diff --git a/core/emitter/x86_matcher.h b/core/emitter/x86_matcher.h
index a1ec8bd41..9bcaa05fc 100644
--- a/core/emitter/x86_matcher.h
+++ b/core/emitter/x86_matcher.h
@@ -20,7 +20,7 @@ enum x86_op_params
 
 encoded_type pg_none = {pg_NONE};
 
-encoded_type __fastcall param_type(x86_Label* lbl)
+encoded_type param_type(x86_Label* lbl)
 {
 	encoded_type rv;
 	//Return pg_MEM_Rel32/pg_MEM_Rel16/pg_MEM_Rel8
@@ -33,7 +33,7 @@ encoded_type __fastcall param_type(x86_Label* lbl)
 	rv.ptr_type=1;
 	return rv;
 }
-encoded_type __fastcall param_type(x86_ptr_imm ptr)
+encoded_type param_type(x86_ptr_imm ptr)
 {
 	encoded_type rv;
 	//Return pg_MEM_Rel32.Due to relocation we cant optimise to 16/8 in one pass ...
@@ -43,7 +43,7 @@ encoded_type __fastcall param_type(x86_ptr_imm ptr)
 	rv.ptr_type=0;
 	return rv;
 }
-encoded_type __fastcall param_type(x86_mrm_t& modrm)
+encoded_type param_type(x86_mrm_t& modrm)
 {
 	encoded_type rv;
 	rv.modrm=modrm;
@@ -51,7 +51,7 @@ encoded_type __fastcall param_type(x86_mrm_t& modrm)
 	rv.type=pg_ModRM;
 	return rv;
 }
-encoded_type __fastcall param_type(x86_reg reg)
+encoded_type param_type(x86_reg reg)
 {
 	encoded_type rv;
 	rv.reg=REG_ID(reg);
@@ -65,7 +65,7 @@ encoded_type __fastcall param_type(x86_reg reg)
 	return rv;
 }
 
-encoded_type __fastcall param_type(u32 imm)
+encoded_type param_type(u32 imm)
 {
 	encoded_type rv;
 	rv.imm=imm;
@@ -82,7 +82,7 @@ encoded_type __fastcall param_type(u32 imm)
 
 	return rv;
 }
-void __fastcall Match_opcode(x86_block* block,const x86_opcode* ops,encoded_type pg1,encoded_type pg2,encoded_type pg3)
+void Match_opcode(x86_block* block,const x86_opcode* ops,encoded_type pg1,encoded_type pg2,encoded_type pg3)
 {
 	block->opcode_count++;
 	const x86_opcode* match=0;
diff --git a/core/emitter/x86_op_encoder.h b/core/emitter/x86_op_encoder.h
index fd6c03244..0f3e9db9c 100644
--- a/core/emitter/x86_op_encoder.h
+++ b/core/emitter/x86_op_encoder.h
@@ -61,7 +61,7 @@
 #include "build.h"
 
 #if BUILD_COMPILER == COMPILER_GCC
-	#define __fastcall 
+	#define __fastcall BALLZZ!!
 #endif
 
 
@@ -173,7 +173,7 @@ struct encoded_type
 
 struct x86_opcode;
 
-typedef void __fastcall x86_opcode_encoderFP(x86_block* block,const x86_opcode* op,encoded_type* p1,encoded_type* p2,u32 p3);
+typedef void x86_opcode_encoderFP(x86_block* block,const x86_opcode* op,encoded_type* p1,encoded_type* p2,u32 p3);
 
 //enc_param_none is alower w/ params set to implicit registers (ie , mov eax,xxxx is enc_imm , pg1:pg_EAX , pg2:pg_imm
 
@@ -190,7 +190,7 @@ struct x86_opcode
 };
 
 //mod|reg|rm
-void __fastcall encode_modrm(x86_block* block,encoded_type* mrm, u32 extra)
+void encode_modrm(x86_block* block,encoded_type* mrm, u32 extra)
 {
 	if (mrm->type != pg_ModRM)
 	{
@@ -213,7 +213,7 @@ void __fastcall encode_modrm(x86_block* block,encoded_type* mrm, u32 extra)
 }
 #ifdef X64
 //x64 stuff
-void __fastcall encode_rex(x86_block* block,encoded_type* mrm,u32 mrm_reg,u32 ofe=0)
+void encode_rex(x86_block* block,encoded_type* mrm,u32 mrm_reg,u32 ofe=0)
 {
 	u32 flags = (ofe>>3) & 1; //opcode field extension
 
