Use ctr like gcc does

Signed-off-by: lizzie <lizzie@eden-emu.dev>

externals/powah/powah_emit.hpp

@@ -368,10 +368,11 @@ struct Context {
 
     void MFLR(GPR const rt) { MFSPR(powah::GPR{8}, rt, powah::GPR{0}); }
     void MTLR(GPR const rt) { MTSPR(powah::GPR{8}, rt, powah::GPR{0}); }
+    void BLR() { base[offset++] = 0x4e800020; } //BCLR(R0, CR0, R0);
 
-    void BLR() {
-        base[offset++] = 0x4e800020; //BCLR(R0, CR0, R0);
-    }
+    void MFCTR(GPR const rt) { MFSPR(powah::GPR{9}, rt, powah::GPR{0}); }
+    void MTCTR(GPR const rt) { MTSPR(powah::GPR{9}, rt, powah::GPR{0}); }
+    void BCTRL() { base[offset++] = 0x4e800421; } //BCCTRL(R0, CR0, R0);
 
     // TODO: PowerPC 11 stuff
     void ISEL(GPR const rd, GPR const ra, GPR const rb, uint32_t d) {

src/dynarmic/src/dynarmic/backend/ppc64/a32_core.h

@@ -23,6 +23,8 @@ struct A32JitState {
     u32 cpsr_nzcv = 0;
     u32 fpscr = 0;
     u8 check_bit = 0;
+    void* run_fn = nullptr;
+
     IR::LocationDescriptor GetLocationDescriptor() const {
         return IR::LocationDescriptor{regs[15] | (u64(upper_location_descriptor) << 32)};
     }

src/dynarmic/src/dynarmic/backend/ppc64/a32_interface.cpp

@@ -74,6 +74,7 @@ struct A32Core final {
         auto const loc = thread_ctx.GetLocationDescriptor();
         auto const entry = process.GetOrEmit(loc);
         using CodeFn = HaltReason (*)(A32AddressSpace*, A32JitState*, volatile u32*, void*);
+        thread_ctx.run_fn = (void*)&A32Core::Run;
         return (CodeFn(entry))(&process, &thread_ctx, halt_reason, reinterpret_cast<void*>(&A32Core::Run));
     }
 };

src/dynarmic/src/dynarmic/backend/ppc64/a64_core.h

@@ -29,6 +29,8 @@ struct A64JitState {
     u32 fpsr = 0;
     volatile u32 halt_reason = 0;
     u8 check_bit = 0;
+    void* run_fn = nullptr;
+
     IR::LocationDescriptor GetLocationDescriptor() const {
         const u64 fpcr_u64 = u64(fpcr & A64::LocationDescriptor::fpcr_mask) << A64::LocationDescriptor::fpcr_shift;
         const u64 pc_u64 = pc & A64::LocationDescriptor::pc_mask;

src/dynarmic/src/dynarmic/backend/ppc64/a64_interface.cpp

@@ -76,6 +76,7 @@ struct A64Core final {
         const auto loc = thread_ctx.GetLocationDescriptor();
         const auto entry = process.GetOrEmit(loc);
         using CodeFn = HaltReason (*)(A64AddressSpace*, A64JitState*, volatile u32*, void*);
+        thread_ctx.run_fn = (void*)&A64Core::Run;
         return (CodeFn(entry))(&process, &thread_ctx, halt_reason, reinterpret_cast<void*>(&A64Core::Run));
     }
 };

src/dynarmic/src/dynarmic/backend/ppc64/abi.h

@@ -8,11 +8,10 @@
 
 namespace Dynarmic::Backend::PPC64 {
 
-// Jit fn signature => (AXXAddressSpace& process, AXXJitState& thread_ctx, volatile u32* halt_reason, void* link_fn)
+// Jit fn signature => (AXXAddressSpace& process, AXXJitState& thread_ctx, volatile u32* halt_reason)
 constexpr powah::GPR RPROCESS = powah::R3;
 constexpr powah::GPR RJIT = powah::R4;
 constexpr powah::GPR RHALTREASON = powah::R5;
-constexpr powah::GPR RLINKFN = powah::R6;
 constexpr powah::GPR RNZCV = powah::R31;
 
 constexpr powah::GPR ABI_PARAM1 = powah::R3;

src/dynarmic/src/dynarmic/backend/ppc64/emit_ppc64.cpp

@@ -21,6 +21,7 @@
 #include "dynarmic/ir/basic_block.h"
 #include "dynarmic/ir/microinstruction.h"
 #include "dynarmic/ir/opcodes.h"
+#include "stack_layout.h"
 
 namespace Dynarmic::Backend::PPC64 {
 
@@ -155,23 +156,28 @@ void EmitTerminal(powah::Context& code, EmitContext& ctx, IR::Term::ReturnToDisp
 }
 
 void EmitTerminal(powah::Context& code, EmitContext& ctx, IR::Term::LinkBlock terminal, IR::LocationDescriptor initial_location, bool) {
+    auto const tmp_lr = ctx.reg_alloc.ScratchGpr();
     auto const tmp = ctx.reg_alloc.ScratchGpr();
-    code.LI(tmp, terminal.next.Value());
+    code.LD(tmp_lr, PPC64::RJIT, offsetof(StackLayout, lr));
+    code.MFCTR(tmp_lr);
     if (ctx.emit_conf.a64_variant) {
+        code.LI(tmp, terminal.next.Value());
         code.STD(tmp, PPC64::RJIT, offsetof(A64JitState, pc));
-        code.MFLR(PPC64::RLINKFN);
+        code.LD(tmp, PPC64::RJIT, offsetof(A64JitState, run_fn));
+        code.MTCTR(tmp);
         for (u32 i = 0; i < 4; ++i)
             code.STD(powah::GPR{3 + i}, powah::R1, -((GPR_ORDER.size() + i) * 8));
         code.ADDIS(powah::R1, powah::R1, -sizeof(StackLayout));
-        code.BLR();
+        code.BCTRL();
         code.ADDI(powah::R1, powah::R1, sizeof(StackLayout));
         for (u32 i = 0; i < 4; ++i)
             code.LD(powah::GPR{3 + i}, powah::R1, -((GPR_ORDER.size() + i) * 8));
-        code.MTLR(PPC64::RLINKFN);
     } else {
+        code.LI(tmp, terminal.next.Value());
         code.STW(tmp, PPC64::RJIT, offsetof(A32JitState, regs) + sizeof(u32) * 15);
         ASSERT(false && "unimp");
     }
+    code.MTCTR(tmp_lr);
 }
 
 void EmitTerminal(powah::Context& code, EmitContext& ctx, IR::Term::LinkBlockFast terminal, IR::LocationDescriptor initial_location, bool) {

src/dynarmic/src/dynarmic/backend/ppc64/stack_layout.h

@@ -15,6 +15,7 @@ struct alignas(16) StackLayout {
     std::array<u64, 18> regs;
     std::array<u64, SpillCount> spill;
     u64 check_bit;
+    void* lr;
 };
 
 static_assert(sizeof(StackLayout) % 16 == 0);