improve ARM feature detection

diff --git a/configure b/configure
index 4cff898..053dba9 100755 (executable)
--- a/configure
+++ b/configure
@@ -231,12 +231,6 @@ if [ "$ARCH" = "arm" ]; then
   if check_define __thumb__; then
     CFLAGS="$CFLAGS -marm"
   fi
-
-  if [ "$have_armv7" = "yes" ]; then
-    ASFLAGS="$ASFLAGS --defsym HAVE_ARMV7=1"
-  else
-    ASFLAGS="$ASFLAGS --defsym HAVE_ARMV7=0"
-  fi
 else
   # dynarec only available on ARM
   enable_dynarec="no"
@@ -439,12 +433,6 @@ echo "PLATFORM = $platform" >> $config_mak
 echo "BUILTIN_GPU = $builtin_gpu" >> $config_mak
 echo "SOUND_DRIVERS = $sound_drivers" >> $config_mak
 echo "PLUGINS = $plugins" >> $config_mak
-if [ "$have_armv6" = "yes" ]; then
-  echo "HAVE_ARMV6 = 1" >> $config_mak
-fi
-if [ "$have_armv7" = "yes" ]; then
-  echo "HAVE_ARMV7 = 1" >> $config_mak
-fi
 if [ "$have_arm_neon" = "yes" ]; then
   echo "HAVE_NEON = 1" >> $config_mak
 fi

diff --git a/include/arm_features.h b/include/arm_features.h
new file mode 100644 (file)
index 0000000..dcdda91
--- /dev/null
+++ b/include/arm_features.h
@@ -0,0 +1,28 @@
+#ifndef __ARM_FEATURES_H__
+#define __ARM_FEATURES_H__
+
+#if defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) \
+ || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) \
+ || defined(__ARM_ARCH_7EM__)
+
+#define HAVE_ARMV7
+#define HAVE_ARMV6
+#define HAVE_ARMV5
+
+#elif defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) \
+   || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) \
+   || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__)
+
+#define HAVE_ARMV6
+#define HAVE_ARMV5
+
+#elif defined(__ARM_ARCH_5__) || defined(__ARM_ARCH_5E__) \
+   || defined(__ARM_ARCH_5T__) || defined(__ARM_ARCH_5TE__) || defined(__ARM_ARCH_5TEJ__)
+
+#define HAVE_ARMV5
+
+#endif
+
+/* no need for HAVE_NEON - GCC defines __ARM_NEON__ consistently */
+
+#endif /* __ARM_FEATURES_H__ */

diff --git a/jni/Android.mk b/jni/Android.mk
index 98f59f8..60db810 100644 (file)
--- a/jni/Android.mk
+++ b/jni/Android.mk
@@ -19,12 +19,6 @@ endif
 ifeq ($(TARGET_ARCH),arm)
    LOCAL_ARM_MODE := arm
 
-   ifeq ($(TARGET_ARCH_EABI),armeabi-v7a)
-      LOCAL_CFLAGS += -DHAVE_ARMV7=1
-   else
-      LOCAL_CFLAGS += -DHAVE_ARMV7=0
-   endif
-
    LOCAL_CFLAGS += -DANDROID_ARM
 
    LOCAL_SRC_FILES += ../libpcsxcore/gte_arm.S

diff --git a/libpcsxcore/gte_arm.s b/libpcsxcore/gte_arm.S
similarity index 97%
rename from libpcsxcore/gte_arm.s
rename to libpcsxcore/gte_arm.S
index 8700f69..e711e82 100644 (file)
--- a/libpcsxcore/gte_arm.s
+++ b/libpcsxcore/gte_arm.S
@@ -5,37 +5,37 @@
  * See the COPYING file in the top-level directory.
  */
 
-/* .equiv HAVE_ARMV7, 1 */
+#include "arm_features.h"
 
