[dynarmic] spill to XMM if possible, use sub/add instead of inc/dec as per recommendation

externals/dynarmic/src/dynarmic/backend/x64/a32_emit_x64.cpp

@@ -233,7 +233,7 @@ void A32EmitX64::GenTerminalHandlers() {
     terminal_handler_pop_rsb_hint = code.getCurr<const void*>();
     calculate_location_descriptor();
     code.mov(eax, dword[code.ABI_JIT_PTR + offsetof(A32JitState, rsb_ptr)]);
-    code.dec(eax);
+    code.sub(eax, 1);
     code.and_(eax, u32(A32JitState::RSBPtrMask));
     code.mov(dword[code.ABI_JIT_PTR + offsetof(A32JitState, rsb_ptr)], eax);
     code.cmp(rbx, qword[code.ABI_JIT_PTR + offsetof(A32JitState, rsb_location_descriptors) + rax * sizeof(u64)]);

externals/dynarmic/src/dynarmic/backend/x64/a64_emit_x64.cpp

@@ -208,7 +208,7 @@ void A64EmitX64::GenTerminalHandlers() {
     terminal_handler_pop_rsb_hint = code.getCurr<const void*>();
     calculate_location_descriptor();
     code.mov(eax, dword[code.ABI_JIT_PTR + offsetof(A64JitState, rsb_ptr)]);
-    code.dec(eax);
+    code.sub(eax, 1);
     code.and_(eax, u32(A64JitState::RSBPtrMask));
     code.mov(dword[code.ABI_JIT_PTR + offsetof(A64JitState, rsb_ptr)], eax);
     code.cmp(rbx, qword[code.ABI_JIT_PTR + offsetof(A64JitState, rsb_location_descriptors) + rax * sizeof(u64)]);

externals/dynarmic/src/dynarmic/backend/x64/emit_x64_floating_point.cpp

@@ -1206,7 +1206,7 @@ static void EmitFPRSqrtEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* i
                     }
 
                     // a > 0 && a < 0x00800000;
-                    code.dec(tmp);
+                    code.sub(tmp, 1);
                     code.cmp(tmp, 0x007FFFFF);
                     code.jb(fallback, code.T_NEAR); //within -127,128
                     needs_fallback = true;

externals/dynarmic/src/dynarmic/backend/x64/emit_x64_vector.cpp

@@ -3326,7 +3326,7 @@ void EmitX64::EmitVectorPolynomialMultiply8(EmitContext& ctx, IR::Inst* inst) {
         code.paddb(mask, mask);
         code.paddb(xmm_a, xmm_a);
         code.pblendvb(result, alternate);
-        code.dec(counter);
+        code.sub(counter, 1);
         code.jnz(loop);
 
         ctx.reg_alloc.DefineValue(inst, result);
@@ -3370,7 +3370,7 @@ void EmitX64::EmitVectorPolynomialMultiplyLong8(EmitContext& ctx, IR::Inst* inst
         code.paddw(mask, mask);
         code.paddw(xmm_a, xmm_a);
         code.pblendvb(result, alternate);
-        code.dec(counter);
+        code.sub(counter, 1);
         code.jnz(loop);
 
         ctx.reg_alloc.DefineValue(inst, result);

externals/dynarmic/src/dynarmic/backend/x64/hostloc.h

@@ -78,16 +78,16 @@ inline bool HostLocIsFlag(HostLoc reg) {
 
 inline HostLoc HostLocRegIdx(int idx) {
     ASSERT(idx >= 0 && idx <= 15);
-    return static_cast<HostLoc>(idx);
+    return HostLoc(idx);
 }
 
 inline HostLoc HostLocXmmIdx(int idx) {
     ASSERT(idx >= 0 && idx <= 15);
-    return static_cast<HostLoc>(static_cast<size_t>(HostLoc::XMM0) + idx);
+    return HostLoc(size_t(HostLoc::XMM0) + idx);
 }
 
 inline HostLoc HostLocSpill(size_t i) {
-    return static_cast<HostLoc>(static_cast<size_t>(HostLoc::FirstSpill) + i);
+    return HostLoc(size_t(HostLoc::FirstSpill) + i);
 }
 
 inline bool HostLocIsSpill(HostLoc reg) {

externals/dynarmic/src/dynarmic/backend/x64/reg_alloc.cpp

@@ -440,10 +440,13 @@ HostLoc RegAlloc::SelectARegister(const boost::container::static_vector<HostLoc,
         // all over the place - it also fixes bugs with high reg pressure
         } else if (*it >= HostLoc::R13 && *it <= HostLoc::R15) {
             // skip, do not touch
+        // Intel recommends to reuse registers as soon as they're overwritable (DO NOT SPILL)
+        } else if (loc_info.IsEmpty()) {
+            it_empty_candidate = it;
+            break;
+        // No empty registers for some reason (very evil) - just do normal LRU
         } else {
             if (loc_info.lru_counter < min_lru_counter) {
-                if (loc_info.IsEmpty())
-                    it_empty_candidate = it;
                 // Otherwise a "quasi"-LRU
                 min_lru_counter = loc_info.lru_counter;
                 if (*it >= HostLoc::R8 && *it <= HostLoc::R15) {
@@ -454,9 +457,6 @@ HostLoc RegAlloc::SelectARegister(const boost::container::static_vector<HostLoc,
                 if (min_lru_counter == 0)
                     break; //early exit
             }
-            // only if not assigned (i.e for failcase of all LRU=0)
-            if (it_empty_candidate == desired_locations.cend() && loc_info.IsEmpty())
-                it_empty_candidate = it;
         }
     }
     // Final resolution goes as follows:
@@ -527,11 +527,10 @@ void RegAlloc::Move(HostLoc to, HostLoc from) noexcept {
 
