This change adjusts the translation exit of `vsetivli`. While preserving
the existing helper, state update flow, and TB boundaries, it uses the
generic `gen_goto_tb()` path to reduce the overhead of repeated successor-TB
lookups.
`vsetivli` is used to set the vector length and type, and both its AVL and
vtype encodings are provided by immediates. Given the CPU configuration and
the source TB execution context, after this instruction completes normally,
the vector state changes relevant to subsequent translation are
deterministic. Therefore, this exit is eligible for direct TB linking.
Measured on an AMD Ryzen AI 9 H 365 (x86-64) host, using qemu-riscv64
with -cpu rv64,v=true,vlen=128,elen=64. A RISC-V GCC 14.2.0 cc1 built for
rv64gcv compiles the packages with -O2 -g0 -march=rv64gcv -mabi=lp64d.
Only cc1 runs under emulation; make, the compiler driver, assembler and
linker run on the host.
Each variant is warmed up once, followed by three paired clean builds in
AB/BA/AB order, using -j1 and fixed host CPU affinity (CPU 7). Median build
wall times are:
Before (s) After (s) Wall time reduction
LZ4 1.10.0 125.017 123.060 1.57%
zlib 1.3.1 71.062 70.310 1.06%
In a separate untimed test, a loop executes two vsetivli instructions per
iteration for 37 iterations. Host-side GDB entry breakpoints show that
helper_lookup_tb_ptr() entries at the two successor PCs fall from 74 to 0.
The guest still executes 74 vsetivli instructions. Initial TB dispatch and
link establishment still occur through the normal dispatcher.
Signed-off-by: WX Chen <[email protected]>
---
target/riscv/tcg/insn_trans/trans_rvv.c.inc | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
diff --git a/target/riscv/tcg/insn_trans/trans_rvv.c.inc
b/target/riscv/tcg/insn_trans/trans_rvv.c.inc
index
a22e2cae6ce8e5a055c106fb76c305c92ac53d3c..b299c82d7a29057433e280b49b5220e6ad71498f
100644
--- a/target/riscv/tcg/insn_trans/trans_rvv.c.inc
+++ b/target/riscv/tcg/insn_trans/trans_rvv.c.inc
@@ -234,8 +234,7 @@ static bool do_vsetivli(DisasContext *s, int rd, TCGv s1,
TCGv s2)
gen_helper_vsetvl(dst, tcg_env, s1, s2, tcg_constant_tl(0));
gen_set_gpr(s, rd, dst);
finalize_rvv_inst(s);
- gen_update_pc(s, s->cur_insn_len);
- lookup_and_goto_ptr(s);
+ gen_goto_tb(s, 0, s->cur_insn_len);
s->base.is_jmp = DISAS_NORETURN;
return true;
---
base-commit: d73149283205645625bfa9fe029cfb68e4c4a9ae
change-id: 20261104-research-vsetivli-direct-chain-20261103-69536f07be37
Best regards,
--
WX Chen <[email protected]>