Sprinter/mame
2
0
Fork 0
mirror of https://github.com/holub/mame synced 2025-04-27 02:33:13 +03:00
Code Issues Actions 15 Packages Projects Releases Wiki Activity

i386: sse opcodes improvements [Samuele Zannoli]

Browse Source 
- add opcodes MOVHLPS MOVLHPS
 - safer implementation of PACKUSWB PACKSSDW SHUFPS UNPCKLPS UNPCKHPS

The safer implementation is needed in cases where the source and destination registers are the same.
This commit is contained in:
yz70s 2015-01-25 22:07:51 +01:00
parent f2a9d4e90a
commit 35947ff256
1 changed files with 80 additions and 46 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

126
src/emu/cpu/i386/pentops.inc
View File

@ -2249,26 +2249,30 @@ void i386_device::mmx_packuswb_r64_rm64() // Opcode 0f 67
MMXPROLOG();
UINT8 modrm = FETCH();
if( modrm >= 0xc0 ) {
MMX_REG ds, sd;
int s,d;
s=modrm & 0x7;
d=(modrm >> 3) & 0x7;
MMX(d).b[0]=SaturatedSignedWordToUnsignedByte(MMX(d).s[0]);
MMX(d).b[1]=SaturatedSignedWordToUnsignedByte(MMX(d).s[1]);
MMX(d).b[2]=SaturatedSignedWordToUnsignedByte(MMX(d).s[2]);
MMX(d).b[3]=SaturatedSignedWordToUnsignedByte(MMX(d).s[3]);
MMX(d).b[4]=SaturatedSignedWordToUnsignedByte(MMX(s).s[0]);
MMX(d).b[5]=SaturatedSignedWordToUnsignedByte(MMX(s).s[1]);
MMX(d).b[6]=SaturatedSignedWordToUnsignedByte(MMX(s).s[2]);
MMX(d).b[7]=SaturatedSignedWordToUnsignedByte(MMX(s).s[3]);
ds.q = MMX(d).q;
sd.q = MMX(s).q;
MMX(d).b[0]=SaturatedSignedWordToUnsignedByte(ds.s[0]);
MMX(d).b[1]=SaturatedSignedWordToUnsignedByte(ds.s[1]);
MMX(d).b[2]=SaturatedSignedWordToUnsignedByte(ds.s[2]);
MMX(d).b[3]=SaturatedSignedWordToUnsignedByte(ds.s[3]);
MMX(d).b[4]=SaturatedSignedWordToUnsignedByte(sd.s[0]);
MMX(d).b[5]=SaturatedSignedWordToUnsignedByte(sd.s[1]);
MMX(d).b[6]=SaturatedSignedWordToUnsignedByte(sd.s[2]);
MMX(d).b[7]=SaturatedSignedWordToUnsignedByte(sd.s[3]);
} else {
MMX_REG s;
MMX_REG s,t;
int d=(modrm >> 3) & 0x7;
UINT32 ea = GetEA(modrm, 0);
READMMX(ea, s);
MMX(d).b[0]=SaturatedSignedWordToUnsignedByte(MMX(d).s[0]);
MMX(d).b[1]=SaturatedSignedWordToUnsignedByte(MMX(d).s[1]);
MMX(d).b[2]=SaturatedSignedWordToUnsignedByte(MMX(d).s[2]);
MMX(d).b[3]=SaturatedSignedWordToUnsignedByte(MMX(d).s[3]);
t.q = MMX(d).q;
MMX(d).b[0]=SaturatedSignedWordToUnsignedByte(t.s[0]);
MMX(d).b[1]=SaturatedSignedWordToUnsignedByte(t.s[1]);
MMX(d).b[2]=SaturatedSignedWordToUnsignedByte(t.s[2]);