     ASSERT(LocInfo(to).IsEmpty() && !LocInfo(from).IsLocked());
     ASSERT(bit_width <= HostLocBitWidth(to));
+    ASSERT_MSG(!LocInfo(from).IsEmpty(), "Mov eliminated");
 
-    if (!LocInfo(from).IsEmpty()) {
+    EmitMove(bit_width, to, from);
-        EmitMove(bit_width, to, from);
+    LocInfo(to) = std::exchange(LocInfo(from), {});
-        LocInfo(to) = std::exchange(LocInfo(from), {});
-    }
 }
 
 void RegAlloc::CopyToScratch(size_t bit_width, HostLoc to, HostLoc from) noexcept {
@@ -565,30 +564,36 @@ void RegAlloc::SpillRegister(HostLoc loc) noexcept {
     ASSERT_MSG(HostLocIsRegister(loc), "Only registers can be spilled");
     ASSERT_MSG(!LocInfo(loc).IsEmpty(), "There is no need to spill unoccupied registers");
     ASSERT_MSG(!LocInfo(loc).IsLocked(), "Registers that have been allocated must not be spilt");
-
+    auto const new_loc = FindFreeSpill(HostLocIsXMM(loc));
-    const HostLoc new_loc = FindFreeSpill();
     Move(new_loc, loc);
 }
 
-HostLoc RegAlloc::FindFreeSpill() const noexcept {
+HostLoc RegAlloc::FindFreeSpill(bool is_xmm) const noexcept {
-    for (size_t i = static_cast<size_t>(HostLoc::FirstSpill); i < hostloc_info.size(); i++) {
+    // Do not spill XMM into other XMM silly
-        const auto loc = static_cast<HostLoc>(i);
+    if (!is_xmm) {
-        if (LocInfo(loc).IsEmpty()) {
+        // TODO(lizzie): Using lower (xmm0 and such) registers results in issues/crashes - INVESTIGATE WHY
-            return loc;
+        // Intel recommends to spill GPR onto XMM registers IF POSSIBLE
-        }
+        for (size_t i = size_t(HostLoc::XMM15); i >= size_t(HostLoc::XMM0); --i)
+            if (const auto loc = HostLoc(i); LocInfo(loc).IsEmpty())
+                return loc;
     }
-
+    // Otherwise go to stack spilling
+    for (size_t i = size_t(HostLoc::FirstSpill); i < hostloc_info.size(); ++i)
+        if (const auto loc = HostLoc(i); LocInfo(loc).IsEmpty())
+            return loc;
     ASSERT_FALSE("All spill locations are full");
-}
+};
-
-inline static Xbyak::RegExp SpillToOpArg_Helper1(HostLoc loc, size_t reserved_stack_space) noexcept {
-    ASSERT(HostLocIsSpill(loc));
-    size_t i = static_cast<size_t>(loc) - static_cast<size_t>(HostLoc::FirstSpill);
-    ASSERT_MSG(i < SpillCount, "Spill index greater than number of available spill locations");
-    return Xbyak::util::rsp + reserved_stack_space + ABI_SHADOW_SPACE + offsetof(StackLayout, spill) + i * sizeof(StackLayout::spill[0]);
-}
 
