X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?p=pcsx_rearmed.git;a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fassem_arm.c;h=ddbfafaa6394073664528eefdbc750ab8e48ea19;hp=57c6e917e4ad3ff6debdd8bf54a1cb5fc9a9dd48;hb=82ed88ebe25a0312aab83623b5a983bd96f3d830;hpb=13e35c04a17506ac0848d7fb089fa817b3d697e9

diff --git a/libpcsxcore/new_dynarec/assem_arm.c b/libpcsxcore/new_dynarec/assem_arm.c
index 57c6e917..ddbfafaa 100644
--- a/libpcsxcore/new_dynarec/assem_arm.c
+++ b/libpcsxcore/new_dynarec/assem_arm.c
@@ -19,6 +19,12 @@
  *   51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.          *
  * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
 
+#ifdef PCSX
+#include "../gte_arm.h"
+#include "../gte_neon.h"
+#include "pcnt.h"
+#endif
+
 extern int cycle_count;
 extern int last_count;
 extern int pcaddr;
@@ -245,6 +251,7 @@ int verify_dirty(int addr)
   #endif
   if((*ptr&0xFF000000)!=0xeb000000) ptr++;
   assert((*ptr&0xFF000000)==0xeb000000); // bl instruction
+#ifndef DISABLE_TLB
   u_int verifier=(int)ptr+((signed int)(*ptr<<8)>>6)+8; // get target of bl
   if(verifier==(u_int)verify_code_vm||verifier==(u_int)verify_code_ds) {
     unsigned int page=source>>12;
@@ -255,6 +262,7 @@ int verify_dirty(int addr)
     }
     source = source+(map_value<<2);
   }
+#endif
   //printf("verify_dirty: %x %x %x\n",source,copy,len);
   return !memcmp((void *)source,(void *)copy,len);
 }
@@ -298,11 +306,13 @@ void get_bounds(int addr,u_int *start,u_int *end)
   #endif
   if((*ptr&0xFF000000)!=0xeb000000) ptr++;
   assert((*ptr&0xFF000000)==0xeb000000); // bl instruction
+#ifndef DISABLE_TLB
   u_int verifier=(int)ptr+((signed int)(*ptr<<8)>>6)+8; // get target of bl
   if(verifier==(u_int)verify_code_vm||verifier==(u_int)verify_code_ds) {
     if(memory_map[source>>12]>=0x80000000) source = 0;
     else source = source+(memory_map[source>>12]<<2);
   }
+#endif
   *start=source;
   *end=source+len;
 }
@@ -1920,6 +1930,16 @@ void emit_movzwl_indexed(int offset, int rs, int rt)
     output_w32(0xe15000b0|rd_rn_rm(rt,rs,0)|(((-offset)<<4)&0xf00)|((-offset)&0xf));
   }
 }
+static void emit_ldrd(int offset, int rs, int rt)
+{
+  assert(offset>-256&&offset<256);
+  assem_debug("ldrd %s,%s+%d\n",regname[rt],regname[rs],offset);
+  if(offset>=0) {
+    output_w32(0xe1c000d0|rd_rn_rm(rt,rs,0)|((offset<<4)&0xf00)|(offset&0xf));
+  }else{
+    output_w32(0xe14000d0|rd_rn_rm(rt,rs,0)|(((-offset)<<4)&0xf00)|((-offset)&0xf));
+  }
+}
 void emit_readword(int addr, int rt)
 {
   u_int offset = addr-(u_int)&dynarec_local;
@@ -2154,6 +2174,14 @@ void emit_shrcc_imm(int rs,u_int imm,int rt)
   output_w32(0x31a00000|rd_rn_rm(rt,0,rs)|0x20|(imm<<7));
 }
 