@@ -237,7 +237,7 @@ void __fastcall encode_rex(x86_block* block,encoded_type* mrm,u32 mrm_reg,u32 of
 
 //Encoding function (partially) specialised by templates to gain speed :)
 template < enc_param enc_1,enc_imm enc_2,u32 sz,x86_operand_size enc_op_size>
-void __fastcall x86_encode_opcode_tmpl(x86_block* block, const x86_opcode* op, encoded_type* p1,encoded_type* p2,u32 p3)
+void x86_encode_opcode_tmpl(x86_block* block, const x86_opcode* op, encoded_type* p1,encoded_type* p2,u32 p3)
 {
 	//printf("Encoding : ");
 
diff --git a/core/hw/arm7/virt_arm.cpp b/core/hw/arm7/virt_arm.cpp
index d55c4dda1..f9aef686b 100644
--- a/core/hw/arm7/virt_arm.cpp
+++ b/core/hw/arm7/virt_arm.cpp
@@ -311,7 +311,7 @@ void virt_arm_init()
 	VARM::virt_arm_init();
 }
 
-u32 __fastcall virt_arm_op(u32 opcode)
+u32 DYNACALL virt_arm_op(u32 opcode)
 {
 	return VARM::virt_arm_op(opcode);
 }
diff --git a/core/hw/sh4/dyna/blockmanager.cpp b/core/hw/sh4/dyna/blockmanager.cpp
index eec9cdd4b..cb84ca4c3 100644
--- a/core/hw/sh4/dyna/blockmanager.cpp
+++ b/core/hw/sh4/dyna/blockmanager.cpp
@@ -87,24 +87,24 @@ blkmap_t blkmap;
 u32 bm_gc_luc,bm_gcf_luc;
 
 
-#define FPCA(x) ((DynarecCodeEntry*&)sh4rcb.fpcb[(x>>1)&(8*1024*1024-1)])
+#define FPCA(x) ((DynarecCodeEntryPtr&)sh4rcb.fpcb[(x>>1)&(8*1024*1024-1)])
 
-DynarecCodeEntry* DYNACALL bm_GetCode(u32 addr)
+DynarecCodeEntryPtr DYNACALL bm_GetCode(u32 addr)
 {
 	//rdv_FailedToFindBlock_pc=addr;
-	DynarecCodeEntry* rv=(DynarecCodeEntry*)FPCA(addr);
+	DynarecCodeEntryPtr rv=(DynarecCodeEntryPtr)FPCA(addr);
 
-	return (DynarecCodeEntry*)rv;
+	return (DynarecCodeEntryPtr)rv;
 }
 
-DynarecCodeEntry* DYNACALL bm_GetCode2(u32 addr)
+DynarecCodeEntryPtr DYNACALL bm_GetCode2(u32 addr)
 {
-	return (DynarecCodeEntry*)bm_GetCode(addr);
+	return (DynarecCodeEntryPtr)bm_GetCode(addr);
 }
 
 RuntimeBlockInfo* DYNACALL bm_GetBlock(u32 addr)
 {
-	DynarecCodeEntry* cde=bm_GetCode(addr);
+	DynarecCodeEntryPtr cde=bm_GetCode(addr);
 
 	if (cde==ngen_FailedToFindBlock)
 		return 0;
diff --git a/core/hw/sh4/dyna/blockmanager.h b/core/hw/sh4/dyna/blockmanager.h
index c8876fc90..8f87dc03a 100644
--- a/core/hw/sh4/dyna/blockmanager.h
+++ b/core/hw/sh4/dyna/blockmanager.h
@@ -5,12 +5,12 @@
 #include "decoder.h"
 #pragma once
 
-typedef void DynarecCodeEntry();
+typedef void (*DynarecCodeEntryPtr)();
 
 struct RuntimeBlockInfo_Core
 {
 	u32 addr;
-	DynarecCodeEntry* code;
+	DynarecCodeEntryPtr code;
 	u32 lookups;
 };
 
@@ -81,13 +81,13 @@ struct CachedBlockInfo: RuntimeBlockInfo_Core
 void bm_WriteBlockMap(const string& file);
 
 
-DynarecCodeEntry* DYNACALL bm_GetCode(u32 addr);
+DynarecCodeEntryPtr DYNACALL bm_GetCode(u32 addr);
 