MMX(d).b[3]=SaturatedSignedWordToUnsignedByte(t.s[3]);
MMX(d).b[4]=SaturatedSignedWordToUnsignedByte(s.s[0]);
MMX(d).b[5]=SaturatedSignedWordToUnsignedByte(s.s[1]);
MMX(d).b[6]=SaturatedSignedWordToUnsignedByte(s.s[2]);
@ -2362,21 +2366,30 @@ void i386_device::mmx_packssdw_r64_rm64() // Opcode 0f 6b
UINT8 modrm = FETCH();
if( modrm >= 0xc0 ) {
int s,d;
INT32 t1, t2, t3, t4;
s=modrm & 0x7;
d=(modrm >> 3) & 0x7;
MMX(d).s[0]=SaturatedSignedDwordToSignedWord(MMX(d).i[0]);
MMX(d).s[1]=SaturatedSignedDwordToSignedWord(MMX(d).i[1]);
MMX(d).s[2]=SaturatedSignedDwordToSignedWord(MMX(s).i[0]);
MMX(d).s[3]=SaturatedSignedDwordToSignedWord(MMX(s).i[1]);
} else {
t1 = MMX(d).i[0];
t2 = MMX(d).i[1];
t3 = MMX(s).i[0];
t4 = MMX(s).i[1];
MMX(d).s[0] = SaturatedSignedDwordToSignedWord(t1);
MMX(d).s[1] = SaturatedSignedDwordToSignedWord(t2);
MMX(d).s[2] = SaturatedSignedDwordToSignedWord(t3);
MMX(d).s[3] = SaturatedSignedDwordToSignedWord(t4);
}
else {
MMX_REG s;
INT32 t1, t2;
int d=(modrm >> 3) & 0x7;
UINT32 ea = GetEA(modrm, 0);
READMMX(ea, s);
MMX(d).s[0]=SaturatedSignedDwordToSignedWord(MMX(d).i[0]);
MMX(d).s[1]=SaturatedSignedDwordToSignedWord(MMX(d).i[1]);
MMX(d).s[2]=SaturatedSignedDwordToSignedWord(s.i[0]);
MMX(d).s[3]=SaturatedSignedDwordToSignedWord(s.i[1]);
t1 = MMX(d).i[0];
t2 = MMX(d).i[1];
MMX(d).s[0] = SaturatedSignedDwordToSignedWord(t1);
MMX(d).s[1] = SaturatedSignedDwordToSignedWord(t2);
MMX(d).s[2] = SaturatedSignedDwordToSignedWord(s.i[0]);
MMX(d).s[3] = SaturatedSignedDwordToSignedWord(s.i[1]);
}
CYCLES(1); // TODO: correct cycle count
}
@ -2711,9 +2724,11 @@ void i386_device::sse_movlps_r128_m64() // Opcode 0f 12
{
UINT8 modrm = FETCH();
if( modrm >= 0xc0 ) {
// unsupported by cpu
// MOVHLPS opcode
XMM((modrm >> 3) & 0x7).q[0] = XMM(modrm & 0x7).q[1];
CYCLES(1); // TODO: correct cycle count
} else {
// MOVLPS opcode
UINT32 ea = GetEA(modrm, 0);
READXMM_LO64(ea, XMM((modrm >> 3) & 0x7));
CYCLES(1); // TODO: correct cycle count
@ -2737,9 +2752,11 @@ void i386_device::sse_movhps_r128_m64() // Opcode 0f 16
{
UINT8 modrm = FETCH();
if( modrm >= 0xc0 ) {
// unsupported by cpu
// MOVLHPS opcode
XMM((modrm >> 3) & 0x7).q[1] = XMM(modrm & 0x7).q[0];
CYCLES(1); // TODO: correct cycle count
} else {
// MOVHPS opcode
UINT32 ea = GetEA(modrm, 0);
READXMM_HI64(ea, XMM((modrm >> 3) & 0x7));
