target/riscv: add 'cpu->cfg.vlenb', remove 'cpu->cfg.vlen'
cover target/riscv: add 'cpu->cfg.vlenb', remove 'cpu->cfg.vlen'
Hi,
When working in the KVM Vector support, adding support for the 'vlenb' register, I noticed that we have a *LOT* of code that uses 'vlenb', in a far greater amount than 'vlen', and we resort to do 'vlen >> 3' every time we need 'vlenb'. The conclusion is that we're storing the wrong state - we should store 'vlenb' instead.
We can't just get rid of the 'vlen' property since it has been exposed to users already, but what we can do is to change our internal representation. Users will keep setting 'vlen' but we'll store it as 'vlenb' internally.
To fully remove 'vlen' we need to either use 'vlenb' to retrieve it (i.e. vlen = vlenb << 3), if we can't help but use 'vlenb', or try to rework the logic to avoid using vlenb. This was the case of patch 12, where we avoid using 'vlen' to calculate 'vlmax' in a condition where vl_eq_vlmax is true.
There was a minor performance gain after these changes, in particular after patch 07. Performance gain isn't our goal here, but at least we have evidence that we're not making the emulation slower with these changes.
Patches based on Alistair's riscv-to-apply.next.
Daniel Henrique Barboza (13):
target/riscv: add 'vlenb' field in cpu->cfg target/riscv/csr.c: use 'vlenb' instead of 'vlen' target/riscv/gdbstub.c: use 'vlenb' instead of shifting 'vlen' target/riscv/insn_trans/trans_rvbf16.c.inc: use cpu->cfg.vlenb target/riscv/insn_trans/trans_rvv.c.inc: use 'vlenb' target/riscv/insn_trans/trans_rvvk.c.inc: use 'vlenb' target/riscv/vector_helper.c: use 'vlenb' target/riscv/vector_helper.c: use vlenb in HELPER(vsetvl) target/riscv/cpu.h: use 'vlenb' in vext_get_vlmax() target/riscv/insn_trans/trans_rvv.c.inc: use 'vlenb' in MAXSZ() trans_rvv.c.inc: remove vlmax arg from vec_element_loadx() trans_rvv.c.inc: use cpu_vl in trans_vrgather_vi() target/riscv/cpu.c: remove cpu->cfg.vlen
target/riscv/cpu.c | 12 +- target/riscv/cpu.h | 3 +- target/riscv/cpu_cfg.h | 2 +- target/riscv/csr.c | 4 +- target/riscv/gdbstub.c | 6 +- target/riscv/insn_trans/trans_rvbf16.c.inc | 12 +- target/riscv/insn_trans/trans_rvv.c.inc | 200 +++++++++++---------- target/riscv/insn_trans/trans_rvvk.c.inc | 16 +- target/riscv/tcg/tcg-cpu.c | 4 +- target/riscv/vector_helper.c | 27 +-- 10 files changed, 155 insertions(+), 131 deletions(-)
Signature · 1 line
2.43.0
PATCH 1/13 target/riscv: add 'vlenb' field in cpu->cfg
Our usage of 'vlenb' is overwhelming superior than the use of 'vlen'. We're using 'vlenb' most of the time, having to do 'vlen >> 3' or 'vlen / 8' in every instance.
In hindsight we would be better if the 'vlenb' property was introduced instead of 'vlen'. That's not what happened, and now we can't easily get rid of it due to user scripts all around. What we can do, however, is to change our internal representation to use 'vlenb'.
Add a 'vlenb' field in cpu->cfg. It'll be set via the existing 'vlen' property, i.e. setting 'vlen' will also set 'vlenb'.
We'll replace all 'vlen >> 3' code to use 'vlenb' directly. Start with the single instance we have in target/riscv/cpu.c.
--- target/riscv/cpu.c | 4 +++- target/riscv/cpu_cfg.h | 1 + 2 files changed, 4 insertions(+), 1 deletion(-)
target/riscv/cpu.c +3 -1
@@ -847,7 +847,7 @@ static void riscv_cpu_dump_state(CPUState *cs, FILE *f, int flags) csr_ops[csrno].name, val); } } -uint16_t vlenb = cpu->cfg.vlen >>3 ; +uint16_t vlenb = cpu->cfg.vlenb;for (i =0 ; i <32 ; i++) {qemu_fprintf (f," %-8s " , riscv_rvv_regnames[i]);
@@ -1314,6 +1314,7 @@ static void riscv_cpu_init(Object *obj)/* Default values for non-bool cpu properties */ cpu->cfg.pmu_mask =MAKE_64BIT_MASK (3 ,16 ); cpu->cfg.vlen =128 ; + cpu->cfg.vlenb =128 >>3 ; cpu->cfg.elen =64 ; cpu->env.vext_ver =VEXT_VERSION_1_00_0 ; }
@@ -1810,6 +1811,7 @@ static void prop_vlen_set(Object *obj, Visitor *v, const char *name,cpu_option_add_user_setting (name, value); cpu->cfg.vlen = value; + cpu->cfg.vlenb = value >>3 ; } static void prop_vlen_get(Object *obj, Visitor *v, const char *name,
target/riscv/cpu_cfg.h +1 -0
@@ -140,6 +140,7 @@ struct RISCVCPUConfig {uint32_t pmu_mask;uint16_t vlen; +uint16_t vlenb;uint16_t elen;uint16_t cbom_blocksize;uint16_t cbop_blocksize;
Signature · 1 line
2.43.0
PATCH 2/13 target/riscv/csr.c: use 'vlenb' instead of 'vlen'
In reply to Daniel Henrique BarbozaAs a bonus, we're being more idiomatic using cpu->cfg.vlenb when reading CSR_VLENB.
--- target/riscv/csr.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
target/riscv/csr.c +2 -2
@@ -683,7 +683,7 @@ static RISCVException read_vl(CPURISCVState *env, int csrno,static int read_vlenb (CPURISCVState *env,int csrno,target_ulong *val) { - *val =riscv_cpu_cfg (env)->vlen >>3 ; + *val =riscv_cpu_cfg (env)->vlenb;return RISCV_EXCP_NONE ; }
@@ -738,7 +738,7 @@ static RISCVException write_vstart(CPURISCVState *env, int csrno,* The vstart CSR is defined to have only enough writable bits * to hold the largest element index, i.e. lg2(VLEN) bits. */ - env->vstart = val & ~(~0ULL <<ctzl (riscv_cpu_cfg (env)->vlen)); + env->vstart = val & ~(~0ULL <<ctzl (riscv_cpu_cfg (env)->vlenb <<3 ));return RISCV_EXCP_NONE ; }
Signature · 1 line
2.43.0
PATCH 3/13 target/riscv/gdbstub.c: use 'vlenb' instead of shifting 'vlen'
In reply to Daniel Henrique Barboza--- target/riscv/gdbstub.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
target/riscv/gdbstub.c +3 -3
@@ -130,7 +130,7 @@ static int riscv_gdb_set_fpu(CPURISCVState *env, uint8_t *mem_buf, int n)static int riscv_gdb_get_vector (CPURISCVState *env,GByteArray *buf,int n) { -uint16_t vlenb =riscv_cpu_cfg (env)->vlen >>3 ; +uint16_t vlenb =riscv_cpu_cfg (env)->vlenb;if (n <32 ) {int i;int cnt =0 ;
@@ -146,7 +146,7 @@ static int riscv_gdb_get_vector(CPURISCVState *env, GByteArray *buf, int n)static int riscv_gdb_set_vector (CPURISCVState *env,uint8_t *mem_buf,int n) { -uint16_t vlenb =riscv_cpu_cfg (env)->vlen >>3 ; +uint16_t vlenb =riscv_cpu_cfg (env)->vlenb;if (n <32 ) {int i;for (i =0 ; i < vlenb; i +=8 ) {
@@ -266,7 +266,7 @@ static int ricsv_gen_dynamic_vector_xml(CPUState *cs, int base_reg)RISCVCPU *cpu =RISCV_CPU (cs);GString *s =g_string_new (NULL );g_autoptr (GString) ts =g_string_new ("" ); -int reg_width = cpu->cfg.vlen; +int reg_width = cpu->cfg.vlenb <<3 ;int num_regs =0 ;int i;
Signature · 1 line
2.43.0
PATCH 4/13 target/riscv/insn_trans/trans_rvbf16.c.inc: use cpu->cfg.vlenb
In reply to Daniel Henrique BarbozaUse ctx->cfg_ptr->vlenb instead of ctx->cfg_ptr->vlen / 8.