 
 #if HOST_OS==OS_LINUX
 extern "C" {
 #endif
-DynarecCodeEntry* DYNACALL bm_GetCode2(u32 addr);
+DynarecCodeEntryPtr DYNACALL bm_GetCode2(u32 addr);
 #if HOST_OS==OS_LINUX
 }
 #endif
diff --git a/core/hw/sh4/dyna/driver.cpp b/core/hw/sh4/dyna/driver.cpp
index 669c39a95..2d22045b7 100644
--- a/core/hw/sh4/dyna/driver.cpp
+++ b/core/hw/sh4/dyna/driver.cpp
@@ -208,7 +208,7 @@ void RuntimeBlockInfo::Setup(u32 rpc,fpscr_t rfpu_cfg)
 	AnalyseBlock(this);
 }
 
-DynarecCodeEntry* rdv_CompilePC()
+DynarecCodeEntryPtr rdv_CompilePC()
 {
 	u32 pc=next_pc;
 
@@ -240,7 +240,7 @@ DynarecCodeEntry* rdv_CompilePC()
 	return rv->code;
 }
 
-DynarecCodeEntry* DYNACALL rdv_FailedToFindBlock(u32 pc)
+DynarecCodeEntryPtr DYNACALL rdv_FailedToFindBlock(u32 pc)
 {
 	//printf("rdv_FailedToFindBlock ~ %08X\n",pc);
 	next_pc=pc;
@@ -269,25 +269,25 @@ u32 DYNACALL rdv_DoInterrupts(void* block_cpde)
 	return next_pc;
 }
 
-DynarecCodeEntry* DYNACALL rdv_BlockCheckFail(u32 pc)
+DynarecCodeEntryPtr DYNACALL rdv_BlockCheckFail(u32 pc)
 {
 	next_pc=pc;
 	recSh4_ClearCache();
 	return rdv_CompilePC();
 }
 
-DynarecCodeEntry* rdv_FindCode()
+DynarecCodeEntryPtr rdv_FindCode()
 {
-	DynarecCodeEntry* rv=bm_GetCode(next_pc);
+	DynarecCodeEntryPtr rv=bm_GetCode(next_pc);
 	if (rv==ngen_FailedToFindBlock)
 		return 0;
 	
 	return rv;
 }
 
-DynarecCodeEntry* rdv_FindOrCompile()
+DynarecCodeEntryPtr rdv_FindOrCompile()
 {
-	DynarecCodeEntry* rv=bm_GetCode(next_pc);
+	DynarecCodeEntryPtr rv=bm_GetCode(next_pc);
 	if (rv==ngen_FailedToFindBlock)
 		rv=rdv_CompilePC();
 	
@@ -324,7 +324,7 @@ void* DYNACALL rdv_LinkBlock(u8* code,u32 dpc)
 			next_pc=rbi->NextBlock;
 	}
 
-	DynarecCodeEntry* rv=rdv_FindOrCompile();
+	DynarecCodeEntryPtr rv=rdv_FindOrCompile();
 
 	bool do_link=bm_GetBlock(code)==rbi;
 
diff --git a/core/hw/sh4/dyna/ngen.h b/core/hw/sh4/dyna/ngen.h
index 1a7592fa8..d41fd60d7 100644
--- a/core/hw/sh4/dyna/ngen.h
+++ b/core/hw/sh4/dyna/ngen.h
@@ -63,15 +63,15 @@ void* emit_GetCCPtr();
 void emit_SetBaseAddr();
 
 //Called from ngen_FailedToFindBlock
-DynarecCodeEntry* DYNACALL rdv_FailedToFindBlock(u32 pc);
+DynarecCodeEntryPtr DYNACALL rdv_FailedToFindBlock(u32 pc);
 //Called when a block check failed, and the block needs to be invalidated
-DynarecCodeEntry* DYNACALL rdv_BlockCheckFail(u32 pc);
+DynarecCodeEntryPtr DYNACALL rdv_BlockCheckFail(u32 pc);
 //Called to compile code @pc
-DynarecCodeEntry* rdv_CompilePC();
+DynarecCodeEntryPtr rdv_CompilePC();
 //Returns 0 if there is no code @pc, code ptr otherwise
-DynarecCodeEntry* rdv_FindCode();
+DynarecCodeEntryPtr rdv_FindCode();
 //Finds or compiles code @pc
-DynarecCodeEntry* rdv_FindOrCompile();
+DynarecCodeEntryPtr rdv_FindOrCompile();
 