 .text
 .align 2
 
 .macro sgnxt16 rd rs
-.if HAVE_ARMV7
+#ifdef HAVE_ARMV7
     sxth     \rd, \rs
-.else
+#else
     lsl      \rd, \rs, #16
     asr      \rd, \rd, #16
-.endif
+#endif
 .endm
 
 @ prepare work reg for ssatx
 @ in: wr reg, bit to saturate to
 .macro ssatx_prep wr bit
-.if !HAVE_ARMV7
+#ifndef HAVE_ARMV7
     mov      \wr, #(1<<(\bit-1))
-.endif
+#endif
 .endm
 
 .macro ssatx rd wr bit
-.if HAVE_ARMV7
+#ifdef HAVE_ARMV7
     ssat     \rd, #\bit, \rd
-.else
+#else
     cmp      \rd, \wr
     subge    \rd, \wr, #1
     cmn      \rd, \wr
     rsblt    \rd, \wr, #0
-.endif
+#endif
 .endm
 
 @ prepare work reg for ssatx0 (sat to 0..2^(bit-1))
@@ -52,17 +52,19 @@
 .endm
 
 .macro usat16_ rd rs
-.if HAVE_ARMV7
+#ifdef HAVE_ARMV7
     usat     \rd, #16, \rs
-.else
+#else
     subs     \rd, \rs, #0
     movlt    \rd, #0
     cmp      \rd, #0x10000
     movge    \rd, #0x0ff00
     orrge    \rd, #0x000ff
-.endif
+#endif
 .endm
 
+#ifdef HAVE_ARMV5
+
 .macro udiv_ rd rm rs
     lsl      \rm, #16
     clz      \rd, \rs
@@ -408,6 +410,7 @@ gteMVMVA_part_cv3sh12_arm:
     bx       lr
     .size    gteMVMVA_part_cv3sh12_arm, .-gteMVMVA_part_cv3sh12_arm
 
+#endif /* HAVE_ARMV5 */
 
 .global gteNCLIP_arm @ r0=CP2 (d,c),
 gteNCLIP_arm:
@@ -430,12 +433,12 @@ gteNCLIP_arm:
     movs        r2, r1, lsl #1
     adc         r5, r5
     cmp         r5, #0
-.if HAVE_ARMV7
+#ifdef HAVE_ARMV7
     movtgt      lr, #((1<<31)|(1<<16))>>16
-.else
+#else
     movgt       lr, #(1<<31)
     orrgt       lr, #(1<<16)
-.endif
+#endif
     cmn         r5, #1
     orrmi       lr, r6
     str         r1, [r0, #4*24]
@@ -460,7 +463,12 @@ gteNCLIP_arm:
     rsblt    r2, r1, #0
 .endif
     str      r2, [r0, #4*9]
+#ifdef HAVE_ARMV5
     ldrd     r2, [r0, #4*26]      @ gteMAC23
+#else
+    ldr      r2, [r0, #4*26]
+    ldr      r3, [r0, #4*27]
+#endif
     orrlt    r12, #(1<<31)|(1<<24)
     cmp      r2, r1
     subge    r2, r1, #1
@@ -486,7 +494,12 @@ gteNCLIP_arm:
     rsblt    r3, r1, #0
 .endif
     orrlt    r12, #1<<22
+#ifdef HAVE_ARMV5
     strd     r2, [r0, #4*10]      @ gteIR23
+#else
+    str      r2, [r0, #4*10]
+    str      r3, [r0, #4*11]
+#endif
     str      r12,[r0, #4*(32+31)] @ gteFLAG
     bx       lr
 .endm

diff --git a/libpcsxcore/new_dynarec/assem_arm.c b/libpcsxcore/new_dynarec/assem_arm.c
index 77cfafa..51cceec 100644 (file)
--- a/libpcsxcore/new_dynarec/assem_arm.c
+++ b/libpcsxcore/new_dynarec/assem_arm.c
@@ -28,6 +28,7 @@
 #include "../gte_neon.h"
 #include "pcnt.h"
 #endif
+#include "arm_features.h"
 
