a32_emit_x64: Use r14 for page_table pointer

src/backend/x64/a32_emit_x64.cpp

@@ -4,6 +4,7 @@
  * General Public License version 2 or any later version.
  */
 
+#include <algorithm>
 #include <optional>
 #include <unordered_map>
 #include <utility>
@@ -85,7 +86,15 @@ A32EmitX64::BlockDescriptor A32EmitX64::Emit(IR::Block& block) {
     // Start emitting.
     EmitCondPrelude(block);
 
-    RegAlloc reg_alloc{code, A32JitState::SpillCount, SpillToOpArg<A32JitState>, any_gpr, any_xmm};
+    static const std::vector<HostLoc> gpr_order = [this]{
+        std::vector<HostLoc> gprs{any_gpr};
+        if (config.page_table) {
+            gprs.erase(std::find(gprs.begin(), gprs.end(), HostLoc::R14));
+        }
+        return gprs;
+    }();
+
+    RegAlloc reg_alloc{code, A32JitState::SpillCount, SpillToOpArg<A32JitState>, gpr_order, any_xmm};
     A32EmitContext ctx{reg_alloc, block};
 
     for (auto iter = block.begin(); iter != block.end(); ++iter) {
@@ -778,21 +787,20 @@ static Xbyak::RegExp EmitVAddrLookup(BlockOfCode& code, RegAlloc& reg_alloc,
                                      Xbyak::Reg64 vaddr,
                                      std::optional<Xbyak::Reg64> arg_scratch = {}) {
     constexpr size_t page_bits = A32::UserConfig::PAGE_BITS;
-    const Xbyak::Reg64 page_table = arg_scratch ? *arg_scratch : reg_alloc.ScratchGpr();
+    const Xbyak::Reg64 page = arg_scratch ? *arg_scratch : reg_alloc.ScratchGpr();
-    const Xbyak::Reg64 tmp = reg_alloc.ScratchGpr();
+    const Xbyak::Reg64 tmp = config.absolute_offset_page_table ? page : reg_alloc.ScratchGpr();
-    code.mov(page_table, reinterpret_cast<u64>(config.page_table));
     code.mov(tmp, vaddr);
     code.shr(tmp, static_cast<int>(page_bits));
-    code.mov(page_table, qword[page_table + tmp * sizeof(void*)]);
+    code.mov(page, qword[r14 + tmp * sizeof(void*)]);
-    code.test(page_table, page_table);
+    code.test(page, page);
     code.jz(abort);
     if (config.absolute_offset_page_table) {
-        return page_table + vaddr;
+        return page + vaddr;
     }
     constexpr size_t page_mask = (1 << page_bits) - 1;
     code.mov(tmp, vaddr);
     code.and_(tmp, static_cast<u32>(page_mask));
-    return page_table + tmp;
+    return page + tmp;
 }
 
 template<std::size_t bitsize>

src/backend/x64/a32_interface.cpp

@@ -4,6 +4,7 @@
  * General Public License version 2 or any later version.
  */
 
+#include <functional>
 #include <memory>
 
 #include <boost/icl/interval_set.hpp>
@@ -19,6 +20,7 @@
 #include "backend/x64/devirtualize.h"
 #include "backend/x64/jitstate_info.h"
 #include "common/assert.h"
+#include "common/cast_util.h"
 #include "common/common_types.h"
 #include "common/llvm_disassemble.h"
 #include "common/scope_exit.h"
@@ -39,9 +41,20 @@ static RunCodeCallbacks GenRunCodeCallbacks(A32::UserCallbacks* cb, CodePtr (*Lo
     };
 }
 
+static std::function<void(BlockOfCode&)> GenRCP(const A32::UserConfig& config) {
+    if (!config.page_table) {
+        return [](BlockOfCode&){};
+    }
+
+    const u64 r14_value = Common::BitCast<u64>(config.page_table);
+    return [r14_value](BlockOfCode& code) {
+        code.mov(code.r14, r14_value);
+    };
+}
+
 struct Jit::Impl {
     Impl(Jit* jit, A32::UserConfig config)
-            : block_of_code(GenRunCodeCallbacks(config.callbacks, &GetCurrentBlockThunk, this), JitStateInfo{jit_state})
+            : block_of_code(GenRunCodeCallbacks(config.callbacks, &GetCurrentBlockThunk, this), JitStateInfo{jit_state}, GenRCP(config))
             , emitter(block_of_code, config, jit)
             , config(std::move(config))
             , jit_interface(jit)

src/backend/x64/a64_interface.cpp

@@ -34,11 +34,15 @@ static RunCodeCallbacks GenRunCodeCallbacks(A64::UserCallbacks* cb, CodePtr (*Lo
     };
 }
 
