Merge pull request #10847 from unknownbrackets/arm64jit

arm64jit: Update rounding mode on thread switch

Core/HLE/sceKernelThread.cpp

@@ -31,6 +31,7 @@
 #include "Core/MIPS/MIPS.h"
 #include "Core/CoreTiming.h"
 #include "Core/MemMapHelpers.h"
+#include "Core/MIPS/JitCommon/JitCommon.h"
 #include "Core/Reporting.h"
 #include "Common/ChunkFile.h"
 
@@ -1460,6 +1461,10 @@ void __KernelLoadContext(ThreadContext *ctx, bool vfpuEnabled)
 	}
 
 	memcpy(currentMIPS->other, ctx->other, sizeof(ctx->other));
+	if (MIPSComp::jit) {
+		// When thread switching, we must update the rounding mode if cached in the jit.
+		MIPSComp::jit->UpdateFCR31();
+	}
 
 	// Reset the llBit, the other thread may have touched memory.
 	currentMIPS->llBit = 0;

Core/MIPS/ARM/ArmAsm.cpp

@@ -123,20 +123,6 @@ void ArmJit::GenerateFixedCode() {
 		POP(2, SCRATCHREG1, R_PC);
 	}
 
-	// Must preserve SCRATCHREG1 (R0), destroys SCRATCHREG2 (LR)
-	updateRoundingMode = AlignCode16(); {
-		PUSH(2, SCRATCHREG1, R_LR);
-		LDR(SCRATCHREG2, CTXREG, offsetof(MIPSState, fcr31));
-		MOVI2R(SCRATCHREG1, 0x1000003);
-		TST(SCRATCHREG2, SCRATCHREG1);
-		FixupBranch skip = B_CC(CC_EQ);  // zero
-		MOVI2R(SCRATCHREG2, 1);
-		MOVP2R(SCRATCHREG1, &js.hasSetRounding);
-		STRB(SCRATCHREG2, SCRATCHREG1, 0);
-		SetJumpTarget(skip);
-		POP(2, SCRATCHREG1, R_PC);
-	}
-
 	FlushLitPool();
 
 	enterDispatcher = AlignCode16();

Core/MIPS/ARM/ArmCompFPU.cpp

@@ -416,8 +416,10 @@ void ArmJit::Comp_mxc1(MIPSOpcode op)
 			// Must clear before setting, since ApplyRoundingMode() assumes it was cleared.
 			RestoreRoundingMode();
 			bool wasImm = gpr.IsImm(rt);
+			u32 immVal = -1;
 			if (wasImm) {
-				gpr.SetImm(MIPS_REG_FPCOND, (gpr.GetImm(rt) >> 23) & 1);
+				immVal = gpr.GetImm(rt);
+				gpr.SetImm(MIPS_REG_FPCOND, (immVal >> 23) & 1);
 				gpr.MapReg(rt);
 			} else {
 				gpr.MapDirtyIn(MIPS_REG_FPCOND, rt);
@@ -433,8 +435,10 @@ void ArmJit::Comp_mxc1(MIPSOpcode op)
 				MOV(SCRATCHREG1, Operand2(gpr.R(rt), ST_LSR, 23));
 				AND(gpr.R(MIPS_REG_FPCOND), SCRATCHREG1, Operand2(1));
 #endif
+				UpdateRoundingMode();
+			} else {
+				UpdateRoundingMode(immVal);
 			}
-			UpdateRoundingMode();
 			ApplyRoundingMode();
 		} else {
 			Comp_Generic(op);

Core/MIPS/ARM/ArmJit.cpp

@@ -106,6 +106,9 @@ void ArmJit::DoState(PointerWrap &p)
 	}
 }
 
+void ArmJit::UpdateFCR31() {
+}
+
 void ArmJit::FlushAll()
 {
 	gpr.FlushAll();
@@ -637,8 +640,12 @@ void ArmJit::ApplyRoundingMode(bool force) {
 }
 
 // Does (must!) not destroy R0 (SCRATCHREG1). Destroys R14 (SCRATCHREG2).
-void ArmJit::UpdateRoundingMode() {
-	QuickCallFunction(R1, updateRoundingMode);
+void ArmJit::UpdateRoundingMode(u32 fcr31) {
+	// We must set js.hasSetRounding at compile time, or this block will use the wrong rounding mode.
+	// The fcr31 parameter is -1 when not known at compile time, so we just assume it was changed.
+	if (fcr31 & 0x01000003) {
+		js.hasSetRounding = true;
+	}
 }
 
