Move OAM RAM to stubs also

Makes accesses more efficient for MIPS. Make accesses also fast for palette
reads.

7 files changed, 49 insertions, 26 deletions

diff --git a/Makefile b/Makefile
index a5f2649..dfc7c37 100644
--- a/Makefile
+++ b/Makefile
@@ -362,6 +362,15 @@ else ifneq (,$(findstring armv,$(platform)))
 	endif
 	LDFLAGS := -Wl,--no-undefined	
 
+# MIPS
+else ifeq ($(platform), mips32)
+	TARGET := $(TARGET_NAME)_libretro.so
+	SHARED := -shared -nostdlib -Wl,--version-script=link.T
+	fpic := -fPIC -DPIC
+	CFLAGS += -fomit-frame-pointer -ffast-math -march=mips32 -mtune=mips32r2 -mhard-float
+	HAVE_DYNAREC := 1
+	CPU_ARCH := mips
+
 # emscripten
 else ifeq ($(platform), emscripten)
 	TARGET := $(TARGET_NAME)_libretro_$(platform).bc
diff --git a/arm/arm_stub.S b/arm/arm_stub.S
index 8e6cc9b..374daba 100644
--- a/arm/arm_stub.S
+++ b/arm/arm_stub.S
@@ -820,6 +820,8 @@ defsymbl(palette_ram)
   .space 0x400
 defsymbl(palette_ram_converted)
   .space 0x400
+defsymbl(oam_ram)
+  .space 0x400
 defsymbl(spsr)
   .space 24
 defsymbl(reg_mode)
diff --git a/cpu.c b/cpu.c
index ea0d69e..badb9c2 100644
--- a/cpu.c
+++ b/cpu.c
@@ -1630,6 +1630,7 @@ void raise_interrupt(irq_type irq_raised)
 
 #ifndef HAVE_DYNAREC
 u8 *memory_map_read [8 * 1024];
+u16 oam_ram[512];
 u16 palette_ram[512];
 u16 palette_ram_converted[512];
 #endif
diff --git a/gba_memory.c b/gba_memory.c
index e647304..b66dce7 100644
--- a/gba_memory.c
+++ b/gba_memory.c
@@ -305,7 +305,6 @@ u32 gamepak_waitstate_sequential[2][3][3] =
   }
 };
 
-u16 oam_ram[512];
 u16 io_registers[1024 * 16];
 u8 ewram[1024 * 256 * 2];
 u8 iwram[1024 * 32 * 2];
diff --git a/psp/mips_emit.h b/psp/mips_emit.h
index d24c174..818b724 100644
--- a/psp/mips_emit.h
+++ b/psp/mips_emit.h
@@ -2630,6 +2630,7 @@ typedef struct {
   bool check_smc;       // Whether the memory can contain code
   bool bus16;           // Whether it can only be accessed at 16bit
   u32 baseptr;          // Memory base address.
+  u32 baseoff;          // Offset from base_reg
 } t_stub_meminfo;
 
 // Generates the stub to access memory for a given region, access type,
@@ -2738,7 +2739,11 @@ static void emit_pmemld_stub(
   } else {
     // Generate upper bits of the addr and do addr mirroring
     // (The address hi16 is rounded up since load uses signed offset)
-    mips_emit_lui(reg_rv, ((base_addr + 0x8000) >> 16));
+    if (!meminfo->baseoff) {
+      mips_emit_lui(reg_rv, ((base_addr + 0x8000) >> 16));
+    } else {
+      base_addr = meminfo->baseoff;
+    }
 
     if (region == 2) {
       // Can't do EWRAM with an `andi` instruction (18 bits mask)
@@ -2761,8 +2766,9 @@ static void emit_pmemld_stub(
       mips_emit_addu(reg_rv, reg_rv, reg_a0);    // addr = base + adjusted offset
     } else {
       // Generate regular (<=32KB) mirroring
-      mips_emit_andi(reg_a0, reg_a0, memmask);   // Clear upper bits (mirroring)
-      mips_emit_addu(reg_rv, reg_rv, reg_a0);    // Adds to base addr
+      mips_reg_number breg = (meminfo->baseoff ? reg_base : reg_rv);
+      mips_emit_andi(reg_temp, reg_a0, memmask); // Clear upper bits (mirroring)
+      mips_emit_addu(reg_rv, breg, reg_temp);    // Adds to base addr
     }
   }
 
@@ -3154,7 +3160,7 @@ static void emit_phand(
     mips_emit_ins(reg_temp, reg_a0, 6, size); // Alignment bits (1 or 2, to bits 6 (and 7)
   }
 
-  unsigned tbloff = 256 + 2048 + 220 + 4 * toff;  // Skip regs and palettes
+  unsigned tbloff = 256 + 3*1024 + 220 + 4 * toff;  // Skip regs and RAMs
   mips_emit_addu(reg_rv, reg_temp, reg_base); // Add to the base_reg the table offset
   mips_emit_lw(reg_rv, reg_rv, tbloff);       // Read addr from table
   mips_emit_sll(reg_temp, reg_rv, 4);         // 26 bit immediate to the MSB
@@ -3229,21 +3235,21 @@ void init_emitter() {
 
