Re: [PATCH v2] RISC-V: Add support for Ztso

[Palmer Dabbelt <palmer@rivosinc.com>]

On Thu, 22 Sep 2022 21:35:26 PDT (-0700), alistair23@gmail.com wrote:
> On Sat, Sep 17, 2022 at 6:12 PM Palmer Dabbelt <palmer@rivosinc.com> wrote:
>>
>> The Ztso extension was recently frozen, this adds it as a CPU property
>> and adds various fences throughout the port in order to allow TSO
>> targets to function on weaker hosts.  We need no fences for AMOs as
>> they're already SC, the placess we need barriers are described.
>>
>> Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
>> ---
>> Like the v1 this has been pretty minimally tested, but I figured I'd
>> just send it.  This is what I would have done the first time had I not
>> read that comment near TCG_GUEST_DEFAULT_MO, I think trying to describe
>> that when I ran into Richard at the Cauldron was probably more confusing
>> than just digging up the code and sending it along.
>
> At a quick glance this looks reasonable. Did I miss anything or is
> this ready to go?

Richard and I ended up chatting about it at the Cauldron: the original 
plan was to have TCG insert these fences, but that would require a lot 
of plumbing for the RISC-V use case where we can have heterogenous harts 
(and likely soon dynamic harts, IIRC there's some HW that already 
supports that).

Maybe Richard can chime in, but I think the rough plan was to go with 
this and just ammend the commit message explaining why it's different 
than the x86-on-arm64 support that's planned.  I think

    The Ztso extension was recently frozen, this adds it as a CPU property
    and adds various fences throughout the port in order to allow TSO
    targets to function on weaker hosts.  We need no fences for AMOs as
    they're already SC, the placess we need barriers are described.  
    These fences are placed in the RISC-V backend rather than TCG as is 
    planned for x86-on-arm64 because RISC-V allows heterogenous (and 
    likely soon dynamic) hart memory models.

should do it.