 // IDEA - could have a WriteDualExit that takes two destinations and two condition flags,

Core/MIPS/ARM/ArmJit.h

@@ -175,6 +175,7 @@ public:
 
 	void ClearCache() override;
 	void InvalidateCacheAt(u32 em_address, int length = 4) override;
+	void UpdateFCR31() override;
 
 	void EatPrefix() override { js.EatPrefix(); }
 
@@ -202,7 +203,7 @@ private:
 	void WriteDownCountR(ArmGen::ARMReg reg);
 	void RestoreRoundingMode(bool force = false);
 	void ApplyRoundingMode(bool force = false);
-	void UpdateRoundingMode();
+	void UpdateRoundingMode(u32 fcr31 = -1);
 	void MovFromPC(ArmGen::ARMReg r);
 	void MovToPC(ArmGen::ARMReg r);
 
@@ -310,7 +311,6 @@ public:
 
 	const u8 *restoreRoundingMode;
 	const u8 *applyRoundingMode;
-	const u8 *updateRoundingMode;
 
 	const u8 *breakpointBailout;
 };

Core/MIPS/ARM64/Arm64Asm.cpp

@@ -177,17 +177,6 @@ void Arm64Jit::GenerateFixedCode(const JitOptions &jo) {
 		ADDI2R(SCRATCH2, SCRATCH2, 4);
 		SetJumpTarget(skip);
 
-		// We need both SCRATCH1 and SCRATCH2 for updating hasSetRounding.
-		PUSH(SCRATCH2);
-		// We can only skip if the rounding mode is zero and flush is not set.
-		CMPI2R(SCRATCH2, 0);
-		FixupBranch skip2 = B(CC_EQ);
-		MOVI2R(SCRATCH2, 1);
-		MOVP2R(SCRATCH1_64, &js.hasSetRounding);
-		STRB(INDEX_UNSIGNED, SCRATCH2, SCRATCH1_64, 0);
-		SetJumpTarget(skip2);
-		POP(SCRATCH2);
-
 		// Let's update js.currentRoundingFunc with the right convertS0ToSCRATCH1 func.
 		MOVP2R(SCRATCH1_64, convertS0ToSCRATCH1);
 		LSL(SCRATCH2, SCRATCH2, 3);

Core/MIPS/ARM64/Arm64CompFPU.cpp

@@ -375,8 +375,10 @@ void Arm64Jit::Comp_mxc1(MIPSOpcode op)
 			// Must clear before setting, since ApplyRoundingMode() assumes it was cleared.
 			RestoreRoundingMode();
 			bool wasImm = gpr.IsImm(rt);
+			u32 immVal = -1;
 			if (wasImm) {
-				gpr.SetImm(MIPS_REG_FPCOND, (gpr.GetImm(rt) >> 23) & 1);
+				immVal = gpr.GetImm(rt);
+				gpr.SetImm(MIPS_REG_FPCOND, (immVal >> 23) & 1);
 				gpr.MapReg(rt);
 			} else {
 				gpr.MapDirtyIn(MIPS_REG_FPCOND, rt);
@@ -387,9 +389,11 @@ void Arm64Jit::Comp_mxc1(MIPSOpcode op)
 			STR(INDEX_UNSIGNED, gpr.R(rt), CTXREG, offsetof(MIPSState, fcr31));
 			if (!wasImm) {
 				UBFX(gpr.R(MIPS_REG_FPCOND), gpr.R(rt), 23, 1);
+				// TODO: We do have the fcr31 value in a register here, could use that in UpdateRoundingMode to avoid reloading it.
+				UpdateRoundingMode();
+			} else {
+				UpdateRoundingMode(immVal);
 			}
-			// TODO: We do have the fcr31 value in a register here, could use that in UpdateRoundingMode to avoid reloading it.
-			UpdateRoundingMode();
 			ApplyRoundingMode();
 		} else {
 			Comp_Generic(op);

Core/MIPS/ARM64/Arm64Jit.cpp

@@ -79,7 +79,7 @@ Arm64Jit::Arm64Jit(MIPSState *mips) : blocks(mips, this), gpr(mips, &js, &jo), f
 	AllocCodeSpace(1024 * 1024 * 16);  // 32MB is the absolute max because that's what an ARM branch instruction can reach, backwards and forwards.
 	GenerateFixedCode(jo);
 	js.startDefaultPrefix = mips_->HasDefaultPrefix();
-	js.currentRoundingFunc = convertS0ToSCRATCH1[0];
+	js.currentRoundingFunc = convertS0ToSCRATCH1[mips_->fcr31 & 3];
 }
 