--- target/riscv/insn_trans/trans_rvbf16.c.inc | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-)
target/riscv/insn_trans/trans_rvbf16.c.inc +6 -6
@@ -83,8 +83,8 @@ static bool trans_vfncvtbf16_f_f_w(DisasContext *ctx, arg_vfncvtbf16_f_f_w *a) data = FIELD_DP32(data, VDATA, VMA, ctx->vma); tcg_gen_gvec_3_ptr(vreg_ofs(ctx, a->rd), vreg_ofs(ctx, 0), vreg_ofs(ctx, a->rs2), tcg_env, - ctx->cfg_ptr->vlen / 8, - ctx->cfg_ptr->vlen / 8, data, + ctx->cfg_ptr->vlenb, + ctx->cfg_ptr->vlenb, data, gen_helper_vfncvtbf16_f_f_w); mark_vs_dirty(ctx); gen_set_label(over);
@@ -112,8 +112,8 @@ static bool trans_vfwcvtbf16_f_f_v(DisasContext *ctx, arg_vfwcvtbf16_f_f_v *a) data = FIELD_DP32(data, VDATA, VMA, ctx->vma); tcg_gen_gvec_3_ptr(vreg_ofs(ctx, a->rd), vreg_ofs(ctx, 0), vreg_ofs(ctx, a->rs2), tcg_env, - ctx->cfg_ptr->vlen / 8, - ctx->cfg_ptr->vlen / 8, data, + ctx->cfg_ptr->vlenb, + ctx->cfg_ptr->vlenb, data, gen_helper_vfwcvtbf16_f_f_v); mark_vs_dirty(ctx); gen_set_label(over);
@@ -143,8 +143,8 @@ static bool trans_vfwmaccbf16_vv(DisasContext *ctx, arg_vfwmaccbf16_vv *a) tcg_gen_gvec_4_ptr(vreg_ofs(ctx, a->rd), vreg_ofs(ctx, 0), vreg_ofs(ctx, a->rs1), vreg_ofs(ctx, a->rs2), tcg_env, - ctx->cfg_ptr->vlen / 8, - ctx->cfg_ptr->vlen / 8, data, + ctx->cfg_ptr->vlenb, + ctx->cfg_ptr->vlenb, data, gen_helper_vfwmaccbf16_vv); mark_vs_dirty(ctx); gen_set_label(over);
Signature · 1 line
2.43.0
PATCH 5/13 target/riscv/insn_trans/trans_rvv.c.inc: use 'vlenb'
In reply to Daniel Henrique BarbozaUse s->cfg_ptr->vlenb instead of "s->cfg_ptr->vlen / 8" and "s->cfg_ptr->vlen >> 3".
--- target/riscv/insn_trans/trans_rvv.c.inc | 140 ++++++++++++------------ 1 file changed, 70 insertions(+), 70 deletions(-)
target/riscv/insn_trans/trans_rvv.c.inc +70 -70
@@ -217,7 +217,7 @@ static bool trans_vsetivli(DisasContext *s, arg_vsetivli *a) /* vector register offset from env */ static uint32_t vreg_ofs(DisasContext *s, int reg) { - return offsetof(CPURISCVState, vreg) + reg * s->cfg_ptr->vlen / 8; + return offsetof(CPURISCVState, vreg) + reg * s->cfg_ptr->vlenb; } /* check functions */
@@ -627,11 +627,11 @@ static bool ldst_us_trans(uint32_t vd, uint32_t rs1, uint32_t data, * As simd_desc supports at most 2048 bytes, and in this implementation, * the max vector group length is 4096 bytes. So split it into two parts. * - * The first part is vlen in bytes, encoded in maxsz of simd_desc. + * The first part is vlen in bytes (vlenb), encoded in maxsz of simd_desc. * The second part is lmul, encoded in data of simd_desc. */ - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, vd)); tcg_gen_addi_ptr(mask, tcg_env, vreg_ofs(s, 0));
@@ -791,8 +791,8 @@ static bool ldst_stride_trans(uint32_t vd, uint32_t rs1, uint32_t rs2, mask = tcg_temp_new_ptr(); base = get_gpr(s, rs1, EXT_NONE); stride = get_gpr(s, rs2, EXT_NONE); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, vd)); tcg_gen_addi_ptr(mask, tcg_env, vreg_ofs(s, 0));
@@ -897,8 +897,8 @@ static bool ldst_index_trans(uint32_t vd, uint32_t rs1, uint32_t vs2, mask = tcg_temp_new_ptr(); index = tcg_temp_new_ptr(); base = get_gpr(s, rs1, EXT_NONE); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, vd)); tcg_gen_addi_ptr(index, tcg_env, vreg_ofs(s, vs2));
@@ -1036,8 +1036,8 @@ static bool ldff_trans(uint32_t vd, uint32_t rs1, uint32_t data, dest = tcg_temp_new_ptr(); mask = tcg_temp_new_ptr(); base = get_gpr(s, rs1, EXT_NONE); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, vd)); tcg_gen_addi_ptr(mask, tcg_env, vreg_ofs(s, 0));
@@ -1086,7 +1086,7 @@ static bool ldst_whole_trans(uint32_t vd, uint32_t rs1, uint32_t nf, uint32_t width, gen_helper_ldst_whole *fn, DisasContext *s, bool is_store) { - uint32_t evl = (s->cfg_ptr->vlen / 8) * nf / width; + uint32_t evl = s->cfg_ptr->vlenb * nf / width; TCGLabel *over = gen_new_label(); tcg_gen_brcondi_tl(TCG_COND_GEU, cpu_vstart, evl, over);
@@ -1096,8 +1096,8 @@ static bool ldst_whole_trans(uint32_t vd, uint32_t rs1, uint32_t nf, uint32_t data = FIELD_DP32(0, VDATA, NF, nf); dest = tcg_temp_new_ptr(); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); base = get_gpr(s, rs1, EXT_NONE); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, vd));
@@ -1199,8 +1199,8 @@ do_opivv_gvec(DisasContext *s, arg_rmrr *a, GVecGen3Fn *gvec_fn, data = FIELD_DP32(data, VDATA, VMA, s->vma); tcg_gen_gvec_4_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), vreg_ofs(s, a->rs1), vreg_ofs(s, a->rs2), - tcg_env, s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data, fn); + tcg_env, s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, fn); } mark_vs_dirty(s); gen_set_label(over);
@@ -1248,8 +1248,8 @@ static bool opivx_trans(uint32_t vd, uint32_t rs1, uint32_t vs2, uint32_t vm, data = FIELD_DP32(data, VDATA, VTA, s->vta); data = FIELD_DP32(data, VDATA, VTA_ALL_1S, s->cfg_vta_all_1s); data = FIELD_DP32(data, VDATA, VMA, s->vma); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, vd)); tcg_gen_addi_ptr(src2, tcg_env, vreg_ofs(s, vs2));
@@ -1410,8 +1410,8 @@ static bool opivi_trans(uint32_t vd, uint32_t imm, uint32_t vs2, uint32_t vm, data = FIELD_DP32(data, VDATA, VTA, s->vta); data = FIELD_DP32(data, VDATA, VTA_ALL_1S, s->cfg_vta_all_1s); data = FIELD_DP32(data, VDATA, VMA, s->vma); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, vd)); tcg_gen_addi_ptr(src2, tcg_env, vreg_ofs(s, vs2));
@@ -1492,8 +1492,8 @@ static bool do_opivv_widen(DisasContext *s, arg_rmrr *a, tcg_gen_gvec_4_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), vreg_ofs(s, a->rs1), vreg_ofs(s, a->rs2), - tcg_env, s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, + tcg_env, s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, fn); mark_vs_dirty(s); gen_set_label(over);
