[iQue] Match handwritten asm files in libultra/os, match C replacements of osGetCount and __osSetCompare (#2413)

* [iQue] Match handwritten asm files in libultra/os, match C replacements of osGetCount and __osSetCompare

* Format

* Adjust some label names in dcache functions

Makefile

@@ -648,6 +648,9 @@ $(BUILD_DIR)/src/libultra/%.o: CFLAGS := $(EGCS_CFLAGS) -mno-abicalls
 $(BUILD_DIR)/src/libultra/%.o: CCASFLAGS := $(EGCS_CCASFLAGS)
 $(BUILD_DIR)/src/libultra/%.o: ASOPTFLAGS := $(EGCS_ASOPTFLAGS)
 
+$(BUILD_DIR)/src/libultra/reg/_%.o: OPTFLAGS := -O0
+$(BUILD_DIR)/src/libultra/reg/_%.o: MIPS_VERSION := -mgp64 -mfp64 -mips3
+
 $(BUILD_DIR)/src/libultra/libc/ll.o: OPTFLAGS := -O0
 $(BUILD_DIR)/src/libultra/libc/llcvt.o: OPTFLAGS := -O0
 

include/ultra64/asm.h

@@ -78,6 +78,9 @@
 #define MTC0(dst, src) \
     .set noreorder; mtc0 dst, src; .set reorder
 
+#define CACHE(op, base)  \
+    .set noreorder; cache op, base; .set reorder
+
 #define CFC1(dst, src) \
     .set noreorder; cfc1 dst, src; .set reorder
 #define CTC1(src, dst) \

src/libultra/os/invaldcache.s

@@ -20,56 +20,50 @@
  *  the entire data cache is invalidated.
  */
 LEAF(osInvalDCache)
-.set noreorder
     /* If the amount to invalidate is less than or equal to 0, return immediately */
     blez    a1, 3f
-     nop
     /* If the amount to invalidate is as large as or larger than
      * the data cache size, invalidate all */
     li      t3, DCACHE_SIZE
     bgeu    a1, t3, 4f
-     nop
     /* Ensure end address doesn't wrap around and end up smaller
      * than the start address */
     move    t0, a0
     addu    t1, a0, a1
     bgeu    t0, t1, 3f
-     nop
     /* Mask start with cache line */
+    addiu  t1, t1, -DCACHE_LINESIZE
     andi    t2, t0, DCACHE_LINEMASK
     /* If mask is not zero, the start is not cache aligned */
     beqz    t2, 1f
-     addiu  t1, t1, -DCACHE_LINESIZE
     /* Subtract mask result to align to cache line */
     subu    t0, t0, t2
     /* Hit-Writeback-Invalidate unaligned part */
-    cache   (CACH_PD | C_HWBINV), (t0)
+    CACHE(  (CACH_PD | C_HWBINV), (t0))
     /* If that's all there is to do, return early */
     bgeu    t0, t1, 3f
-     nop
     addiu   t0, t0, DCACHE_LINESIZE
 1:
     /* Mask end with cache line */
     andi    t2, t1, DCACHE_LINEMASK
     /* If mask is not zero, the end is not cache aligned */
-    beqz    t2, 1f
-     nop
+    beqz    t2, 2f
     /* Subtract mask result to align to cache line */
     subu    t1, t1, t2
     /* Hit-Writeback-Invalidate unaligned part */
-    cache   (CACH_PD | C_HWBINV), DCACHE_LINESIZE(t1)
+    CACHE(  (CACH_PD | C_HWBINV), DCACHE_LINESIZE(t1))
     /* If that's all there is to do, return early */
     bltu    t1, t0, 3f
-     nop
     /* Invalidate the rest */
-1:
+2:
     /* Hit-Invalidate */
-    cache   (CACH_PD | C_HINV), (t0)
-    bltu    t0, t1, 1b
+    CACHE(  (CACH_PD | C_HINV), (t0))
+.set noreorder
+    bltu    t0, t1, 2b
      addiu  t0, t0, DCACHE_LINESIZE
+.set reorder
 3:
     jr      ra
-     nop
 