 #if !BASE_ADDR_FIXED
 char translation_cache[1 << TARGET_SIZE_2] __attribute__((aligned(4096)));
@@ -223,7 +224,7 @@ int get_pointer(void *stub)
 u_int get_clean_addr(int addr)
 {
   int *ptr=(int *)addr;
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   ptr+=4;
   #else
   ptr+=6;
@@ -240,7 +241,7 @@ u_int get_clean_addr(int addr)
 int verify_dirty(int addr)
 {
   u_int *ptr=(u_int *)addr;
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   // get from literal pool
   assert((*ptr&0xFFFF0000)==0xe59f0000);
   u_int offset=*ptr&0xfff;
@@ -279,7 +280,7 @@ int verify_dirty(int addr)
 // guarantees that it's not dirty
 int isclean(int addr)
 {
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   int *ptr=((u_int *)addr)+4;
   #else
   int *ptr=((u_int *)addr)+6;
@@ -296,7 +297,7 @@ int isclean(int addr)
 void get_bounds(int addr,u_int *start,u_int *end)
 {
   u_int *ptr=(u_int *)addr;
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   // get from literal pool
   assert((*ptr&0xFFFF0000)==0xe59f0000);
   u_int offset=*ptr&0xfff;
@@ -1005,7 +1006,7 @@ void emit_movimm(u_int imm,u_int rt)
     assem_debug("mvn %s,#%d\n",regname[rt],imm);
     output_w32(0xe3e00000|rd_rn_rm(rt,0,0)|armval);
   }else if(imm<65536) {
-    #ifdef ARMv5_ONLY
+    #ifndef HAVE_ARMV7
     assem_debug("mov %s,#%d\n",regname[rt],imm&0xFF00);
     output_w32(0xe3a00000|rd_rn_imm_shift(rt,0,imm>>8,8));
     assem_debug("add %s,%s,#%d\n",regname[rt],regname[rt],imm&0xFF);
@@ -1014,7 +1015,7 @@ void emit_movimm(u_int imm,u_int rt)
     emit_movw(imm,rt);
     #endif
   }else{
-    #ifdef ARMv5_ONLY
+    #ifndef HAVE_ARMV7
     emit_loadlp(imm,rt);
     #else
     emit_movw(imm&0x0000FFFF,rt);
@@ -1278,7 +1279,7 @@ void emit_andimm(int rs,int imm,int rt)
     assem_debug("bic %s,%s,#%d\n",regname[rt],regname[rs],imm);
     output_w32(0xe3c00000|rd_rn_rm(rt,rs,0)|armval);
   }else if(imm==65535) {
-    #ifdef ARMv5_ONLY
+    #ifndef HAVE_ARMV7
     assem_debug("bic %s,%s,#FF000000\n",regname[rt],regname[rs]);
     output_w32(0xe3c00000|rd_rn_rm(rt,rs,0)|0x4FF);
     assem_debug("bic %s,%s,#00FF0000\n",regname[rt],regname[rt]);
@@ -1289,7 +1290,7 @@ void emit_andimm(int rs,int imm,int rt)
     #endif
   }else{
     assert(imm>0&&imm<65535);
-    #ifdef ARMv5_ONLY
+    #ifndef HAVE_ARMV7
     assem_debug("mov r14,#%d\n",imm&0xFF00);
     output_w32(0xe3a00000|rd_rn_imm_shift(HOST_TEMPREG,0,imm>>8,8));
     assem_debug("add r14,r14,#%d\n",imm&0xFF);
@@ -1353,6 +1354,14 @@ void emit_lsls_imm(int rs,int imm,int rt)
   output_w32(0xe1b00000|rd_rn_rm(rt,0,rs)|(imm<<7));
 }
 