@@ -1568,8 +1568,8 @@ static bool do_opiwv_widen(DisasContext *s, arg_rmrr *a, tcg_gen_gvec_4_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), vreg_ofs(s, a->rs1), vreg_ofs(s, a->rs2), - tcg_env, s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data, fn); + tcg_env, s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, fn); mark_vs_dirty(s); gen_set_label(over); return true;
@@ -1639,8 +1639,8 @@ static bool opivv_trans(uint32_t vd, uint32_t vs1, uint32_t vs2, uint32_t vm, data = FIELD_DP32(data, VDATA, VTA_ALL_1S, s->cfg_vta_all_1s); data = FIELD_DP32(data, VDATA, VMA, s->vma); tcg_gen_gvec_4_ptr(vreg_ofs(s, vd), vreg_ofs(s, 0), vreg_ofs(s, vs1), - vreg_ofs(s, vs2), tcg_env, s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data, fn); + vreg_ofs(s, vs2), tcg_env, s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, fn); mark_vs_dirty(s); gen_set_label(over); return true;
@@ -1831,8 +1831,8 @@ static bool trans_##NAME(DisasContext *s, arg_rmrr *a) \ tcg_gen_gvec_4_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), \ vreg_ofs(s, a->rs1), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, \ - s->cfg_ptr->vlen / 8, data, \ + s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, data, \ fns[s->sew]); \ mark_vs_dirty(s); \ gen_set_label(over); \
@@ -2036,8 +2036,8 @@ static bool trans_vmv_v_v(DisasContext *s, arg_vmv_v_v *a) tcg_gen_brcond_tl(TCG_COND_GEU, cpu_vstart, cpu_vl, over); tcg_gen_gvec_2_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, a->rs1), - tcg_env, s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data, + tcg_env, s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, fns[s->sew]); gen_set_label(over); }
@@ -2082,8 +2082,8 @@ static bool trans_vmv_v_x(DisasContext *s, arg_vmv_v_x *a) }; tcg_gen_ext_tl_i64(s1_i64, s1); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, a->rd)); fns[s->sew](dest, s1_i64, tcg_env, desc); }
@@ -2121,8 +2121,8 @@ static bool trans_vmv_v_i(DisasContext *s, arg_vmv_v_i *a) s1 = tcg_constant_i64(simm); dest = tcg_temp_new_ptr(); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, a->rd)); fns[s->sew](dest, s1, tcg_env, desc);
@@ -2275,8 +2275,8 @@ static bool trans_##NAME(DisasContext *s, arg_rmrr *a) \ tcg_gen_gvec_4_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), \ vreg_ofs(s, a->rs1), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, \ - s->cfg_ptr->vlen / 8, data, \ + s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, data, \ fns[s->sew - 1]); \ mark_vs_dirty(s); \ gen_set_label(over); \
@@ -2303,8 +2303,8 @@ static bool opfvf_trans(uint32_t vd, uint32_t rs1, uint32_t vs2, dest = tcg_temp_new_ptr(); mask = tcg_temp_new_ptr(); src2 = tcg_temp_new_ptr(); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, vd)); tcg_gen_addi_ptr(src2, tcg_env, vreg_ofs(s, vs2));
@@ -2391,8 +2391,8 @@ static bool trans_##NAME(DisasContext *s, arg_rmrr *a) \ tcg_gen_gvec_4_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), \ vreg_ofs(s, a->rs1), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, \ - s->cfg_ptr->vlen / 8, data, \ + s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, data, \ fns[s->sew - 1]); \ mark_vs_dirty(s); \ gen_set_label(over); \
@@ -2465,8 +2465,8 @@ static bool trans_##NAME(DisasContext *s, arg_rmrr *a) \ tcg_gen_gvec_4_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), \ vreg_ofs(s, a->rs1), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, \ - s->cfg_ptr->vlen / 8, data, \ + s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, data, \ fns[s->sew - 1]); \ mark_vs_dirty(s); \ gen_set_label(over); \
@@ -2581,8 +2581,8 @@ static bool do_opfv(DisasContext *s, arg_rmr *a, data = FIELD_DP32(data, VDATA, VMA, s->vma); tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), vreg_ofs(s, a->rs2), tcg_env, - s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data, fn); + s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, fn); mark_vs_dirty(s); gen_set_label(over); return true;
@@ -2691,8 +2691,8 @@ static bool trans_vfmv_v_f(DisasContext *s, arg_vfmv_v_f *a) do_nanbox(s, t1, cpu_fpr[a->rs1]); dest = tcg_temp_new_ptr(); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, a->rd)); fns[s->sew - 1](dest, t1, tcg_env, desc);
@@ -2770,8 +2770,8 @@ static bool trans_##NAME(DisasContext *s, arg_rmr *a) \ data = FIELD_DP32(data, VDATA, VMA, s->vma); \ tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, \ - s->cfg_ptr->vlen / 8, data, \ + s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, data, \ fns[s->sew - 1]); \ mark_vs_dirty(s); \ gen_set_label(over); \
@@ -2821,8 +2821,8 @@ static bool trans_##NAME(DisasContext *s, arg_rmr *a) \ data = FIELD_DP32(data, VDATA, VMA, s->vma); \ tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, \ - s->cfg_ptr->vlen / 8, data, \ + s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, data, \ fns[s->sew]); \ mark_vs_dirty(s); \ gen_set_label(over); \
@@ -2888,8 +2888,8 @@ static bool trans_##NAME(DisasContext *s, arg_rmr *a) \ data = FIELD_DP32(data, VDATA, VMA, s->vma); \ tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, \ - s->cfg_ptr->vlen / 8, data, \ + s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, data, \ fns[s->sew - 1]); \ mark_vs_dirty(s); \ gen_set_label(over); \
@@ -2937,8 +2937,8 @@ static bool trans_##NAME(DisasContext *s, arg_rmr *a) \ data = FIELD_DP32(data, VDATA, VMA, s->vma); \ tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, \ - s->cfg_ptr->vlen / 8, data, \ + s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, data, \ fns[s->sew]); \ mark_vs_dirty(s); \ gen_set_label(over); \
@@ -3027,8 +3027,8 @@ static bool trans_##NAME(DisasContext *s, arg_r *a) \ tcg_gen_gvec_4_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), \ vreg_ofs(s, a->rs1), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, \ - s->cfg_ptr->vlen / 8, data, fn); \ + s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, data, fn); \ mark_vs_dirty(s); \ gen_set_label(over); \ return true; \