 4:
     li      t0, K0BASE
@@ -77,10 +71,10 @@ LEAF(osInvalDCache)
     addiu   t1, t1, -DCACHE_LINESIZE
 5:
     /* Index-Writeback-Invalidate */
-    cache   (CACH_PD | C_IWBINV), (t0)
+    CACHE(  (CACH_PD | C_IWBINV), (t0))
+.set noreorder
     bltu    t0, t1, 5b
      addiu  t0, DCACHE_LINESIZE
-    jr      ra
-     nop
 .set reorder
+    jr      ra
 END(osInvalDCache)

src/libultra/os/invalicache.s

@@ -5,42 +5,40 @@
 .text
 
 LEAF(osInvalICache)
-.set noreorder
     /* If the amount to invalidate is less than or equal to 0, return immediately */
     blez    a1, 2f
-     nop
     /* If the amount to invalidate is as large as or larger than */
     /* the instruction cache size, invalidate all */
     li      t3, ICACHE_SIZE
     bgeu    a1, t3, 3f
-     nop
     /* ensure end address doesn't wrap around and end up smaller */
     /* than the start address */
     move    t0, a0
     addu    t1, a0, a1
     bgeu    t0, t1, 2f
-     nop
     /* Mask and subtract to align to cache line */
-    andi    t2, t0, ICACHE_LINEMASK
     addiu   t1, t1, -ICACHE_LINESIZE
+    andi    t2, t0, ICACHE_LINEMASK
     subu    t0, t0, t2
 1:
-    cache   (CACH_PI | C_HINV), (t0)
+    CACHE(   (CACH_PI | C_HINV), (t0))
+.set noreorder
     bltu    t0, t1, 1b
      addiu  t0, t0, ICACHE_LINESIZE
+.set reorder
 2:
     jr      ra
-     nop
 
 3:
     li      t0, K0BASE
     addu    t1, t0, t3
     addiu   t1, t1, -ICACHE_LINESIZE
 4:
-    cache   (CACH_PI | C_IINV), (t0)
+    CACHE(   (CACH_PI | C_IINV), (t0))
+.set noreorder
     bltu    t0, t1, 4b
      addiu  t0, ICACHE_LINESIZE
+.set reorder
     jr      ra
-     nop
 .set reorder
 END(osInvalICache)

src/libultra/os/setintmask.s

@@ -116,7 +116,7 @@ LEAF(osSetIntMask)
     la      t0, __OSGlobalIntMask
     lw      t3, (t0)
      /* Bitwise-OR in the disabled CPU bits of __OSGlobalIntMask */
-    xor     t0, t3, ~0
+    xor     t0, t3, 0xFFFFFFFF
     and     t0, t0, SR_IMASK
     or      v0, v0, t0
      /* Fetch MI_INTR_MASK_REG */
@@ -125,7 +125,7 @@ LEAF(osSetIntMask)
     beqz    t2, 1f
      srl    t1, t3, RCP_IMASKSHIFT
     /* Bitwise-OR in the disabled RCP bits of __OSGlobalIntMask */
-    xor     t1, t1, ~0
+    xor     t1, t1, 0xFFFFFFFF
     and     t1, t1, (RCP_IMASK >> RCP_IMASKSHIFT)
     or      t2, t2, t1
 1:

src/libultra/os/writebackdcache.s

@@ -13,41 +13,41 @@
  */
 LEAF(osWritebackDCache)
     /* If the amount to write back is less than or equal to 0, return immediately */
-    blez    a1, .ret
+    blez    a1, 2f
 
     /* If the amount to write back is as large as or larger than */
     /* the data cache size, write back all */
     li      t3, DCACHE_SIZE
-    bgeu    a1, t3, .all
+    bgeu    a1, t3, 3f
 
