From: Palmer Dabbelt <palmer@dabbelt.com>
To: liuyang22@iscas.ac.cn
Cc: gdb-patches@sourceware.org, Andrew Burgess <aburgess@redhat.com>,
simon.marchi@polymtl.ca
Subject: Re: [PATCH] gdb: RISC-V: Refine lr/sc sequence support
Date: Thu, 02 Nov 2023 08:58:53 -0700 (PDT) [thread overview]
Message-ID: <mhng-e5f786bc-2a32-4288-958c-0743887d3d2f@palmer-ri-x1c9> (raw)
In-Reply-To: <8bd610eb-2713-4f1a-b928-ad9fa3020dd0@iscas.ac.cn>
On Thu, 02 Nov 2023 08:21:16 PDT (-0700), liuyang22@iscas.ac.cn wrote:
>
> On 2023/11/2 22:03, Yang Liu wrote:
>> Per RISC-V spec, the lr/sc sequence can consist of up to 16 instructions, and we
>> cannot insert breakpoints in the middle of this sequence. Before this, we only
>> detected a specific pattern (the most common one). This patch improves this part
>> and now supports more complex pattern detection.
Thanks. The general idea here seems pretty reasonable, but I haven't
looked that closely at the code. I'm super behind on Linux stuff right
now, so I might be slower that usual...
>> gdb/ChangeLog:
>>
>> * gdb/riscv-tdep.c (class riscv_insn): Add more needed opcode enums.
>> (riscv_insn::decode): Decode newly added opcodes.
>> (riscv_insn_is_non_cti_and_allowed_in_atomic_sequence): New.
>> (riscv_insn_is_direct_branch): New.
>> (riscv_next_pc_atomic_sequence): Removed.
>> (riscv_deal_with_atomic_sequence): Rename from riscv_next_pc_atomic_sequence.
>> (riscv_software_single_step): Adjust to use the renamed one.
>>
>> Signed-off-by: Yang Liu <liuyang22@iscas.ac.cn>
>> ---
>> gdb/riscv-tdep.c | 266 ++++++++++++++++++++++++++++++++++++++++-------
>> 1 file changed, 226 insertions(+), 40 deletions(-)
>>
>> diff --git a/gdb/riscv-tdep.c b/gdb/riscv-tdep.c
>> index 3a2891c2c92..faff1aab420 100644
>> --- a/gdb/riscv-tdep.c
>> +++ b/gdb/riscv-tdep.c
>> @@ -1578,8 +1578,34 @@ class riscv_insn
>> BLTU,
>> BGEU,
>> /* These are needed for stepping over atomic sequences. */
>> - LR,
>> - SC,
>> + SLTI,
>> + SLTIU,
>> + XORI,
>> + ORI,
>> + ANDI,
>> + SLLI,
>> + SLLIW,
>> + SRLI,
>> + SRLIW,
>> + SRAI,
>> + SRAIW,
>> + SUB,
>> + SUBW,
>> + SLL,
>> + SLLW,
>> + SLT,
>> + SLTU,
>> + XOR,
>> + SRL,
>> + SRLW,
>> + SRA,
>> + SRAW,
>> + OR,
>> + AND,
>> + LR_W,
>> + LR_D,
>> + SC_W,
>> + SC_D,
>> /* This instruction is used to do a syscall. */
>> ECALL,
>>
>> @@ -1768,6 +1794,13 @@ class riscv_insn
>> m_imm.s = EXTRACT_CBTYPE_IMM (ival);
>> }
>>
>> + void decode_ca_type_insn (enum opcode opcode, ULONGEST ival)
>> + {
>> + m_opcode = opcode;
>> + m_rs1 = decode_register_index_short (ival, OP_SH_CRS1S);
>> + m_rs2 = decode_register_index_short (ival, OP_SH_CRS2S);
>> + }
>> +
>> /* Fetch instruction from target memory at ADDR, return the content of
>> the instruction, and update LEN with the instruction length. */
>> static ULONGEST fetch_instruction (struct gdbarch *gdbarch,
>> @@ -1882,14 +1915,62 @@ riscv_insn::decode (struct gdbarch *gdbarch, CORE_ADDR pc)
>> decode_b_type_insn (BLTU, ival);
>> else if (is_bgeu_insn (ival))
>> decode_b_type_insn (BGEU, ival);
>> + else if (is_slti_insn(ival))
>> + decode_i_type_insn (SLTI, ival);
>> + else if (is_sltiu_insn(ival))
>> + decode_i_type_insn (SLTIU, ival);
>> + else if (is_xori_insn(ival))
>> + decode_i_type_insn (XORI, ival);
>> + else if (is_ori_insn(ival))
>> + decode_i_type_insn (ORI, ival);
>> + else if (is_andi_insn(ival))
