Revert "[common/core/dynarmic] Optimize page table allocations (#4219)"

This commit is contained in:
CamilleLaVey
2026-09-20 14:18:28 -04:00
parent d82a9c3ba5
commit ffb75c2b68
27 changed files with 394 additions and 737 deletions
@@ -371,10 +371,8 @@ EmitConfig A32AddressSpace::GetEmitConfig() {
.page_table_pointer = std::bit_cast<u64>(conf.page_table),
.page_table_address_space_bits = 32,
.page_table_pointer_mask = conf.page_table_pointer_mask,
.page_table_pointer_mask_bits = conf.page_table_pointer_mask_bits,
.page_table_log2_stride = conf.page_table_log2_stride,
.page_table_marked_bit = conf.page_table_marked_bit,
.page_table_sign_extension = conf.page_table_sign_extension,
.silently_mirror_page_table = true,
.absolute_offset_page_table = conf.absolute_offset_page_table,
.detect_misaligned_access_via_page_table = conf.detect_misaligned_access_via_page_table,
@@ -545,10 +545,8 @@ EmitConfig A64AddressSpace::GetEmitConfig() {
.page_table_pointer = std::bit_cast<u64>(conf.page_table),
.page_table_address_space_bits = conf.page_table_address_space_bits,
.page_table_pointer_mask = conf.page_table_pointer_mask,
.page_table_pointer_mask_bits = conf.page_table_pointer_mask_bits,
.page_table_log2_stride = conf.page_table_log2_stride,
.page_table_marked_bit = conf.page_table_marked_bit,
.page_table_sign_extension = conf.page_table_sign_extension,
.silently_mirror_page_table = conf.silently_mirror_page_table,
.absolute_offset_page_table = conf.absolute_offset_page_table,
.detect_misaligned_access_via_page_table = conf.detect_misaligned_access_via_page_table,
@@ -128,10 +128,8 @@ struct EmitConfig {
// Page table
u64 page_table_pointer;
std::size_t page_table_address_space_bits;
u64 page_table_pointer_mask;
int page_table_pointer_mask_bits;
std::size_t page_table_log2_stride;
std::optional<std::uint8_t> page_table_marked_bit;
std::optional<std::uint8_t> page_table_sign_extension;
bool silently_mirror_page_table;
bool absolute_offset_page_table;
u8 detect_misaligned_access_via_page_table;
@@ -273,18 +273,9 @@ std::pair<oaknut::XReg, oaknut::XReg> InlinePageTableEmitVAddrLookup(oaknut::Cod
// load x0 = *<(u8*)pagetable + index>
code.LDR(Xscratch0, Xpagetable, Xscratch0);
if (ctx.conf.page_table_marked_bit) {
code.TST(Xscratch0, 1ULL << *ctx.conf.page_table_marked_bit);
code.B(NE, *fallback);
}
if (ctx.conf.page_table_pointer_mask != 0) {
code.AND(Xscratch0, Xscratch0, ctx.conf.page_table_pointer_mask);
}
// TODO: combine this with page_table_pointer_mask
if (ctx.conf.page_table_sign_extension) {
code.SBFM(Xscratch0, Xscratch0, 0, *ctx.conf.page_table_sign_extension);
if (ctx.conf.page_table_pointer_mask_bits != 0) {
const u64 mask = u64(~u64(0)) << ctx.conf.page_table_pointer_mask_bits;
code.AND(Xscratch0, Xscratch0, mask);
}
code.CBZ(Xscratch0, *fallback);
@@ -9,7 +9,6 @@
#pragma once
#include <bit>
#include <utility>
#include "dynarmic/backend/x64/xbyak.h"
#include "dynarmic/backend/x64/a32_emit_x64.h"
@@ -79,46 +78,27 @@ Xbyak::RegExp EmitVAddrLookup(BlockOfCode& code, EmitContext& ctx, size_t bitsiz
template<>
[[maybe_unused]] Xbyak::RegExp EmitVAddrLookup<A32EmitContext>(BlockOfCode& code, A32EmitContext& ctx, size_t bitsize, Xbyak::Label& abort, Xbyak::Reg64 vaddr) {
const Xbyak::Reg64 page = ctx.reg_alloc.ScratchGpr(code);
const Xbyak::Reg64 tmp = ctx.conf.absolute_offset_page_table && ctx.conf.page_table_pointer_mask == 0 ? page : ctx.reg_alloc.ScratchGpr(code);
const Xbyak::Reg32 tmp = ctx.conf.absolute_offset_page_table ? page.cvt32() : ctx.reg_alloc.ScratchGpr(code).cvt32();
EmitDetectMisalignedVAddr(code, ctx, bitsize, abort, vaddr, tmp);
EmitDetectMisalignedVAddr(code, ctx, bitsize, abort, vaddr, tmp.cvt64());
code.mov(tmp, vaddr);
// TODO: This code assumes vaddr has been zext from 32-bits to 64-bits.
