From 198df76f1e6e57343ec7e17d46cc45395bf225e1 Mon Sep 17 00:00:00 2001 From: notaz Date: Wed, 22 Jun 2011 18:52:24 +0300 Subject: drc: merge Ari64's patch: 18_loop_preload_fix --- libpcsxcore/new_dynarec/new_dynarec.c | 39 ++++++++++++++++++++++++----------- 1 file changed, 27 insertions(+), 12 deletions(-) (limited to 'libpcsxcore') diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c index 74b64b7..b0c0b4c 100644 --- a/libpcsxcore/new_dynarec/new_dynarec.c +++ b/libpcsxcore/new_dynarec/new_dynarec.c @@ -9880,7 +9880,7 @@ int new_recompile_block(int addr) // If a register is allocated during a loop, try to allocate it for the // entire loop, if possible. This avoids loading/storing registers // inside of the loop. - + signed char f_regmap[HOST_REGS]; clear_all_regs(f_regmap); for(i=0;i>2; if(t>0&&(itype[t-1]!=UJUMP&&itype[t-1]!=RJUMP&&itype[t-1]!=CJUMP&&itype[t-1]!=SJUMP&&itype[t-1]!=FJUMP)) // loop_preload can't handle jumps into delay slots - if(t<2||(itype[t-2]!=UJUMP)) // call/ret assumes no registers allocated + if(t<2||(itype[t-2]!=UJUMP&&itype[t-2]!=RJUMP)||rt1[t-2]!=31) // call/ret assumes no registers allocated for(hr=0;hr64) { @@ -9953,7 +9953,7 @@ int new_recompile_block(int addr) // a mov, which is of negligible benefit. So such cases are // skipped below. if(f_regmap[hr]>0) { - if(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0) { + if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) { int r=f_regmap[hr]; for(j=t;j<=i;j++) { @@ -9992,7 +9992,7 @@ int new_recompile_block(int addr) break; } // call/ret fast path assumes no registers allocated - if(k>2&&(itype[k-3]==UJUMP||itype[k-3]==RJUMP)) { + if(k>2&&(itype[k-3]==UJUMP||itype[k-3]==RJUMP)&&rt1[k-3]==31) { break; } if(r>63) { @@ -10140,7 +10140,7 @@ int new_recompile_block(int addr) } } }else{ - int count=0; + // Non branch or undetermined branch target for(hr=0;hr>2; if(t=earliest_available[hr]) { - // Score a point for hoisting loop invariant - if(t1&&itype[t-2]!=UJUMP&&itype[t-2]!=RJUMP)||(t>1&&rt1[t-2]!=31)) { // call/ret assumes no registers allocated + // Score a point for hoisting loop invariant + if(t>2; - if(t==loop_start[maxscore]) regs[t].regmap_entry[maxscore]=reg; + if(t==loop_start[maxscore]) { + if(t==1||(t>1&&itype[t-2]!=UJUMP&&itype[t-2]!=RJUMP)||(t>1&&rt1[t-2]!=31)) // call/ret assumes no registers allocated + regs[t].regmap_entry[maxscore]=reg; + } } else { @@ -10439,6 +10450,7 @@ int new_recompile_block(int addr) } } } + // Preload target address for load instruction (non-constant) if(itype[i+1]==LOAD&&rs1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) { if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0) { @@ -10455,6 +10467,7 @@ int new_recompile_block(int addr) } } } + // Load source into target register if(lt1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) { if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0) { @@ -10471,6 +10484,7 @@ int new_recompile_block(int addr) } } } + // Preload map address #ifndef HOST_IMM_ADDR32 if(itype[i+1]==LOAD||itype[i+1]==LOADLR||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS||itype[i+1]==C2LS) { hr=get_reg(regs[i+1].regmap,TLREG); @@ -10510,6 +10524,7 @@ int new_recompile_block(int addr) } } #endif + // Address for store instruction (non-constant) if(itype[i+1]==STORE||itype[i+1]==STORELR ||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2 if(get_reg(regs[i+1].regmap,rs1[i+1])<0) { -- cgit v1.2.3