diff options
| -rwxr-xr-x | src/dynarec/arm64_emitter.h | 5 | ||||
| -rwxr-xr-x | src/dynarec/arm64_printer.c | 49 | ||||
| -rwxr-xr-x | src/dynarec/dynarec_arm64_660f.c | 23 |
3 files changed, 74 insertions, 3 deletions
diff --git a/src/dynarec/arm64_emitter.h b/src/dynarec/arm64_emitter.h index ea7a8ae9..364dda4d 100755 --- a/src/dynarec/arm64_emitter.h +++ b/src/dynarec/arm64_emitter.h @@ -1316,4 +1316,9 @@ #define UADDLV_16(Rd, Rn) EMIT(ADDLV_vector(0, 1, 0b01, Rn, Rd)) #define UADDLV_32(Rd, Rn) EMIT(ADDLV_vector(0, 1, 0b10, Rn, Rd)) +// MOV Immediate +#define MOVI_vector(Q, op, abc, cmode, defgh, Rd) ((Q)<<30 | (op)<<29 | 0b0111100000<<19 | (abc)<<16 | (cmode)<<12 | 1<<10 | (defgh)<<5 | (Rd)) +#define MOVIQ_8(Rd, imm8) EMIT(MOVI_vector(1, 0, (((imm8)>>5)&0b111), 0b1110, ((imm8)&0b11111), Rd)) +#define MOVI_8(Rd, imm8) EMIT(MOVI_vector(0, 0, (((imm8)>>5)&0b111), 0b1110, ((imm8)&0b11111), Rd)) + #endif //__ARM64_EMITTER_H__ diff --git a/src/dynarec/arm64_printer.c b/src/dynarec/arm64_printer.c index 606271af..fd87692d 100755 --- a/src/dynarec/arm64_printer.c +++ b/src/dynarec/arm64_printer.c @@ -1149,7 +1149,7 @@ const char* arm64_print(uint32_t opcode, uintptr_t addr) const char* Y[] = {"8B", "16B", "4H", "8H", "2S", "4S", "??", "???"}; const char* Vd = Y[(sf<<1) | a.Q]; const char* Z[] = {"8H", "4S", "2D", "?"}; - const char* Va = Y[(sf<<1)]; + const char* Va = Z[(sf<<1)]; snprintf(buff, sizeof(buff), "%cAB%cL%s V%d.%s, V%d.%s, V%d.%s", a.U?'U':'S', a.c?'A':'D', a.Q?"2":"", Rd, Va, Rn, Vd, Rm, Vd); return buff; } @@ -1167,6 +1167,53 @@ const char* arm64_print(uint32_t opcode, uintptr_t addr) snprintf(buff, sizeof(buff), "%cADDLV V%d.%s, V%d.%s", a.U?'U':'S', Rd, Z[sf], Rn, Vd); return buff; } + + // MOV immediate + if(isMask(opcode, "0Q00111100000iii111001iiiiiddddd", &a)) { + const char* Y[] = {"8B", "16B"}; + const char* Vd = Y[a.Q]; + snprintf(buff, sizeof(buff), "MOVI V%d.%s, #0x%x", Rd, Vd, imm); + return buff; + } + + // LD1/ST1 single structure + if(isMask(opcode, "0Q0011010L000000cc0Sffnnnnnttttt", &a)) { + int scale = a.c; + int idx = 0; + const char* Y[] = {"B", "H", "S", "D"}; + switch(scale) { + case 3: scale = sf; /* rep = 1; */ break; + case 0: idx = (a.Q<<3) | (a.S<<2) | sf; break; + case 1: idx = (a.Q<<2) | (a.S<<1) | (sf>>1); break; + case 2: if(!(sf&1)) + idx = (a.Q<<1) | a.S; + else { + scale = 3; + idx = a.Q; + } + break; + } + snprintf(buff, sizeof(buff), "%s1 {V%d.%s}[%d], %s", a.L?"LD":"ST", Rd, Y[scale], idx, XtSp[Rt]); + return buff; + } + + // (S/U)QXT(U)N + if(isMask(opcode, "0Q101110ff100001001010nnnnnddddd", &a)) { + const char* Y[] = {"8B", "16B", "4H", "8H", "2S", "4S", "?", "??"}; + const char* Vd = Y[(sf<<1) | a.Q]; + const char* Z[] = {"8H", "4S", "2D", "?"}; + const char* Va = Z[sf]; + snprintf(buff, sizeof(buff), "SQXTUN%s V%d.%s, V%d.%s", a.Q?"2":"", Rd, Vd, Rn, Va); + return buff; + } + if(isMask(opcode, "0QU01110ff100001010010nnnnnddddd", &a)) { + const char* Y[] = {"8B", "16B", "4H", "8H", "2S", "4S", "?", "??"}; + const char* Vd = Y[(sf<<1) | a.Q]; + const char* Z[] = {"8H", "4S", "2D", "?"}; + const char* Va = Z[sf]; + snprintf(buff, sizeof(buff), "%cQXTN%s V%d.%s, V%d.%s", a.U?'U':'S', a.Q?"2":"", Rd, Vd, Rn, Va); + return buff; + } snprintf(buff, sizeof(buff), "%08X ???", __builtin_bswap32(opcode)); diff --git a/src/dynarec/dynarec_arm64_660f.c b/src/dynarec/dynarec_arm64_660f.c index 0d0ffdae..98f44979 100755 --- a/src/dynarec/dynarec_arm64_660f.c +++ b/src/dynarec/dynarec_arm64_660f.c @@ -220,6 +220,25 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n FCOMI(x1, x2); break; + case 0x38: // SSSE3 opcodes + nextop = F8; + switch(nextop) { + case 0x00: + INST_NAME("PSHUFB Gx, Ex"); + nextop = F8; + GETGX(q0); + GETEX(q1, 0); + d0 = fpu_get_scratch(dyn); + MOVIQ_8(d0, 0b10001111); + VANDQ(d0, d0, q1); // mask the index + VTBLQ1_8(q0, q0, d0); + break; + + default: + DEFAULT; + } + break; + #define GO(GETFLAGS, NO, YES, F) \ READFLAGS(F); \ GETFLAGS; \ @@ -348,11 +367,11 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n nextop = F8; GETGX(v0); GETEX(v1, 0); - SQXTUN_16(v0, v0); + SQXTUN_8(v0, v0); if(v0==v1) { VMOVeD(v0, 1, v0, 0); } else { - SQXTUN2_16(v0, v1); + SQXTUN2_8(v0, v1); } break; case 0x68: |