>> + decode_i_type_insn (ANDI, ival);
>> + else if (is_slli_insn(ival))
>> + decode_i_type_insn (SLLI, ival);
>> + else if (is_slliw_insn(ival))
>> + decode_i_type_insn (SLLIW, ival);
>> + else if (is_srli_insn(ival))
>> + decode_i_type_insn (SRLI, ival);
>> + else if (is_srliw_insn(ival))
>> + decode_i_type_insn (SRLIW, ival);
>> + else if (is_srai_insn(ival))
>> + decode_i_type_insn (SRAI, ival);
>> + else if (is_sraiw_insn(ival))
>> + decode_i_type_insn (SRAIW, ival);
>> + else if (is_sub_insn(ival))
>> + decode_r_type_insn (SUB, ival);
>> + else if (is_subw_insn(ival))
>> + decode_r_type_insn (SUBW, ival);
>> + else if (is_sll_insn(ival))
>> + decode_r_type_insn (SLL, ival);
>> + else if (is_sllw_insn(ival))
>> + decode_r_type_insn (SLLW, ival);
>> + else if (is_slt_insn(ival))
>> + decode_r_type_insn (SLT, ival);
>> + else if (is_sltu_insn(ival))
>> + decode_r_type_insn (SLTU, ival);
>> + else if (is_xor_insn(ival))
>> + decode_r_type_insn (XOR, ival);
>> + else if (is_srl_insn(ival))
>> + decode_r_type_insn (SRL, ival);
>> + else if (is_srlw_insn(ival))
>> + decode_r_type_insn (SRLW, ival);
>> + else if (is_sra_insn(ival))
>> + decode_r_type_insn (SRA, ival);
>> + else if (is_sraw_insn(ival))
>> + decode_r_type_insn (SRAW, ival);
>> + else if (is_or_insn(ival))
>> + decode_r_type_insn (OR, ival);
>> + else if (is_and_insn(ival))
>> + decode_r_type_insn (AND, ival);
>> else if (is_lr_w_insn (ival))
>> - decode_r_type_insn (LR, ival);
>> + decode_r_type_insn (LR_W, ival);
>> else if (is_lr_d_insn (ival))
>> - decode_r_type_insn (LR, ival);
>> + decode_r_type_insn (LR_D, ival);
>> else if (is_sc_w_insn (ival))
>> - decode_r_type_insn (SC, ival);
>> + decode_r_type_insn (SC_W, ival);
>> else if (is_sc_d_insn (ival))
>> - decode_r_type_insn (SC, ival);
>> + decode_r_type_insn (SC_D, ival);
>> else if (is_ecall_insn (ival))
>> decode_i_type_insn (ECALL, ival);
>> else if (is_ld_insn (ival))
>> @@ -1944,6 +2025,24 @@ riscv_insn::decode (struct gdbarch *gdbarch, CORE_ADDR pc)
>> m_rd = decode_register_index (ival, OP_SH_CRS1S);
>> m_imm.s = EXTRACT_CITYPE_LUI_IMM (ival);
>> }
>> + else if (is_c_srli_insn (ival))
>> + decode_cb_type_insn (SRLI, ival);
>> + else if (is_c_srai_insn (ival))
>> + decode_cb_type_insn (SRAI, ival);
>> + else if (is_c_andi_insn (ival))
>> + decode_cb_type_insn (ANDI, ival);
>> + else if (is_c_sub_insn (ival))
>> + decode_ca_type_insn (SUB, ival);
>> + else if (is_c_xor_insn (ival))
>> + decode_ca_type_insn (XOR, ival);
>> + else if (is_c_or_insn (ival))
>> + decode_ca_type_insn (OR, ival);
>> + else if (is_c_and_insn (ival))
>> + decode_ca_type_insn (AND, ival);
>> + else if (is_c_subw_insn (ival))
>> + decode_ca_type_insn (SUBW, ival);
>> + else if (is_c_addw_insn (ival))
>> + decode_ca_type_insn (ADDW, ival);
>> else if (is_c_li_insn (ival))
>> decode_ci_type_insn (LI, ival);
>> /* C_SD and C_FSW have the same opcode. C_SD is RV64 and RV128 only,
>> @@ -4404,51 +4503,138 @@ riscv_next_pc (struct regcache *regcache, CORE_ADDR pc)
>> return next_pc;
>> }
>>
>> +/* Return true if INSN is not a control transfer instruction and is allowed to
>> + appear in the middle of the lr/sc sequence. */
>> +
>> +static bool
>> +riscv_insn_is_non_cti_and_allowed_in_atomic_sequence(const struct riscv_insn &insn)
>> +{
>> + switch (insn.opcode ())
>> + {
>> + case riscv_insn::LUI:
>> + case riscv_insn::AUIPC:
>> + case riscv_insn::ADDI:
>> + case riscv_insn::ADDIW:
>> + case riscv_insn::SLTI:
>> + case riscv_insn::SLTIU:
>> + case riscv_insn::XORI:
>> + case riscv_insn::ORI:
>> + case riscv_insn::ANDI:
>> + case riscv_insn::SLLI:
>> + case riscv_insn::SLLIW:
>> + case riscv_insn::SRLI:
>> + case riscv_insn::SRLIW:
>> + case riscv_insn::SRAI:
>> + case riscv_insn::ADD:
>> + case riscv_insn::ADDW:
>> + case riscv_insn::SRAIW:
>> + case riscv_insn::SUB:
>> + case riscv_insn::SUBW:
>> + case riscv_insn::SLL:
>> + case riscv_insn::SLLW:
>> + case riscv_insn::SLT:
>> + case riscv_insn::SLTU:
>> + case riscv_insn::XOR:
>> + case riscv_insn::SRL:
>> + case riscv_insn::SRLW:
>> + case riscv_insn::SRA:
>> + case riscv_insn::SRAW:
>> + case riscv_insn::OR:
>> + case riscv_insn::AND:
>> + return true;
>> + }
>> +
>> + return false;
>> +}
>> +
>> +/* Return true if INSN is a direct branch insctruction. */
>> +
>> +static bool
>> +riscv_insn_is_direct_branch(const struct riscv_insn &insn)
>> +{
>> + switch (insn.opcode ())
>> + {
>> + case riscv_insn::BEQ:
>> + case riscv_insn::BNE:
>> + case riscv_insn::BLT:
>> + case riscv_insn::BGE:
>> + case riscv_insn::BLTU:
>> + case riscv_insn::BGEU:
>> + case riscv_insn::JAL:
>> + return true;
>> + }
>> +
>> + return false;
>> +}
>> +
>> /* We can't put a breakpoint in the middle of a lr/sc atomic sequence, so look
>> for the end of the sequence and put the breakpoint there. */
>>
>> -static bool
>> -riscv_next_pc_atomic_sequence (struct regcache *regcache, CORE_ADDR pc,
>> - CORE_ADDR *next_pc)
>> +static std::vector<CORE_ADDR>
>> +riscv_deal_with_atomic_sequence (struct regcache *regcache, CORE_ADDR pc)
>> {
>> struct gdbarch *gdbarch = regcache->arch ();
>> struct riscv_insn insn;
>> - CORE_ADDR cur_step_pc = pc;
>> - CORE_ADDR last_addr = 0;
>> + CORE_ADDR cur_step_pc = pc, next_pc;
>> + std::vector<CORE_ADDR> next_pcs;
>> + const int atomic_sequence_length = 16;
>> + bool found_valid_atomic_sequence = false;
>> + enum riscv_insn::opcode lr_opcode;
>>
>> /* First instruction has to be a load reserved. */
>> insn.decode (gdbarch, cur_step_pc);
>> - if (insn.opcode () != riscv_insn::LR)
>> - return false;
>> - cur_step_pc = cur_step_pc + insn.length ();
>> + lr_opcode = insn.opcode ();
>> + if (lr_opcode != riscv_insn::LR_D && lr_opcode != riscv_insn::LR_W)
>> + return {};
>>
>> - /* Next instruction should be branch to exit. */
>> - insn.decode (gdbarch, cur_step_pc);
>> - if (insn.opcode () != riscv_insn::BNE)
>> - return false;
>> - last_addr = cur_step_pc + insn.imm_signed ();
>> - cur_step_pc = cur_step_pc + insn.length ();
>> + /* A lr/sc sequence comprise at most 16 instructions placed sequentially in memory. */
>> + for (int insn_count = 0; insn_count < atomic_sequence_length; ++insn_count)
>> + {
>> + cur_step_pc += insn.length ();
>> + insn.decode (gdbarch, cur_step_pc);
>>
>> - /* Next instruction should be store conditional. */
>> - insn.decode (gdbarch, cur_step_pc);
>> - if (insn.opcode () != riscv_insn::SC)
>> - return false;
>> - cur_step_pc = cur_step_pc + insn.length ();
>> + /* The dynamic code executed between lr/sc can only contain instructions
>> + from the base I instruction set, excluding loads, stores, backward jumps,
>> + taken backward branches, JALR, FENCE, FENCE.I, and SYSTEM instructions.
