Retro/pcem
Watch
1
0
Fork
You've already forked pcem
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Pack decoded ModR/M data into a single word - minor recompiler speedup.

Browse source
This commit is contained in:
SarahW 2016-08-14 16:29:54 +01:00
commit 535604b760
34 changed files with 1263 additions and 1270 deletions
Download patch file Download diff file Expand all files Collapse all files

208
src/x86_ops_mmx_pack.h
View file

@ -3,9 +3,9 @@ static int opPUNPCKLDQ_a16(uint32_t fetchdat)
MMX_ENTER();
fetch_ea_16(fetchdat);
if (mod == 3)
if (cpu_mod == 3)
{
MM[reg].l[1] = MM[rm].l[0];
MM[cpu_reg].l[1] = MM[cpu_rm].l[0];
CLOCK_CYCLES(1);
}
else
@ -13,7 +13,7 @@ static int opPUNPCKLDQ_a16(uint32_t fetchdat)
uint32_t src;
src = readmeml(easeg, eaaddr); if (abrt) return 0;
MM[reg].l[1] = src;
MM[cpu_reg].l[1] = src;
CLOCK_CYCLES(2);
}
@ -24,9 +24,9 @@ static int opPUNPCKLDQ_a32(uint32_t fetchdat)
MMX_ENTER();
fetch_ea_32(fetchdat);
if (mod == 3)
if (cpu_mod == 3)
{
MM[reg].l[1] = MM[rm].l[0];
MM[cpu_reg].l[1] = MM[cpu_rm].l[0];
CLOCK_CYCLES(1);
}
else
@ -34,7 +34,7 @@ static int opPUNPCKLDQ_a32(uint32_t fetchdat)
uint32_t src;
src = readmeml(easeg, eaaddr); if (abrt) return 0;
MM[reg].l[1] = src;
MM[cpu_reg].l[1] = src;
CLOCK_CYCLES(2);
}
@ -49,8 +49,8 @@ static int opPUNPCKHDQ_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat);
MMX_GETSRC();
MM[reg].l[0] = MM[reg].l[1];
MM[reg].l[1] = src.l[1];
MM[cpu_reg].l[0] = MM[cpu_reg].l[1];
MM[cpu_reg].l[1] = src.l[1];
return 0;
}
@ -62,8 +62,8 @@ static int opPUNPCKHDQ_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat);
MMX_GETSRC();
MM[reg].l[0] = MM[reg].l[1];
MM[reg].l[1] = src.l[1];
MM[cpu_reg].l[0] = MM[cpu_reg].l[1];
MM[cpu_reg].l[1] = src.l[1];
return 0;
}
@ -76,14 +76,14 @@ static int opPUNPCKLBW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat);
MMX_GETSRC();
MM[reg].b[7] = src.b[3];
MM[reg].b[6] = MM[reg].b[3];
MM[reg].b[5] = src.b[2];
MM[reg].b[4] = MM[reg].b[2];
MM[reg].b[3] = src.b[1];
MM[reg].b[2] = MM[reg].b[1];
MM[reg].b[1] = src.b[0];
MM[reg].b[0] = MM[reg].b[0];
MM[cpu_reg].b[7] = src.b[3];
MM[cpu_reg].b[6] = MM[cpu_reg].b[3];
MM[cpu_reg].b[5] = src.b[2];
MM[cpu_reg].b[4] = MM[cpu_reg].b[2];
MM[cpu_reg].b[3] = src.b[1];
MM[cpu_reg].b[2] = MM[cpu_reg].b[1];
MM[cpu_reg].b[1] = src.b[0];
MM[cpu_reg].b[0] = MM[cpu_reg].b[0];
return 0;
}
