bug fix in comm poll fifo, and back to -O3

diff --git a/Makefile b/Makefile
index c3bd689..a3b2c96 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -3,7 +3,7 @@ DEBUG ?= 0
 CFLAGS += -Wall -ggdb -ffunction-sections -fdata-sections
 CFLAGS += -I.
 ifeq "$(DEBUG)" "0"
-CFLAGS += -O2 -finline-functions -DNDEBUG
+CFLAGS += -O3 -DNDEBUG
 endif
 
 # This is actually needed, bevieve me.

diff --git a/README.md b/README.md
index 13ff159..aa0466d 100644 (file)
--- a/README.md
+++ b/README.md
@@ -29,11 +29,11 @@ assuming $TC points to the appropriate cross compile toolchain directory:
 
 platform|toolchain|configure command
 --------|---------|-----------------
-gp2x,wiz,caanoo|open2x|CROSS_COMPILE=arm-open2x-linux- CFLAGS="-I$TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux/include -fno-stack-protector -fno-common -finline-limit=42" LDFLAGS="--sysroot $TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux -L$TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux/lib" ./configure --platform=gp2x
-gp2x,wiz,caanoo|open2x with ubuntu arm gcc 4.7|CROSS_COMPILE=arm-linux-gnueabi- CFLAGS="-I$TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux/include -fno-stack-protector -fno-common -finline-limit=42 -fipa-pta" LDFLAGS="-B$TC/gcc-4.1.1-glibc-2.3.6/lib/gcc/arm-open2x-linux/4.1.1 -B$TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux/lib -L$TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux/lib" ./configure --platform=gp2x
-opendingux|opendingux|CROSS_COMPILE=mipsel-linux- CFLAGS="-I$TC/usr/include -I$TC/usr/include/SDL -fno-stack-protector -fno-common -finline-limit=42 -fipa-pta" LDFLAGS="--sysroot $TC -L$TC/lib" ./configure --platform=opendingux
-opendingux|opendingux with ubuntu mips gcc 5.4|CROSS_COMPILE=mipsel-linux-gnu- CFLAGS="-I$TC/usr/include -I$TC/usr/include/SDL -fno-stack-protector -fno-common -finline-limit=42 -fipa-pta" LDFLAGS="-B$TC/usr/lib -B$TC/lib -Wl,-rpath-link=$TC/usr/lib -Wl,-rpath-link=$TC/lib" ./configure --platform=opendingux
-gcw0|gcw0|CROSS_COMPILE=mipsel-gcw0-linux-uclibc- CFLAGS="-I$TC/usr/mipsel-gcw0-linux-uclibc/sysroot/usr/include -I$TC/usr/mipsel-gcw0-linux-uclibc/sysroot/usr/include/SDL -fno-stack-protector -fno-common -finline-limit=42 -fipa-pta" LDFLAGS="--sysroot $TC/usr/mipsel-gcw0-linux-uclibc/sysroot" ./configure --platform=gcw0
+gp2x,wiz,caanoo|open2x|CROSS_COMPILE=arm-open2x-linux- CFLAGS="-I$TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux/include -finline-limit=42 -fno-unroll-loops -fno-stack-protector -fno-common" LDFLAGS="--sysroot $TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux -L$TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux/lib" ./configure --platform=gp2x
+gp2x,wiz,caanoo|open2x with ubuntu arm gcc 4.7|CROSS_COMPILE=arm-linux-gnueabi- CFLAGS="-I$TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux/include -finline-limit=42 -fno-unroll-loops -fno-stack-protector -fno-common" LDFLAGS="-B$TC/gcc-4.1.1-glibc-2.3.6/lib/gcc/arm-open2x-linux/4.1.1 -B$TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux/lib -L$TC/gcc-4.1.1-glibc-2.3.6/arm-open2x-linux/lib" ./configure --platform=gp2x
+opendingux|opendingux|CROSS_COMPILE=mipsel-linux- CFLAGS="-I$TC/usr/include -I$TC/usr/include/SDL" LDFLAGS="--sysroot $TC -L$TC/lib" ./configure --platform=opendingux
+opendingux|opendingux with ubuntu mips gcc 5.4|CROSS_COMPILE=mipsel-linux-gnu- CFLAGS="-I$TC/usr/include -I$TC/usr/include/SDL" LDFLAGS="-B$TC/usr/lib -B$TC/lib -Wl,-rpath-link=$TC/usr/lib -Wl,-rpath-link=$TC/lib" ./configure --platform=opendingux
+gcw0|gcw0|CROSS_COMPILE=mipsel-gcw0-linux-uclibc- CFLAGS="-I$TC/usr/mipsel-gcw0-linux-uclibc/sysroot/usr/include -I$TC/usr/mipsel-gcw0-linux-uclibc/sysroot/usr/include/SDL" LDFLAGS="--sysroot $TC/usr/mipsel-gcw0-linux-uclibc/sysroot" ./configure --platform=gcw0
 
 For gp2x, wiz, and caanoo you may need to compile libpng first, and additionally
 cyclone_gp2x.patch may need to be applied to the cpu/cyclone submodule:
@@ -67,7 +67,7 @@ You need to install the resulting binary onto your device manually.
 For opendingux and gcw0, copy the opk to your SD card.
 For gp2x, wiz and caanoo, the easiest way is to unpack
 [PicoDrive_191.zip](http://notaz.gp2x.de/releases/PicoDrive/PicoDrive_191.zip)
-on you SD card and replace the PicoDrive binary.
+on your SD card and replace the PicoDrive binary.
 