 //code -> pointer to code of block, dpc -> if dynamic block, pc. if cond, 0 for next, 1 for branch
 void* DYNACALL rdv_LinkBlock(u8* code,u32 dpc);
diff --git a/core/hw/sh4/sh4_mem.h b/core/hw/sh4/sh4_mem.h
index 5b3a051ca..f9f132f90 100644
--- a/core/hw/sh4/sh4_mem.h
+++ b/core/hw/sh4/sh4_mem.h
@@ -4,8 +4,6 @@
 //main system mem
 extern VArray2 mem_b;
 
-#define MEMCALL __fastcall
-
 #include "hw/mem/_vmem.h"
 #include "modules/mmu.h"
 
diff --git a/core/hw/sh4/sh4_mmr.h b/core/hw/sh4/sh4_mmr.h
index 740c9969f..2bfead711 100644
--- a/core/hw/sh4/sh4_mmr.h
+++ b/core/hw/sh4/sh4_mmr.h
@@ -31,7 +31,7 @@ void WriteMem_P4(u32 addr,u32 data,u32 sz);
 //Area7
 u32 ReadMem_area7(u32 addr,u32 sz);
 void WriteMem_area7(u32 addr,u32 data,u32 sz);
-void __fastcall WriteMem_sq_32(u32 address,u32 data);*/
+void DYNACALL WriteMem_sq_32(u32 address,u32 data);*/
 
 //Init/Res/Term
 void sh4_mmr_init();
diff --git a/core/linux/common.cpp b/core/linux/common.cpp
index cdd536c2b..98a08b5ff 100644
--- a/core/linux/common.cpp
+++ b/core/linux/common.cpp
@@ -47,6 +47,9 @@ struct sigcontext uc_mcontext;
 #define GET_PC_FROM_CONTEXT(c) (((ucontext_t *)(c))->uc_mcontext.eip)
 #else
 #define GET_PC_FROM_CONTEXT(c) (((ucontext_t *)(c))->uc_mcontext.gregs[REG_EIP])
+#define GET_ESP_FROM_CONTEXT(c) (((ucontext_t *)(c))->uc_mcontext.gregs[REG_ESP])
+#define GET_EAX_FROM_CONTEXT(c) (((ucontext_t *)(c))->uc_mcontext.gregs[REG_EAX])
+#define GET_ECX_FROM_CONTEXT(c) (((ucontext_t *)(c))->uc_mcontext.gregs[REG_ECX])
 #endif
 #else
 #error fix ->pc support
@@ -54,6 +57,7 @@ struct sigcontext uc_mcontext;
 
 #include "hw/sh4/dyna/ngen.h"
 
+bool ngen_Rewrite(unat& addr,unat retadr,unat acc);
 u32* ngen_readm_fail_v2(u32* ptr,u32* regs,u32 saddr);
 bool VramLockedWrite(u8* address);
 bool BM_LockedWrite(u8* address);
@@ -68,12 +72,24 @@ void fault_handler (int sn, siginfo_t * si, void *ctxr)
 	