@ -95,14 +95,14 @@ static int opPUNPCKLBW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat);
MMX_GETSRC();
MM[reg].b[7] = src.b[3];
MM[reg].b[6] = MM[reg].b[3];
MM[reg].b[5] = src.b[2];
MM[reg].b[4] = MM[reg].b[2];
MM[reg].b[3] = src.b[1];
MM[reg].b[2] = MM[reg].b[1];
MM[reg].b[1] = src.b[0];
MM[reg].b[0] = MM[reg].b[0];
MM[cpu_reg].b[7] = src.b[3];
MM[cpu_reg].b[6] = MM[cpu_reg].b[3];
MM[cpu_reg].b[5] = src.b[2];
MM[cpu_reg].b[4] = MM[cpu_reg].b[2];
MM[cpu_reg].b[3] = src.b[1];
MM[cpu_reg].b[2] = MM[cpu_reg].b[1];
MM[cpu_reg].b[1] = src.b[0];
MM[cpu_reg].b[0] = MM[cpu_reg].b[0];
return 0;
}
@ -115,14 +115,14 @@ static int opPUNPCKHBW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat);
MMX_GETSRC();
MM[reg].b[0] = MM[reg].b[4];
MM[reg].b[1] = src.b[4];
MM[reg].b[2] = MM[reg].b[5];
MM[reg].b[3] = src.b[5];
MM[reg].b[4] = MM[reg].b[6];
MM[reg].b[5] = src.b[6];
MM[reg].b[6] = MM[reg].b[7];
MM[reg].b[7] = src.b[7];
MM[cpu_reg].b[0] = MM[cpu_reg].b[4];
MM[cpu_reg].b[1] = src.b[4];
MM[cpu_reg].b[2] = MM[cpu_reg].b[5];
MM[cpu_reg].b[3] = src.b[5];
MM[cpu_reg].b[4] = MM[cpu_reg].b[6];
MM[cpu_reg].b[5] = src.b[6];
MM[cpu_reg].b[6] = MM[cpu_reg].b[7];
MM[cpu_reg].b[7] = src.b[7];
return 0;
}
@ -134,14 +134,14 @@ static int opPUNPCKHBW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat);
MMX_GETSRC();
MM[reg].b[0] = MM[reg].b[4];
MM[reg].b[1] = src.b[4];
MM[reg].b[2] = MM[reg].b[5];
MM[reg].b[3] = src.b[5];
MM[reg].b[4] = MM[reg].b[6];
MM[reg].b[5] = src.b[6];
MM[reg].b[6] = MM[reg].b[7];
MM[reg].b[7] = src.b[7];
MM[cpu_reg].b[0] = MM[cpu_reg].b[4];
MM[cpu_reg].b[1] = src.b[4];
MM[cpu_reg].b[2] = MM[cpu_reg].b[5];
MM[cpu_reg].b[3] = src.b[5];
MM[cpu_reg].b[4] = MM[cpu_reg].b[6];
MM[cpu_reg].b[5] = src.b[6];
MM[cpu_reg].b[6] = MM[cpu_reg].b[7];
MM[cpu_reg].b[7] = src.b[7];
return 0;
}
@ -154,10 +154,10 @@ static int opPUNPCKLWD_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat);
MMX_GETSRC();
MM[reg].w[3] = src.w[1];
MM[reg].w[2] = MM[reg].w[1];
MM[reg].w[1] = src.w[0];
MM[reg].w[0] = MM[reg].w[0];
MM[cpu_reg].w[3] = src.w[1];
MM[cpu_reg].w[2] = MM[cpu_reg].w[1];
MM[cpu_reg].w[1] = src.w[0];
MM[cpu_reg].w[0] = MM[cpu_reg].w[0];
return 0;
}
@ -169,10 +169,10 @@ static int opPUNPCKLWD_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat);
MMX_GETSRC();
MM[reg].w[3] = src.w[1];
MM[reg].w[2] = MM[reg].w[1];