code.mov(tmp, vaddr.cvt32());
code.shr(tmp, int(page_table_const_bits));
if (ctx.conf.page_table_log2_stride > 3) {
code.shl(tmp, int(ctx.conf.page_table_log2_stride));
code.mov(page, qword[r14 + tmp.cvt64()]);
} else {
code.mov(page, qword[r14 + tmp.cvt64() * int(ctx.conf.page_table_log2_stride)]);
}
// check for marked bit, use as unmapped if marked
if (ctx.conf.page_table_marked_bit) {
code.bt(page, *ctx.conf.page_table_marked_bit);
code.jc(abort, code.T_NEAR);
}
// mask away attributes
if (ctx.conf.page_table_pointer_mask == 0) {
code.shl(tmp, int(ctx.conf.page_table_log2_stride));
code.mov(page, qword[r14 + tmp.cvt64()]);
if (ctx.conf.page_table_pointer_mask_bits == 0) {
code.test(page, page);
} else if (std::in_range<s32>(ctx.conf.page_table_pointer_mask)) {
code.and_(page, ctx.conf.page_table_pointer_mask);
} else {
code.mov(tmp, ctx.conf.page_table_pointer_mask);
code.and_(page, tmp);
code.and_(page, ~u32(0) << ctx.conf.page_table_pointer_mask_bits);
}
// check for sign bit, apply sign extension as needed
if (ctx.conf.page_table_sign_extension) {
code.shl(page, 63 - int(*ctx.conf.page_table_sign_extension));
code.sar(page, 63 - int(*ctx.conf.page_table_sign_extension));
}
code.jz(abort, code.T_NEAR);
if (ctx.conf.absolute_offset_page_table) {
return page + vaddr;
}
code.mov(tmp, vaddr);
code.and_(tmp, u32(page_table_const_mask));
code.mov(tmp, vaddr.cvt32());
code.and_(tmp, static_cast<u32>(page_table_const_mask));
return page + tmp.cvt64();
}
@@ -128,7 +108,7 @@ template<>
const size_t unused_top_bits = 64 - ctx.conf.page_table_address_space_bits;
const Xbyak::Reg64 page = ctx.reg_alloc.ScratchGpr(code);
const Xbyak::Reg64 tmp = ctx.conf.absolute_offset_page_table && ctx.conf.page_table_pointer_mask == 0 ? page : ctx.reg_alloc.ScratchGpr(code);
const Xbyak::Reg64 tmp = ctx.conf.absolute_offset_page_table ? page : ctx.reg_alloc.ScratchGpr(code);
EmitDetectMisalignedVAddr(code, ctx, bitsize, abort, vaddr, tmp);
@@ -163,26 +143,11 @@ template<>
code.shl(tmp, int(ctx.conf.page_table_log2_stride));
code.mov(page, qword[r14 + tmp]);
// check for marked bit, use as unmapped if marked
if (ctx.conf.page_table_marked_bit) {
code.bt(page, *ctx.conf.page_table_marked_bit);
code.jc(abort, code.T_NEAR);
}
// mask away attributes
if (ctx.conf.page_table_pointer_mask == 0) {
if (ctx.conf.page_table_pointer_mask_bits == 0) {
code.test(page, page);
} else if (std::in_range<s32>(ctx.conf.page_table_pointer_mask)) {
code.and_(page, ctx.conf.page_table_pointer_mask);
} else {
code.mov(tmp, ctx.conf.page_table_pointer_mask);
code.and_(page, tmp);
code.and_(page, ~u32(0) << ctx.conf.page_table_pointer_mask_bits);
}
if (ctx.conf.page_table_sign_extension) {
code.shl(page, *ctx.conf.page_table_sign_extension);
code.sar(page, *ctx.conf.page_table_sign_extension);
}
code.jz(abort, code.T_NEAR);
if (ctx.conf.absolute_offset_page_table) {
return page + vaddr;
@@ -159,23 +159,14 @@ struct UserConfig {
/// Maximum size is limited by the maximum length of a x86_64 / arm64 jump.