@@ -3061,8 +3061,8 @@ static bool trans_vcpop_m(DisasContext *s, arg_rmr *a) mask = tcg_temp_new_ptr(); src2 = tcg_temp_new_ptr(); dst = dest_gpr(s, a->rd); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(src2, tcg_env, vreg_ofs(s, a->rs2)); tcg_gen_addi_ptr(mask, tcg_env, vreg_ofs(s, 0));
@@ -3090,8 +3090,8 @@ static bool trans_vfirst_m(DisasContext *s, arg_rmr *a) mask = tcg_temp_new_ptr(); src2 = tcg_temp_new_ptr(); dst = dest_gpr(s, a->rd); - desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data)); + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data)); tcg_gen_addi_ptr(src2, tcg_env, vreg_ofs(s, a->rs2)); tcg_gen_addi_ptr(mask, tcg_env, vreg_ofs(s, 0));
@@ -3128,8 +3128,8 @@ static bool trans_##NAME(DisasContext *s, arg_rmr *a) \ data = FIELD_DP32(data, VDATA, VMA, s->vma); \ tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), \ vreg_ofs(s, 0), vreg_ofs(s, a->rs2), \ - tcg_env, s->cfg_ptr->vlen / 8, \ - s->cfg_ptr->vlen / 8, \ + tcg_env, s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, \ data, fn); \ mark_vs_dirty(s); \ gen_set_label(over); \
@@ -3171,8 +3171,8 @@ static bool trans_viota_m(DisasContext *s, arg_viota_m *a) }; tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), vreg_ofs(s, a->rs2), tcg_env, - s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data, fns[s->sew]); + s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, fns[s->sew]); mark_vs_dirty(s); gen_set_label(over); return true;
@@ -3200,8 +3200,8 @@ static bool trans_vid_v(DisasContext *s, arg_vid_v *a) gen_helper_vid_v_w, gen_helper_vid_v_d, }; tcg_gen_gvec_2_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), - tcg_env, s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, + tcg_env, s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, fns[s->sew]); mark_vs_dirty(s); gen_set_label(over);
@@ -3620,8 +3620,8 @@ static bool trans_vcompress_vm(DisasContext *s, arg_r *a) data = FIELD_DP32(data, VDATA, VTA, s->vta); tcg_gen_gvec_4_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), vreg_ofs(s, a->rs1), vreg_ofs(s, a->rs2), - tcg_env, s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data, + tcg_env, s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, fns[s->sew]); mark_vs_dirty(s); gen_set_label(over);
@@ -3641,7 +3641,7 @@ static bool trans_##NAME(DisasContext *s, arg_##NAME * a) \ vext_check_isa_ill(s) && \ QEMU_IS_ALIGNED(a->rd, LEN) && \ QEMU_IS_ALIGNED(a->rs2, LEN)) { \ - uint32_t maxsz = (s->cfg_ptr->vlen >> 3) * LEN; \ + uint32_t maxsz = s->cfg_ptr->vlenb * LEN; \ if (s->vstart_eq_zero) { \ tcg_gen_gvec_mov(s->sew, vreg_ofs(s, a->rd), \ vreg_ofs(s, a->rs2), maxsz, maxsz); \
@@ -3723,8 +3723,8 @@ static bool int_ext_op(DisasContext *s, arg_rmr *a, uint8_t seq) tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), vreg_ofs(s, a->rs2), tcg_env, - s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data, fn); + s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, fn); mark_vs_dirty(s); gen_set_label(over);
Signature · 1 line
2.43.0
PATCH 6/13 target/riscv/insn_trans/trans_rvvk.c.inc: use 'vlenb'
In reply to Daniel Henrique BarbozaUse s->cfg_ptr->vlenb instead of s->cfg_ptr->vlen / 8.
--- target/riscv/insn_trans/trans_rvvk.c.inc | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-)
target/riscv/insn_trans/trans_rvvk.c.inc +8 -8
@@ -174,7 +174,7 @@ GEN_OPIVX_GVEC_TRANS_CHECK(vandn_vx, andcs, zvkb_vx_check) data = FIELD_DP32(data, VDATA, VMA, s->vma); \ tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, s->cfg_ptr->vlen / 8, \ + s->cfg_ptr->vlenb, s->cfg_ptr->vlenb, \ data, fns[s->sew]); \ mark_vs_dirty(s); \ gen_set_label(over); \
@@ -267,7 +267,7 @@ GEN_OPIVI_WIDEN_TRANS(vwsll_vi, IMM_ZX, vwsll_vx, vwsll_vx_check) rd_v = tcg_temp_new_ptr(); \ rs2_v = tcg_temp_new_ptr(); \ desc = tcg_constant_i32( \ - simd_desc(s->cfg_ptr->vlen / 8, s->cfg_ptr->vlen / 8, data)); \ + simd_desc(s->cfg_ptr->vlenb, s->cfg_ptr->vlenb, data)); \ tcg_gen_addi_ptr(rd_v, tcg_env, vreg_ofs(s, a->rd)); \ tcg_gen_addi_ptr(rs2_v, tcg_env, vreg_ofs(s, a->rs2)); \ gen_helper_##NAME(rd_v, rs2_v, tcg_env, desc); \
@@ -345,7 +345,7 @@ GEN_V_UNMASKED_TRANS(vaesem_vs, vaes_check_vs, ZVKNED_EGS) rs2_v = tcg_temp_new_ptr(); \ uimm_v = tcg_constant_i32(a->rs1); \ desc = tcg_constant_i32( \ - simd_desc(s->cfg_ptr->vlen / 8, s->cfg_ptr->vlen / 8, data)); \ + simd_desc(s->cfg_ptr->vlenb, s->cfg_ptr->vlenb, data)); \ tcg_gen_addi_ptr(rd_v, tcg_env, vreg_ofs(s, a->rd)); \ tcg_gen_addi_ptr(rs2_v, tcg_env, vreg_ofs(s, a->rs2)); \ gen_helper_##NAME(rd_v, rs2_v, uimm_v, tcg_env, desc); \
@@ -413,7 +413,7 @@ GEN_VI_UNMASKED_TRANS(vaeskf2_vi, vaeskf2_check, ZVKNED_EGS) \ tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, a->rs1), \ vreg_ofs(s, a->rs2), tcg_env, \ - s->cfg_ptr->vlen / 8, s->cfg_ptr->vlen / 8, \ + s->cfg_ptr->vlenb, s->cfg_ptr->vlenb, \ data, gen_helper_##NAME); \ \ mark_vs_dirty(s); \
@@ -466,8 +466,8 @@ static bool trans_vsha2cl_vv(DisasContext *s, arg_rmrr *a) data = FIELD_DP32(data, VDATA, VMA, s->vma); tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, a->rs1), - vreg_ofs(s, a->rs2), tcg_env, s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data, + vreg_ofs(s, a->rs2), tcg_env, s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, s->sew == MO_32 ? gen_helper_vsha2cl32_vv : gen_helper_vsha2cl64_vv);
@@ -500,8 +500,8 @@ static bool trans_vsha2ch_vv(DisasContext *s, arg_rmrr *a) data = FIELD_DP32(data, VDATA, VMA, s->vma); tcg_gen_gvec_3_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, a->rs1), - vreg_ofs(s, a->rs2), tcg_env, s->cfg_ptr->vlen / 8, - s->cfg_ptr->vlen / 8, data, + vreg_ofs(s, a->rs2), tcg_env, s->cfg_ptr->vlenb, + s->cfg_ptr->vlenb, data, s->sew == MO_32 ? gen_helper_vsha2ch32_vv : gen_helper_vsha2ch64_vv);
Signature · 1 line
2.43.0
PATCH 7/13 target/riscv/vector_helper.c: use 'vlenb'
In reply to Daniel Henrique BarbozaUse 'cpu->cfg.vlenb' instead of 'cpu->cfg.vlen >> 3'.
