[v2,01/28] target/i386: Add tcg/access.[ch]

Provide a method to amortize page lookup across large blocks.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/i386/tcg/access.h    |  40 +++++++++
 target/i386/tcg/access.c    | 160 ++++++++++++++++++++++++++++++++++++
 target/i386/tcg/meson.build |   1 +
 3 files changed, 201 insertions(+)
 create mode 100644 target/i386/tcg/access.h
 create mode 100644 target/i386/tcg/access.c

On 4/9/24 07:02, Richard Henderson wrote:
> Provide a method to amortize page lookup across large blocks.
> 
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>   target/i386/tcg/access.h    |  40 +++++++++
>   target/i386/tcg/access.c    | 160 ++++++++++++++++++++++++++++++++++++
>   target/i386/tcg/meson.build |   1 +
>   3 files changed, 201 insertions(+)
>   create mode 100644 target/i386/tcg/access.h
>   create mode 100644 target/i386/tcg/access.c
> 
> diff --git a/target/i386/tcg/access.h b/target/i386/tcg/access.h
> new file mode 100644
> index 0000000000..d70808a3a3
> --- /dev/null
> +++ b/target/i386/tcg/access.h
> @@ -0,0 +1,40 @@
> +/* SPDX-License-Identifier: GPL-2.0-or-later */
> +/* Access guest memory in blocks. */
> +
> +#ifndef X86_TCG_ACCESS_H
> +#define X86_TCG_ACCESS_H
> +
> +/* An access covers at most sizeof(X86XSaveArea), at most 2 pages. */
> +typedef struct X86Access {
> +    target_ulong vaddr;
> +    void *haddr1;
> +    void *haddr2;
> +    uint16_t size;
> +    uint16_t size1;
> +    /*
> +     * If we can't access the host page directly, we'll have to do I/O access
> +     * via ld/st helpers. These are internal details, so we store the rest
> +     * to do the access here instead of passing it around in the helpers.
> +     */
> +    int mmu_idx;
> +    CPUX86State *env;
> +    uintptr_t ra;
> +} X86Access;
> +
> +void access_prepare_mmu(X86Access *ret, CPUX86State *env,
> +                        vaddr vaddr, unsigned size,
> +                        MMUAccessType type, int mmu_idx, uintptr_t ra);
> +void access_prepare(X86Access *ret, CPUX86State *env, vaddr vaddr,
> +                    unsigned size, MMUAccessType type, uintptr_t ra);
> +
> +uint8_t  access_ldb(X86Access *ac, vaddr addr);
> +uint16_t access_ldw(X86Access *ac, vaddr addr);
> +uint32_t access_ldl(X86Access *ac, vaddr addr);
> +uint64_t access_ldq(X86Access *ac, vaddr addr);
> +
> +void access_stb(X86Access *ac, vaddr addr, uint8_t val);
> +void access_stw(X86Access *ac, vaddr addr, uint16_t val);
> +void access_stl(X86Access *ac, vaddr addr, uint32_t val);
> +void access_stq(X86Access *ac, vaddr addr, uint64_t val);
> +
> +#endif
> diff --git a/target/i386/tcg/access.c b/target/i386/tcg/access.c
> new file mode 100644
> index 0000000000..8b70f3244b
> --- /dev/null
> +++ b/target/i386/tcg/access.c
> @@ -0,0 +1,160 @@
> +/* SPDX-License-Identifier: GPL-2.0-or-later */
> +/* Access guest memory in blocks. */
> +
> +#include "qemu/osdep.h"
> +#include "cpu.h"
> +#include "exec/cpu_ldst.h"
> +#include "exec/exec-all.h"
> +#include "access.h"
> +
> +
> +void access_prepare_mmu(X86Access *ret, CPUX86State *env,
> +                        vaddr vaddr, unsigned size,
> +                        MMUAccessType type, int mmu_idx, uintptr_t ra)
> +{
> +    int size1, size2;
> +    void *haddr1, *haddr2;
> +
> +    assert(size > 0 && size <= TARGET_PAGE_SIZE);
> +
> +    size1 = MIN(size, -(vaddr | TARGET_PAGE_MASK)),
> +    size2 = size - size1;
> +
> +    memset(ret, 0, sizeof(*ret));
> +    ret->vaddr = vaddr;
> +    ret->size = size;
> +    ret->size1 = size1;
> +    ret->mmu_idx = mmu_idx;
> +    ret->env = env;
> +    ret->ra = ra;
> +
> +    haddr1 = probe_access(env, vaddr, size1, type, mmu_idx, ra);
> +    ret->haddr1 = haddr1;
> +
> +    if (unlikely(size2)) {
> +        haddr2 = probe_access(env, vaddr + size1, size2, type, mmu_idx, ra);
> +        if (haddr2 == haddr1 + size1) {
> +            ret->size1 = size;
> +        } else {
> +            ret->haddr2 = haddr2;
> +        }
> +    }

Should there be an assert(!ret->haddr2) here for the CONFIG_USER_ONLY 
case, or alternatively a g_assert_unreachable() in the "else" above?