std::uint32_t code_cache_size = 128 * 1024 * 1024; // bytes
/// Applies a bit mask to the bits in host pointers from the page table.
/// Masks out the first N bits in host pointers from the page table.
/// The intention behind this is to allow users of Dynarmic to pack attributes in the
/// same integer and update the pointer attribute pair atomically.
/// If the configured value is ~(0b111ULL), all pointers will be forcefully aligned to 8 bytes.
std::uint64_t page_table_pointer_mask = 0;
/// If the configured value is 3, all pointers will be forcefully aligned to 8 bytes.
std::int32_t page_table_pointer_mask_bits = 0;
/// Log2 of the size per page entry, value should be either 3 or 4
std::uint32_t page_table_log2_stride = 3;
/// Setting this value has Dynarmic check the specified bit of the page pointer provided by page table.
/// If the bit is set to 1, Dynarmic will treat it as unmapped.
/// This bit should be included as part of `page_table_pointer_mask_bits`.
std::optional<std::uint8_t> page_table_marked_bit = std::nullopt;
/// If this value is set, Dynarmic will sign extend the page table pointer by this bit.
/// Useful for compacting bits into the page table and should be used as part of `page_table_pointer_mask`.
std::optional<std::uint8_t> page_table_sign_extension = std::nullopt;
// Log2 of the size per page entry, value should be either 3 or 4
std::size_t page_table_log2_stride = 3;
/// Select the architecture version to use.
/// There are minor behavioural differences between versions.
@@ -173,23 +173,14 @@ struct UserConfig {
/// This is only used if page_table is not nullptr.
std::uint32_t page_table_address_space_bits = 36;
/// Applies a bit mask to the bits in host pointers from the page table.
/// Masks out the first N bits in host pointers from the page table.
/// The intention behind this is to allow users of Dynarmic to pack attributes in the
/// same integer and update the pointer attribute pair atomically.
/// If the configured value is ~(0b111ULL), all pointers will be forcefully aligned to 8 bytes.
std::uint64_t page_table_pointer_mask = 0;
/// If the configured value is 3, all pointers will be forcefully aligned to 8 bytes.
std::int32_t page_table_pointer_mask_bits = 0;
/// Log2 of the size per page entry, value should be either 3 or 4
std::uint32_t page_table_log2_stride = 3;
/// Setting this value has Dynarmic check the specified bit of the page pointer provided by page table.
/// If the bit is set to 1, Dynarmic will treat it as unmapped.
/// This bit should be included as part of `page_table_pointer_mask`.
std::optional<std::uint8_t> page_table_marked_bit = std::nullopt;
/// If this value is set, Dynarmic will sign extend the page table pointer by this bit.
/// Useful for compacting bits into the page table and should be used as part of `page_table_pointer_mask`.
std::optional<std::uint8_t> page_table_sign_extension = std::nullopt;
// Log2 of the size per page entry, value should be either 3 or 4
std::size_t page_table_log2_stride = 3;
/// Counter-timer frequency register. The value of the register is not interpreted by
/// dynarmic.