@@ -1387,3 +1387,25 @@ DEF_HELPER_5(th_vse_v_w, void, ptr, ptr, tl, env, i32)
DEF_HELPER_5(th_vse_v_w_mask, void, ptr, ptr, tl, env, i32)
DEF_HELPER_5(th_vse_v_d, void, ptr, ptr, tl, env, i32)
DEF_HELPER_5(th_vse_v_d_mask, void, ptr, ptr, tl, env, i32)
+DEF_HELPER_6(th_vlxb_v_b, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxb_v_h, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxb_v_w, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxb_v_d, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxh_v_h, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxh_v_w, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxh_v_d, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxw_v_w, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxw_v_d, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxe_v_b, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxe_v_h, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxe_v_w, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxe_v_d, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxbu_v_b, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxbu_v_h, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxbu_v_w, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxbu_v_d, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxhu_v_h, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxhu_v_w, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxhu_v_d, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxwu_v_w, void, ptr, ptr, tl, ptr, env, i32)
+DEF_HELPER_6(th_vlxwu_v_d, void, ptr, ptr, tl, ptr, env, i32)
@@ -179,6 +179,20 @@ static bool th_check_nf(DisasContext *s, uint32_t vd, uint32_t nf)
return size <= 8 && vd + size <= 32;
}
+/*
+ * The destination vector register group cannot overlap a source vector register
+ * group of a different element width.
+ *
+ * The overlap check rule is different from RVV1.0. The function
+ * "require_noover" describes the check rule in RVV1.0. In general, in some
+ * situations, the destination vector register group can overlap the source
+ * vector register group of a different element width. While XTheadVector not.
+ */
+static inline bool th_check_overlap_group(int rd, int dlen, int rs, int slen)
+{
+ return ((rd >= rs + slen) || (rs >= rd + dlen));
+}
+
/*
* common translation macro
*
@@ -453,6 +467,76 @@ GEN_TH_TRANS(th_vsh_v, 1, r2nfvm, st_us_op_th, st_us_check_th)
GEN_TH_TRANS(th_vsw_v, 2, r2nfvm, st_us_op_th, st_us_check_th)
GEN_TH_TRANS(th_vse_v, 3, r2nfvm, st_us_op_th, st_us_check_th)
+/*
+ * index load and store
+ */
+
+#define gen_helper_ldst_index_th gen_helper_ldst_index
+
+/*
+ * This function is almost the copy of ld_index_op, except:
+ * 1) different data encoding
+ * 2) XTheadVector has more insns to handle zero/sign-extended.
+ */
+static bool ld_index_op_th(DisasContext *s, arg_rnfvm *a, uint8_t seq)
+{
+ uint32_t data = 0;
+ gen_helper_ldst_index_th *fn;
+ static gen_helper_ldst_index_th * const fns[7][4] = {
+ { gen_helper_th_vlxb_v_b, gen_helper_th_vlxb_v_h,
+ gen_helper_th_vlxb_v_w, gen_helper_th_vlxb_v_d },
+ { NULL, gen_helper_th_vlxh_v_h,
+ gen_helper_th_vlxh_v_w, gen_helper_th_vlxh_v_d },
+ { NULL, NULL,
+ gen_helper_th_vlxw_v_w, gen_helper_th_vlxw_v_d },
+ { gen_helper_th_vlxe_v_b, gen_helper_th_vlxe_v_h,
+ gen_helper_th_vlxe_v_w, gen_helper_th_vlxe_v_d },
+ { gen_helper_th_vlxbu_v_b, gen_helper_th_vlxbu_v_h,
+ gen_helper_th_vlxbu_v_w, gen_helper_th_vlxbu_v_d },
+ { NULL, gen_helper_th_vlxhu_v_h,
+ gen_helper_th_vlxhu_v_w, gen_helper_th_vlxhu_v_d },
+ { NULL, NULL,
+ gen_helper_th_vlxwu_v_w, gen_helper_th_vlxwu_v_d },
+ };
+
+ fn = fns[seq][s->sew];
+ if (fn == NULL) {
+ return false;
+ }
+
+ data = FIELD_DP32(data, VDATA_TH, MLEN, s->mlen);
+ data = FIELD_DP32(data, VDATA_TH, VM, a->vm);
+ data = FIELD_DP32(data, VDATA_TH, LMUL, s->lmul);
+ data = FIELD_DP32(data, VDATA_TH, NF, a->nf);
+ return ldst_index_trans(a->rd, a->rs1, a->rs2, data, fn, s);
+}
+
+/*
+ * For vector indexed segment loads, the destination vector register
+ * groups cannot overlap the source vector register group (specified by
+ * `vs2`), else an illegal instruction exception is raised.