+void emit_shrne_imm(int rs,u_int imm,int rt)
+{
+  assert(imm>0);
+  assert(imm<32);
+  assem_debug("lsrne %s,%s,#%d\n",regname[rt],regname[rs],imm);
+  output_w32(0x11a00000|rd_rn_rm(rt,0,rs)|0x20|(imm<<7));
+}
+
 void emit_negmi(int rs, int rt)
 {
   assem_debug("rsbmi %s,%s,#0\n",regname[rt],regname[rs]);
@@ -2556,34 +2584,40 @@ void emit_jno_unlikely(int a)
   output_w32(0x72800000|rd_rn_rm(15,15,0));
 }
 
-// Save registers before function call
-void save_regs(u_int reglist)
+static void save_regs_all(u_int reglist)
 {
-  reglist&=0x100f; // only save the caller-save registers, r0-r3, r12
+  int i;
   if(!reglist) return;
   assem_debug("stmia fp,{");
-  if(reglist&1) assem_debug("r0, ");
-  if(reglist&2) assem_debug("r1, ");
-  if(reglist&4) assem_debug("r2, ");
-  if(reglist&8) assem_debug("r3, ");
-  if(reglist&0x1000) assem_debug("r12");
+  for(i=0;i<16;i++)
+    if(reglist&(1<<i))
+      assem_debug("r%d,",i);
   assem_debug("}\n");
   output_w32(0xe88b0000|reglist);
 }
-// Restore registers after function call
-void restore_regs(u_int reglist)
+static void restore_regs_all(u_int reglist)
 {
-  reglist&=0x100f; // only restore the caller-save registers, r0-r3, r12
+  int i;
   if(!reglist) return;
   assem_debug("ldmia fp,{");
-  if(reglist&1) assem_debug("r0, ");
-  if(reglist&2) assem_debug("r1, ");
-  if(reglist&4) assem_debug("r2, ");
-  if(reglist&8) assem_debug("r3, ");
-  if(reglist&0x1000) assem_debug("r12");
+  for(i=0;i<16;i++)
+    if(reglist&(1<<i))
+      assem_debug("r%d,",i);
   assem_debug("}\n");
   output_w32(0xe89b0000|reglist);
 }
+// Save registers before function call
+static void save_regs(u_int reglist)
+{
+  reglist&=0x100f; // only save the caller-save registers, r0-r3, r12
+  save_regs_all(reglist);
+}
+// Restore registers after function call
+static void restore_regs(u_int reglist)
+{
+  reglist&=0x100f; // only restore the caller-save registers, r0-r3, r12
+  restore_regs_all(reglist);
+}
 
 // Write back consts using r14 so we don't disturb the other registers
 void wb_consts(signed char i_regmap[],uint64_t i_is32,u_int i_dirty,int i)
@@ -2625,12 +2659,24 @@ void literal_pool(int n)
   int i;
   for(i=0;i<literalcount;i++)
   {
+    u_int l_addr=(u_int)out;
+    int j;
+    for(j=0;j<i;j++) {
+      if(literals[j][1]==literals[i][1]) {
+        //printf("dup %08x\n",literals[i][1]);
+        l_addr=literals[j][0];
+        break;
+      }
+    }
     ptr=(u_int *)literals[i][0];
-    u_int offset=(u_int)out-(u_int)ptr-8;
+    u_int offset=l_addr-(u_int)ptr-8;
     assert(offset<4096);
     assert(!(offset&3));
     *ptr|=offset;
-    output_w32(literals[i][1]);
+    if(l_addr==(u_int)out) {
+      literals[i][0]=l_addr; // remember for dupes
+      output_w32(literals[i][1]);
+    }
   }
   literalcount=0;
 }
@@ -2707,6 +2753,23 @@ static void pass_args(int a0, int a1)
   }
 }
 