+void emit_lslpls_imm(int rs,int imm,int rt)
+{
+  assert(imm>0);
+  assert(imm<32);
+  assem_debug("lslpls %s,%s,#%d\n",regname[rt],regname[rs],imm);
+  output_w32(0x51b00000|rd_rn_rm(rt,0,rs)|(imm<<7));
+}
+
 void emit_shrimm(int rs,u_int imm,int rt)
 {
   assert(imm>0);
@@ -1403,7 +1412,7 @@ void emit_shrdimm(int rs,int rs2,u_int imm,int rt)
 
 void emit_signextend16(int rs,int rt)
 {
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   emit_shlimm(rs,16,rt);
   emit_sarimm(rt,16,rt);
   #else
@@ -1414,7 +1423,7 @@ void emit_signextend16(int rs,int rt)
 
 void emit_signextend8(int rs,int rt)
 {
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   emit_shlimm(rs,24,rt);
   emit_sarimm(rt,24,rt);
   #else
@@ -1502,20 +1511,12 @@ void emit_cmpimm(int rs,int imm)
     output_w32(0xe3700000|rd_rn_rm(0,rs,0)|armval);
   }else if(imm>0) {
     assert(imm<65536);
-    #ifdef ARMv5_ONLY
     emit_movimm(imm,HOST_TEMPREG);
-    #else
-    emit_movw(imm,HOST_TEMPREG);
-    #endif
     assem_debug("cmp %s,r14\n",regname[rs]);
     output_w32(0xe1500000|rd_rn_rm(0,rs,HOST_TEMPREG));
   }else{
     assert(imm>-65536);
-    #ifdef ARMv5_ONLY
     emit_movimm(-imm,HOST_TEMPREG);
-    #else
-    emit_movw(-imm,HOST_TEMPREG);
-    #endif
     assem_debug("cmn %s,r14\n",regname[rs]);
     output_w32(0xe1700000|rd_rn_rm(0,rs,HOST_TEMPREG));
   }
@@ -2295,7 +2296,7 @@ void emit_cmov2imm_e_ne_compact(int imm1,int imm2,u_int rt)
     output_w32(0x12400000|rd_rn_rm(rt,rt,0)|armval);
   }
   else {
-    #ifdef ARMv5_ONLY
+    #ifndef HAVE_ARMV7
     emit_movimm(imm1,rt);
     add_literal((int)out,imm2);
     assem_debug("ldrne %s,pc+? [=%x]\n",regname[rt],imm2);
@@ -2586,6 +2587,14 @@ void emit_andne_imm(int rs,int imm,int rt)
   output_w32(0x12000000|rd_rn_rm(rt,rs,0)|armval);
 }
 