     /* ensure end address doesn't wrap around and end up smaller */
     /* than the start address */
     move    t0, a0
     addu    t1, a0, a1
-    bgeu    t0, t1, .ret
+    bgeu    t0, t1, 2f
 
     /* Mask and subtract to align to cache line */
-    andi    t2, t0, DCACHE_LINEMASK
     addiu   t1, t1, -DCACHE_LINESIZE
+    andi    t2, t0, DCACHE_LINEMASK
     subu    t0, t0, t2
 1:
+    CACHE(  (CACH_PD | C_HWB), (t0))
 .set noreorder
-    cache   (CACH_PD | C_HWB), (t0)
     bltu    t0, t1, 1b
      addiu  t0, t0, DCACHE_LINESIZE
 .set reorder
-.ret:
+2:
     jr      ra
 
 /* same as osWritebackDCacheAll in operation */
-.all:
+3:
     li      t0, K0BASE
     addu    t1, t0, t3
     addiu   t1, t1, -DCACHE_LINESIZE
-1:
+4:
+    CACHE(  (CACH_PD | C_IWBINV), (t0))
 .set noreorder
-    cache   (CACH_PD | C_IWBINV), (t0)
-    bltu    t0, t1, 1b
+    bltu    t0, t1, 4b
      addiu  t0, DCACHE_LINESIZE
 .set reorder
     jr      ra

src/libultra/reg/_getcount.c

@@ -0,0 +1,32 @@
+#ifdef BBPLAYER
+#include "ultra64.h"
+
+u32 __osBbLastRCount;
+u32 __osBbRCountWraps;
+u32 __osBbLastVCount;
+u32 __osBbVCountWraps;
+
+u32 osGetCount(void) {
+    u32 count;
+    u32 mask = __osDisableInt();
+
+    __asm__("mfc0 %0, $%1" : "=r"(count) : "i"(C0_COUNT));
+
+    if (count < __osBbLastRCount) {
+        __osBbRCountWraps++;
+    }
+    __osBbLastRCount = count;
+
+    count = (((u64)__osBbRCountWraps << 32) | count) * 125ull / 192ull;
+
+    if (count < __osBbLastVCount) {
+        __osBbVCountWraps++;
+    }
+    __osBbLastVCount = count;
+
+    __osRestoreInt(mask);
+
+    return count;
+}
+
+#endif

src/libultra/reg/_setcompare.c

@@ -0,0 +1,20 @@
+#ifdef BBPLAYER
+#include "ultra64.h"
+
+extern u32 __osBbLastRCount;
+extern u32 __osBbRCountWraps;
+extern u32 __osBbLastVCount;
+extern u32 __osBbVCountWraps;
+
+void __osSetCompare(u32 v) {
+    if (v != 0) {
+        u32 mask = __osDisableInt();
+        u32 wraps = (v < __osBbLastVCount) ? (__osBbVCountWraps + 1) : __osBbVCountWraps;
+
+        v = (((u64)wraps << 32) | v) * 192ull / 125ull;
+
+        __osRestoreInt(mask);
+    }
+    __asm__("mtc0 %0, $%1" ::"r"(v), "i"(C0_COMPARE));
+}
+#endif

tools/disasm/ique-cn/files_boot.csv

@@ -59,8 +59,8 @@ B50,80000FA0,src/boot/zlib
 6F10,80007360,src/libultra/os/seteventmesg
 6FE0,80007430,src/libultra/os/getcause
 6FF0,80007440,src/libultra/os/setwatchlo
-7000,80007450,src/libultra/os/getcount
-7100,80007550,src/libultra/os/setcompare
+7000,80007450,src/libultra/reg/_getcount
+7100,80007550,src/libultra/reg/_setcompare
 71B0,80007600,src/libultra/io/sirawread
 7200,80007650,src/libultra/io/sirawwrite
 7250,800076A0,src/libultra/os/destroythread