+ */
+static bool ld_index_check_th(DisasContext *s, arg_rnfvm* a)
+{
+ return (require_xtheadvector(s) &&
+ vext_check_isa_ill(s) &&
+ th_check_overlap_mask(s, a->rd, a->vm, false) &&
+ th_check_reg(s, a->rd, false) &&
+ th_check_reg(s, a->rs2, false) &&
+ th_check_nf(s, a->rd, a->nf) &&
+ ((a->nf == 1) ||
+ th_check_overlap_group(a->rd, a->nf << s->lmul,
+ a->rs2, 1 << s->lmul)));
+}
+
+GEN_TH_TRANS(th_vlxb_v, 0, rnfvm, ld_index_op_th, ld_index_check_th)
+GEN_TH_TRANS(th_vlxh_v, 1, rnfvm, ld_index_op_th, ld_index_check_th)
+GEN_TH_TRANS(th_vlxw_v, 2, rnfvm, ld_index_op_th, ld_index_check_th)
+GEN_TH_TRANS(th_vlxe_v, 3, rnfvm, ld_index_op_th, ld_index_check_th)
+GEN_TH_TRANS(th_vlxbu_v, 4, rnfvm, ld_index_op_th, ld_index_check_th)
+GEN_TH_TRANS(th_vlxhu_v, 5, rnfvm, ld_index_op_th, ld_index_check_th)
+GEN_TH_TRANS(th_vlxwu_v, 6, rnfvm, ld_index_op_th, ld_index_check_th)
+
#define TH_TRANS_STUB(NAME) \
static bool trans_##NAME(DisasContext *s, arg_##NAME *a) \
{ \
@@ -466,13 +550,6 @@ TH_TRANS_STUB(th_vleff_v)
TH_TRANS_STUB(th_vlbuff_v)
TH_TRANS_STUB(th_vlhuff_v)
TH_TRANS_STUB(th_vlwuff_v)
-TH_TRANS_STUB(th_vlxb_v)
-TH_TRANS_STUB(th_vlxh_v)
-TH_TRANS_STUB(th_vlxw_v)
-TH_TRANS_STUB(th_vlxe_v)
-TH_TRANS_STUB(th_vlxbu_v)
-TH_TRANS_STUB(th_vlxhu_v)
-TH_TRANS_STUB(th_vlxwu_v)
TH_TRANS_STUB(th_vsxb_v)
TH_TRANS_STUB(th_vsxh_v)
TH_TRANS_STUB(th_vsxw_v)
@@ -360,8 +360,8 @@ typedef target_ulong vext_get_index_addr(target_ulong base,
uint32_t idx, void *vs2);
#define GEN_VEXT_GET_INDEX_ADDR(NAME, ETYPE, H) \
-static target_ulong NAME(target_ulong base, \
- uint32_t idx, void *vs2) \
+target_ulong NAME(target_ulong base, \
+ uint32_t idx, void *vs2) \
{ \
return (base + *((ETYPE *)vs2 + H(idx))); \
}
@@ -239,4 +239,8 @@ static inline target_ulong adjust_addr(CPURISCVState *env, target_ulong addr)
return (addr & ~env->cur_pmmask) | env->cur_pmbase;
}
+target_ulong idx_b(target_ulong base, uint32_t idx, void *vs2);
+target_ulong idx_h(target_ulong base, uint32_t idx, void *vs2);
+target_ulong idx_w(target_ulong base, uint32_t idx, void *vs2);
+target_ulong idx_d(target_ulong base, uint32_t idx, void *vs2);
#endif /* TARGET_RISCV_VECTOR_INTERNALS_H */
@@ -436,3 +436,85 @@ GEN_TH_ST_US(th_vse_v_b, int8_t, int8_t , ste_b)
GEN_TH_ST_US(th_vse_v_h, int16_t, int16_t, ste_h)
GEN_TH_ST_US(th_vse_v_w, int32_t, int32_t, ste_w)
GEN_TH_ST_US(th_vse_v_d, int64_t, int64_t, ste_d)
+
+/*
+ * index: access vector element from indexed memory
+ */
+typedef target_ulong th_get_index_addr(target_ulong base,
+ uint32_t idx, void *vs2);
+
+/*
+ * This function is almost the copy of vext_ldst_index, except:
+ * 1) different mask layout
+ * 2) different data encoding
+ * 3) different mask/tail elements process policy
+ */
+static inline void
+th_ldst_index(void *vd, void *v0, target_ulong base,
+ void *vs2, CPURISCVState *env, uint32_t desc,
+ th_get_index_addr get_index_addr,
+ th_ldst_elem_fn *ldst_elem,
+ clear_fn *clear_elem,
+ uint32_t esz, uint32_t msz, uintptr_t ra)
+{
+ uint32_t i, k;
+ uint32_t nf = th_nf(desc);
+ uint32_t vm = th_vm(desc);
+ uint32_t mlen = th_mlen(desc);
+ uint32_t vlmax = th_maxsz(desc) / esz;
+
+ VSTART_CHECK_EARLY_EXIT(env);
+
+ /* load bytes from guest memory */
+ for (i = env->vstart; i < env->vl; env->vstart = ++i) {
+ k = 0;
+ if (!vm && !th_elem_mask(v0, mlen, i)) {
+ continue;
+ }
+ while (k < nf) {
+ abi_ptr addr = get_index_addr(base, i, vs2) + k * msz;
+ ldst_elem(env, adjust_addr(env, addr), i + k * vlmax, vd, ra);
+ k++;
+ }
+ }
+ env->vstart = 0;
+ /* clear tail elements */
+ if (clear_elem) {
+ for (k = 0; k < nf; k++) {
+ clear_elem(vd, env->vl + k * vlmax, env->vl * esz, vlmax * esz);
+ }
+ }
+}
+
+/* Similar to GEN_VEXT_LD_INDEX */
+#define GEN_TH_LD_INDEX(NAME, MTYPE, ETYPE, INDEX_FN, LOAD_FN, CLEAR_FN) \
+void HELPER(NAME)(void *vd, void *v0, target_ulong base, \
+ void *vs2, CPURISCVState *env, uint32_t desc) \
+{ \
+ th_ldst_index(vd, v0, base, vs2, env, desc, INDEX_FN, \
+ LOAD_FN, CLEAR_FN, sizeof(ETYPE), sizeof(MTYPE), \
+ GETPC()); \
+}
+
+GEN_TH_LD_INDEX(th_vlxb_v_b, int8_t, int8_t, idx_b, ldb_b, clearb_th)
+GEN_TH_LD_INDEX(th_vlxb_v_h, int8_t, int16_t, idx_h, ldb_h, clearh_th)
+GEN_TH_LD_INDEX(th_vlxb_v_w, int8_t, int32_t, idx_w, ldb_w, clearl_th)
+GEN_TH_LD_INDEX(th_vlxb_v_d, int8_t, int64_t, idx_d, ldb_d, clearq_th)
+GEN_TH_LD_INDEX(th_vlxh_v_h, int16_t, int16_t, idx_h, ldh_h, clearh_th)
+GEN_TH_LD_INDEX(th_vlxh_v_w, int16_t, int32_t, idx_w, ldh_w, clearl_th)
+GEN_TH_LD_INDEX(th_vlxh_v_d, int16_t, int64_t, idx_d, ldh_d, clearq_th)
+GEN_TH_LD_INDEX(th_vlxw_v_w, int32_t, int32_t, idx_w, ldw_w, clearl_th)
