Re: [PATCH 23/30] tcg/loongarch: Add softmmu load/store helpers, implement qemu_ld/qemu_st ops

[Richard Henderson]

On 9/20/21 1:04 AM, WANG Xuerui wrote:
> +static void tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
> +{
> +    tcg_out_opc_b(s, 0);
> +    bool ok = reloc_sd10k16(s->code_ptr - 1, target);
> +    tcg_debug_assert(ok);
> +}

Hmm.  This is an existing bug in tcg/riscv/.  We should have no asserts on relocations 
being in range.  We should always be able to tell our caller that the relocation failed, 
and we'll try again with a smaller TB.

In this case, return the result of reloc_sd10k16 and ...

> +
> +    tcg_out_goto(s, l->raddr);
> +    return true;
> +}

... return the result of tcg_out_goto.


> +static void tcg_out_tlb_load(TCGContext *s, TCGReg addrl, TCGMemOpIdx oi,
> +                             tcg_insn_unit **label_ptr, bool is_load)
> +{
> +    MemOp opc = get_memop(oi);
> +    unsigned s_bits = opc & MO_SIZE;
> +    unsigned a_bits = get_alignment_bits(opc);
> +    tcg_target_long compare_mask;
> +    int mem_index = get_mmuidx(oi);
> +    int fast_ofs = TLB_MASK_TABLE_OFS(mem_index);
> +    int mask_ofs = fast_ofs + offsetof(CPUTLBDescFast, mask);
> +    int table_ofs = fast_ofs + offsetof(CPUTLBDescFast, table);
> +    TCGReg mask_base = TCG_AREG0, table_base = TCG_AREG0;
> +
> +    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, mask_base, mask_ofs);
> +    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, table_base, table_ofs);

I think we can eliminate the mask_base and table_base variables now.  This dates from the 
TCG_TARGET_TLB_DISPLACEMENT_BITS thing, where we would need to compute an intermediate 
offset, and adjust these base registers.

> +    /* Clear the non-page, non-alignment bits from the address.  */
> +    compare_mask = (tcg_target_long)TARGET_PAGE_MASK | ((1 << a_bits) - 1);
> +    if (compare_mask == sextreg(compare_mask, 0, 12)) {
> +        tcg_out_opc_andi(s, TCG_REG_TMP1, addrl, compare_mask);

LoongArch uses an unsigned mask for andi, not signed.  The immediate case will never match 
for LoongArch.

> +static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
> +{
> +    TCGMemOpIdx oi = l->oi;
> +    MemOp opc = get_memop(oi);
> +    TCGReg a0 = tcg_target_call_iarg_regs[0];
> +    TCGReg a1 = tcg_target_call_iarg_regs[1];
> +    TCGReg a2 = tcg_target_call_iarg_regs[2];
> +    TCGReg a3 = tcg_target_call_iarg_regs[3];

Drop these, since you've already named TCG_REG_A0 etc.

> +
> +    /* We don't support oversize guests */
> +    if (TCG_TARGET_REG_BITS < TARGET_LONG_BITS) {
> +        g_assert_not_reached();
> +    }

This is redundant with TCG_TARGET_REG_BITS == 64.

> +    tcg_out_call(s, qemu_ld_helpers[opc & MO_SSIZE]);
> +    tcg_out_mov(s, (opc & MO_SIZE) == MO_64, l->datalo_reg, a0);

Because you have single-insn sign-extend instructions, it's better to always call the 
unsigned load function, and then sign-extend here.  See the aarch64 version.

> +    tcg_out_mov(s, TCG_TYPE_PTR, a2, l->datalo_reg);
> +    switch (s_bits) {
> +    case MO_8:
> +        tcg_out_ext8u(s, a2, a2);
> +        break;
> +    case MO_16:
> +        tcg_out_ext16u(s, a2, a2);
> +        break;
> +    default:
> +        break;
> +    }

Do you have a pointer to the LoongArch ABI?  Do 32-bit values need to be sign- or 
zero-extended in the call arguments?

Anyway, merge the move and extend.

> +    tcg_out_movi(s, TCG_TYPE_PTR, a4, (tcg_target_long)l->raddr);

Oh, just FYI, this is another case where movi wants to handle pc-relative 
addresses.

> +    if (guest_base == 0) {
> +        tcg_out_opc_add_d(s, base, addr_regl, TCG_REG_ZERO);

Don't add zero.  RISCV bug that's been fixed recently.

> +static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg lo,
> +                                   TCGReg base, MemOp opc)
> +{
> +    /* Byte swapping is left to middle-end expansion.  */
> +    tcg_debug_assert((opc & MO_BSWAP) == 0);
> +
> +    switch (opc & (MO_SSIZE)) {

MO_SIZE here.

> +static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is_64)

Shouldn't need is_64 argument for store.  It's only relevant to load so that TCG_TYPE_I32 
values are always sign-extended in the host register.


r~