From: Alistair <alistair23@gmail.com>
To: Richard Henderson <richard.henderson@linaro.org>, qemu-devel@nongnu.org
Cc: cota@braap.org, alex.bennee@linaro.org
Subject: Re: [Qemu-devel] [PATCH 08/13] tcg/riscv: enable dynamic TLB sizing
Date: Fri, 25 Jan 2019 14:16:14 -0800 [thread overview]
Message-ID: <ec37df14-eee4-9220-43b5-76bbcad547e9@gmail.com> (raw)
In-Reply-To: <20190123225705.28963-9-richard.henderson@linaro.org>
On 1/23/19 2:57 PM, Richard Henderson wrote:
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alistair Francis <alistair.francis@wdc.com>
Alistair
> ---
> tcg/riscv/tcg-target.h | 2 +-
> tcg/riscv/tcg-target.inc.c | 126 ++++++++++++++++---------------------
> 2 files changed, 56 insertions(+), 72 deletions(-)
>
> diff --git a/tcg/riscv/tcg-target.h b/tcg/riscv/tcg-target.h
> index 1eb032626c..83b123ca03 100644
> --- a/tcg/riscv/tcg-target.h
> +++ b/tcg/riscv/tcg-target.h
> @@ -33,7 +33,7 @@
>
> #define TCG_TARGET_INSN_UNIT_SIZE 4
> #define TCG_TARGET_TLB_DISPLACEMENT_BITS 20
> -#define TCG_TARGET_IMPLEMENTS_DYN_TLB 0
> +#define TCG_TARGET_IMPLEMENTS_DYN_TLB 1
> #define TCG_TARGET_NB_REGS 32
>
> typedef enum {
> diff --git a/tcg/riscv/tcg-target.inc.c b/tcg/riscv/tcg-target.inc.c
> index 6cf8de32b5..b785f4acb7 100644
> --- a/tcg/riscv/tcg-target.inc.c
> +++ b/tcg/riscv/tcg-target.inc.c
> @@ -958,6 +958,17 @@ static void * const qemu_st_helpers[16] = {
> [MO_BEQ] = helper_be_stq_mmu,
> };
>
> +/* We don't support oversize guests */
> +QEMU_BUILD_BUG_ON(TCG_TARGET_REG_BITS < TARGET_LONG_BITS);
> +
> +/* We expect tlb_mask to be before tlb_table. */
> +QEMU_BUILD_BUG_ON(offsetof(CPUArchState, tlb_table) <
> + offsetof(CPUArchState, tlb_mask));
> +
> +/* We expect tlb_mask to be "near" tlb_table. */
> +QEMU_BUILD_BUG_ON(offsetof(CPUArchState, tlb_table) -
> + offsetof(CPUArchState, tlb_mask) >= 0x800);
> +
> static void tcg_out_tlb_load(TCGContext *s, TCGReg addrl,
> TCGReg addrh, TCGMemOpIdx oi,
> tcg_insn_unit **label_ptr, bool is_load)
> @@ -965,94 +976,67 @@ static void tcg_out_tlb_load(TCGContext *s, TCGReg addrl,
> TCGMemOp opc = get_memop(oi);
> unsigned s_bits = opc & MO_SIZE;
> unsigned a_bits = get_alignment_bits(opc);
> - target_ulong mask;
> + tcg_target_long compare_mask;
> int mem_index = get_mmuidx(oi);
> - int cmp_off
> - = (is_load
> - ? offsetof(CPUArchState, tlb_table[mem_index][0].addr_read)
> - : offsetof(CPUArchState, tlb_table[mem_index][0].addr_write));
> - int add_off = offsetof(CPUArchState, tlb_table[mem_index][0].addend);
> - RISCVInsn load_cmp_op = (TARGET_LONG_BITS == 64 ? OPC_LD :
> - TCG_TARGET_REG_BITS == 64 ? OPC_LWU : OPC_LW);
> - RISCVInsn load_add_op = TCG_TARGET_REG_BITS == 64 ? OPC_LD : OPC_LW;
> - TCGReg base = TCG_AREG0;
> + int mask_off, table_off;
> + TCGReg mask_base = TCG_AREG0, table_base = TCG_AREG0;
>
> - /* We don't support oversize guests */
> - if (TCG_TARGET_REG_BITS < TARGET_LONG_BITS) {
> - g_assert_not_reached();
> + mask_off = offsetof(CPUArchState, tlb_mask[mem_index]);
> + table_off = offsetof(CPUArchState, tlb_table[mem_index]);
> + if (table_off > 0x7ff) {
> + int mask_hi = mask_off - sextreg(mask_off, 0, 12);
