[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-arm] [PATCH v2 52/67] target/arm: Implement SVE store vector/p
From: |
Peter Maydell |
Subject: |
Re: [Qemu-arm] [PATCH v2 52/67] target/arm: Implement SVE store vector/predicate register |
Date: |
Tue, 27 Feb 2018 14:21:06 +0000 |
On 17 February 2018 at 18:23, Richard Henderson
<address@hidden> wrote:
> Signed-off-by: Richard Henderson <address@hidden>
> ---
> target/arm/translate-sve.c | 101
> +++++++++++++++++++++++++++++++++++++++++++++
> target/arm/sve.decode | 6 +++
> 2 files changed, 107 insertions(+)
>
> diff --git a/target/arm/translate-sve.c b/target/arm/translate-sve.c
> index b000a2482e..9c724980a0 100644
> --- a/target/arm/translate-sve.c
> +++ b/target/arm/translate-sve.c
> @@ -3501,6 +3501,95 @@ static void do_ldr(DisasContext *s, uint32_t vofs,
> uint32_t len,
> tcg_temp_free_i64(t0);
> }
>
> +/* Similarly for stores. */
> +static void do_str(DisasContext *s, uint32_t vofs, uint32_t len,
> + int rn, int imm)
> +{
> + uint32_t len_align = QEMU_ALIGN_DOWN(len, 8);
> + uint32_t len_remain = len % 8;
> + uint32_t nparts = len / 8 + ctpop8(len_remain);
> + int midx = get_mem_index(s);
> + TCGv_i64 addr, t0;
> +
> + addr = tcg_temp_new_i64();
> + t0 = tcg_temp_new_i64();
> +
> + /* Note that unpredicated load/store of vector/predicate registers
> + * are defined as a stream of bytes, which equates to little-endian
> + * operations on larger quantities. There is no nice way to force
> + * a little-endian load for aarch64_be-linux-user out of line.
"store" in this case, I assume.
> + *
> + * Attempt to keep code expansion to a minimum by limiting the
> + * amount of unrolling done.
> + */
> + if (nparts <= 4) {
> + int i;
> +
> + for (i = 0; i < len_align; i += 8) {
> + tcg_gen_ld_i64(t0, cpu_env, vofs + i);
> + tcg_gen_addi_i64(addr, cpu_reg_sp(s, rn), imm + i);
> + tcg_gen_qemu_st_i64(t0, addr, midx, MO_LEQ);
> + }
> + } else {
> + TCGLabel *loop = gen_new_label();
> + TCGv_ptr i = TCGV_NAT_TO_PTR(glue(tcg_const_local_, ptr)(0));
> + TCGv_ptr src;
> +
> + gen_set_label(loop);
> +
> + src = tcg_temp_new_ptr();
> + tcg_gen_add_ptr(src, cpu_env, i);
> + tcg_gen_ld_i64(t0, src, vofs);
> +
> + /* Minimize the number of local temps that must be re-read from
> + * the stack each iteration. Instead, re-compute values other
> + * than the loop counter.
> + */
> + tcg_gen_addi_ptr(src, i, imm);
> +#if UINTPTR_MAX == UINT32_MAX
> + tcg_gen_extu_i32_i64(addr, TCGV_PTR_TO_NAT(src));
> + tcg_gen_add_i64(addr, addr, cpu_reg_sp(s, rn));
> +#else
> + tcg_gen_add_i64(addr, TCGV_PTR_TO_NAT(src), cpu_reg_sp(s, rn));
> +#endif
We should be able to avoid the ifdef with more support for tcg_*_ptr
ops (similar to an earlier patch in the series).
> + tcg_temp_free_ptr(src);
> +
> + tcg_gen_qemu_st_i64(t0, addr, midx, MO_LEQ);
> +
> + tcg_gen_addi_ptr(i, i, 8);
> +
> + glue(tcg_gen_brcondi_, ptr)(TCG_COND_LTU, TCGV_PTR_TO_NAT(i),
> + len_align, loop);
> + tcg_temp_free_ptr(i);
> + }
> +
> + /* Predicate register stores can be any multiple of 2. */
> + if (len_remain) {
> + tcg_gen_ld_i64(t0, cpu_env, vofs + len_align);
> + tcg_gen_addi_i64(addr, cpu_reg_sp(s, rn), imm + len_align);
> +
> + switch (len_remain) {
> + case 2:
> + case 4:
> + case 8:
> + tcg_gen_qemu_st_i64(t0, addr, midx, MO_LE | ctz32(len_remain));
> + break;
> +
> + case 6:
> + tcg_gen_qemu_st_i64(t0, addr, midx, MO_LEUL);
> + tcg_gen_addi_i64(addr, addr, 4);
> + tcg_gen_shri_i64(addr, addr, 32);
> + tcg_gen_qemu_st_i64(t0, addr, midx, MO_LEUW);
> + break;
> +
> + default:
> + g_assert_not_reached();
> + }
> + }
> + tcg_temp_free_i64(addr);
> + tcg_temp_free_i64(t0);
> +}
> +
> #undef ptr
thanks
-- PMM
- [Qemu-arm] [PATCH v2 47/67] target/arm: Implement SVE integer convert to floating-point, (continued)
- [Qemu-arm] [PATCH v2 47/67] target/arm: Implement SVE integer convert to floating-point, Richard Henderson, 2018/02/17
- [Qemu-arm] [PATCH v2 48/67] target/arm: Implement SVE floating-point arithmetic (predicated), Richard Henderson, 2018/02/17
- [Qemu-arm] [PATCH v2 49/67] target/arm: Implement SVE FP Multiply-Add Group, Richard Henderson, 2018/02/17
- [Qemu-arm] [PATCH v2 50/67] target/arm: Implement SVE Floating Point Accumulating Reduction Group, Richard Henderson, 2018/02/17
- [Qemu-arm] [PATCH v2 52/67] target/arm: Implement SVE store vector/predicate register, Richard Henderson, 2018/02/17
- Re: [Qemu-arm] [PATCH v2 52/67] target/arm: Implement SVE store vector/predicate register,
Peter Maydell <=
- [Qemu-arm] [PATCH v2 51/67] target/arm: Implement SVE load and broadcast element, Richard Henderson, 2018/02/17
- [Qemu-arm] [PATCH v2 53/67] target/arm: Implement SVE scatter stores, Richard Henderson, 2018/02/17
- [Qemu-arm] [PATCH v2 54/67] target/arm: Implement SVE prefetches, Richard Henderson, 2018/02/17
- [Qemu-arm] [PATCH v2 55/67] target/arm: Implement SVE gather loads, Richard Henderson, 2018/02/17
- [Qemu-arm] [PATCH v2 56/67] target/arm: Implement SVE scatter store vector immediate, Richard Henderson, 2018/02/17