+static void mov_loadtype_adj(int type,int rs,int rt)
+{
+  switch(type) {
+    case LOADB_STUB:  emit_signextend8(rs,rt); break;
+    case LOADBU_STUB: emit_andimm(rs,0xff,rt); break;
+    case LOADH_STUB:  emit_signextend16(rs,rt); break;
+    case LOADHU_STUB: emit_andimm(rs,0xffff,rt); break;
+    case LOADW_STUB:  if(rs!=rt) emit_mov(rs,rt); break;
+    default: assert(0);
+  }
+}
+
+#ifdef PCSX
+#include "pcsxmem.h"
+#include "pcsxmem_inline.c"
+#endif
+
 do_readstub(int n)
 {
   assem_debug("do_readstub %x\n",start+stubs[n][3]*4);
@@ -2780,16 +2843,10 @@ do_readstub(int n)
   int cc=get_reg(i_regmap,CCREG);
   if(cc<0)
     emit_loadreg(CCREG,2);
-  emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*stubs[n][6]+2,2);
+  emit_addimm(cc<0?2:cc,CLOCK_ADJUST((int)stubs[n][6]+1),2);
   emit_call(handler);
   if(itype[i]==C1LS||itype[i]==C2LS||(rt>=0&&rt1[i]!=0)) {
-    switch(type) {
-      case LOADB_STUB:  emit_signextend8(0,rt); break;
-      case LOADBU_STUB: emit_andimm(0,0xff,rt); break;
-      case LOADH_STUB:  emit_signextend16(0,rt); break;
-      case LOADHU_STUB: emit_andimm(0,0xffff,rt); break;
-      case LOADW_STUB:  if(rt!=0) emit_mov(0,rt); break;
-    }
+    mov_loadtype_adj(type,0,rt);
   }
   if(restore_jump)
     set_jump_target(restore_jump,(int)out);
@@ -2914,7 +2971,10 @@ inline_readstub(int type, int i, u_int addr, signed char regmap[], int target, i
   if(rs<0) rs=get_reg(regmap,-1);
   assert(rs>=0);
 #ifdef PCSX
-  u_int handler,host_addr=0;
+  u_int handler,host_addr=0,is_dynamic,far_call=0;
+  int cc=get_reg(regmap,CCREG);
+  if(pcsx_direct_read(type,addr,CLOCK_ADJUST(adj+1),cc,target?rs:-1,rt))
+    return;
   handler=get_direct_memhandler(mem_rtab,addr,type,&host_addr);
   if (handler==0) {
     if(rt<0)
@@ -2931,6 +2991,15 @@ inline_readstub(int type, int i, u_int addr, signed char regmap[], int target, i
     }
     return;
   }
+  is_dynamic=pcsxmem_is_handler_dynamic(addr);
+  if(is_dynamic) {
+    if(type==LOADB_STUB||type==LOADBU_STUB)
+      handler=(int)jump_handler_read8;
+    if(type==LOADH_STUB||type==LOADHU_STUB)
+      handler=(int)jump_handler_read16;
+    if(type==LOADW_STUB)
+      handler=(int)jump_handler_read32;
+  }
 
