32 mf[4][4];
 
     guPerspectiveF(mf, perspNorm, fovy, aspect, near, far, scale);
-    guMtxF2L((MtxF*)mf, m);
+
+    guMtxF2L(mf, m);
 }

src/libultra/gu/position.c

@@ -50,5 +50,5 @@ void guPosition(Mtx* m, f32 rot, f32 pitch, f32 yaw, f32 scale, f32 x, f32 y, f3
 
     guPositionF(mf, rot, pitch, yaw, scale, x, y, z);
 
-    guMtxF2L((MtxF*)mf, m);
+    guMtxF2L(mf, m);
 }

src/libultra/gu/rotate.c

@@ -8,9 +8,11 @@ void guRotateF(f32 m[4][4], f32 a, f32 x, f32 y, f32 z) {
     f32 bc;
     f32 ca;
     f32 t;
+#if !PLATFORM_N64
     f32 xs;
     f32 ys;
     f32 zs;
+#endif
 
     guNormalize(&x, &y, &z);
 
@@ -25,9 +27,15 @@ void guRotateF(f32 m[4][4], f32 a, f32 x, f32 y, f32 z) {
 
     guMtxIdentF(m);
 
+#if PLATFORM_N64
+#define xs (x * sine)
+#define ys (y * sine)
+#define zs (z * sine)
+#else
     xs = x * sine;
     ys = y * sine;
     zs = z * sine;
+#endif
 
     t = x * x;
     m[0][0] = (1 - t) * cosine + t;
@@ -47,5 +55,6 @@ void guRotate(Mtx* m, f32 a, f32 x, f32 y, f32 z) {
     f32 mf[4][4];
 
     guRotateF(mf, a, x, y, z);
-    guMtxF2L((MtxF*)mf, m);
+
+    guMtxF2L(mf, m);
 }

src/libultra/gu/scale.c

@@ -0,0 +1,18 @@
+#include "global.h"
+
+void guScaleF(f32 m[4][4], f32 x, f32 y, f32 z) {
+    guMtxIdentF(m);
+
+    m[0][0] = x;
+    m[1][1] = y;
+    m[2][2] = z;
+    m[3][3] = 1;
+}
+
+void guScale(Mtx* m, f32 x, f32 y, f32 z) {
+    f32 mf[4][4];
+
+    guScaleF(mf, x, y, z);
+
+    guMtxF2L(mf, m);
+}

src/libultra/gu/translate.c

@@ -0,0 +1,17 @@
+#include "global.h"
+
+void guTranslateF(f32 m[4][4], f32 x, f32 y, f32 z) {
+    guMtxIdentF(m);
+
+    m[3][0] = x;
+    m[3][1] = y;
+    m[3][2] = z;
+}
+
+void guTranslate(Mtx* m, f32 x, f32 y, f32 z) {
+    f32 mf[4][4];
+
+    guTranslateF(mf, x, y, z);
+
+    guMtxF2L(mf, m);
+}

src/libultra/io/contsetch.c

@@ -7,6 +7,8 @@
  * direct SI devices.
  */
 s32 osContSetCh(u8 ch) {
+    s32 ret = 0;
+
     __osSiGetAccess();
 
     if (ch > MAXCONTROLLERS) {
@@ -17,5 +19,5 @@ s32 osContSetCh(u8 ch) {
 
     __osContLastCmd = CONT_CMD_END;
     __osSiRelAccess();
-    return 0;
+    return ret;
 }

src/libultra/io/epidma.c

@@ -2,7 +2,7 @@
 #include "ultra64/internal.h"
 
 s32 osEPiStartDma(OSPiHandle* handle, OSIoMesg* mb, s32 direction) {
-    s32 ret;
+    register s32 ret;
 
     if (!__osPiDevMgr.active) {
         return -1;

src/libultra/io/sptaskyielded.c

@@ -1,8 +1,8 @@
 #include "global.h"
 
 u32 osSpTaskYielded(OSTask* task) {
-    u32 ret;
     u32 status = __osSpGetStatus();
+    u32 ret;
 
     if (status & SP_STATUS_YIELDED) {
         ret = OS_TASK_YIELDED;

src/libultra/libc/xldtob.c

@@ -4,8 +4,8 @@
 
 #define BUFF_LEN 0x20
 
-short _Ldunscale(short*, _Pft*);
-void _Genld(_Pft*, char, char*, short, short);
+static short _Ldunscale(short*, _Pft*);
+static void _Genld(_Pft*, char, char*, short, short);
 
 static const double pows[] = { 10e0L, 10e1L, 10e3L, 10e7L, 10e15L, 10e31L, 10e63L, 10e127L, 10e255L };
 
@@ -156,7 +156,7 @@ void _Ldtob(_Pft* args, char code) {
     _Genld((_Pft*)args, code, ptr, nsig, exp);
 }
 
-short _Ldunscale(short* pex, _Pft* px) {
+static short _Ldunscale(short* pex, _Pft* px) {
     unsigned short* ps = (unsigned short*)px;
     short xchar = (ps[_D0] & _DMASK) >> _DOFF;
 
@@ -176,7 +176,7 @@ short _Ldunscale(short* pex, _Pft* px) {
     }
 }
 
-void _Genld(_Pft* px, char code, char* p, short nsig, short xexp) {
+static void _Genld(_Pft* px, char code, char* p, short nsig, short xexp) {
     const char point = '.';
 
     if (nsig <= 0) {