 Arm64Jit::~Arm64Jit() {
@@ -98,13 +98,14 @@ void Arm64Jit::DoState(PointerWrap &p) {
 		js.hasSetRounding = 1;
 	}
 
-	if (p.GetMode() == PointerWrap::MODE_READ) {
-		js.currentRoundingFunc = convertS0ToSCRATCH1[(mips_->fcr31) & 3];
-	}
+	// Note: we can't update the currentRoundingFunc here because fcr31 wasn't loaded yet.
 }
 
-void Arm64Jit::FlushAll()
-{
+void Arm64Jit::UpdateFCR31() {
+	js.currentRoundingFunc = convertS0ToSCRATCH1[mips_->fcr31 & 3];
+}
+
+void Arm64Jit::FlushAll() {
 	gpr.FlushAll();
 	fpr.FlushAll();
 	FlushPrefixV();
@@ -601,7 +602,12 @@ void Arm64Jit::ApplyRoundingMode(bool force) {
 }
 
 // Destroys SCRATCH1 and SCRATCH2
-void Arm64Jit::UpdateRoundingMode() {
+void Arm64Jit::UpdateRoundingMode(u32 fcr31) {
+	// We must set js.hasSetRounding at compile time, or this block will use the wrong rounding mode.
+	// The fcr31 parameter is -1 when not known at compile time, so we just assume it was changed.
+	if (fcr31 & 0x01000003) {
+		js.hasSetRounding = true;
+	}
 	QuickCallFunction(SCRATCH2_64, updateRoundingMode);
 }
 

Core/MIPS/ARM64/Arm64Jit.h

@@ -176,6 +176,7 @@ public:
 
 	void ClearCache() override;
 	void InvalidateCacheAt(u32 em_address, int length = 4) override;
+	void UpdateFCR31() override;
 
 	void EatPrefix() override { js.EatPrefix(); }
 
@@ -201,7 +202,7 @@ private:
 	void WriteDownCountR(Arm64Gen::ARM64Reg reg, bool updateFlags = true);
 	void RestoreRoundingMode(bool force = false);
 	void ApplyRoundingMode(bool force = false);
-	void UpdateRoundingMode();
+	void UpdateRoundingMode(u32 fcr31 = -1);
 	void MovFromPC(Arm64Gen::ARM64Reg r);
 	void MovToPC(Arm64Gen::ARM64Reg r);
 

Core/MIPS/IR/IRFrontend.cpp

@@ -200,6 +200,8 @@ void IRFrontend::ApplyRoundingMode(bool force) {
 
 // Destroys SCRATCH1 and SCRATCH2
 void IRFrontend::UpdateRoundingMode() {
+	// We must set js.hasSetRounding at compile time, or this block will use the wrong rounding mode.
+	js.hasSetRounding = true;
 	ir.Write(IROp::UpdateRoundingMode);
 }
 

Core/MIPS/IR/IRJit.cpp

@@ -55,6 +55,9 @@ void IRJit::DoState(PointerWrap &p) {
 	frontend_.DoState(p);
 }
 
+void IRJit::UpdateFCR31() {
+}
+
 void IRJit::ClearCache() {
 	ILOG("IRJit: Clearing the cache!");
 	blocks_.Clear();

Core/MIPS/IR/IRJit.h

@@ -158,6 +158,7 @@ public:
 
 	void ClearCache() override;
 	void InvalidateCacheAt(u32 em_address, int length = 4) override;
+	void UpdateFCR31() override;
 
 	const u8 *GetDispatcher() const override { return nullptr; }
 

Core/MIPS/JitCommon/JitCommon.h

@@ -131,6 +131,7 @@ namespace MIPSComp {
 		virtual void Compile(u32 em_address) = 0;
 		virtual void CompileFunction(u32 start_address, u32 length) { }
 		virtual void ClearCache() = 0;
+		virtual void UpdateFCR31() = 0;
 		virtual MIPSOpcode GetOriginalOp(MIPSOpcode op) = 0;
 
 		// No jit operations may be run between these calls.