   // call a memhandler
   if(rt>=0)
@@ -2940,22 +3009,30 @@ inline_readstub(int type, int i, u_int addr, signed char regmap[], int target, i
     emit_movimm(addr,0);
   else if(rs!=0)
     emit_mov(rs,0);
-  int cc=get_reg(regmap,CCREG);
-  if(cc<0)
-    emit_loadreg(CCREG,2);
-  emit_readword((int)&last_count,3);
-  emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*(adj+1),2);
-  emit_add(2,3,3);
-  emit_writeword(3,(int)&Count);
-
   int offset=(int)handler-(int)out-8;
   if(offset<-33554432||offset>=33554432) {
     // unreachable memhandler, a plugin func perhaps
-    emit_movimm(handler,1);
-    emit_callreg(1);
+    emit_movimm(handler,12);
+    far_call=1;
+  }
+  if(cc<0)
+    emit_loadreg(CCREG,2);
+  if(is_dynamic) {
+    emit_movimm(((u_int *)mem_rtab)[addr>>12]<<1,1);
+    emit_addimm(cc<0?2:cc,CLOCK_ADJUST(adj+1),2);
+  }
+  else {
+    emit_readword((int)&last_count,3);
+    emit_addimm(cc<0?2:cc,CLOCK_ADJUST(adj+1),2);
+    emit_add(2,3,2);
+    emit_writeword(2,(int)&Count);
   }
+
+  if(far_call)
+    emit_callreg(12);
   else
     emit_call(handler);
+
   if(rt>=0) {
     switch(type) {
       case LOADB_STUB:  emit_signextend8(0,rt); break;
@@ -3005,7 +3082,7 @@ inline_readstub(int type, int i, u_int addr, signed char regmap[], int target, i
   //emit_movimm(ftable,0);
   emit_movimm(((u_int *)ftable)[addr>>16],0);
   //emit_readword((int)&last_count,12);
-  emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*(adj+1),2);
+  emit_addimm(cc<0?2:cc,CLOCK_ADJUST(adj+1),2);
 #ifndef PCSX
   if((signed int)addr>=(signed int)0xC0000000) {
     // Pagefault address
@@ -3022,7 +3099,7 @@ inline_readstub(int type, int i, u_int addr, signed char regmap[], int target, i
   // but not doing so causes random crashes...
   emit_readword((int)&Count,HOST_TEMPREG);
   emit_readword((int)&next_interupt,2);
-  emit_addimm(HOST_TEMPREG,-CLOCK_DIVIDER*(adj+1),HOST_TEMPREG);
+  emit_addimm(HOST_TEMPREG,-CLOCK_ADJUST(adj+1),HOST_TEMPREG);
   emit_writeword(2,(int)&last_count);
   emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc);
   if(cc<0) {
@@ -3122,10 +3199,10 @@ do_writestub(int n)
   int cc=get_reg(i_regmap,CCREG);
   if(cc<0)
     emit_loadreg(CCREG,2);
-  emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*stubs[n][6]+2,2);
+  emit_addimm(cc<0?2:cc,CLOCK_ADJUST((int)stubs[n][6]+1),2);
   // returns new cycle_count
   emit_call(handler);
-  emit_addimm(0,-CLOCK_DIVIDER*stubs[n][6]-2,cc<0?2:cc);
+  emit_addimm(0,-CLOCK_ADJUST((int)stubs[n][6]+1),cc<0?2:cc);
   if(cc<0)