 Send bug reports, fixes etc to <derkub@gmail.com>
 Kai-Uwe Bloem

diff --git a/pico/32x/memory.c b/pico/32x/memory.c
index 3e11cbc..7f494e7 100644 (file)
--- a/pico/32x/memory.c
+++ b/pico/32x/memory.c
@@ -191,7 +191,7 @@ static NOINLINE u32 sh2_poll_read(u32 a, u32 d, unsigned int cycles, SH2* sh2)
   int hix = (a >> 1) % PFIFO_CNT;
   struct sh2_poll_fifo *fifo = sh2_poll_fifo[hix];
   struct sh2_poll_fifo *p;
-  int cpu = sh2 ? sh2->is_slave+1 : 0;
+  int cpu = sh2 ? sh2->is_slave : -1;
   unsigned idx;
 
   a &= ~0x20000000; // ignore writethrough bit
@@ -204,7 +204,7 @@ static NOINLINE u32 sh2_poll_read(u32 a, u32 d, unsigned int cycles, SH2* sh2)
     if (cpu != p->cpu) {
       if (CYCLES_GT(cycles, p->cycles+80)) {
         // drop older fifo stores that may cause synchronisation problems.
-        sh2_poll_rd[hix] = idx;
+        p->a = -1;
       } else if (p->a == a) {
         // replace current data with fifo value and discard fifo entry
         d = p->d;
@@ -221,24 +221,37 @@ static NOINLINE void sh2_poll_write(u32 a, u32 d, unsigned int cycles, SH2 *sh2)
   int hix = (a >> 1) % PFIFO_CNT;
   struct sh2_poll_fifo *fifo = sh2_poll_fifo[hix];
   struct sh2_poll_fifo *q = &fifo[(sh2_poll_wr[hix]-1) % PFIFO_SZ];
-  int cpu = sh2 ? sh2->is_slave+1 : 0;
+  int cpu = sh2 ? sh2->is_slave : -1;
+  unsigned rd = sh2_poll_rd[hix], wr = sh2_poll_wr[hix];
+  unsigned idx, nrd;
 
   a &= ~0x20000000; // ignore writethrough bit
+
+  // throw out any values written by other cpus, plus heading cancelled stuff
+  for (idx = nrd = wr; idx != rd; ) {
+    idx = (idx-1) % PFIFO_SZ;
+    if (fifo[idx].a == a && fifo[idx].cpu != cpu)      { fifo[idx].a = -1; }
+    if (fifo[idx].a != -1)                             { nrd = idx; }
+  }
+  rd = nrd;
+
   // fold 2 consecutive writes to the same address to avoid reading of
   // intermediate values that may cause synchronisation problems.
   // NB this can take an eternity on m68k: mov.b <addr1.l>,<addr2.l> needs
   // 28 m68k-cycles (~80 sh2-cycles) to complete (observed in Metal Head)
-  if (q->a == a && sh2_poll_wr[hix] != sh2_poll_rd[hix] && !CYCLES_GT(cycles,q->cycles+30)) {
+  if (q->a == a && rd != wr && !CYCLES_GT(cycles,q->cycles+30)) {
     q->d = d;
   } else {
     // store write to poll address in fifo
-    fifo[sh2_poll_wr[hix]] =
+    fifo[wr] =
         (struct sh2_poll_fifo){ .cycles = cycles, .a = a, .d = d, .cpu = cpu };
-    sh2_poll_wr[hix] = (sh2_poll_wr[hix]+1) % PFIFO_SZ;
-    if (sh2_poll_wr[hix] == sh2_poll_rd[hix])
+    wr = (wr+1) % PFIFO_SZ;
+    if (wr == rd)
       // fifo overflow, discard oldest value
-      sh2_poll_rd[hix] = (sh2_poll_rd[hix]+1) % PFIFO_SZ;
+      rd = (rd+1) % PFIFO_SZ;
   }
+
+  sh2_poll_rd[hix] = rd; sh2_poll_wr[hix] = wr;
 }
 
 u32 REGPARM(3) p32x_sh2_poll_memory8(unsigned int a, u32 d, SH2 *sh2)
@@ -2367,6 +2380,7 @@ void PicoMemSetup32x(void)
   sh2_drc_mem_setup(&ssh2);
   memset(sh2_poll_rd, 0, sizeof(sh2_poll_rd));
   memset(sh2_poll_wr, 0, sizeof(sh2_poll_wr));
+  memset(sh2_poll_fifo, -1, sizeof(sh2_poll_fifo));
 
   // z80 hack
   z80_map_set(z80_write_map, 0x8000, 0xffff, z80_md_bank_write_32x, 1);