> + int table_hi = table_off - sextreg(table_off, 0, 12);
> +
> + if (likely(mask_hi == table_hi)) {
> + mask_base = table_base = TCG_REG_TMP1;
> + tcg_out_opc_upper(s, OPC_LUI, mask_base, mask_hi);
> + tcg_out_opc_reg(s, OPC_ADD, mask_base, mask_base, TCG_AREG0);
> + mask_off -= mask_hi;
> + table_off -= mask_hi;
> + } else {
> + mask_base = TCG_REG_TMP0;
> + table_base = TCG_REG_TMP1;
> + tcg_out_opc_upper(s, OPC_LUI, mask_base, mask_hi);
> + tcg_out_opc_reg(s, OPC_ADD, mask_base, mask_base, TCG_AREG0);
> + table_off -= mask_off;
> + mask_off -= mask_hi;
> + tcg_out_opc_imm(s, OPC_ADDI, table_base, mask_base, mask_off);
> + }
> }
>
> + tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, mask_base, mask_off);
> + tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, table_base, table_off);
> +
> + tcg_out_opc_imm(s, OPC_SRLI, TCG_REG_TMP2, addrl,
> + TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
> + tcg_out_opc_reg(s, OPC_AND, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP0);
> + tcg_out_opc_reg(s, OPC_ADD, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP1);
> +
> + /* Load the tlb comparator and the addend. */
> + tcg_out_ld(s, TCG_TYPE_TL, TCG_REG_TMP0, TCG_REG_TMP2,
> + is_load ? offsetof(CPUTLBEntry, addr_read)
> + : offsetof(CPUTLBEntry, addr_write));
> + tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP2, TCG_REG_TMP2,
> + offsetof(CPUTLBEntry, addend));
> +
> /* We don't support unaligned accesses. */
> if (a_bits < s_bits) {
> a_bits = s_bits;
> }
> - mask = (target_ulong)TARGET_PAGE_MASK | ((1 << a_bits) - 1);
> -
> -
> - /* Compensate for very large offsets. */
> - if (add_off >= 0x1000) {
> - int adj;
> - base = TCG_REG_TMP2;
> - if (cmp_off <= 2 * 0xfff) {
> - adj = 0xfff;
> - tcg_out_opc_imm(s, OPC_ADDI, base, TCG_AREG0, adj);
> - } else {
> - adj = cmp_off - sextreg(cmp_off, 0, 12);
> - tcg_debug_assert(add_off - adj >= -0x1000
> - && add_off - adj < 0x1000);
> -
> - tcg_out_opc_upper(s, OPC_LUI, base, adj);
> - tcg_out_opc_reg(s, OPC_ADD, base, base, TCG_AREG0);
> - }
> - add_off -= adj;
> - cmp_off -= adj;
> - }
> -
> - /* Extract the page index. */
> - if (CPU_TLB_BITS + CPU_TLB_ENTRY_BITS < 12) {
> - tcg_out_opc_imm(s, OPC_SRLI, TCG_REG_TMP0, addrl,
> - TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
> - tcg_out_opc_imm(s, OPC_ANDI, TCG_REG_TMP0, TCG_REG_TMP0,
> - MAKE_64BIT_MASK(CPU_TLB_ENTRY_BITS, CPU_TLB_BITS));
> - } else if (TARGET_PAGE_BITS >= 12) {
> - tcg_out_opc_upper(s, OPC_LUI, TCG_REG_TMP0,
> - MAKE_64BIT_MASK(TARGET_PAGE_BITS, CPU_TLB_BITS));
> - tcg_out_opc_reg(s, OPC_AND, TCG_REG_TMP0, TCG_REG_TMP0, addrl);
> - tcg_out_opc_imm(s, OPC_SRLI, TCG_REG_TMP0, TCG_REG_TMP0,
> - CPU_TLB_BITS - CPU_TLB_ENTRY_BITS);
> - } else {
> - tcg_out_opc_imm(s, OPC_SRLI, TCG_REG_TMP0, addrl, TARGET_PAGE_BITS);
> - tcg_out_opc_imm(s, OPC_ANDI, TCG_REG_TMP0, TCG_REG_TMP0,
> - MAKE_64BIT_MASK(0, CPU_TLB_BITS));
> - tcg_out_opc_imm(s, OPC_SLLI, TCG_REG_TMP0, TCG_REG_TMP0,
> - CPU_TLB_ENTRY_BITS);
> - }
> -
> - /* Add that to the base address to index the tlb. */
> - tcg_out_opc_reg(s, OPC_ADD, TCG_REG_TMP2, base, TCG_REG_TMP0);
> - base = TCG_REG_TMP2;
> -