CYCLES(1); // TODO: correct cycle count
@ -3367,7 +3384,7 @@ void i386_device::sse_ucomiss_r128_r128m32() // Opcode 0f 2e
CYCLES(1); // TODO: correct cycle count
}
void i386_device::sse_shufps() // Opcode 0f 67
void i386_device::sse_shufps() // Opcode 0f c6
{
UINT8 modrm = FETCH();
UINT8 sel = FETCH();
@ -3380,20 +3397,24 @@ void i386_device::sse_shufps() // Opcode 0f 67
s=modrm & 0x7;
d=(modrm >> 3) & 0x7;
if( modrm >= 0xc0 ) {
UINT32 t;
t=XMM(d).d[m1];
XMM(d).d[1]=XMM(d).d[m2];
XMM(d).d[0]=t;
XMM(d).d[2]=XMM(s).d[m3];
XMM(d).d[3]=XMM(s).d[m4];
UINT32 t1,t2,t3,t4;
t1=XMM(d).d[m1];
t2=XMM(d).d[m2];
t3=XMM(s).d[m3];
t4=XMM(s).d[m4];
XMM(d).d[0]=t1;
XMM(d).d[1]=t2;
XMM(d).d[2]=t3;
XMM(d).d[3]=t4;
} else {
UINT32 t;
UINT32 t1,t2;
XMM_REG src;
UINT32 ea = GetEA(modrm, 0);
READXMM(ea, src);
t=XMM(d).d[m1];
XMM(d).d[1]=XMM(d).d[m2];
XMM(d).d[0]=t;
t1=XMM(d).d[m1];
t2=XMM(d).d[m2];
XMM(d).d[0]=t1;
XMM(d).d[1]=t2;
XMM(d).d[2]=src.d[m3];
XMM(d).d[3]=src.d[m4];
}
@ -3404,19 +3425,25 @@ void i386_device::sse_unpcklps_r128_rm128() // Opcode 0f 14
{
UINT8 modrm = FETCH();
int s,d;
UINT32 t1, t2, t3, t4;
s=modrm & 0x7;
d=(modrm >> 3) & 0x7;
if( modrm >= 0xc0 ) {
XMM(d).d[3]=XMM(s).d[1];
XMM(d).d[2]=XMM(d).d[1];
XMM(d).d[1]=XMM(s).d[0];
//XMM(d).d[0]=XMM(d).d[0];
t1 = XMM(s).d[1];
t2 = XMM(d).d[1];
t3 = XMM(s).d[0];
t4 = XMM(d).d[0];
XMM(d).d[3]=t1;
XMM(d).d[2]=t2;
XMM(d).d[1]=t3;
XMM(d).d[0]=t4;
} else {
XMM_REG src;
UINT32 ea = GetEA(modrm, 0);
READXMM(ea, src);
t2 = XMM(d).d[1];
XMM(d).d[3]=src.d[1];
XMM(d).d[2]=XMM(d).d[1];
XMM(d).d[2]=t2;
XMM(d).d[1]=src.d[0];
}
CYCLES(1); // TODO: correct cycle count
@ -3426,20 +3453,27 @@ void i386_device::sse_unpckhps_r128_rm128() // Opcode 0f 15
{
UINT8 modrm = FETCH();
int s,d;
UINT32 t1, t2, t3, t4;
s=modrm & 0x7;
d=(modrm >> 3) & 0x7;
if( modrm >= 0xc0 ) {
XMM(d).d[0]=XMM(d).d[2];
XMM(d).d[1]=XMM(s).d[2];
XMM(d).d[2]=XMM(d).d[3];
XMM(d).d[3]=XMM(s).d[3];
t1 = XMM(d).d[2];
t2 = XMM(s).d[2];
t3 = XMM(d).d[3];
t4 = XMM(s).d[3];
XMM(d).d[0]=t1;
XMM(d).d[1]=t2;
XMM(d).d[2]=t3;
XMM(d).d[3]=t4;
} else {
XMM_REG src;
UINT32 ea = GetEA(modrm, 0);
READXMM(ea, src);
XMM(d).d[0]=XMM(d).d[2];
t1 = XMM(d).d[2];
t3 = XMM(d).d[3];
XMM(d).d[0]=t1;
XMM(d).d[1]=src.d[2];
XMM(d).d[2]=XMM(d).d[3];
XMM(d).d[2]=t3;
XMM(d).d[3]=src.d[3];
}
CYCLES(1); // TODO: correct cycle count