MM[reg].w[1] = src.w[0];
MM[reg].w[0] = MM[reg].w[0];
MM[cpu_reg].w[3] = src.w[1];
MM[cpu_reg].w[2] = MM[cpu_reg].w[1];
MM[cpu_reg].w[1] = src.w[0];
MM[cpu_reg].w[0] = MM[cpu_reg].w[0];
return 0;
}
@ -185,10 +185,10 @@ static int opPUNPCKHWD_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat);
MMX_GETSRC();
MM[reg].w[0] = MM[reg].w[2];
MM[reg].w[1] = src.w[2];
MM[reg].w[2] = MM[reg].w[3];
MM[reg].w[3] = src.w[3];
MM[cpu_reg].w[0] = MM[cpu_reg].w[2];
MM[cpu_reg].w[1] = src.w[2];
MM[cpu_reg].w[2] = MM[cpu_reg].w[3];
MM[cpu_reg].w[3] = src.w[3];
return 0;
}
@ -200,10 +200,10 @@ static int opPUNPCKHWD_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat);
MMX_GETSRC();
MM[reg].w[0] = MM[reg].w[2];
MM[reg].w[1] = src.w[2];
MM[reg].w[2] = MM[reg].w[3];
MM[reg].w[3] = src.w[3];
MM[cpu_reg].w[0] = MM[cpu_reg].w[2];
MM[cpu_reg].w[1] = src.w[2];
MM[cpu_reg].w[2] = MM[cpu_reg].w[3];
MM[cpu_reg].w[3] = src.w[3];
return 0;
}
@ -215,16 +215,16 @@ static int opPACKSSWB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat);
MMX_GETSRC();
dst = MM[reg];
dst = MM[cpu_reg];
MM[reg].sb[0] = SSATB(dst.sw[0]);
MM[reg].sb[1] = SSATB(dst.sw[1]);
MM[reg].sb[2] = SSATB(dst.sw[2]);
MM[reg].sb[3] = SSATB(dst.sw[3]);
MM[reg].sb[4] = SSATB(src.sw[0]);
MM[reg].sb[5] = SSATB(src.sw[1]);
MM[reg].sb[6] = SSATB(src.sw[2]);
MM[reg].sb[7] = SSATB(src.sw[3]);
MM[cpu_reg].sb[0] = SSATB(dst.sw[0]);
MM[cpu_reg].sb[1] = SSATB(dst.sw[1]);
MM[cpu_reg].sb[2] = SSATB(dst.sw[2]);
MM[cpu_reg].sb[3] = SSATB(dst.sw[3]);
MM[cpu_reg].sb[4] = SSATB(src.sw[0]);
MM[cpu_reg].sb[5] = SSATB(src.sw[1]);
MM[cpu_reg].sb[6] = SSATB(src.sw[2]);
MM[cpu_reg].sb[7] = SSATB(src.sw[3]);
return 0;
}
@ -235,16 +235,16 @@ static int opPACKSSWB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat);
MMX_GETSRC();
dst = MM[reg];
dst = MM[cpu_reg];
MM[reg].sb[0] = SSATB(dst.sw[0]);
MM[reg].sb[1] = SSATB(dst.sw[1]);
MM[reg].sb[2] = SSATB(dst.sw[2]);
MM[reg].sb[3] = SSATB(dst.sw[3]);
MM[reg].sb[4] = SSATB(src.sw[0]);
MM[reg].sb[5] = SSATB(src.sw[1]);
MM[reg].sb[6] = SSATB(src.sw[2]);
MM[reg].sb[7] = SSATB(src.sw[3]);
MM[cpu_reg].sb[0] = SSATB(dst.sw[0]);
MM[cpu_reg].sb[1] = SSATB(dst.sw[1]);
MM[cpu_reg].sb[2] = SSATB(dst.sw[2]);
MM[cpu_reg].sb[3] = SSATB(dst.sw[3]);
MM[cpu_reg].sb[4] = SSATB(src.sw[0]);
MM[cpu_reg].sb[5] = SSATB(src.sw[1]);
MM[cpu_reg].sb[6] = SSATB(src.sw[2]);
MM[cpu_reg].sb[7] = SSATB(src.sw[3]);