> - /* Load the tlb comparator and the addend. */
> - tcg_out_ldst(s, load_cmp_op, TCG_REG_TMP0, base, cmp_off);
> - tcg_out_ldst(s, load_add_op, TCG_REG_TMP2, base, add_off);
> -
> /* Clear the non-page, non-alignment bits from the address. */
> - if (mask == sextreg(mask, 0, 12)) {
> - tcg_out_opc_imm(s, OPC_ANDI, TCG_REG_TMP1, addrl, mask);
> + compare_mask = (tcg_target_long)TARGET_PAGE_MASK | ((1 << a_bits) - 1);
> + if (compare_mask == sextreg(compare_mask, 0, 12)) {
> + tcg_out_opc_imm(s, OPC_ANDI, TCG_REG_TMP1, addrl, compare_mask);
> } else {
> - tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_TMP1, mask);
> + tcg_out_movi(s, TCG_TYPE_TL, TCG_REG_TMP1, compare_mask);
> tcg_out_opc_reg(s, OPC_AND, TCG_REG_TMP1, TCG_REG_TMP1, addrl);
> - }
> + }
>
> /* Compare masked address with the TLB entry. */
> label_ptr[0] = s->code_ptr;
> tcg_out_opc_branch(s, OPC_BNE, TCG_REG_TMP0, TCG_REG_TMP1, 0);
> /* NOP to allow patching later */
> tcg_out_opc_imm(s, OPC_ADDI, TCG_REG_ZERO, TCG_REG_ZERO, 0);
> - /* TODO: Move this out of line
> - * see:
> - * https://lists.nongnu.org/archive/html/qemu-devel/2018-11/msg02234.html
> - */
>
> /* TLB Hit - translate address using addend. */
> if (TCG_TARGET_REG_BITS > TARGET_LONG_BITS) {
>
next prev parent reply other threads:[~2019-01-25 22:16 UTC|newest]
Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-01-23 22:56 [Qemu-devel] [PATCH 00/13] Dynamic TLB sizing, backends Richard Henderson
2019-01-23 22:56 ` [Qemu-devel] [PATCH 01/13] cputlb: do not evict empty entries to the vtlb Richard Henderson
2019-01-23 22:56 ` [Qemu-devel] [PATCH 02/13] tcg: introduce dynamic TLB sizing Richard Henderson
2019-01-23 22:56 ` [Qemu-devel] [PATCH 03/13] tcg/i386: enable " Richard Henderson
2019-01-23 22:56 ` [Qemu-devel] [PATCH 04/13] tcg/aarch64: " Richard Henderson
2019-01-25 19:12 ` Alex Bennée
2019-01-25 20:09 ` Richard Henderson
2019-01-23 22:56 ` [Qemu-devel] [PATCH 05/13] tcg/ppc: " Richard Henderson
2019-01-23 22:56 ` [Qemu-devel] [PATCH 06/13] tcg/sparc: " Richard Henderson
2019-01-23 22:56 ` [Qemu-devel] [PATCH 07/13] tcg/s390: " Richard Henderson
2019-01-23 22:57 ` [Qemu-devel] [PATCH 08/13] tcg/riscv: " Richard Henderson
2019-01-25 22:16 ` Alistair [this message]
2019-01-23 22:57 ` [Qemu-devel] [PATCH 09/13] tcg/arm: " Richard Henderson
2019-01-23 22:57 ` [Qemu-devel] [PATCH 10/13] tcg/mips: Fix tcg_out_qemu_ld_slow_path Richard Henderson
2019-01-23 22:57 ` [Qemu-devel] [PATCH 11/13] tcg/mips: enable dynamic TLB sizing Richard Henderson
2019-01-23 22:57 ` [Qemu-devel] [PATCH 12/13] tcg/tci: " Richard Henderson
2019-01-23 22:57 ` [Qemu-devel] [PATCH 13/13] cputlb: Remove static tlb sizing Richard Henderson
2019-01-25 22:17 ` Alistair
2019-01-31 17:58 ` [Qemu-devel] [PATCH 00/13] Dynamic TLB sizing, backends no-reply
2019-02-01 22:16 ` no-reply
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=ec37df14-eee4-9220-43b5-76bbcad547e9@gmail.com \
--to=alistair23@gmail.com \
--cc=alex.bennee@linaro.org \
--cc=cota@braap.org \
--cc=qemu-devel@nongnu.org \
--cc=richard.henderson@linaro.org \
/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;
as well as URLs for NNTP newsgroup(s).