--- target/riscv/vector_helper.c | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-)
target/riscv/vector_helper.c +9 -9
@@ -558,7 +558,7 @@ vext_ldst_whole(void *vd, target_ulong base, CPURISCVState *env, uint32_t desc, {uint32_t i, k, off, pos;uint32_t nf =vext_nf (desc); -uint32_t vlenb =riscv_cpu_cfg (env)->vlen >>3 ; +uint32_t vlenb =riscv_cpu_cfg (env)->vlenb;uint32_t max_elems = vlenb >> log2_esz; k = env->vstart / max_elems;
@@ -929,7 +929,7 @@ void HELPER(NAME)(void *vd, void *v0, void *vs1, void *vs2, \ { \uint32_t vl = env->vl; \uint32_t vm =vext_vm (desc); \ -uint32_t total_elems =riscv_cpu_cfg (env)->vlen; \ +uint32_t total_elems =riscv_cpu_cfg (env)->vlenb <<3 ; \uint32_t vta_all_1s =vext_vta_all_1s (desc); \uint32_t i; \ \
@@ -967,7 +967,7 @@ void HELPER(NAME)(void *vd, void *v0, target_ulong s1, \ { \uint32_t vl = env->vl; \uint32_t vm =vext_vm (desc); \ -uint32_t total_elems =riscv_cpu_cfg (env)->vlen; \ +uint32_t total_elems =riscv_cpu_cfg (env)->vlenb <<3 ; \uint32_t vta_all_1s =vext_vta_all_1s (desc); \uint32_t i; \ \
@@ -1171,7 +1171,7 @@ void HELPER(NAME)(void *vd, void *v0, void *vs1, void *vs2, \ { \uint32_t vm =vext_vm (desc); \uint32_t vl = env->vl; \ -uint32_t total_elems =riscv_cpu_cfg (env)->vlen; \ +uint32_t total_elems =riscv_cpu_cfg (env)->vlenb <<3 ; \uint32_t vta_all_1s =vext_vta_all_1s (desc); \uint32_t vma =vext_vma (desc); \uint32_t i; \
@@ -1236,7 +1236,7 @@ void HELPER(NAME)(void *vd, void *v0, target_ulong s1, void *vs2, \ { \uint32_t vm =vext_vm (desc); \uint32_t vl = env->vl; \ -uint32_t total_elems =riscv_cpu_cfg (env)->vlen; \ +uint32_t total_elems =riscv_cpu_cfg (env)->vlenb <<3 ; \uint32_t vta_all_1s =vext_vta_all_1s (desc); \uint32_t vma =vext_vma (desc); \uint32_t i; \
@@ -3971,7 +3971,7 @@ void HELPER(NAME)(void *vd, void *v0, void *vs1, void *vs2, \ { \uint32_t vm =vext_vm (desc); \uint32_t vl = env->vl; \ -uint32_t total_elems =riscv_cpu_cfg (env)->vlen; \ +uint32_t total_elems =riscv_cpu_cfg (env)->vlenb <<3 ; \uint32_t vta_all_1s =vext_vta_all_1s (desc); \uint32_t vma =vext_vma (desc); \uint32_t i; \
@@ -4011,7 +4011,7 @@ void HELPER(NAME)(void *vd, void *v0, uint64_t s1, void *vs2, \ { \uint32_t vm =vext_vm (desc); \uint32_t vl = env->vl; \ -uint32_t total_elems =riscv_cpu_cfg (env)->vlen; \ +uint32_t total_elems =riscv_cpu_cfg (env)->vlenb <<3 ; \uint32_t vta_all_1s =vext_vta_all_1s (desc); \uint32_t vma =vext_vma (desc); \uint32_t i; \
@@ -4528,7 +4528,7 @@ void HELPER(NAME)(void *vd, void *v0, void *vs1, \ uint32_t desc) \ { \uint32_t vl = env->vl; \ -uint32_t total_elems =riscv_cpu_cfg (env)->vlen; \ +uint32_t total_elems =riscv_cpu_cfg (env)->vlenb <<3 ;\uint32_t vta_all_1s =vext_vta_all_1s (desc); \uint32_t i; \int a, b; \
@@ -4615,7 +4615,7 @@ static void vmsetm(void *vd, void *v0, void *vs2, CPURISCVState *env, {uint32_t vm =vext_vm (desc);uint32_t vl = env->vl; -uint32_t total_elems =riscv_cpu_cfg (env)->vlen; +uint32_t total_elems =riscv_cpu_cfg (env)->vlenb <<3 ;uint32_t vta_all_1s =vext_vta_all_1s (desc);uint32_t vma =vext_vma (desc);int i;
Signature · 1 line
2.43.0
PATCH 8/13 target/riscv/vector_helper.c: use vlenb in HELPER(vsetvl)
In reply to Daniel Henrique BarbozaUse the new 'vlenb' CPU config to validate fractional LMUL. The original comparison is done with 'vlen' and 'sew', both in bits. Use sew/8, or sew in bytes, to do a direct comparison with vlenb.
--- target/riscv/vector_helper.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)
target/riscv/vector_helper.c +6 -3
@@ -45,9 +45,12 @@ target_ulong HELPER(vsetvl)(CPURISCVState *env, target_ulong s1, xlen - 1 - R_VTYPE_RESERVED_SHIFT);if (lmul &4 ) { -/* Fractional LMUL - check LMUL * VLEN >= SEW */ -if (lmul ==4 || - cpu->cfg.vlen >> (8 - lmul) < sew) { +/* +* Fractional LMUL - check LMUL * VLEN >= SEW, or +* vlen >> (8 - lmul) < sew. We have 'vlenb', so +* compare it with sew in bytes. +*/ +if (lmul ==4 || cpu->cfg.vlenb >> (8 - lmul) < (sew <<3 )) { vill = true; } }
Signature · 1 line
2.43.0
Re: [PATCH 08/13] target/riscv/vector_helper.c: use vlenb in HELPER(vsetvl)
Quoted 16 lines · Daniel Henrique Barboza
Use the new 'vlenb' CPU config to validate fractional LMUL. The original comparison is done with 'vlen' and 'sew', both in bits. Use sew/8, or sew in bytes, to do a direct comparison with vlenb.
Signed-off-by Daniel Henrique Barboza <[email protected]>--- target/riscv/vector_helper.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)target/riscv/vector_helper.c +0
-0@@ -45,9 +45,12 @@ target_ulong HELPER(vsetvl)(CPURISCVState *env, target_ulong s1, xlen - 1 - R_VTYPE_RESERVED_SHIFT);
target/riscv/vector_helper.c target_ulong HELPER(vsetvl)(CPURISCVState *env, target_ulong s1, · line 53
Quoted 10 lines
target/riscv/vector_helper.c +6
-3if (lmul &4 ) { -/* Fractional LMUL - check LMUL * VLEN >= SEW */ -if (lmul ==4 || - cpu->cfg.vlen >> (8 - lmul) < sew) { +/* +* Fractional LMUL - check LMUL * VLEN >= SEW, or +* vlen >> (8 - lmul) < sew. We have 'vlenb', so +* compare it with sew in bytes. +*/ +if (lmul ==4 || cpu->cfg.vlenb >> (8 - lmul) < (sew <<3 )) {
You're shifting sew the wrong direction. But better to simply adjust the vlenb shift:
- cpu->cfg.vlen >> (8 - lmul) < sew
+ cpu->cfg.vlenb >> (8 - 3 - lmul) < sew
r~
PATCH 9/13 target/riscv/cpu.h: use 'vlenb' in vext_get_vlmax()
In reply to Daniel Henrique Barboza--- target/riscv/cpu.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
target/riscv/cpu.h +2 -1
@@ -692,7 +692,8 @@ static inline uint32_t vext_get_vlmax(RISCVCPU *cpu, target_ulong vtype) {uint8_t sew =FIELD_EX64 (vtype,VTYPE ,VSEW );int8_t lmul =sextract32 (FIELD_EX64 (vtype,VTYPE ,VLMUL ),0 ,3 ); -return cpu->cfg.vlen >> (sew +3 - lmul); +/* vlen = vlenb << 3 */ +return (cpu->cfg.vlenb <<3 ) >> (sew +3 - lmul); } void cpu_get_tb_cpu_state(CPURISCVState *env, vaddr *pc,
Signature · 1 line
2.43.0
Re: [PATCH 09/13] target/riscv/cpu.h: use 'vlenb' in vext_get_vlmax()
Quoted 10 lines · Daniel Henrique Barboza
Signed-off-by Daniel Henrique Barboza <[email protected]>--- target/riscv/cpu.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)target/riscv/cpu.h +0
-0
target/riscv/cpu.h static inline uint32_t vext_get_vlmax(RISCVCPU *cpu, target_ulong vtype) · line 697
Quoted 7 lines
target/riscv/cpu.h +2
-1@@ -692,7 +692,8 @@ static inline uint32_t vext_get_vlmax(RISCVCPU *cpu, target_ulong vtype) {uint8_t sew =FIELD_EX64 (vtype,VTYPE ,VSEW );int8_t lmul =sextract32 (FIELD_EX64 (vtype,VTYPE ,VLMUL ),0 ,3 ); -return cpu->cfg.vlen >> (sew +3 - lmul); +/* vlen = vlenb << 3 */ +return (cpu->cfg.vlenb <<3 ) >> (sew +3 - lmul); }
Better to simplify the arithmetic:
- return cpu->cfg.vlen >> (sew + 3 - lmul);
+ return cpu->cfg.vlenb >> (sew - lmul);
r~
PATCH 10/13 target/riscv/insn_trans/trans_rvv.c.inc: use 'vlenb' in MAXSZ()
In reply to Daniel Henrique BarbozaMAXSZ() returns the value in bytes. Now that we have access to vlenb we don't need to compensate with a '-3' in the scale to use 'vlen'.
MAXSZ() now works as follows:
LMUL lmul=scale MAXSZ (vlenb >> -scale) 1 0 vlenb 2 1 vlenb << 1 4 2 vlenb << 2 8 3 vlenb << 3 1/8 -3 vlenb >> 3 1/4 -2 vlenb >> 2 1/2 -1 vlenb >> 1
--- target/riscv/insn_trans/trans_rvv.c.inc | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)
target/riscv/insn_trans/trans_rvv.c.inc +2 -3
@@ -1160,12 +1160,11 @@ GEN_LDST_WHOLE_TRANS(vs8r_v, 8, 1, true) /* * MAXSZ returns the maximum vector size can be operated in bytes, * which is used in GVEC IR when vl_eq_vlmax flag is set to true - * to accerlate vector operation. + * to accelerate vector operation. */ static inline uint32_t MAXSZ(DisasContext *s) { - int scale = s->lmul - 3; - return s->cfg_ptr->vlen >> -scale; + return s->cfg_ptr->vlenb >> -s->lmul; } static bool opivv_check(DisasContext *s, arg_rmrr *a)
Signature · 1 line
2.43.0
Re: [PATCH 10/13] target/riscv/insn_trans/trans_rvv.c.inc: use 'vlenb' in MAXSZ()
Quoted 10 lines · Daniel Henrique Barboza
MAXSZ()returns the value in bytes. Now that we have access to vlenb we don't need to compensate with a '-3' in the scale to use 'vlen'.
MAXSZ()now works as follows:LMUL lmul=scale MAXSZ (vlenb >> -scale)
1 0 vlenb 2 1 vlenb << 1 4 2 vlenb << 2 8 3 vlenb << 3
Um, what?
target/riscv/insn_trans/trans_rvv.c.inc GEN_LDST_WHOLE_TRANS(vs8r_v, 8, 1, true) · line 1167
Quoted 5 lines
static inline uint32_t MAXSZ(DisasContext *s) {target/riscv/insn_trans/trans_rvv.c.inc +1
-2- int scale = s->lmul - 3; - return s->cfg_ptr->vlen >> -scale; + return s->cfg_ptr->vlenb >> -s->lmul;
There's no left-shift here, either before or after. Bug?
r~
Re: [PATCH 10/13] target/riscv/insn_trans/trans_rvv.c.inc: use 'vlenb' in MAXSZ()
Quoted 14 lines · Richard Henderson
Quoted 10 lines · Daniel Henrique Barboza
MAXSZ()returns the value in bytes. Now that we have access to vlenb we don't need to compensate with a '-3' in the scale to use 'vlen'.
MAXSZ()now works as follows:LMUL lmul=scale MAXSZ (vlenb >> -scale)
1 0 vlenb
2 1 vlenb << 1
4 2 vlenb << 2
8 3 vlenb << 3Um, what?
Quoted 8 lines
Quoted 5 lines
static inline
uint32_tMAXSZ(DisasContext *s)
{
- int scale =s->lmul- 3;
- returns->cfg_ptr->vlen>> -scale;
+ returns->cfg_ptr->vlenb>> -s->lmul;There's no left-shift here, either before or after. Bug?
Yeah, I got lost in that explanation. I'll simplify it to be just a simple case of adjusting the right shift to use vlenb like in patches 8 and 9.
Thanks,
Daniel
Quoted 3 lines
r~
PATCH 11/13 trans_rvv.c.inc: remove vlmax arg from vec_element_loadx()
In reply to Daniel Henrique Barboza--- target/riscv/insn_trans/trans_rvv.c.inc | 26 +++++++++++++++++-------- 1 file changed, 18 insertions(+), 8 deletions(-)
target/riscv/insn_trans/trans_rvv.c.inc +18 -8
@@ -3265,21 +3265,28 @@ static void endian_adjust(TCGv_i32 ofs, int sew) #endif } -/* Load idx >= VLMAX ? 0 : vreg[idx] */ +/* + * Load idx >= VLMAX ? 0 : vreg[idx] + * + * This function assumes ctx->vl_eq_vlmax = true. + */ static void vec_element_loadx(DisasContext *s, TCGv_i64 dest, - int vreg, TCGv idx, int vlmax) + int vreg, TCGv idx) { TCGv_i32 ofs = tcg_temp_new_i32(); + TCGv_i32 vlmax = tcg_temp_new_i32(); TCGv_ptr base = tcg_temp_new_ptr(); TCGv_i64 t_idx = tcg_temp_new_i64(); - TCGv_i64 t_vlmax, t_zero; + TCGv_i64 t_zero, t_vlmax = tcg_temp_new_i64(); /* * Mask the index to the length so that we do * not produce an out-of-range load. */ + tcg_gen_trunc_tl_i32(vlmax, cpu_vl); + tcg_gen_sub_i32(vlmax, vlmax, tcg_constant_i32(1)); tcg_gen_trunc_tl_i32(ofs, idx); - tcg_gen_andi_i32(ofs, ofs, vlmax - 1); + tcg_gen_and_i32(ofs, ofs, vlmax); /* Convert the index to an offset. */ endian_adjust(ofs, s->sew);
@@ -3294,10 +3301,15 @@ static void vec_element_loadx(DisasContext *s, TCGv_i64 dest, vreg_ofs(s, vreg), s->sew, false); /* Flush out-of-range indexing to zero. */ - t_vlmax = tcg_constant_i64(vlmax); t_zero = tcg_constant_i64(0); tcg_gen_extu_tl_i64(t_idx, idx); +#ifdef TARGET_RISCV64 + tcg_gen_mov_i64(t_vlmax, cpu_vl); +#else + tcg_gen_extu_tl_i64(t_vlmax, cpu_vl); +#endif + tcg_gen_movcond_i64(TCG_COND_LTU, dest, t_idx, t_vlmax, dest, t_zero); }
@@ -3534,14 +3546,12 @@ static bool trans_vrgather_vx(DisasContext *s, arg_rmrr *a) } if (a->vm && s->vl_eq_vlmax && !(s->vta && s->lmul < 0)) { - int scale = s->lmul - (s->sew + 3); - int vlmax = s->cfg_ptr->vlen >> -scale; TCGv_i64 dest = tcg_temp_new_i64(); if (a->rs1 == 0) { vec_element_loadi(s, dest, a->rs2, 0, false); } else { - vec_element_loadx(s, dest, a->rs2, cpu_gpr[a->rs1], vlmax); + vec_element_loadx(s, dest, a->rs2, cpu_gpr[a->rs1]); } tcg_gen_gvec_dup_i64(s->sew, vreg_ofs(s, a->rd),
Signature · 1 line
2.43.0
Re: [PATCH 11/13] trans_rvv.c.inc: remove vlmax arg from vec_element_loadx()
Quoted 12 lines · Daniel Henrique Barboza
Signed-off-by Daniel Henrique Barboza <[email protected]>--- target/riscv/insn_trans/trans_rvv.c.inc | 26 +++++++++++++++++-------- 1 file changed, 18 insertions(+), 8 deletions(-)target/riscv/insn_trans/trans_rvv.c.inc +0
-0@@ -3265,21 +3265,28 @@ static void endian_adjust(TCGv_i32 ofs, int sew) #endif }
target/riscv/insn_trans/trans_rvv.c.inc static void endian_adjust(TCGv_i32 ofs, int sew) · line 3274
Quoted 10 lines
target/riscv/insn_trans/trans_rvv.c.inc +6
-2-/* Load idx >= VLMAX ? 0 : vreg[idx] */ +/* + * Load idx >= VLMAX ? 0 : vreg[idx] + * + * This function assumes ctx->vl_eq_vlmax = true. + */ static void vec_element_loadx(DisasContext *s, TCGv_i64 dest, - int vreg, TCGv idx, int vlmax) + int vreg, TCGv idx)
I think removing the cpu configuration constant is a mistake. Compile-time constants are always better than computation...
target/riscv/insn_trans/trans_rvv.c.inc static void vec_element_loadx(DisasContext *s, TCGv_i64 dest, · line 3311
Quoted 5 lines
target/riscv/insn_trans/trans_rvv.c.inc +5
-0+#ifdef TARGET_RISCV64 + tcg_gen_mov_i64(t_vlmax, cpu_vl); +#else + tcg_gen_extu_tl_i64(t_vlmax, cpu_vl); +#endif
That said, no ifdef required -- the second statement should always work.
r~
Re: [PATCH 11/13] trans_rvv.c.inc: remove vlmax arg from vec_element_loadx()
Quoted 17 lines · Richard Henderson
Quoted 16 lines · Daniel Henrique Barboza
Signed-off-by Daniel Henrique Barboza <[email protected]>---
target/riscv/insn_trans/trans_rvv.c.inc | 26 +++++++++++++++++--------1 file changed, 18 insertions(+), 8 deletions(-)target/riscv/insn_trans/trans_rvv.c.inc +3
-1@@ -3265,21 +3265,28 @@ static void endian_adjust(TCGv_i32 ofs, int sew) #endif } -/* Load idx >= VLMAX ? 0 : vreg[idx] */ +/* + * Load idx >= VLMAX ? 0 : vreg[idx] + *
Quoted 8 lines
target/riscv/insn_trans/trans_rvv.c.inc static void endian_adjust(TCGv_i32 ofs, int sew) · line 3274
Quoted 5 lines
target/riscv/insn_trans/trans_rvv.c.inc +3
-1+ * This function assumes ctx->vl_eq_vlmax = true. + */ static void vec_element_loadx(DisasContext *s, TCGv_i64 dest, - int vreg, TCGv idx, int vlmax) + int vreg, TCGv idx)I think removing the cpu configuration constant is a mistake. Compile-time constants are always better than computation...
Apparently my commit msg is AWOL ...
The 'vlmax' used in vec_element_loadx() is being calculated here, in trans_vrgather_vx():
- int scale = s->lmul - (s->sew + 3);
- int vlmax = s->cfg_ptr->vlen >> -scale;
My idea was to eliminate the use of 'vlen' since, in this block, 'vl_eq_vlmax' is true and we have 'vl' in the TCG global 'cpu_vl'.
I didn't find a way of reading 'cpu_vl' into an int variable and passing it as 'vlmax' to vec_element_loadx(), but inside vec_element_loadx() we can operate 'cpu_vl' normally since we're using TCG ops there. This is the reasoning behind this change.
I am now wondering if this is worth the trouble, and we should instead do:
+ int vlmax = cpu->cfg.vlenb >> (s->sew - s->lmul);
Like we're already doing in patch 9. Patch 12 would be a similar case.
Thanks,
Daniel
Quoted 12 lines
target/riscv/insn_trans/trans_rvv.c.inc static void vec_element_loadx(DisasContext *s, TCGv_i64 dest, · line 3311
Quoted 6 lines
target/riscv/insn_trans/trans_rvv.c.inc +5
-0+#ifdef TARGET_RISCV64 + tcg_gen_mov_i64(t_vlmax, cpu_vl); +#else + tcg_gen_extu_tl_i64(t_vlmax, cpu_vl); +#endifThat said, no ifdef required -- the second statement should always work.
r~
Re: [PATCH 11/13] trans_rvv.c.inc: remove vlmax arg from vec_element_loadx()
Quoted 5 lines · Daniel Henrique Barboza
I am now wondering if this is worth the trouble, and we should instead do:
+ int vlmax =
cpu->cfg.vlenb>> (s->sew-s->lmul);Like we're already doing in patch 9. Patch 12 would be a similar case.
This is more like what I expected to see.
r~
PATCH 12/13 trans_rvv.c.inc: use cpu_vl in trans_vrgather_vi()
In reply to Daniel Henrique BarbozaThere's no need to calculate 'vlmax'. We're garanteeing that 'vl_eq_vlmax' is true, and we can retrieve 'vl' via 'cpu_vl'.
--- target/riscv/insn_trans/trans_rvv.c.inc | 29 ++++++++++++++++--------- 1 file changed, 19 insertions(+), 10 deletions(-)
target/riscv/insn_trans/trans_rvv.c.inc +19 -10
@@ -3575,16 +3575,25 @@ static bool trans_vrgather_vi(DisasContext *s, arg_rmrr *a) } if (a->vm && s->vl_eq_vlmax && !(s->vta && s->lmul < 0)) { - int scale = s->lmul - (s->sew + 3); - int vlmax = s->cfg_ptr->vlen >> -scale; - if (a->rs1 >= vlmax) { - tcg_gen_gvec_dup_imm(MO_64, vreg_ofs(s, a->rd), - MAXSZ(s), MAXSZ(s), 0); - } else { - tcg_gen_gvec_dup_mem(s->sew, vreg_ofs(s, a->rd), - endian_ofs(s, a->rs2, a->rs1), - MAXSZ(s), MAXSZ(s)); - } + TCGLabel *dup_imm = gen_new_label(); + TCGLabel *done = gen_new_label(); + TCGv_i32 vlmax = tcg_temp_new_i32(); + + /* a->rs1 >= vlmax, cpu_vl = vlmax */ + tcg_gen_trunc_tl_i32(vlmax, cpu_vl); + tcg_gen_brcond_i32(TCG_COND_GEU, tcg_constant_i32(a->rs1), vlmax, + dup_imm); + tcg_gen_gvec_dup_mem(s->sew, vreg_ofs(s, a->rd), + endian_ofs(s, a->rs2, a->rs1), + MAXSZ(s), MAXSZ(s)); + tcg_gen_br(done); + + gen_set_label(dup_imm); + tcg_gen_gvec_dup_imm(MO_64, vreg_ofs(s, a->rd), + MAXSZ(s), MAXSZ(s), 0); + + gen_set_label(done); + mark_vs_dirty(s); } else { static gen_helper_opivx * const fns[4] = {
Signature · 1 line
2.43.0
PATCH 13/13 target/riscv/cpu.c: remove cpu->cfg.vlen
In reply to Daniel Henrique BarbozaThere is no need to keep both 'vlen' and 'vlenb'. All existing code that requires 'vlen' is retrieving it via 'vlenb << 3'.
--- target/riscv/cpu.c | 8 +++----- target/riscv/cpu_cfg.h | 1 - target/riscv/tcg/tcg-cpu.c | 4 +++- 3 files changed, 6 insertions(+), 7 deletions(-)
target/riscv/cpu.c +3 -5
@@ -1313,7 +1313,6 @@ static void riscv_cpu_init(Object *obj)/* Default values for non-bool cpu properties */ cpu->cfg.pmu_mask =MAKE_64BIT_MASK (3 ,16 ); - cpu->cfg.vlen =128 ; cpu->cfg.vlenb =128 >>3 ; cpu->cfg.elen =64 ; cpu->env.vext_ver =VEXT_VERSION_1_00_0 ;
@@ -1802,22 +1801,21 @@ static void prop_vlen_set(Object *obj, Visitor *v, const char *name,return ; } -if (value != cpu->cfg.vlen &&riscv_cpu_is_vendor (obj)) { +if (value != cpu->cfg.vlenb &&riscv_cpu_is_vendor (obj)) {cpu_set_prop_err (cpu, name, errp);error_append_hint (errp,"Current '%s' val: %u\n" , - name, cpu->cfg.vlen); + name, cpu->cfg.vlenb <<3 );return ; }cpu_option_add_user_setting (name, value); - cpu->cfg.vlen = value; cpu->cfg.vlenb = value >>3 ; }static void prop_vlen_get (Object *obj,Visitor *v,const char *name,void *opaque,Error **errp) { -uint16_t value =RISCV_CPU (obj)->cfg.vlen; +uint16_t value =RISCV_CPU (obj)->cfg.vlenb <<3 ;visit_type_uint16 (v, name, &value, errp); }
target/riscv/cpu_cfg.h +0 -1
@@ -139,7 +139,6 @@ struct RISCVCPUConfig {bool ext_XVentanaCondOps;uint32_t pmu_mask; -uint16_t vlen;uint16_t vlenb;uint16_t elen;uint16_t cbom_blocksize;
target/riscv/tcg/tcg-cpu.c +3 -1
@@ -298,7 +298,9 @@ static void riscv_cpu_validate_misa_mxl(RISCVCPU *cpu, Error **errp)static void riscv_cpu_validate_v (CPURISCVState *env,RISCVCPUConfig *cfg,Error **errp) { -if (cfg->vlen >RV_VLEN_MAX || cfg->vlen <128 ) { +uint32_t vlen = cfg->vlenb <<3 ; + +if (vlen >RV_VLEN_MAX || vlen <128 ) {error_setg (errp,"Vector extension implementation only supports VLEN " "in the range [128, %d]" ,RV_VLEN_MAX );
Signature · 1 line
2.43.0