return 0;
}
@ -256,16 +256,16 @@ static int opPACKUSWB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat);
MMX_GETSRC();
dst = MM[reg];
dst = MM[cpu_reg];
MM[reg].b[0] = USATB(dst.sw[0]);
MM[reg].b[1] = USATB(dst.sw[1]);
MM[reg].b[2] = USATB(dst.sw[2]);
MM[reg].b[3] = USATB(dst.sw[3]);
MM[reg].b[4] = USATB(src.sw[0]);
MM[reg].b[5] = USATB(src.sw[1]);
MM[reg].b[6] = USATB(src.sw[2]);
MM[reg].b[7] = USATB(src.sw[3]);
MM[cpu_reg].b[0] = USATB(dst.sw[0]);
MM[cpu_reg].b[1] = USATB(dst.sw[1]);
MM[cpu_reg].b[2] = USATB(dst.sw[2]);
MM[cpu_reg].b[3] = USATB(dst.sw[3]);
MM[cpu_reg].b[4] = USATB(src.sw[0]);
MM[cpu_reg].b[5] = USATB(src.sw[1]);
MM[cpu_reg].b[6] = USATB(src.sw[2]);
MM[cpu_reg].b[7] = USATB(src.sw[3]);
return 0;
}
@ -276,16 +276,16 @@ static int opPACKUSWB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat);
MMX_GETSRC();
dst = MM[reg];
dst = MM[cpu_reg];
MM[reg].b[0] = USATB(dst.sw[0]);
MM[reg].b[1] = USATB(dst.sw[1]);
MM[reg].b[2] = USATB(dst.sw[2]);
MM[reg].b[3] = USATB(dst.sw[3]);
MM[reg].b[4] = USATB(src.sw[0]);
MM[reg].b[5] = USATB(src.sw[1]);
MM[reg].b[6] = USATB(src.sw[2]);
MM[reg].b[7] = USATB(src.sw[3]);
MM[cpu_reg].b[0] = USATB(dst.sw[0]);
MM[cpu_reg].b[1] = USATB(dst.sw[1]);
MM[cpu_reg].b[2] = USATB(dst.sw[2]);
MM[cpu_reg].b[3] = USATB(dst.sw[3]);
MM[cpu_reg].b[4] = USATB(src.sw[0]);
MM[cpu_reg].b[5] = USATB(src.sw[1]);
MM[cpu_reg].b[6] = USATB(src.sw[2]);
MM[cpu_reg].b[7] = USATB(src.sw[3]);
return 0;
}
@ -297,12 +297,12 @@ static int opPACKSSDW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat);
MMX_GETSRC();
dst = MM[reg];
dst = MM[cpu_reg];
MM[reg].sw[0] = SSATW(dst.sl[0]);
MM[reg].sw[1] = SSATW(dst.sl[1]);
MM[reg].sw[2] = SSATW(src.sl[0]);
MM[reg].sw[3] = SSATW(src.sl[1]);
MM[cpu_reg].sw[0] = SSATW(dst.sl[0]);
MM[cpu_reg].sw[1] = SSATW(dst.sl[1]);
MM[cpu_reg].sw[2] = SSATW(src.sl[0]);
MM[cpu_reg].sw[3] = SSATW(src.sl[1]);
return 0;
}
@ -313,12 +313,12 @@ static int opPACKSSDW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat);
MMX_GETSRC();
dst = MM[reg];
dst = MM[cpu_reg];
MM[reg].sw[0] = SSATW(dst.sl[0]);
MM[reg].sw[1] = SSATW(dst.sl[1]);
MM[reg].sw[2] = SSATW(src.sl[0]);
MM[reg].sw[3] = SSATW(src.sl[1]);
MM[cpu_reg].sw[0] = SSATW(dst.sl[0]);
MM[cpu_reg].sw[1] = SSATW(dst.sl[1]);
MM[cpu_reg].sw[2] = SSATW(src.sl[0]);
MM[cpu_reg].sw[3] = SSATW(src.sl[1]);
return 0;
}