     emit_storereg(CCREG,2);
   if(restore_jump)
@@ -3237,11 +3314,11 @@ inline_writestub(int type, int i, u_int addr, signed char regmap[], int target,
   int cc=get_reg(regmap,CCREG);
   if(cc<0)
     emit_loadreg(CCREG,2);
-  emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*(adj+1),2);
+  emit_addimm(cc<0?2:cc,CLOCK_ADJUST(adj+1),2);
   emit_movimm(handler,3);
   // returns new cycle_count
   emit_call((int)jump_handler_write_h);
-  emit_addimm(0,-CLOCK_DIVIDER*(adj+1),cc<0?2:cc);
+  emit_addimm(0,-CLOCK_ADJUST(adj+1),cc<0?2:cc);
   if(cc<0)
     emit_storereg(CCREG,2);
   restore_regs(reglist);
@@ -3298,7 +3375,7 @@ inline_writestub(int type, int i, u_int addr, signed char regmap[], int target,
   //emit_movimm(ftable,0);
   emit_movimm(((u_int *)ftable)[addr>>16],0);
   //emit_readword((int)&last_count,12);
-  emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*(adj+1),2);
+  emit_addimm(cc<0?2:cc,CLOCK_ADJUST(adj+1),2);
 #ifndef PCSX
   if((signed int)addr>=(signed int)0xC0000000) {
     // Pagefault address
@@ -3312,7 +3389,7 @@ inline_writestub(int type, int i, u_int addr, signed char regmap[], int target,
   emit_call((int)&indirect_jump);
   emit_readword((int)&Count,HOST_TEMPREG);
   emit_readword((int)&next_interupt,2);
-  emit_addimm(HOST_TEMPREG,-CLOCK_DIVIDER*(adj+1),HOST_TEMPREG);
+  emit_addimm(HOST_TEMPREG,-CLOCK_ADJUST(adj+1),HOST_TEMPREG);
   emit_writeword(2,(int)&last_count);
   emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc);
   if(cc<0) {
@@ -3351,9 +3428,9 @@ do_unalignedwritestub(int n)
   int cc=get_reg(i_regmap,CCREG);
   if(cc<0)
     emit_loadreg(CCREG,2);
-  emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*stubs[n][6]+2,2);
+  emit_addimm(cc<0?2:cc,CLOCK_ADJUST((int)stubs[n][6]+1),2);
   emit_call((int)(opcode[i]==0x2a?jump_handle_swl:jump_handle_swr));
-  emit_addimm(0,-CLOCK_DIVIDER*stubs[n][6]-2,cc<0?2:cc);
+  emit_addimm(0,-CLOCK_ADJUST((int)stubs[n][6]+1),cc<0?2:cc);
   if(cc<0)
     emit_storereg(CCREG,2);
   restore_regs(reglist);
@@ -3504,10 +3581,12 @@ do_cop1stub(int n)
   wb_dirtys(i_regs->regmap_entry,i_regs->was32,i_regs->wasdirty);
   if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
   emit_movimm(start+(i-ds)*4,EAX); // Get PC
-  emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); // CHECK: is this right?  There should probably be an extra cycle...
+  emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right?  There should probably be an extra cycle...
   emit_jmp(ds?(int)fp_exception_ds:(int)fp_exception);
 }
 