   // Generate memory handlers
   const t_stub_meminfo ldinfo [] = {
-    { emit_pmemld_stub,  0, 0x4000, false, false, (u32)bios_rom },
+    { emit_pmemld_stub,  0, 0x4000, false, false, (u32)bios_rom, 0},
     // 1 Open load / Ignore store
-    { emit_pmemld_stub,  2, 0x8000, true,  false, (u32)ewram },      // memsize wrong on purpose
-    { emit_pmemld_stub,  3, 0x8000, true,  false, (u32)&iwram[0x8000] },
-    { emit_pmemld_stub,  4,  0x400, false, false, (u32)io_registers },
-    { emit_pmemld_stub,  5,  0x400, false, true,  (u32)palette_ram },
-    { emit_pmemld_stub,  6,    0x0, false, true,  (u32)vram },             // same, vram is a special case
-    { emit_pmemld_stub,  7,  0x400, false, true,  (u32)oam_ram },
-    { emit_pmemld_stub,  8, 0x8000, false, false,  0 },
-    { emit_pmemld_stub,  9, 0x8000, false, false,  0 },
-    { emit_pmemld_stub, 10, 0x8000, false, false,  0 },
-    { emit_pmemld_stub, 11, 0x8000, false, false,  0 },
-    { emit_pmemld_stub, 12, 0x8000, false, false,  0 },
+    { emit_pmemld_stub,  2, 0x8000, true,  false, (u32)ewram, 0 },      // memsize wrong on purpose
+    { emit_pmemld_stub,  3, 0x8000, true,  false, (u32)&iwram[0x8000], 0 },
+    { emit_pmemld_stub,  4,  0x400, false, false, (u32)io_registers, 0 },
+    { emit_pmemld_stub,  5,  0x400, false, true,  (u32)palette_ram, 0x100 },
+    { emit_pmemld_stub,  6,    0x0, false, true,  (u32)vram, 0 },             // same, vram is a special case
+    { emit_pmemld_stub,  7,  0x400, false, true,  (u32)oam_ram, 0x900 },
+    { emit_pmemld_stub,  8, 0x8000, false, false,  0, 0 },
+    { emit_pmemld_stub,  9, 0x8000, false, false,  0, 0 },
+    { emit_pmemld_stub, 10, 0x8000, false, false,  0, 0 },
+    { emit_pmemld_stub, 11, 0x8000, false, false,  0, 0 },
+    { emit_pmemld_stub, 12, 0x8000, false, false,  0, 0 },
     // 13 is EEPROM mapped already (a bit special)
-    { emit_pmemld_stub, 14,      0, false, false,  0 },                    // Mapped via function call
+    { emit_pmemld_stub, 14,      0, false, false,  0, 0 },                    // Mapped via function call
     // 15 Open load / Ignore store
   };
 
@@ -3267,12 +3273,12 @@ void init_emitter() {
   }
 
   const t_stub_meminfo stinfo [] = {
-    { emit_pmemst_stub, 2, 0x8000, true,  false, (u32)ewram },
-    { emit_pmemst_stub, 3, 0x8000, true,  false, (u32)&iwram[0x8000] },
+    { emit_pmemst_stub, 2, 0x8000, true,  false, (u32)ewram, 0 },
+    { emit_pmemst_stub, 3, 0x8000, true,  false, (u32)&iwram[0x8000], 0 },
     // I/O is special and mapped with a function call
-    { emit_palette_hdl, 5,  0x400, false, true,  (u32)palette_ram },
-    { emit_pmemst_stub, 6,    0x0, false, true,  (u32)vram },             // same, vram is a special case
-    { emit_pmemst_stub, 7,  0x400, false, true,  (u32)oam_ram },
+    { emit_palette_hdl, 5,  0x400, false, true,  (u32)palette_ram, 0x100 },
+    { emit_pmemst_stub, 6,    0x0, false, true,  (u32)vram, 0 },          // same, vram is a special case
+    { emit_pmemst_stub, 7,  0x400, false, true,  (u32)oam_ram, 0x900 },
   };
 
   // Store only for "regular"-ish mem regions
diff --git a/psp/mips_stub.S b/psp/mips_stub.S
index cc3a220..1c4ad4b 100644
--- a/psp/mips_stub.S
+++ b/psp/mips_stub.S
@@ -40,6 +40,7 @@
 .global reg_check
 .global palette_ram
 .global palette_ram_converted
+.global oam_ram
 .global init_emitter
 .global mips_lookup_pc
 .global smc_write
@@ -120,11 +121,12 @@
 .equ OAM_UPDATED,         (33 * 4)
 .equ GP_SAVE,             (34 * 4)
 
-.equ SPSR_BASE,           (0x900)
-.equ REGMODE_BASE,        (0x900 + 24)
+.equ SPSR_BASE,           (0x100 + 0x400 * 3)
+.equ REGMODE_BASE,        (SPSR_BASE + 24)
 .equ SUPERVISOR_SPSR,     (3 * 4 + SPSR_BASE)
 .equ SUPERVISOR_LR,       ((3 * (7 * 4)) + (6 * 4) + REGMODE_BASE)
-.equ FNPTRS_BASE,         (0x900 + 220 + 960)
+.equ FNPTRS_MEMOPS,       (REGMODE_BASE + 196)
+.equ FNPTRS_BASE,         (FNPTRS_MEMOPS + 960)
 
 .set noat
 .set noreorder
@@ -625,6 +627,8 @@ palette_ram:
   .space 0x400
 palette_ram_converted:
   .space 0x400
+oam_ram:
+  .space 0x400
 spsr:
   .space 24     # u32[6]
 reg_mode:
diff --git a/x86/x86_stub.S b/x86/x86_stub.S
index 948572c..333c8fd 100644
--- a/x86/x86_stub.S
+++ b/x86/x86_stub.S
@@ -538,6 +538,8 @@ defsymbl(palette_ram)
   .space 0x400
 defsymbl(palette_ram_converted)
   .space 0x400
+defsymbl(oam_ram)
+  .space 0x400
 defsymbl(spsr)
   .space 24
 defsymbl(reg_mode)