+void emit_addpl_imm(int rs,int imm,int rt)
+{
+  u_int armval;
+  genimm_checked(imm,&armval);
+  assem_debug("addpl %s,%s,#%d\n",regname[rt],regname[rs],imm);
+  output_w32(0x52800000|rd_rn_rm(rt,rs,0)|armval);
+}
+
 void emit_jno_unlikely(int a)
 {
   //emit_jno(a);
@@ -3565,7 +3574,7 @@ int do_dirty_stub(int i)
   addr=(u_int)source;
   #endif
   // Careful about the code output here, verify_dirty needs to parse it.
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   emit_loadlp(addr,1);
   emit_loadlp((int)copy,2);
   emit_loadlp(slen*4,3);
@@ -3588,7 +3597,7 @@ int do_dirty_stub(int i)
 void do_dirty_stub_ds()
 {
   // Careful about the code output here, verify_dirty needs to parse it.
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   emit_loadlp((int)start<(int)0xC0000000?(int)source:(int)start,1);
   emit_loadlp((int)copy,2);
   emit_loadlp(slen*4,3);
@@ -4391,7 +4400,16 @@ static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
     case 30:
       emit_movs(sl,temp);
       emit_mvnmi(temp,temp);
+#ifdef HAVE_ARMV5
       emit_clz(temp,temp);
+#else
+      emit_movs(temp,HOST_TEMPREG);
+      emit_movimm(0,temp);
+      emit_jeq((int)out+4*4);
+      emit_addpl_imm(temp,1,temp);
+      emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
+      emit_jns((int)out-2*4);
+#endif
       emit_writeword(sl,(int)&reg_cop2d[30]);
       emit_writeword(temp,(int)&reg_cop2d[31]);
       break;
@@ -4513,6 +4531,7 @@ static void c2op_assemble(int i,struct regstat *i_regs)
     int lm = (source[i] >> 10) & 1;
     switch(c2op) {
 #ifndef DRC_DBG
+#ifdef HAVE_ARMV5
       case GTE_MVMVA: {
         int v  = (source[i] >> 15) & 3;
         int cv = (source[i] >> 13) & 3;
@@ -4555,6 +4574,7 @@ static void c2op_assemble(int i,struct regstat *i_regs)
 #endif
         break;
       }
+#endif /* HAVE_ARMV5 */
       case GTE_OP:
         c2op_prologue(c2op,reglist);
         emit_call((int)(shift?gteOP_part_shift:gteOP_part_noshift));
@@ -5293,8 +5313,15 @@ void multdiv_assemble_arm(int i,struct regstat *i_regs)
         emit_movs(d2,HOST_TEMPREG);
         emit_jeq((int)out+52); // Division by zero
         emit_negmi(HOST_TEMPREG,HOST_TEMPREG);
+#ifdef HAVE_ARMV5
         emit_clz(HOST_TEMPREG,quotient);
         emit_shl(HOST_TEMPREG,quotient,HOST_TEMPREG);
+#else
+        emit_movimm(0,quotient);
+        emit_addpl_imm(quotient,1,quotient);
+        emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
+        emit_jns((int)out-2*4);
+#endif
         emit_orimm(quotient,1<<31,quotient);
         emit_shr(quotient,quotient,quotient);
         emit_cmp(remainder,HOST_TEMPREG);
@@ -5321,9 +5348,17 @@ void multdiv_assemble_arm(int i,struct regstat *i_regs)
         emit_movimm(0xffffffff,quotient); // div0 case
         emit_test(d2,d2);
         emit_jeq((int)out+40); // Division by zero
+#ifdef HAVE_ARMV5
         emit_clz(d2,HOST_TEMPREG);
         emit_movimm(1<<31,quotient);
         emit_shl(d2,HOST_TEMPREG,d2);
+#else
+        emit_movimm(0,HOST_TEMPREG);
+        emit_addpl_imm(d2,1,d2);
+        emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
+        emit_jns((int)out-2*4);
+        emit_movimm(1<<31,quotient);
+#endif
         emit_shr(quotient,HOST_TEMPREG,quotient);
         emit_cmp(remainder,d2);
         emit_subcs(remainder,d2,remainder);
@@ -5554,7 +5589,7 @@ void do_miniht_jump(int rs,int rh,int ht) {
 }
 
 void do_miniht_insert(u_int return_address,int rt,int temp) {
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   emit_movimm(return_address,rt); // PC into link register
   add_to_linker((int)out,return_address,1);
   emit_pcreladdr(temp);

diff --git a/libpcsxcore/new_dynarec/assem_arm.h b/libpcsxcore/new_dynarec/assem_arm.h
index f4e36a9..2254638 100644 (file)
--- a/libpcsxcore/new_dynarec/assem_arm.h
+++ b/libpcsxcore/new_dynarec/assem_arm.h
@@ -17,7 +17,6 @@
 #define RAM_SIZE 0x200000
 
 #ifndef __ARM_ARCH_7A__
-#define ARMv5_ONLY
 //#undef CORTEX_A8_BRANCH_PREDICTION_HACK
 //#undef USE_MINI_HT
 #endif

diff --git a/libpcsxcore/new_dynarec/emu_if.c b/libpcsxcore/new_dynarec/emu_if.c
index 588bc63..b8e9883 100644 (file)
--- a/libpcsxcore/new_dynarec/emu_if.c
+++ b/libpcsxcore/new_dynarec/emu_if.c
@@ -268,9 +268,9 @@ static int ari64_init()
                if (psxCP2[i] != psxNULL)
                        gte_handlers[i] = psxCP2[i];
 
