diff options
| author | Yang Liu <liuyang22@iscas.ac.cn> | 2024-09-17 15:35:15 +0800 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-09-17 09:35:15 +0200 |
| commit | f674953b875fd5423a3c5a02aaaa4f428f32cc41 (patch) | |
| tree | 2f8a5263803293e03c9775a4b01cbe89a514dc9b /src | |
| parent | b8798c580f570c9b75bcaddc285d0a449e2c67a0 (diff) | |
| download | box64-f674953b875fd5423a3c5a02aaaa4f428f32cc41.tar.gz box64-f674953b875fd5423a3c5a02aaaa4f428f32cc41.zip | |
[RV64_DYNAREC] Added more 66 0F opcodes for vector (#1831)
Diffstat (limited to 'src')
| -rw-r--r-- | src/dynarec/rv64/dynarec_rv64_660f_vector.c | 90 |
1 files changed, 90 insertions, 0 deletions
diff --git a/src/dynarec/rv64/dynarec_rv64_660f_vector.c b/src/dynarec/rv64/dynarec_rv64_660f_vector.c index 5618e0b2..27384f5e 100644 --- a/src/dynarec/rv64/dynarec_rv64_660f_vector.c +++ b/src/dynarec/rv64/dynarec_rv64_660f_vector.c @@ -381,10 +381,100 @@ uintptr_t dynarec64_660F_vector(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t i SET_ELEMENT_WIDTH(x1, VECTOR_SEW64, 1); VMV_V_V(q0, v0); break; + case 0x39: + INST_NAME("PMINSD Gx, Ex"); + nextop = F8; + SET_ELEMENT_WIDTH(x1, VECTOR_SEW32, 1); + GETGX_vector(q0, 1, VECTOR_SEW32); + GETEX_vector(q1, 0, 0, VECTOR_SEW32); + VMIN_VV(q0, q0, q1, VECTOR_UNMASKED); + break; + case 0x3D: + INST_NAME("PMAXSD Gx, Ex"); + nextop = F8; + SET_ELEMENT_WIDTH(x1, VECTOR_SEW32, 1); + GETGX_vector(q0, 1, VECTOR_SEW32); + GETEX_vector(q1, 0, 0, VECTOR_SEW32); + VMAX_VV(q0, q0, q1, VECTOR_UNMASKED); + break; default: DEFAULT_VECTOR; } break; + case 0x3A: // these are some more SSSE3+ opcodes + opcode = F8; + switch (opcode) { + case 0x0E: + INST_NAME("PBLENDW Gx, Ex, Ib"); + nextop = F8; + SET_ELEMENT_WIDTH(x1, VECTOR_SEW16, 1); + GETGX_vector(q0, 1, VECTOR_SEW16); + GETEX_vector(q1, 0, 0, VECTOR_SEW16); + u8 = F8; + ADDI(x4, xZR, u8); + VMV_V_X(VMASK, x4); + VADD_VI(q0, 0, q1, VECTOR_MASKED); + break; + case 0x0F: + INST_NAME("PALIGNR Gx, Ex, Ib"); + nextop = F8; + SET_ELEMENT_WIDTH(x1, VECTOR_SEW8, 1); + GETGX_vector(q0, 1, VECTOR_SEW8); + GETEX_vector(q1, 0, 0, VECTOR_SEW8); + u8 = F8; + if (u8 > 31) { + VXOR_VV(q0, q0, q0, VECTOR_UNMASKED); + } else if (u8 > 16) { + v0 = fpu_get_scratch(dyn); + VXOR_VV(q0, q0, q0, VECTOR_UNMASKED); + VSLIDEDOWN_VI(v0, u8 - 16, q0, VECTOR_UNMASKED); + VMV_V_V(q0, v0); + } else if (u8 == 16) { + // nop + } else if (u8 > 0) { + v0 = fpu_get_scratch(dyn); + v1 = fpu_get_scratch(dyn); + VXOR_VV(v0, v0, v0, VECTOR_UNMASKED); + VXOR_VV(v1, v1, v1, VECTOR_UNMASKED); + VSLIDEUP_VI(v0, 16 - u8, q0, VECTOR_UNMASKED); + VSLIDEDOWN_VI(v1, u8, q1, VECTOR_UNMASKED); + VOR_VV(q0, v0, v1, VECTOR_UNMASKED); + } else { + if (q0 != q1) VMV_V_V(q0, q1); + } + break; + default: DEFAULT_VECTOR; + } + break; + case 0x50: + INST_NAME("PMOVMSKD Gd, Ex"); + nextop = F8; + SET_ELEMENT_WIDTH(x1, VECTOR_SEW64, 1); + GETGD; + GETEX_vector(q0, 0, 0, VECTOR_SEW64); + v0 = fpu_get_scratch(dyn); + ADDI(x4, xZR, 63); + VSRL_VX(v0, x4, q0, VECTOR_UNMASKED); + VMSNE_VX(VMASK, xZR, v0, VECTOR_UNMASKED); + VMV_X_S(gd, VMASK); + ANDI(gd, gd, 0b11); + break; + case 0x51: + INST_NAME("SQRTPD Gx, Ex"); + nextop = F8; + SET_ELEMENT_WIDTH(x1, VECTOR_SEW64, 1); + GETGX_vector(q0, 1, VECTOR_SEW64); + GETEX_vector(q1, 0, 0, VECTOR_SEW64); + if (!box64_dynarec_fastnan) { + v0 = fpu_get_scratch(dyn); + VXOR_VV(v0, v0, v0, VECTOR_UNMASKED); + VMFLT_VV(VMASK, v0, q1, VECTOR_UNMASKED); + } + VFSQRT_V(q0, q1, VECTOR_UNMASKED); + if (!box64_dynarec_fastnan) { + VFSGNJN_VV(q0, q0, q0, VECTOR_MASKED); + } + break; case 0x61: INST_NAME("PUNPCKLWD Gx, Ex"); nextop = F8; |