+static std::function<void(BlockOfCode&)> GenRCP(const A64::UserConfig&) {
+    return [](BlockOfCode&){};
+}
+
 struct Jit::Impl final {
 public:
     Impl(Jit* jit, UserConfig conf)
         : conf(conf)
-        , block_of_code(GenRunCodeCallbacks(conf.callbacks, &GetCurrentBlockThunk, this), JitStateInfo{jit_state})
+        , block_of_code(GenRunCodeCallbacks(conf.callbacks, &GetCurrentBlockThunk, this), JitStateInfo{jit_state}, GenRCP(conf))
         , emitter(block_of_code, conf, jit)
     {
         ASSERT(conf.page_table_address_space_bits >= 12 && conf.page_table_address_space_bits <= 64);

src/backend/x64/block_of_code.cpp

@@ -75,14 +75,14 @@ void ProtectMemory(const void* base, size_t size, bool is_executable) {
 
 } // anonymous namespace
 
-BlockOfCode::BlockOfCode(RunCodeCallbacks cb, JitStateInfo jsi)
+BlockOfCode::BlockOfCode(RunCodeCallbacks cb, JitStateInfo jsi, std::function<void(BlockOfCode&)> rcp)
         : Xbyak::CodeGenerator(TOTAL_CODE_SIZE, nullptr, &s_allocator)
         , cb(std::move(cb))
         , jsi(jsi)
         , constant_pool(*this, CONSTANT_POOL_SIZE)
 {
     EnableWriting();
-    GenRunCode();
+    GenRunCode(rcp);
 }
 
 void BlockOfCode::PreludeComplete() {
@@ -155,7 +155,7 @@ void BlockOfCode::ForceReturnFromRunCode(bool mxcsr_already_exited) {
     jmp(return_from_run_code[index]);
 }
 
-void BlockOfCode::GenRunCode() {
+void BlockOfCode::GenRunCode(std::function<void(BlockOfCode&)> rcp) {
     Xbyak::Label loop, enter_mxcsr_then_loop;
 
     align();
@@ -168,14 +168,16 @@ void BlockOfCode::GenRunCode() {
     ABI_PushCalleeSaveRegistersAndAdjustStack(*this);
 
     mov(r15, ABI_PARAM1);
-    mov(r14, ABI_PARAM2); // save temporarily in non-volatile register
+    mov(rbx, ABI_PARAM2); // save temporarily in non-volatile register
 
     cb.GetTicksRemaining->EmitCall(*this);
     mov(qword[r15 + jsi.offsetof_cycles_to_run], ABI_RETURN);
     mov(qword[r15 + jsi.offsetof_cycles_remaining], ABI_RETURN);
 
+    rcp(*this);
+
     SwitchMxcsrOnEntry();
-    jmp(r14);
+    jmp(rbx);
 
     align();
     step_code = getCurr<RunCodeFuncType>();
@@ -187,6 +189,8 @@ void BlockOfCode::GenRunCode() {
     mov(qword[r15 + jsi.offsetof_cycles_to_run], 1);
     mov(qword[r15 + jsi.offsetof_cycles_remaining], 1);
 
+    rcp(*this);
+
     SwitchMxcsrOnEntry();
     jmp(ABI_PARAM2);
 

src/backend/x64/block_of_code.h

@@ -7,6 +7,7 @@
 #pragma once
 
 #include <array>
+#include <functional>
 #include <memory>
 #include <type_traits>
 
@@ -31,7 +32,7 @@ struct RunCodeCallbacks {
 
 class BlockOfCode final : public Xbyak::CodeGenerator {
 public:
-    BlockOfCode(RunCodeCallbacks cb, JitStateInfo jsi);
+    BlockOfCode(RunCodeCallbacks cb, JitStateInfo jsi, std::function<void(BlockOfCode&)> rcp);
     BlockOfCode(const BlockOfCode&) = delete;
 
     /// Call when external emitters have finished emitting their preludes.
@@ -162,7 +163,7 @@ private:
     static constexpr size_t MXCSR_ALREADY_EXITED = 1 << 0;
     static constexpr size_t FORCE_RETURN = 1 << 1;
     std::array<const void*, 4> return_from_run_code;
-    void GenRunCode();
+    void GenRunCode(std::function<void(BlockOfCode&)> rcp);
 
     Xbyak::util::Cpu cpu_info;
 };