+GEN_TH_LD_INDEX(th_vlxw_v_d, int32_t, int64_t, idx_d, ldw_d, clearq_th)
+GEN_TH_LD_INDEX(th_vlxe_v_b, int8_t, int8_t, idx_b, lde_b, clearb_th)
+GEN_TH_LD_INDEX(th_vlxe_v_h, int16_t, int16_t, idx_h, lde_h, clearh_th)
+GEN_TH_LD_INDEX(th_vlxe_v_w, int32_t, int32_t, idx_w, lde_w, clearl_th)
+GEN_TH_LD_INDEX(th_vlxe_v_d, int64_t, int64_t, idx_d, lde_d, clearq_th)
+GEN_TH_LD_INDEX(th_vlxbu_v_b, uint8_t, uint8_t, idx_b, ldbu_b, clearb_th)
+GEN_TH_LD_INDEX(th_vlxbu_v_h, uint8_t, uint16_t, idx_h, ldbu_h, clearh_th)
+GEN_TH_LD_INDEX(th_vlxbu_v_w, uint8_t, uint32_t, idx_w, ldbu_w, clearl_th)
+GEN_TH_LD_INDEX(th_vlxbu_v_d, uint8_t, uint64_t, idx_d, ldbu_d, clearq_th)
+GEN_TH_LD_INDEX(th_vlxhu_v_h, uint16_t, uint16_t, idx_h, ldhu_h, clearh_th)
+GEN_TH_LD_INDEX(th_vlxhu_v_w, uint16_t, uint32_t, idx_w, ldhu_w, clearl_th)
+GEN_TH_LD_INDEX(th_vlxhu_v_d, uint16_t, uint64_t, idx_d, ldhu_d, clearq_th)
+GEN_TH_LD_INDEX(th_vlxwu_v_w, uint32_t, uint32_t, idx_w, ldwu_w, clearl_th)
+GEN_TH_LD_INDEX(th_vlxwu_v_d, uint32_t, uint64_t, idx_d, ldwu_d, clearq_th)
XTheadVector indexed load instructions diff from RVV1.0 in the following points: 1. Different mask reg layout. 2. Different access width. XTheadVector has 7 instructions, th.vlx{b,h,w}.v, th.vlx{b,h,w}u.v and th.vlxe.v. Their index element width and reg data element width are all SEW-bit. The difference between b,h,w,e is memory access width. {b,h,w,e} stands for 8/16/32/SEW-bit. Therefore, it leads to the difference of zero and sign extend. "u" stands for zero-extened, the lack of "u" stands for the other. RVV1.0 instructions vlxei{8,16,32,64}.v. {8,16,32,64} are the width of index element, while the memory and data reg element width are both SEW-bit. 3. Different tail/masked elements process policy. 4. Different check policy. In RVV1.0, in some situations, the destination vector register group can overlap the source vector register group of a different element width. While XTheadVector not. Signed-off-by: Huang Tao <eric.huang@linux.alibaba.com> --- target/riscv/helper.h | 22 +++++ .../riscv/insn_trans/trans_xtheadvector.c.inc | 91 +++++++++++++++++-- target/riscv/vector_helper.c | 4 +- target/riscv/vector_internals.h | 4 + target/riscv/xtheadvector_helper.c | 82 +++++++++++++++++ 5 files changed, 194 insertions(+), 9 deletions(-)