>> + If the C extension is supported, then compressed forms of the aforementioned
>> + I instructions are also permitted. */
That ends up with some odd interactions with the Zc extensions. I
opened https://github.com/riscv/riscv-isa-manual/issues/1157 to check.
>> +
>> + if (riscv_insn_is_non_cti_and_allowed_in_atomic_sequence (insn))
>> + {
>> + continue;
>> + }
>> + /* Look for a conditional branch instruction, check if it's taken forward or not. */
>> + else if (riscv_insn_is_direct_branch (insn))
>> + {
>> + if (insn.imm_signed () > 0)
>> + next_pc = cur_step_pc + insn.imm_signed ();
>
> Oh I forgot to do a push_back here and my simple tests didn't catch
> this. I'll send a V2 patch soon. Sorry.
I also think the branch detection logic here isn't quite right: the
LR/SC sequence only becomes unconstrained if the backwards branch is
taken at runtime. I can't find any examples of people doing that, but
one could imagine some sort of spin-type patterns with an early retry
that do this sort of thing.
The same logic applies to disallowed instructions, but I don't think
it's super likely those end up conditionally executed in LR/SC sequences
so maybe we can ignore that?
> Best regards,
>
> Yang
>
>> + else
>> + break;
>> + }
>> + /* Look for a paired SC instruction which closes the atomic sequence. */
>> + else if ((insn.opcode () == riscv_insn::SC_D && lr_opcode == riscv_insn::LR_D)
>> + || (insn.opcode () == riscv_insn::SC_W && lr_opcode == riscv_insn::LR_W))
>> + {
>> + found_valid_atomic_sequence = true;
>> + }
>> + else
>> + break;
>> + }
>> +
>> + if (!found_valid_atomic_sequence)
>> + return {};
>>
>> /* Next instruction should be branch to start. */
>> insn.decode (gdbarch, cur_step_pc);
>> if (insn.opcode () != riscv_insn::BNE)
>> - return false;
>> + return {};
>> if (pc != (cur_step_pc + insn.imm_signed ()))
>> - return false;
>> - cur_step_pc = cur_step_pc + insn.length ();
>> -
>> - /* We should now be at the end of the sequence. */
>> - if (cur_step_pc != last_addr)
>> - return false;
>> + return {};
>> + cur_step_pc += insn.length ();
>>
>> - *next_pc = cur_step_pc;
>> - return true;
>> + next_pc = cur_step_pc;
>> + next_pcs.push_back (next_pc);
>> + return next_pcs;
>> }
>>
>> /* This is called just before we want to resume the inferior, if we want to
>> @@ -4458,14 +4644,14 @@ riscv_next_pc_atomic_sequence (struct regcache *regcache, CORE_ADDR pc,
>> std::vector<CORE_ADDR>
>> riscv_software_single_step (struct regcache *regcache)
>> {
>> - CORE_ADDR pc, next_pc;
>> -
>> - pc = regcache_read_pc (regcache);
>> + CORE_ADDR cur_pc = regcache_read_pc (regcache), next_pc;
>> + std::vector<CORE_ADDR> next_pcs
>> + = riscv_deal_with_atomic_sequence (regcache, cur_pc);
>>
>> - if (riscv_next_pc_atomic_sequence (regcache, pc, &next_pc))
>> - return {next_pc};
>> + if (!next_pcs.empty ())
>> + return next_pcs;
>>
>> - next_pc = riscv_next_pc (regcache, pc);
>> + next_pc = riscv_next_pc (regcache, cur_pc);
>>
>> return {next_pc};
>> }
next prev parent reply other threads:[~2023-11-02 15:59 UTC|newest]
Thread overview: 4+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-11-02 14:03 Yang Liu
2023-11-02 15:21 ` Yang Liu
2023-11-02 15:58 ` Palmer Dabbelt [this message]
2023-11-02 17:54 ` Yang Liu
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=mhng-e5f786bc-2a32-4288-958c-0743887d3d2f@palmer-ri-x1c9 \
--to=palmer@dabbelt.com \
--cc=aburgess@redhat.com \
--cc=gdb-patches@sourceware.org \
--cc=liuyang22@iscas.ac.cn \
--cc=simon.marchi@polymtl.ca \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox