From ffb0b9e0f051789f97f5efdcfab0b261e62688f9 Mon Sep 17 00:00:00 2001 From: notaz Date: Sat, 8 Oct 2011 01:55:01 +0300 Subject: [PATCH] drc: implement memory access speculation try to guess RAM mirror we are going to hit --- libpcsxcore/new_dynarec/assem_arm.c | 190 +++++++++++++++++++++++++- libpcsxcore/new_dynarec/new_dynarec.c | 57 +++----- libpcsxcore/psxmem.c | 7 +- 3 files changed, 208 insertions(+), 46 deletions(-) diff --git a/libpcsxcore/new_dynarec/assem_arm.c b/libpcsxcore/new_dynarec/assem_arm.c index 71c397c0..e29f6c3f 100644 --- a/libpcsxcore/new_dynarec/assem_arm.c +++ b/libpcsxcore/new_dynarec/assem_arm.c @@ -1134,7 +1134,6 @@ void emit_addimm(u_int rs,int imm,u_int rt) assert(rs<16); assert(rt<16); if(imm!=0) { - assert(imm>-65536&&imm<65536); u_int armval; if(genimm(imm,&armval)) { assem_debug("add %s,%s,#%d\n",regname[rt],regname[rs],imm); @@ -1143,11 +1142,13 @@ void emit_addimm(u_int rs,int imm,u_int rt) assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rs],imm); output_w32(0xe2400000|rd_rn_rm(rt,rs,0)|armval); }else if(imm<0) { + assert(imm>-65536); assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rs],(-imm)&0xFF00); assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rt],(-imm)&0xFF); output_w32(0xe2400000|rd_rn_imm_shift(rt,rs,(-imm)>>8,8)); output_w32(0xe2400000|rd_rn_imm_shift(rt,rt,(-imm)&0xff,0)); }else{ + assert(imm<65536); assem_debug("add %s,%s,#%d\n",regname[rt],regname[rs],imm&0xFF00); assem_debug("add %s,%s,#%d\n",regname[rt],regname[rt],imm&0xFF); output_w32(0xe2800000|rd_rn_imm_shift(rt,rs,imm>>8,8)); @@ -3698,6 +3699,182 @@ void shift_assemble_arm(int i,struct regstat *i_regs) } } } + +#ifdef PCSX +static void speculate_mov(int rs,int rt) +{ + if(rt!=0) { + smrv_strong_next|=1<>rs1[i])&1) speculate_mov(rs1[i],rt1[i]); + else if((smrv_strong>>rs2[i])&1) speculate_mov(rs2[i],rt1[i]); + else if((smrv_weak>>rs1[i])&1) speculate_mov_weak(rs1[i],rt1[i]); + else if((smrv_weak>>rs2[i])&1) speculate_mov_weak(rs2[i],rt1[i]); + else { + smrv_strong_next&=~(1<=0) { + if(get_final_value(hr,i,&value)) + smrv[rt1[i]]=value; + else smrv[rt1[i]]=constmap[i][hr]; + smrv_strong_next|=1<>rs1[i])&1) speculate_mov(rs1[i],rt1[i]); + else if((smrv_weak>>rs1[i])&1) speculate_mov_weak(rs1[i],rt1[i]); + } + break; + case LOAD: + if(start<0x2000&&(rt1[i]==26||(smrv[rt1[i]]>>24)==0xa0)) { + // special case for BIOS + smrv[rt1[i]]=0xa0000000; + smrv_strong_next|=1<>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst); +#endif +} + +enum { + MTYPE_8000 = 0, + MTYPE_8020, + MTYPE_0000, + MTYPE_A000, + MTYPE_1F80, +}; + +static int get_ptr_mem_type(u_int a) +{ + if(a < 0x00200000) { + if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0)) + // return wrong, must use memhandler for BIOS self-test to pass + // 007 does similar stuff from a00 mirror, weird stuff + return MTYPE_8000; + return MTYPE_0000; + } + if(0x1f800000 <= a && a < 0x1f801000) + return MTYPE_1F80; + if(0x80200000 <= a && a < 0x80800000) + return MTYPE_8020; + if(0xa0000000 <= a && a < 0xa0200000) + return MTYPE_A000; + return MTYPE_8000; +} +#endif + +static int emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override) +{ + int jaddr,type=0; + +#ifdef PCSX + int mr=rs1[i]; + if(((smrv_strong|smrv_weak)>>mr)&1) { + type=get_ptr_mem_type(smrv[mr]); + //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type); + } + else { + // use the mirror we are running on + type=get_ptr_mem_type(start); + //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type); + } + + if(type==MTYPE_8020) { // RAM 80200000+ mirror + emit_andimm(addr,~0x00e00000,HOST_TEMPREG); + addr=*addr_reg_override=HOST_TEMPREG; + type=0; + } + else if(type==MTYPE_0000) { // RAM 0 mirror + emit_orimm(addr,0x80000000,HOST_TEMPREG); + addr=*addr_reg_override=HOST_TEMPREG; + type=0; + } + else if(type==MTYPE_A000) { // RAM A mirror + emit_andimm(addr,~0x20000000,HOST_TEMPREG); + addr=*addr_reg_override=HOST_TEMPREG; + type=0; + } + else if(type==MTYPE_1F80) { // scratchpad + emit_addimm(addr,-0x1f800000,HOST_TEMPREG); + emit_cmpimm(HOST_TEMPREG,0x1000); + jaddr=(int)out; + emit_jc(0); + } +#endif + + if(type==0) + { + emit_cmpimm(addr,RAM_SIZE); + jaddr=(int)out; + #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK + // Hint to branch predictor that the branch is unlikely to be taken + if(rs1[i]>=28) + emit_jno_unlikely(0); + else + #endif + emit_jno(0); + } + + return jaddr; +} + #define shift_assemble shift_assemble_arm void loadlr_assemble_arm(int i,struct regstat *i_regs) @@ -3706,6 +3883,7 @@ void loadlr_assemble_arm(int i,struct regstat *i_regs) int offset; int jaddr=0; int memtarget=0,c=0; + int fastload_reg_override=0; u_int hr,reglist=0; th=get_reg(i_regs->regmap,rt1[i]|64); tl=get_reg(i_regs->regmap,rt1[i]); @@ -3740,9 +3918,7 @@ void loadlr_assemble_arm(int i,struct regstat *i_regs) }else{ emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR } - emit_cmpimm(addr,RAM_SIZE); - jaddr=(int)out; - emit_jno(0); + jaddr=emit_fastpath_cmp_jump(i,temp2,&fastload_reg_override); } else { if (opcode[i]==0x22||opcode[i]==0x26) { @@ -3775,8 +3951,10 @@ void loadlr_assemble_arm(int i,struct regstat *i_regs) } if (opcode[i]==0x22||opcode[i]==0x26) { // LWL/LWR if(!c||memtarget) { + int a=temp2; + if(fastload_reg_override) a=fastload_reg_override; //emit_readword_indexed((int)rdram-0x80000000,temp2,temp2); - emit_readword_indexed_tlb(0,temp2,map,temp2); + emit_readword_indexed_tlb(0,a,map,temp2); if(jaddr) add_stub(LOADW_STUB,jaddr,(int)out,i,temp2,(int)i_regs,ccadj[i],reglist); } else @@ -3803,7 +3981,7 @@ void loadlr_assemble_arm(int i,struct regstat *i_regs) //emit_storereg(rt1[i],tl); // DEBUG } if (opcode[i]==0x1A||opcode[i]==0x1B) { // LDL/LDR - // FIXME: little endian + // FIXME: little endian, fastload_reg_override int temp2h=get_reg(i_regs->regmap,FTEMP|64); if(!c||memtarget) { //if(th>=0) emit_readword_indexed((int)rdram-0x80000000,temp2,temp2h); diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c index 610f86ae..2c9130b3 100644 --- a/libpcsxcore/new_dynarec/new_dynarec.c +++ b/libpcsxcore/new_dynarec/new_dynarec.c @@ -89,6 +89,11 @@ struct ll_entry static uint64_t gte_rt[MAXBLOCK]; static uint64_t gte_unneeded[MAXBLOCK]; static int gte_reads_flags; // gte flag read encountered + static u_int smrv[32]; // speculated MIPS register values + static u_int smrv_strong; // mask or regs that are likely to have correct values + static u_int smrv_weak; // same, but somewhat less likely + static u_int smrv_strong_next; // same, but after current insn executes + static u_int smrv_weak_next; int imm[MAXBLOCK]; u_int ba[MAXBLOCK]; char likely[MAXBLOCK]; @@ -135,7 +140,6 @@ struct ll_entry #else static const u_int using_tlb=0; #endif - static u_int sp_in_mirror; int new_dynarec_did_compile; u_int stop_after_jal; extern u_char restore_candidate[512]; @@ -2893,23 +2897,7 @@ void load_assemble(int i,struct regstat *i_regs) if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE) #endif { - #ifdef PCSX - if(sp_in_mirror&&rs1[i]==29) { - emit_andimm(addr,~0x00e00000,HOST_TEMPREG); - emit_cmpimm(HOST_TEMPREG,RAM_SIZE); - fastload_reg_override=HOST_TEMPREG; - } - else - #endif - emit_cmpimm(addr,RAM_SIZE); - jaddr=(int)out; - #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK - // Hint to branch predictor that the branch is unlikely to be taken - if(rs1[i]>=28) - emit_jno_unlikely(0); - else - #endif - emit_jno(0); + jaddr=emit_fastpath_cmp_jump(i,addr,&fastload_reg_override); } } }else{ // using tlb @@ -3194,14 +3182,7 @@ void store_assemble(int i,struct regstat *i_regs) else addr=s; if(!using_tlb) { if(!c) { - #ifdef PCSX - if(sp_in_mirror&&rs1[i]==29) { - emit_andimm(addr,~0x00e00000,HOST_TEMPREG); - emit_cmpimm(HOST_TEMPREG,RAM_SIZE); - faststore_reg_override=HOST_TEMPREG; - } - else - #endif + #ifndef PCSX #ifdef R29_HACK // Strmnnrmn's speed hack if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE) @@ -3224,6 +3205,9 @@ void store_assemble(int i,struct regstat *i_regs) #endif emit_jno(0); } + #else + jaddr=emit_fastpath_cmp_jump(i,addr,&faststore_reg_override); + #endif } }else{ // using tlb int x=0; @@ -3815,6 +3799,7 @@ void c2ls_assemble(int i,struct regstat *i_regs) int memtarget=0,c=0; int jaddr2=0,jaddr3,type; int agr=AGEN1+(i&1); + int fastio_reg_override=0; u_int hr,reglist=0; u_int copr=(source[i]>>16)&0x1f; s=get_reg(i_regs->regmap,rs1[i]); @@ -3856,22 +3841,24 @@ void c2ls_assemble(int i,struct regstat *i_regs) } else { if(!c) { - emit_cmpimm(offset||c||s<0?ar:s,RAM_SIZE); - jaddr2=(int)out; - emit_jno(0); + jaddr2=emit_fastpath_cmp_jump(i,ar,&fastio_reg_override); } if (opcode[i]==0x32) { // LWC2 #ifdef HOST_IMM_ADDR32 if(c) emit_readword_tlb(constmap[i][s]+offset,-1,tl); else #endif - emit_readword_indexed(0,ar,tl); + int a=ar; + if(fastio_reg_override) a=fastio_reg_override; + emit_readword_indexed(0,a,tl); } if (opcode[i]==0x3a) { // SWC2 #ifdef DESTRUCTIVE_SHIFT if(!offset&&!c&&s>=0) emit_mov(s,ar); #endif - emit_writeword_indexed(tl,0,ar); + int a=ar; + if(fastio_reg_override) a=fastio_reg_override; + emit_writeword_indexed(tl,0,a); } } if(jaddr2) @@ -3976,6 +3963,7 @@ void intcall_assemble(int i,struct regstat *i_regs) void ds_assemble(int i,struct regstat *i_regs) { + speculate_register_values(i); is_delayslot=1; switch(itype[i]) { case ALU: @@ -7914,7 +7902,6 @@ void new_dynarec_clear_full() #ifndef DISABLE_TLB using_tlb=0; #endif - sp_in_mirror=0; for(n=0;n<524288;n++) // 0 .. 0x7FFFFFFF memory_map[n]=-1; for(n=524288;n<526336;n++) // 0x80000000 .. 0x807FFFFF @@ -8025,11 +8012,6 @@ int new_recompile_block(int addr) //assert(((u_int)addr&1)==0); new_dynarec_did_compile=1; #ifdef PCSX - if(!sp_in_mirror&&(signed int)(psxRegs.GPR.n.sp&0xffe00000)>0x80200000&& - 0x10000<=psxRegs.GPR.n.sp&&(psxRegs.GPR.n.sp&~0xe0e00000)>16)!=0x1000)) { diff --git a/libpcsxcore/psxmem.c b/libpcsxcore/psxmem.c index 8c7c37b6..f0a88419 100644 --- a/libpcsxcore/psxmem.c +++ b/libpcsxcore/psxmem.c @@ -68,18 +68,19 @@ int psxMemInit() { memset(psxMemRLUT, 0, 0x10000 * sizeof(void *)); memset(psxMemWLUT, 0, 0x10000 * sizeof(void *)); - psxM = mmap((void *)0x80000000, 0x00220000, + psxM = mmap((void *)0x80000000, 0x00210000, PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); psxP = &psxM[0x200000]; - psxH = &psxM[0x210000]; + psxH = mmap((void *)0x1f800000, 0x00010000, + PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); psxR = mmap((void *)0x9fc00000, 0x80000, PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); if (psxMemRLUT == NULL || psxMemWLUT == NULL || psxM != (void *)0x80000000 || psxR != (void *)0x9fc00000 || - psxP == NULL || psxH == NULL) { + psxP == NULL || psxH != (void *)0x1f800000) { SysMessage(_("Error allocating memory!")); return -1; } -- 2.39.5