> +}
> +
> +void access_prepare(X86Access *ret, CPUX86State *env, vaddr vaddr,
> +                    unsigned size, MMUAccessType type, uintptr_t ra)
> +{
> +    int mmu_idx = cpu_mmu_index(env_cpu(env), false);
> +    access_prepare_mmu(ret, env, vaddr, size, type, mmu_idx, ra);
> +}
> +
> +static void *access_ptr(X86Access *ac, vaddr addr, unsigned len)
> +{
> +    vaddr offset = addr - ac->vaddr;
> +
> +    assert(addr >= ac->vaddr);
> +
> +#ifdef CONFIG_USER_ONLY
> +    assert(offset <= ac->size1 - len);
> +    return ac->haddr1 + offset;
> +#else
> +    if (likely(offset <= ac->size1 - len)) {
> +        return ac->haddr1;
> +    }
> +    assert(offset <= ac->size - len);
> +    if (likely(offset >= ac->size1)) {
> +        return ac->haddr2;
> +    }

I think the returns should be (respectively) ac->haddr1 + offset and 
ac->haddr2 + (offset - ac->size1)?

Also I would add a comment above the second "if", like

     /*
      * If the address is not naturally aligned, it might span
      * both pages.  Only return ac->haddr2 if the area is
      * entirely within the second page, otherwise fall back
      * to slow accesses.
      */

Paolo

> +uint8_t access_ldb(X86Access *ac, vaddr addr)
> +{
> +    void *p = access_ptr(ac, addr, sizeof(uint8_t));
> +
> +    if (test_ptr(p)) {
> +        return ldub_p(p);
> +    }
> +    return cpu_ldub_mmuidx_ra(ac->env, addr, ac->mmu_idx, ac->ra);
> +}
> +
> +uint16_t access_ldw(X86Access *ac, vaddr addr)
> +{
> +    void *p = access_ptr(ac, addr, sizeof(uint16_t));
> +
> +    if (test_ptr(p)) {
> +        return lduw_le_p(p);
> +    }
> +    return cpu_lduw_le_mmuidx_ra(ac->env, addr, ac->mmu_idx, ac->ra);
> +}
> +
> +uint32_t access_ldl(X86Access *ac, vaddr addr)
> +{
> +    void *p = access_ptr(ac, addr, sizeof(uint32_t));
> +
> +    if (test_ptr(p)) {
> +        return ldl_le_p(p);
> +    }
> +    return cpu_ldl_le_mmuidx_ra(ac->env, addr, ac->mmu_idx, ac->ra);
> +}
> +
> +uint64_t access_ldq(X86Access *ac, vaddr addr)
> +{
> +    void *p = access_ptr(ac, addr, sizeof(uint64_t));
> +
> +    if (test_ptr(p)) {
> +        return ldq_le_p(p);
> +    }
> +    return cpu_ldq_le_mmuidx_ra(ac->env, addr, ac->mmu_idx, ac->ra);
> +}
> +
> +void access_stb(X86Access *ac, vaddr addr, uint8_t val)
> +{
> +    void *p = access_ptr(ac, addr, sizeof(uint8_t));
> +
> +    if (test_ptr(p)) {
> +        stb_p(p, val);
> +    } else {
> +        cpu_stb_mmuidx_ra(ac->env, addr, val, ac->mmu_idx, ac->ra);
> +    }
> +}
> +
> +void access_stw(X86Access *ac, vaddr addr, uint16_t val)
> +{
> +    void *p = access_ptr(ac, addr, sizeof(uint16_t));
> +
> +    if (test_ptr(p)) {
> +        stw_le_p(p, val);
> +    } else {
> +        cpu_stw_le_mmuidx_ra(ac->env, addr, val, ac->mmu_idx, ac->ra);
> +    }
> +}
> +
> +void access_stl(X86Access *ac, vaddr addr, uint32_t val)
> +{
> +    void *p = access_ptr(ac, addr, sizeof(uint32_t));
> +
> +    if (test_ptr(p)) {
> +        stl_le_p(p, val);
> +    } else {
> +        cpu_stl_le_mmuidx_ra(ac->env, addr, val, ac->mmu_idx, ac->ra);
> +    }
> +}
> +
> +void access_stq(X86Access *ac, vaddr addr, uint64_t val)
> +{
> +    void *p = access_ptr(ac, addr, sizeof(uint64_t));
> +
> +    if (test_ptr(p)) {
> +        stq_le_p(p, val);
> +    } else {
> +        cpu_stq_le_mmuidx_ra(ac->env, addr, val, ac->mmu_idx, ac->ra);
> +    }
> +}
> diff --git a/target/i386/tcg/meson.build b/target/i386/tcg/meson.build
> index f9110e890c..1105b35d92 100644
> --- a/target/i386/tcg/meson.build
> +++ b/target/i386/tcg/meson.build
> @@ -1,4 +1,5 @@
>   i386_ss.add(when: 'CONFIG_TCG', if_true: files(
> +  'access.c',
>     'bpt_helper.c',
>     'cc_helper.c',
>     'excp_helper.c',