-#if !defined(DRC_DBG)
-#ifdef __arm__
+#if defined(__arm__) && !defined(DRC_DBG)
        gte_handlers[0x06] = gteNCLIP_arm;
+#ifdef HAVE_ARMV5
        gte_handlers_nf[0x01] = gteRTPS_nf_arm;
        gte_handlers_nf[0x30] = gteRTPT_nf_arm;
 #endif

diff --git a/libpcsxcore/new_dynarec/linkage_arm.S b/libpcsxcore/new_dynarec/linkage_arm.S
index 4748078..5b70745 100644 (file)
--- a/libpcsxcore/new_dynarec/linkage_arm.S
+++ b/libpcsxcore/new_dynarec/linkage_arm.S
@@ -19,6 +19,9 @@
  *   51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.          *
  * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
 
+#include "arm_features.h"
+
+
        .global dynarec_local
        .global reg
        .global hi
@@ -163,6 +166,13 @@ FCR31 = align0
        .type   FCR31, %object
        .size   FCR31, 4
 
+#ifndef HAVE_ARMV5
+.macro blx rd
+       mov     lr, pc
+       bx      \rd
+.endm
+#endif
+
 .macro load_varadr reg var
 #if defined(__ARM_ARCH_7A__) && !defined(__PIC__)
        movw    \reg, #:lower16:\var

diff --git a/plugins/dfsound/arm_utils.S b/plugins/dfsound/arm_utils.S
index 22e5844..f4ef243 100644 (file)
--- a/plugins/dfsound/arm_utils.S
+++ b/plugins/dfsound/arm_utils.S
@@ -8,6 +8,7 @@
  * See the COPYING file in the top-level directory.
  */
 
+#include "arm_features.h"
 
 .text
 .align 2
@@ -100,7 +101,7 @@ mcr_finish:
     vstmiage    r3!, {d8}
     bx          lr
 
-#else
+#elif defined(HAVE_ARMV5)
 
 .global mix_chan @ (int start, int count, int lv, int rv)
 mix_chan:

diff --git a/plugins/dfsound/spu.c b/plugins/dfsound/spu.c
index 45a7886..2466569 100644 (file)
--- a/plugins/dfsound/spu.c
+++ b/plugins/dfsound/spu.c
@@ -25,6 +25,7 @@
 #include "externals.h"
 #include "registers.h"
 #include "out.h"
+#include "arm_features.h"
 
 #ifdef ENABLE_NLS
 #include <libintl.h>
@@ -615,7 +616,7 @@ static int do_samples_noise(int ch, int ns, int ns_to)
  return ret;
 }
 
-#ifdef __arm__
+#ifdef HAVE_ARMV5
 // asm code; lv and rv must be 0-3fff
 extern void mix_chan(int start, int count, int lv, int rv);
 extern void mix_chan_rvb(int start, int count, int lv, int rv);

diff --git a/plugins/gpu_unai/gpu_fixedpoint.h b/plugins/gpu_unai/gpu_fixedpoint.h
index 03f07d2..e72fda1 100644 (file)
--- a/plugins/gpu_unai/gpu_fixedpoint.h
+++ b/plugins/gpu_unai/gpu_fixedpoint.h
@@ -21,6 +21,8 @@
 #ifndef FIXED_H
 #define FIXED_H
 
+#include "arm_features.h"
+
 typedef s32 fixed;
 
 #ifdef GPU_TABLE_10_BITS
@@ -55,7 +57,7 @@ INLINE u32 Log2(u32 _a)
 }
 */
 
-#ifdef __arm__
+#ifdef HAVE_ARMV5
 INLINE u32 Log2(u32 x) { u32 res; asm("clz %0,%1" : "=r" (res) : "r" (x)); return 32-res; }
 #else
 INLINE u32 Log2(u32 x) { u32 i = 0; for ( ; x > 0; ++i, x >>= 1); return i - 1; }