> Alistair
>
>> ---
>>  target/riscv/cpu.c                      |  3 +++
>>  target/riscv/cpu.h                      |  1 +
>>  target/riscv/insn_trans/trans_rva.c.inc | 11 ++++++++---
>>  target/riscv/insn_trans/trans_rvi.c.inc | 16 ++++++++++++++--
>>  target/riscv/insn_trans/trans_rvv.c.inc | 20 ++++++++++++++++++++
>>  target/riscv/translate.c                |  3 +++
>>  6 files changed, 49 insertions(+), 5 deletions(-)
>>
>> diff --git a/target/riscv/cpu.c b/target/riscv/cpu.c
>> index ac6f82ebd0..d66169efa5 100644
>> --- a/target/riscv/cpu.c
>> +++ b/target/riscv/cpu.c
>> @@ -919,6 +919,8 @@ static Property riscv_cpu_extensions[] = {
>>      DEFINE_PROP_BOOL("zhinx", RISCVCPU, cfg.ext_zhinx, false),
>>      DEFINE_PROP_BOOL("zhinxmin", RISCVCPU, cfg.ext_zhinxmin, false),
>>
>> +    DEFINE_PROP_BOOL("ztso", RISCVCPU, cfg.ext_ztso, false),
>> +
>>      /* Vendor-specific custom extensions */
>>      DEFINE_PROP_BOOL("xventanacondops", RISCVCPU, cfg.ext_XVentanaCondOps, false),
>>
>> @@ -1094,6 +1096,7 @@ static void riscv_isa_string_ext(RISCVCPU *cpu, char **isa_str, int max_str_len)
>>          ISA_EDATA_ENTRY(zksed, ext_zksed),
>>          ISA_EDATA_ENTRY(zksh, ext_zksh),
>>          ISA_EDATA_ENTRY(zkt, ext_zkt),
>> +        ISA_EDATA_ENTRY(ztso, ext_ztso),
>>          ISA_EDATA_ENTRY(zve32f, ext_zve32f),
>>          ISA_EDATA_ENTRY(zve64f, ext_zve64f),
>>          ISA_EDATA_ENTRY(zhinx, ext_zhinx),
>> diff --git a/target/riscv/cpu.h b/target/riscv/cpu.h
>> index 5c7acc055a..c64fd4e258 100644
>> --- a/target/riscv/cpu.h
>> +++ b/target/riscv/cpu.h
>> @@ -433,6 +433,7 @@ struct RISCVCPUConfig {
>>      bool ext_zve32f;
>>      bool ext_zve64f;
>>      bool ext_zmmul;
>> +    bool ext_ztso;
>>      bool rvv_ta_all_1s;
>>
>>      uint32_t mvendorid;
>> diff --git a/target/riscv/insn_trans/trans_rva.c.inc b/target/riscv/insn_trans/trans_rva.c.inc
>> index 45db82c9be..9066e1bde3 100644
>> --- a/target/riscv/insn_trans/trans_rva.c.inc
>> +++ b/target/riscv/insn_trans/trans_rva.c.inc
>> @@ -26,7 +26,11 @@ static bool gen_lr(DisasContext *ctx, arg_atomic *a, MemOp mop)
>>          tcg_gen_mb(TCG_MO_ALL | TCG_BAR_STRL);
>>      }
>>      tcg_gen_qemu_ld_tl(load_val, src1, ctx->mem_idx, mop);
>> -    if (a->aq) {
>> +    /*
>> +     * TSO defines AMOs as acquire+release-RCsc, but does not define LR/SC as
>> +     * AMOs.  Instead treat them like loads.
>> +     */
>> +    if (a->aq || ctx->ztso) {
>>          tcg_gen_mb(TCG_MO_ALL | TCG_BAR_LDAQ);
>>      }
>>
>> @@ -61,9 +65,10 @@ static bool gen_sc(DisasContext *ctx, arg_atomic *a, MemOp mop)
>>      gen_set_label(l1);
>>      /*
>>       * Address comparison failure.  However, we still need to
>> -     * provide the memory barrier implied by AQ/RL.
>> +     * provide the memory barrier implied by AQ/RL/TSO.
>>       */
>> -    tcg_gen_mb(TCG_MO_ALL + a->aq * TCG_BAR_LDAQ + a->rl * TCG_BAR_STRL);
>> +    TCGBar bar_strl = (ctx->ztso || a->rl) ? TCG_BAR_STRL : 0;
>> +    tcg_gen_mb(TCG_MO_ALL + a->aq * TCG_BAR_LDAQ + bar_strl);
>>      gen_set_gpr(ctx, a->rd, tcg_constant_tl(1));
>>
>>      gen_set_label(l2);
>> diff --git a/target/riscv/insn_trans/trans_rvi.c.inc b/target/riscv/insn_trans/trans_rvi.c.inc
>> index ca8e3d1ea1..9bef42a3e5 100644
>> --- a/target/riscv/insn_trans/trans_rvi.c.inc
>> +++ b/target/riscv/insn_trans/trans_rvi.c.inc
>> @@ -261,11 +261,19 @@ static bool gen_load_i128(DisasContext *ctx, arg_lb *a, MemOp memop)
>>
>>  static bool gen_load(DisasContext *ctx, arg_lb *a, MemOp memop)
>>  {
>> +    bool out;
>> +
>>      if (get_xl(ctx) == MXL_RV128) {
>> -        return gen_load_i128(ctx, a, memop);
>> +        out = gen_load_i128(ctx, a, memop);
>>      } else {
>> -        return gen_load_tl(ctx, a, memop);
>> +        out = gen_load_tl(ctx, a, memop);
>> +    }
>> +
>> +    if (ctx->ztso) {
>> +        tcg_gen_mb(TCG_MO_ALL | TCG_BAR_LDAQ);
>>      }
>> +
>> +    return out;
>>  }
>>
>>  static bool trans_lb(DisasContext *ctx, arg_lb *a)
>> @@ -322,6 +330,10 @@ static bool gen_store_tl(DisasContext *ctx, arg_sb *a, MemOp memop)
>>      TCGv addr = get_address(ctx, a->rs1, a->imm);
>>      TCGv data = get_gpr(ctx, a->rs2, EXT_NONE);
>>
>> +    if (ctx->ztso) {
>> +        tcg_gen_mb(TCG_MO_ALL | TCG_BAR_STRL);
>> +    }
>> +
>>      tcg_gen_qemu_st_tl(data, addr, ctx->mem_idx, memop);
>>      return true;
>>  }
>> diff --git a/target/riscv/insn_trans/trans_rvv.c.inc b/target/riscv/insn_trans/trans_rvv.c.inc
>> index 6c091824b6..1994b38035 100644
>> --- a/target/riscv/insn_trans/trans_rvv.c.inc
>> +++ b/target/riscv/insn_trans/trans_rvv.c.inc
>> @@ -671,8 +671,28 @@ static bool ldst_us_trans(uint32_t vd, uint32_t rs1, uint32_t data,
>>      tcg_gen_addi_ptr(dest, cpu_env, vreg_ofs(s, vd));
>>      tcg_gen_addi_ptr(mask, cpu_env, vreg_ofs(s, 0));
>>
>> +    /*
>> +     * According to the specification
>> +     *
>> +     *   Additionally, if the Ztso extension is implemented, then vector memory
>> +     *   instructions in the V extension and Zve family of extensions follow
>> +     *   RVTSO at the instruction level.  The Ztso extension does not
>> +     *   strengthen the ordering of intra-instruction element accesses.
>> +     *
>> +     * as a result neither ordered nor unordered accesses from the V
>> +     * instructions need ordering within the loop but we do still need barriers
>> +     * around the loop.
>> +     */
>> +    if (is_store && s->ztso) {
>> +      tcg_gen_mb(TCG_MO_ALL | TCG_BAR_STRL);
>> +    }
>> +
>>      fn(dest, mask, base, cpu_env, desc);
>>
>> +    if (!is_store && s->ztso) {
>> +      tcg_gen_mb(TCG_MO_ALL | TCG_BAR_LDAQ);
>> +    }
>> +
>>      tcg_temp_free_ptr(dest);
>>      tcg_temp_free_ptr(mask);
>>
>> diff --git a/target/riscv/translate.c b/target/riscv/translate.c
>> index 63b04e8a94..c7c574b09f 100644
>> --- a/target/riscv/translate.c
>> +++ b/target/riscv/translate.c
>> @@ -109,6 +109,8 @@ typedef struct DisasContext {
>>      /* PointerMasking extension */
>>      bool pm_mask_enabled;
>>      bool pm_base_enabled;
>> +    /* Ztso */
>> +    bool ztso;
>>      /* TCG of the current insn_start */
>>      TCGOp *insn_start;
>>  } DisasContext;
>> @@ -1109,6 +1111,7 @@ static void riscv_tr_init_disas_context(DisasContextBase *dcbase, CPUState *cs)
>>      memset(ctx->ftemp, 0, sizeof(ctx->ftemp));
>>      ctx->pm_mask_enabled = FIELD_EX32(tb_flags, TB_FLAGS, PM_MASK_ENABLED);
>>      ctx->pm_base_enabled = FIELD_EX32(tb_flags, TB_FLAGS, PM_BASE_ENABLED);
>> +    ctx->ztso = cpu->cfg.ext_ztso;
>>      ctx->zero = tcg_constant_tl(0);
>>  }
>>
>> --
>> 2.34.1
>>
>>