This patch uses gen_goto_tb() to reduce the overhead of the unconditional lookup_and_goto_ptr() exit currently generated for vsetvli.
When translating vsetvli, the immediate vtype provides information that can be used to optimize the exit generated by do_vsetvl(). For the vsetvli forms optimized by this patch, the immediate vtype and existing state update flow determine the other relevant translation state. Only VL_EQ_VLMAX can vary with AVL. This patch therefore uses separate jump slots depending on whether the returned VL equals VLMAX, enabling direct TB linking after vsetvli. The register-vtype vsetvl instruction retains the existing lookup path, as its vtype is not known at translation time. The package benchmarks measure build time with the RISC-V cc1 compiler running under QEMU, while the dedicated workload repeatedly executes two vsetvli instructions, alternating between VL = VLMAX and VL < VLMAX. benchmark result: Dedicated workload: | Workload | QEMU | Runs | Median (s) | Min–max (s) | Wall change (s) | Wall change | Wall / baseline | | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | | vsetvli-chain | baseline | 5 | 2.154520 | 2.145543–2.173475 | 0.000000 | 0.00% | 1.000x | | vsetvli-chain | candidate | 5 | 0.492811 | 0.489481–0.499191 | -1.661710 | -77.13% | 0.229x | Perf of dedicated workload: | Workload | QEMU | Event | Samples | Symbol | Symbol samples | Overhead | | --- | --- | --- | ---: | --- | ---: | ---: | | vsetvli-chain | baseline | cycles:u | 2151 | `helper_lookup_tb_ptr` | 2007 | 93.58%| | vsetvli-chain | candidate | cycles:u | 482 | `helper_lookup_tb_ptr` | 0 | 0.00% | Package build benchmark | Package | QEMU | Runs | Median wall (s) | Min–max (s) | Stddev (s) | Wall change | Wall / baseline | | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | | lz4 | candidate | 3 | 79.506 | 79.000–80.063 | 0.532 | -1.25% | 0.987x | | zlib | candidate | 3 | 45.474 | 45.426–45.476 | 0.028 | -0.44% | 0.996x | | sqlite | candidate | 3 | 278.771 | 278.470–278.962 | 0.248 | -1.07% | 0.989x | Signed-off-by: WX Chen <[email protected]> --- v3: - Expand the commit message to explain the optimization rationale, following feedback from Sashiko AI review. - Add a benchmark result. - No code changes. v2: The overall structure is quite the same as v1 except that v2 considers the condition that QEMU forces `VL_EQ_VLMAX = false` when `maxsz < 8`, which reduces the number of `goto_tb` slots from two to one in some cases. For `vsetvli` with `rd != x0 && rs1 == x0`, v2 also uses a single `goto_tb` slot because the successor TB lookup key is fixed for a given source TB. Both changes follow Chao Liu's suggestions. v2 patch link: https://lore.kernel.org/qemu-devel/20261110-research-vsetvli-direct-chain-20261104-v2-1-8e65aadfb...@gmail.com/ benchmark link: https://github.com/zhenbaii/qemu-riscv-bench To run the benchmark from the benchmark repository, pass the baseline and patched QEMU commit IDs to the vsetvli-v2 recipe: ./reproduce.sh BASELINE_COMMIT PATCHED_COMMIT --recipe vsetvli-v2 --- target/riscv/tcg/insn_trans/trans_rvv.c.inc | 30 ++++++++++++++++++++++++----- 1 file changed, 25 insertions(+), 5 deletions(-) diff --git a/target/riscv/tcg/insn_trans/trans_rvv.c.inc b/target/riscv/tcg/insn_trans/trans_rvv.c.inc index 7e1295f33e7d4b7a2ff26cbf428f6b5d9eb682ff..17e8eb5ecaede655b58b5bdc9f10ff9284efc5b6 100644 --- a/target/riscv/tcg/insn_trans/trans_rvv.c.inc +++ b/target/riscv/tcg/insn_trans/trans_rvv.c.inc @@ -191,7 +191,8 @@ static bool require_noover(const int8_t dst, const int8_t dst_lmul, return !is_overlapped(dst, dst_size, src, src_size); } -static bool do_vsetvl(DisasContext *s, int rd, int rs1, TCGv s2) +static bool do_vsetvl(DisasContext *s, int rd, int rs1, TCGv s2, + uint32_t vtype, bool vtype_is_imm) { TCGv s1, dst; @@ -215,8 +216,27 @@ static bool do_vsetvl(DisasContext *s, int rd, int rs1, TCGv s2) gen_set_gpr(s, rd, dst); finalize_rvv_inst(s); - gen_update_pc(s, s->cur_insn_len); - lookup_and_goto_ptr(s); + if (!vtype_is_imm || (rd == 0 && rs1 == 0)) { + gen_update_pc(s, s->cur_insn_len); + lookup_and_goto_ptr(s); + } else { + uint32_t vsew = FIELD_EX32(vtype, VTYPE, VSEW); + int8_t lmul = sextract32(FIELD_EX32(vtype, VTYPE, VLMUL), 0, 3); + uint32_t vlmax = vext_get_vlmax(s->cfg_ptr->vlenb, vsew, lmul); + + if (rs1 == 0 || (vlmax << vsew) < 8) { + gen_goto_tb(s, 0, s->cur_insn_len); + } else { + target_ulong pc_save = s->pc_save; + TCGLabel *not_vlmax = gen_new_label(); + + tcg_gen_brcondi_tl(TCG_COND_NE, dst, vlmax, not_vlmax); + gen_goto_tb(s, 0, s->cur_insn_len); + s->pc_save = pc_save; + gen_set_label(not_vlmax); + gen_goto_tb(s, 1, s->cur_insn_len); + } + } s->base.is_jmp = DISAS_NORETURN; return true; } @@ -244,13 +264,13 @@ static bool do_vsetivli(DisasContext *s, int rd, TCGv s1, TCGv s2) static bool trans_vsetvl(DisasContext *s, arg_vsetvl *a) { TCGv s2 = get_gpr(s, a->rs2, EXT_ZERO); - return do_vsetvl(s, a->rd, a->rs1, s2); + return do_vsetvl(s, a->rd, a->rs1, s2, 0, false); } static bool trans_vsetvli(DisasContext *s, arg_vsetvli *a) { TCGv s2 = tcg_constant_tl(a->zimm); - return do_vsetvl(s, a->rd, a->rs1, s2); + return do_vsetvl(s, a->rd, a->rs1, s2, a->zimm, true); } static bool trans_vsetivli(DisasContext *s, arg_vsetivli *a) --- base-commit: a3565937b39779d4e3704298095f566e7559ad53 change-id: 20261104-research-vsetvli-direct-chain-20261104-206bc2104ddb Best regards, -- WX Chen <[email protected]>