 void RegAlloc::EmitMove(const size_t bit_width, const HostLoc to, const HostLoc from) noexcept {
+    auto const spill_to_op_arg_helper = [&](HostLoc loc, size_t reserved_stack_space) {
+        ASSERT(HostLocIsSpill(loc));
+        size_t i = size_t(loc) - size_t(HostLoc::FirstSpill);
+        ASSERT_MSG(i < SpillCount, "Spill index greater than number of available spill locations");
+        return Xbyak::util::rsp + reserved_stack_space + ABI_SHADOW_SPACE + offsetof(StackLayout, spill) + i * sizeof(StackLayout::spill[0]);
+    };
+    auto const spill_xmm_to_op = [&](const HostLoc loc) {
+        return Xbyak::util::xword[spill_to_op_arg_helper(loc, reserved_stack_space)];
+    };
     if (HostLocIsXMM(to) && HostLocIsXMM(from)) {
         MAYBE_AVX(movaps, HostLocToXmm(to), HostLocToXmm(from));
     } else if (HostLocIsGPR(to) && HostLocIsGPR(from)) {
@@ -613,7 +618,7 @@ void RegAlloc::EmitMove(const size_t bit_width, const HostLoc to, const HostLoc
             MAYBE_AVX(movd, HostLocToReg64(to).cvt32(), HostLocToXmm(from));
         }
     } else if (HostLocIsXMM(to) && HostLocIsSpill(from)) {
-        const Xbyak::Address spill_addr = SpillToOpArg(from);
+        const Xbyak::Address spill_addr = spill_xmm_to_op(from);
         ASSERT(spill_addr.getBit() >= bit_width);
         switch (bit_width) {
         case 128:
@@ -631,7 +636,7 @@ void RegAlloc::EmitMove(const size_t bit_width, const HostLoc to, const HostLoc
             UNREACHABLE();
         }
     } else if (HostLocIsSpill(to) && HostLocIsXMM(from)) {
-        const Xbyak::Address spill_addr = SpillToOpArg(to);
+        const Xbyak::Address spill_addr = spill_xmm_to_op(to);
         ASSERT(spill_addr.getBit() >= bit_width);
         switch (bit_width) {
         case 128:
@@ -651,16 +656,16 @@ void RegAlloc::EmitMove(const size_t bit_width, const HostLoc to, const HostLoc
     } else if (HostLocIsGPR(to) && HostLocIsSpill(from)) {
         ASSERT(bit_width != 128);
         if (bit_width == 64) {
-            code->mov(HostLocToReg64(to), Xbyak::util::qword[SpillToOpArg_Helper1(from, reserved_stack_space)]);
+            code->mov(HostLocToReg64(to), Xbyak::util::qword[spill_to_op_arg_helper(from, reserved_stack_space)]);
         } else {
-            code->mov(HostLocToReg64(to).cvt32(), Xbyak::util::dword[SpillToOpArg_Helper1(from, reserved_stack_space)]);
+            code->mov(HostLocToReg64(to).cvt32(), Xbyak::util::dword[spill_to_op_arg_helper(from, reserved_stack_space)]);
         }
     } else if (HostLocIsSpill(to) && HostLocIsGPR(from)) {
         ASSERT(bit_width != 128);
         if (bit_width == 64) {
-            code->mov(Xbyak::util::qword[SpillToOpArg_Helper1(to, reserved_stack_space)], HostLocToReg64(from));
+            code->mov(Xbyak::util::qword[spill_to_op_arg_helper(to, reserved_stack_space)], HostLocToReg64(from));
         } else {
-            code->mov(Xbyak::util::dword[SpillToOpArg_Helper1(to, reserved_stack_space)], HostLocToReg64(from).cvt32());
+            code->mov(Xbyak::util::dword[spill_to_op_arg_helper(to, reserved_stack_space)], HostLocToReg64(from).cvt32());
         }
     } else {
         ASSERT_FALSE("Invalid RegAlloc::EmitMove");
@@ -677,8 +682,4 @@ void RegAlloc::EmitExchange(const HostLoc a, const HostLoc b) noexcept {
     }
 }
 
-Xbyak::Address RegAlloc::SpillToOpArg(const HostLoc loc) noexcept {
-    return Xbyak::util::xword[SpillToOpArg_Helper1(loc, reserved_stack_space)];
-}
-
 }  // namespace Dynarmic::Backend::X64

externals/dynarmic/src/dynarmic/backend/x64/reg_alloc.h

@@ -243,7 +243,7 @@ private:
     void MoveOutOfTheWay(HostLoc reg) noexcept;
 
     void SpillRegister(HostLoc loc) noexcept;
-    HostLoc FindFreeSpill() const noexcept;
+    HostLoc FindFreeSpill(bool is_xmm) const noexcept;
     
     inline HostLocInfo& LocInfo(const HostLoc loc) noexcept {
         ASSERT(loc != HostLoc::RSP && loc != ABI_JIT_PTR);
@@ -256,7 +256,6 @@ private:
 
     void EmitMove(const size_t bit_width, const HostLoc to, const HostLoc from) noexcept;
     void EmitExchange(const HostLoc a, const HostLoc b) noexcept;
-    Xbyak::Address SpillToOpArg(const HostLoc loc) noexcept;
 
 //data
     alignas(64) boost::container::static_vector<HostLoc, 28> gpr_order;

externals/dynarmic/src/dynarmic/backend/x64/verbose_debugging_output.cpp

@@ -22,7 +22,7 @@ void PrintVerboseDebuggingOutputLine(RegisterData& reg_data, HostLoc hostloc, si
         } else if (HostLocIsXMM(hostloc)) {
             return reg_data.xmms[HostLocToXmm(hostloc).getIdx()];
         } else if (HostLocIsSpill(hostloc)) {
-            return (*reg_data.spill)[static_cast<size_t>(hostloc) - static_cast<size_t>(HostLoc::FirstSpill)];
+            return (*reg_data.spill)[size_t(hostloc) - size_t(HostLoc::FirstSpill)];
         } else {
             fmt::print("invalid hostloc! ");
             return {0, 0};