| Message ID | ZQHLcL6E5uNvjkaN@p100 |
|---|---|
| State | New |
| Headers | show |
| Series | target/hppa: Optimize ldcw/ldcd instruction translation | expand |
On 9/13/23 07:47, Helge Deller wrote: > + haddr = (uint32_t *)((uintptr_t)vaddr); > + old = *haddr; This is horribly incorrect, both for user-only and system mode. > + /* if already zero, do not write 0 again to reduce memory presssure */ > + if (old == 0) { > + return 0; > + } > + old = qatomic_xchg(haddr, (uint32_t) 0); You're also dropping the required host memory barrier. Frankly, the current tcg_gen_atomic_xchg_reg is as optimized as you'll be able to do. I really doubt the "avoid write 0" is measurable at all. r~
On 9/13/23 18:55, Richard Henderson wrote: > On 9/13/23 07:47, Helge Deller wrote: >> + haddr = (uint32_t *)((uintptr_t)vaddr); >> + old = *haddr; > > This is horribly incorrect, both for user-only and system mode. Richard, thank you for the review! But would you mind explaining why this is incorrect? I thought the "vaddr = probe_access()" calculates the host address, so shouldn't it be the right address? >> + /* if already zero, do not write 0 again to reduce memory presssure */ >> + if (old == 0) { >> + return 0; >> + } >> + old = qatomic_xchg(haddr, (uint32_t) 0); > > You're also dropping the required host memory barrier. ? > Frankly, the current tcg_gen_atomic_xchg_reg is as optimized as you'll be able to do. I really doubt the "avoid write 0" is measurable at all. Helge
On 9/13/23 10:19, Helge Deller wrote: > On 9/13/23 18:55, Richard Henderson wrote: >> On 9/13/23 07:47, Helge Deller wrote: >>> + haddr = (uint32_t *)((uintptr_t)vaddr); >>> + old = *haddr; >> >> This is horribly incorrect, both for user-only and system mode. > > Richard, thank you for the review! > But would you mind explaining why this is incorrect? > I thought the "vaddr = probe_access()" calculates the host address, so > shouldn't it be the right address? The vaddr name is confusing (since it implies virtual address, which the return from probe_access is not) as are the casts, which are unnecessary. > >>> + /* if already zero, do not write 0 again to reduce memory presssure */ >>> + if (old == 0) { >>> + return 0; >>> + } >>> + old = qatomic_xchg(haddr, (uint32_t) 0); >> >> You're also dropping the required host memory barrier. > > ? The path through the read+test+return, without the qatomic_xchg, has no host memory barrier to provide sequential consistency of the entire operation. r~
Hi Richard, On 9/13/23 22:30, Richard Henderson wrote: > On 9/13/23 10:19, Helge Deller wrote: >> On 9/13/23 18:55, Richard Henderson wrote: >>> On 9/13/23 07:47, Helge Deller wrote: >>>> + haddr = (uint32_t *)((uintptr_t)vaddr); >>>> + old = *haddr; >>> >>> This is horribly incorrect, both for user-only and system mode. >> >> Richard, thank you for the review! >> But would you mind explaining why this is incorrect? >> I thought the "vaddr = probe_access()" calculates the host address, so >> shouldn't it be the right address? > > The vaddr name is confusing (since it implies virtual address, which > the return from probe_access is not) as are the casts, which are > unnecessary. Still, I think my code isn't as wrong as you said. But I tend to agree with you on this: > Frankly, the current tcg_gen_atomic_xchg_reg is as optimized as > you'll be able to do. tcg_gen_atomic_xchg_reg() seems to generate on x86-64: 0000000000525160 <helper_atomic_xchgl_be>: 525160: 53 push %rbx 525161: 4c 8b 44 24 08 mov 0x8(%rsp),%r8 525166: 89 d3 mov %edx,%ebx 525168: 89 ca mov %ecx,%edx 52516a: b9 04 00 00 00 mov $0x4,%ecx 52516f: e8 1c a6 ff ff call 51f790 <atomic_mmu_lookup> 525174: 48 89 c2 mov %rax,%rdx 525177: 89 d8 mov %ebx,%eax 525179: 0f c8 bswap %eax 52517b: 87 02 xchg %eax,(%rdx) 52517d: 5b pop %rbx 52517e: 0f c8 bswap %eax 525180: c3 ret and atomic_mmu_lookup() is basically the same as probe_access(), so there is probably no gain in my patch. Please ignore my patch. Thank you! Helge