+#ifndef DISABLE_TLB
+
 /* TLB */
 
 int do_tlb_r(int s,int ar,int map,int x,int a,int shift,int c,u_int addr)
@@ -3591,6 +3670,17 @@ generate_map_const(u_int addr,int reg) {
   emit_movimm((addr>>12)+(((u_int)memory_map-(u_int)&dynarec_local)>>2),reg);
 }
 
+#else
+
+static int do_tlb_r() { return 0; }
+static int do_tlb_r_branch() { return 0; }
+static int gen_tlb_addr_r() { return 0; }
+static int do_tlb_w() { return 0; }
+static int do_tlb_w_branch() { return 0; }
+static int gen_tlb_addr_w() { return 0; }
+
+#endif // DISABLE_TLB
+
 /* Special assem */
 
 void shift_assemble_arm(int i,struct regstat *i_regs)
@@ -4046,7 +4136,7 @@ void cop0_assemble(int i,struct regstat *i_regs)
         emit_readword((int)&last_count,ECX);
         emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
         emit_add(HOST_CCREG,ECX,HOST_CCREG);
-        emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG);
+        emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
         emit_writeword(HOST_CCREG,(int)&Count);
       }
       emit_call((int)MFC0);
@@ -4061,19 +4151,21 @@ void cop0_assemble(int i,struct regstat *i_regs)
     signed char s=get_reg(i_regs->regmap,rs1[i]);
     char copr=(source[i]>>11)&0x1f;
     assert(s>=0);
+#ifdef MUPEN64
     emit_writeword(s,(int)&readmem_dword);
     wb_register(rs1[i],i_regs->regmap,i_regs->dirty,i_regs->is32);
-#ifdef MUPEN64
     emit_addimm(FP,(int)&fake_pc-(int)&dynarec_local,0);
     emit_movimm((source[i]>>11)&0x1f,1);
     emit_writeword(0,(int)&PC);
     emit_writebyte(1,(int)&(fake_pc.f.r.nrd));
+#else
+    wb_register(rs1[i],i_regs->regmap,i_regs->dirty,i_regs->is32);
 #endif
     if(copr==9||copr==11||copr==12||copr==13) {
-      emit_readword((int)&last_count,ECX);
+      emit_readword((int)&last_count,HOST_TEMPREG);
       emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
-      emit_add(HOST_CCREG,ECX,HOST_CCREG);
-      emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG);
+      emit_add(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
+      emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
       emit_writeword(HOST_CCREG,(int)&Count);
     }
     // What a mess.  The status register (12) can enable interrupts,
@@ -4089,19 +4181,23 @@ void cop0_assemble(int i,struct regstat *i_regs)
         emit_writeword(HOST_CCREG,(int)&last_count);
         emit_movimm(0,HOST_CCREG);
         emit_storereg(CCREG,HOST_CCREG);
+        if(s!=1)
+          emit_mov(s,1);
         emit_movimm(copr,0);
         emit_call((int)pcsx_mtc0_ds);
         return;
       }
 #endif
-      emit_movimm(start+i*4+4,0);
-      emit_movimm(0,1);
-      emit_writeword(0,(int)&pcaddr);
-      emit_writeword(1,(int)&pending_exception);
+      emit_movimm(start+i*4+4,HOST_TEMPREG);
+      emit_writeword(HOST_TEMPREG,(int)&pcaddr);
+      emit_movimm(0,HOST_TEMPREG);
+      emit_writeword(HOST_TEMPREG,(int)&pending_exception);
     }
     //else if(copr==12&&is_delayslot) emit_call((int)MTC0_R12);
     //else
 #ifdef PCSX
+    if(s!=1)
+      emit_mov(s,1);
     emit_movimm(copr,0);
     emit_call((int)pcsx_mtc0);
 #else
@@ -4110,7 +4206,7 @@ void cop0_assemble(int i,struct regstat *i_regs)
     if(copr==9||copr==11||copr==12||copr==13) {
       emit_readword((int)&Count,HOST_CCREG);
       emit_readword((int)&next_interupt,ECX);
-      emit_addimm(HOST_CCREG,-CLOCK_DIVIDER*ccadj[i],HOST_CCREG);
+      emit_addimm(HOST_CCREG,-CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
       emit_sub(HOST_CCREG,ECX,HOST_CCREG);
       emit_writeword(ECX,(int)&last_count);
       emit_storereg(CCREG,HOST_CCREG);
@@ -4142,7 +4238,7 @@ void cop0_assemble(int i,struct regstat *i_regs)
       emit_readword((int)&last_count,ECX);
       if(i_regs->regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
       emit_add(HOST_CCREG,ECX,HOST_CCREG);
-      emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG);
+      emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
       emit_writeword(HOST_CCREG,(int)&Count);
       emit_call((int)TLBWR_new);
     }
@@ -4163,7 +4259,7 @@ void cop0_assemble(int i,struct regstat *i_regs)
     {
       int count=ccadj[i];
       if(i_regs->regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
-      emit_addimm(HOST_CCREG,CLOCK_DIVIDER*count,HOST_CCREG); // TODO: Should there be an extra cycle here?
+      emit_addimm(HOST_CCREG,CLOCK_ADJUST(count),HOST_CCREG); // TODO: Should there be an extra cycle here?
       emit_jmp((int)jump_eret);
     }
 #endif
@@ -4311,36 +4407,104 @@ void cop2_assemble(int i,struct regstat *i_regs)
   }
 }
 
-void c2op_assemble(int i,struct regstat *i_regs)
+static void c2op_prologue(u_int op,u_int reglist)
+{
+  save_regs_all(reglist);
+#ifdef PCNT
+  emit_movimm(op,0);
+  emit_call((int)pcnt_gte_start);
+#endif
+  emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0); // cop2 regs
+}
+
+static void c2op_epilogue(u_int op,u_int reglist)
+{
+#ifdef PCNT
+  emit_movimm(op,0);
+  emit_call((int)pcnt_gte_end);
+#endif
+  restore_regs_all(reglist);
+}
+
+static void c2op_assemble(int i,struct regstat *i_regs)
 {
   signed char temp=get_reg(i_regs->regmap,-1);
   u_int c2op=source[i]&0x3f;
   u_int hr,reglist=0;
-  int need_flags;
+  int need_flags,need_ir;
   for(hr=0;hr<HOST_REGS;hr++) {
     if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
   }
-  if(i==0||itype[i-1]!=C2OP)
-    save_regs(reglist);
 