 	if (VramLockedWrite((u8*)si->si_addr) || BM_LockedWrite((u8*)si->si_addr))
 		return;
-#if !defined( HOST_NO_REC) && HOST_CPU==CPU_ARM
-	else if (dyna_cde)
-	{
-		GET_PC_FROM_CONTEXT(ctxr)=(u32)ngen_readm_fail_v2((u32*)GET_PC_FROM_CONTEXT(ctxr),(u32*)&(ctx->uc_mcontext.arm_r0),(unat)si->si_addr);
-	}
-#endif
+	#if !defined(HOST_NO_REC)
+		#if HOST_CPU==CPU_ARM
+			else if (dyna_cde)
+			{
+				GET_PC_FROM_CONTEXT(ctxr)=(u32)ngen_readm_fail_v2((u32*)GET_PC_FROM_CONTEXT(ctxr),(u32*)&(ctx->uc_mcontext.arm_r0),(unat)si->si_addr);
+			}
+		#elif HOST_CPU==CPU_X86
+			else if ( ngen_Rewrite((unat&)GET_PC_FROM_CONTEXT(ctxr),*(unat*)GET_ESP_FROM_CONTEXT(ctxr),GET_EAX_FROM_CONTEXT(ctxr)) )
+			{
+				//remove the call from call stack
+				GET_ESP_FROM_CONTEXT(ctxr)+=4;
+				//restore the addr from eax to ecx so its valid again
+				GET_ECX_FROM_CONTEXT(ctxr)=GET_EAX_FROM_CONTEXT(ctxr);
+			}
+		#else
+			#error JIT: Not supported arch
+		#endif
+	#endif
 	else
 	{
 		printf("SIGSEGV @ fault_handler+0x%08X ... %08X -> was not in vram\n",GET_PC_FROM_CONTEXT(ctxr)-(u32)fault_handler,si->si_addr);
diff --git a/core/rec-ARM/arm_dyna.cpp b/core/rec-ARM/arm_dyna.cpp
index 7cfff89cd..45325e326 100755
--- a/core/rec-ARM/arm_dyna.cpp
+++ b/core/rec-ARM/arm_dyna.cpp
@@ -1979,7 +1979,7 @@ __default:
 void ngen_Compile(RuntimeBlockInfo* block,bool force_checks, bool reset, bool staging,bool optimise)
 {
 	//printf("Compile: %08X, %d, %d\n",block->addr,staging,optimise);
-	block->code=(DynarecCodeEntry*)EMIT_GET_PTR();
+	block->code=(DynarecCodeEntryPtr)EMIT_GET_PTR();
 
 	//StoreImms(r0,r1,(u32)&last_run_block,(u32)code); //useful when code jumps to random locations ...
 	++blockno;
diff --git a/core/rec-x86/win86_driver.cpp b/core/rec-x86/win86_driver.cpp
index 4c85b6b24..a7c3d6e99 100644
--- a/core/rec-x86/win86_driver.cpp
+++ b/core/rec-x86/win86_driver.cpp
@@ -301,7 +301,7 @@ void ngen_Compile(RuntimeBlockInfo* block,bool force_checks, bool reset, bool st
 	x86e->x86_size=emit_FreeSpace();
 	x86e->do_realloc=false;
 
-	block->code=(DynarecCodeEntry*)emit_GetCCPtr();
+	block->code=(DynarecCodeEntryPtr)emit_GetCCPtr();
 
 	x86e->Emit(op_add32,&memops_t,block->memops);
 	x86e->Emit(op_add32,&memops_l,block->linkedmemops);
diff --git a/core/types.h b/core/types.h
index 62d8ee578..e46a604f5 100644
--- a/core/types.h
+++ b/core/types.h
@@ -13,10 +13,8 @@
 
 #if BUILD_COMPILER==COMPILER_VC
 #define DYNACALL  __fastcall
-#define DYNACALL_T
 #else
-#define DYNACALL
-#define DYNACALL_T __attribute__((fastcall))
+#define DYNACALL __attribute__((fastcall))
 #endif
 
 #if BUILD_COMPILER==COMPILER_VC
@@ -511,7 +509,7 @@ void os_DebugBreak();
 #define stricmp strcasecmp
 #endif
 
-//#define __fastcall <nothing useful is here "" must not happen ever>
+#define __fastcall <nothing useful is here "" must not happen ever>
 #ifndef STRIP_TEXT
 #define verify(x) if((x)==false){ msgboxf("Verify Failed  : " #x "\n in %s -> %s : %d \n",MBX_ICONERROR,(__FUNCTION__),(__FILE__),__LINE__); dbgbreak;}
 #define die(reason) { msgboxf("Fatal error : %s\n in %s -> %s : %d \n",MBX_ICONERROR,(reason),(__FUNCTION__),(__FILE__),__LINE__); dbgbreak;}
diff --git a/core/windows/win86_driver.cpp b/core/windows/win86_driver.cpp
index 41efb2630..384df8f51 100644
--- a/core/windows/win86_driver.cpp
+++ b/core/windows/win86_driver.cpp
@@ -299,7 +299,7 @@ void ngen_Compile(RuntimeBlockInfo* block,bool force_checks, bool reset, bool st
 	x86e->x86_size=emit_FreeSpace();
 	x86e->do_realloc=false;
 
-	block->code=(DynarecCodeEntry*)emit_GetCCPtr();
+	block->code=(DynarecCodeEntryPtr)emit_GetCCPtr();
 
 	x86e->Emit(op_add32,&memops_t,block->memops);
 	x86e->Emit(op_add32,&memops_l,block->linkedmemops);