diff --git a/target/hppa/helper.h b/target/hppa/helper.h index c7e35ce8c7..26d022c714 100644 --- a/target/hppa/helper.h +++ b/target/hppa/helper.h @@ -16,7 +16,7 @@ DEF_HELPER_FLAGS_3(stby_b_parallel, TCG_CALL_NO_WG, void, env, tl, tr) DEF_HELPER_FLAGS_3(stby_e, TCG_CALL_NO_WG, void, env, tl, tr) DEF_HELPER_FLAGS_3(stby_e_parallel, TCG_CALL_NO_WG, void, env, tl, tr) -DEF_HELPER_FLAGS_1(ldc_check, TCG_CALL_NO_RWG, void, tl) +DEF_HELPER_FLAGS_3(ldc, TCG_CALL_NO_WG, tr, env, tl, i32) DEF_HELPER_FLAGS_4(probe, TCG_CALL_NO_WG, tr, env, tl, i32, i32) diff --git a/target/hppa/op_helper.c b/target/hppa/op_helper.c index f25a5a72aa..d61f067e71 100644 --- a/target/hppa/op_helper.c +++ b/target/hppa/op_helper.c @@ -155,13 +155,63 @@ void HELPER(stby_e_parallel)(CPUHPPAState *env, target_ulong addr, do_stby_e(env, addr, val, true, GETPC()); } -void HELPER(ldc_check)(target_ulong addr) +target_ureg HELPER(ldc)(CPUHPPAState *env, target_ulong addr, uint32_t size) { + uintptr_t ra = GETPC(); + int mmu_idx = cpu_mmu_index(env, 0); + void *vaddr; + + /* + * For hppa1.1, LDCW is undefined unless aligned mod 16. + * However actual hardware succeeds with aligned mod 4. + * Detect this case and log a GUEST_ERROR. + * + * TODO: HPPA64 relaxes the over-alignment requirement + * with the ,co completer. + */ if (unlikely(addr & 0xf)) { qemu_log_mask(LOG_GUEST_ERROR, "Undefined ldc to unaligned address mod 16: " TARGET_FMT_lx "\n", addr); } + + vaddr = probe_access(env, addr, size, MMU_DATA_STORE, mmu_idx, ra); + if (vaddr == NULL) { + cpu_loop_exit_restore(env_cpu(env), ra); + } + + if (size == 4) { + /* 32-bit ldcw */ + uint32_t old, *haddr; + + haddr = (uint32_t *)((uintptr_t)vaddr); + old = *haddr; + + /* if already zero, do not write 0 again to reduce memory presssure */ + if (old == 0) { + return 0; + } + old = qatomic_xchg(haddr, (uint32_t) 0); + return be32_to_cpu(old); + } else { + /* 64-bit ldcd */ +#ifdef TARGET_HPPA64 + uint64_t old, *haddr; + + haddr = (uint64_t *)((uintptr_t)vaddr); + old = *haddr; + + /* if already zero, do not write 0 again to reduce memory presssure */ + if (old == 0) { + return 0; + } + old = qatomic_xchg(haddr, (uint64_t) 0); + return be64_to_cpu(old); +#else + hppa_dynamic_excp(env, EXCP_ILL, ra); + return 0; +#endif + } } target_ureg HELPER(probe)(CPUHPPAState *env, target_ulong addr, diff --git a/target/hppa/translate.c b/target/hppa/translate.c index c04dc15228..c96691ab62 100644 --- a/target/hppa/translate.c +++ b/target/hppa/translate.c @@ -2857,9 +2857,9 @@ static bool trans_st(DisasContext *ctx, arg_ldst *a) static bool trans_ldc(DisasContext *ctx, arg_ldst *a) { - MemOp mop = MO_TE | MO_ALIGN | a->size; - TCGv_reg zero, dest, ofs; + TCGv_reg dest, ofs; TCGv_tl addr; + TCGv_i32 sz; nullify_over(ctx); @@ -2874,18 +2874,8 @@ static bool trans_ldc(DisasContext *ctx, arg_ldst *a) form_gva(ctx, &addr, &ofs, a->b, a->x, a->scale ? a->size : 0, a->disp, a->sp, a->m, ctx->mmu_idx == MMU_PHYS_IDX); - /* - * For hppa1.1, LDCW is undefined unless aligned mod 16. - * However actual hardware succeeds with aligned mod 4. - * Detect this case and log a GUEST_ERROR. - * - * TODO: HPPA64 relaxes the over-alignment requirement - * with the ,co completer. - */ - gen_helper_ldc_check(addr); - - zero = tcg_constant_reg(0); - tcg_gen_atomic_xchg_reg(dest, addr, zero, ctx->mmu_idx, mop); + sz = tcg_constant_i32((a->size == MO_32) ? 4 : 8); + gen_helper_ldc(dest, cpu_env, addr, sz); if (a->m) { save_gpr(ctx, a->b, ofs);
ldcw (load word and clear) is the only atomic memory instruction of the hppa architecture and thus is heavily used by the Linux and HP/UX kernel to implement locks. Since ldcw always writes a zero, optimize it to not write zero again if the memory already contained zero (as the lock was already locked). This reduces the memory contention. Furthermore, use a native qatomic_xchg() call to write the zero and thus give a small performace improvement and avoid additional atomic locking in the target. Signed-off-by: Helge Deller <deller@gmx.de>