   if (gte_handlers[c2op]!=NULL) {
-    int cc=get_reg(i_regs->regmap,CCREG);
-    emit_movimm(source[i],1); // opcode
-    if (cc>=0&&gte_cycletab[c2op])
-      emit_addimm(cc,gte_cycletab[c2op]/2,cc); // XXX: could just adjust ccadj?
-    emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0); // cop2 regs
-    emit_writeword(1,(int)&psxRegs.code);
     need_flags=!(gte_unneeded[i+1]>>63); // +1 because of how liveness detection works
-    assem_debug("gte unneeded %016llx, need_flags %d\n",gte_unneeded[i+1],need_flags);
+    need_ir=(gte_unneeded[i+1]&0xe00)!=0xe00;
+    assem_debug("gte unneeded %016llx, need_flags %d, need_ir %d\n",
+      gte_unneeded[i+1],need_flags,need_ir);
 #ifdef ARMv5_ONLY
     // let's take more risk here
     need_flags=need_flags&&gte_reads_flags;
 #endif
-    emit_call((int)(need_flags?gte_handlers[c2op]:gte_handlers_nf[c2op]));
-  }
+    switch(c2op) {
+      case GTE_MVMVA: {
+        int shift = (source[i] >> 19) & 1;
+        int v  = (source[i] >> 15) & 3;
+        int cv = (source[i] >> 13) & 3;
+        int mx = (source[i] >> 17) & 3;
+        int lm = (source[i] >> 10) & 1;
+        reglist&=0x10ff; // +{r4-r7}
+        c2op_prologue(c2op,reglist);
+        /* r4,r5 = VXYZ(v) packed; r6 = &MX11(mx); r7 = &CV1(cv) */
+        if(v<3)
+          emit_ldrd(v*8,0,4);
+        else {
+          emit_movzwl_indexed(9*4,0,4);  // gteIR
+          emit_movzwl_indexed(10*4,0,6);
+          emit_movzwl_indexed(11*4,0,5);
+          emit_orrshl_imm(6,16,4);
+        }
+        if(mx<3)
+          emit_addimm(0,32*4+mx*8*4,6);
+        else
+          emit_readword((int)&zeromem_ptr,6);
+        if(cv<3)
+          emit_addimm(0,32*4+(cv*8+5)*4,7);
+        else
+          emit_readword((int)&zeromem_ptr,7);
+#ifdef __ARM_NEON__
+        emit_movimm(source[i],1); // opcode
+        emit_call((int)gteMVMVA_part_neon);
+        if(need_flags) {
+          emit_movimm(lm,1);
+          emit_call((int)gteMACtoIR_flags_neon);
+        }
+#else
+        if(cv==3&&shift)
+          emit_call((int)gteMVMVA_part_cv3sh12_arm);
+        else {
+          emit_movimm(shift,1);
+          emit_call((int)(need_flags?gteMVMVA_part_arm:gteMVMVA_part_nf_arm));
+        }
+        if(need_flags||need_ir) {
+          if(need_flags)
+            emit_call((int)(lm?gteMACtoIR_lm1:gteMACtoIR_lm0));
+          else
+            emit_call((int)(lm?gteMACtoIR_lm1_nf:gteMACtoIR_lm0_nf)); // lm0 borked
+        }
+#endif
+        break;
+      }
 
-  if(i>=slen-1||itype[i+1]!=C2OP)
-    restore_regs(reglist);
+      default:
+        reglist&=0x100f;
+        c2op_prologue(c2op,reglist);
+        emit_movimm(source[i],1); // opcode
+        emit_writeword(1,(int)&psxRegs.code);
+        emit_call((int)(need_flags?gte_handlers[c2op]:gte_handlers_nf[c2op]));
+        break;
+    }
+    c2op_epilogue(c2op,reglist);
+  }
 }
 
 void cop1_unusable(int i,struct regstat *i_regs)