Core/MIPS/MIPS.cpp

@@ -293,6 +293,11 @@ void MIPSState::DoState(PointerWrap &p) {
 	p.Do(inDelaySlot);
 	p.Do(llBit);
 	p.Do(debugCount);
+
+	if (p.mode == p.MODE_READ && MIPSComp::jit) {
+		// Now that we've loaded fcr31, update any jit state associated.
+		MIPSComp::jit->UpdateFCR31();
+	}
 }
 
 void MIPSState::SingleStep() {

Core/MIPS/MIPS/MipsJit.cpp

@@ -66,6 +66,9 @@ void MipsJit::DoState(PointerWrap &p)
 	}
 }
 
+void MipsJit::UpdateFCR31() {
+}
+
 void MipsJit::FlushAll()
 {
 	//gpr.FlushAll();

Core/MIPS/MIPS/MipsJit.h

@@ -136,6 +136,7 @@ public:
 
 	void ClearCache() override;
 	void InvalidateCacheAt(u32 em_address, int length = 4) override;
+	void UpdateFCR31() override;
 
 	void EatPrefix() override { js.EatPrefix(); }
 

Core/MIPS/MIPSInt.cpp

@@ -528,6 +528,10 @@ namespace MIPSInt
 				if (fs == 31) {
 					currentMIPS->fcr31 = value & 0x0181FFFF;
 					currentMIPS->fpcond = (value >> 23) & 1;
+					if (MIPSComp::jit) {
+						// In case of DISABLE, we need to tell jit we updated FCR31.
+						MIPSComp::jit->UpdateFCR31();
+					}
 				} else {
 					WARN_LOG_REPORT(CPU, "WriteFCR: Unexpected reg %d (value %08x)", fs, value);
 				}

Core/MIPS/x86/Asm.cpp

@@ -109,18 +109,6 @@ void Jit::GenerateFixedCode(JitOptions &jo) {
 		RET();
 	}
 
-	updateRoundingMode = AlignCode16(); {
-		// If it's only ever 0, we don't actually bother applying or restoring it.
-		// This is the most common situation.
-		TEST(32, MIPSSTATE_VAR(fcr31), Imm32(0x01000003));
-		FixupBranch skip = J_CC(CC_Z);
-		// TODO: Move the hasSetRounding flag somewhere we can reach it through the context pointer, or something.
-		MOV(PTRBITS, R(RAX), ImmPtr(&js.hasSetRounding));
-		MOV(8, MatR(RAX), Imm8(1));
-		SetJumpTarget(skip);
-		RET();
-	}
-
 	enterDispatcher = AlignCode16();
 	ABI_PushAllCalleeSavedRegsAndAdjustStack();
 #ifdef _M_X64

Core/MIPS/x86/CompFPU.cpp

@@ -426,7 +426,7 @@ void Jit::Comp_mxc1(MIPSOpcode op) {
 				if ((gpr.GetImm(rt) & 0x1000003) == 0) {
 					// Default nearest / no-flush mode, just leave it cleared.
 				} else {
-					UpdateRoundingMode();
+					UpdateRoundingMode(gpr.GetImm(rt));
 					ApplyRoundingMode();
 				}
 			} else {

Core/MIPS/x86/Jit.cpp

@@ -142,6 +142,9 @@ void Jit::DoState(PointerWrap &p) {
 	CBreakPoints::SetSkipFirst(0);
 }
 
+void Jit::UpdateFCR31() {
+}
+
 void Jit::GetStateAndFlushAll(RegCacheState &state) {
 	gpr.GetState(state.gpr);
 	fpr.GetState(state.fpr);
@@ -195,8 +198,12 @@ void Jit::ApplyRoundingMode(bool force) {
 	}
 }
 
-void Jit::UpdateRoundingMode() {
-	CALL(updateRoundingMode);
+void Jit::UpdateRoundingMode(u32 fcr31) {
+	// We must set js.hasSetRounding at compile time, or this block will use the wrong rounding mode.
+	// The fcr31 parameter is -1 when not known at compile time, so we just assume it was changed.
+	if (fcr31 & 0x01000003) {
+		js.hasSetRounding = true;
+	}
 }
 
 void Jit::ClearCache()
@@ -415,8 +422,6 @@ void Jit::AddContinuedBlock(u32 dest) {
 bool Jit::DescribeCodePtr(const u8 *ptr, std::string &name) {
 	if (ptr == applyRoundingMode)
 		name = "applyRoundingMode";
-	else if (ptr == updateRoundingMode)
-		name = "updateRoundingMode";
 	else if (ptr == dispatcher)
 		name = "dispatcher";
 	else if (ptr == dispatcherInEAXNoCheck)

Core/MIPS/x86/Jit.h

@@ -155,7 +155,7 @@ public:
 
 	void RestoreRoundingMode(bool force = false);
 	void ApplyRoundingMode(bool force = false);
-	void UpdateRoundingMode();
+	void UpdateRoundingMode(u32 fcr31 = -1);
 
 	JitBlockCache *GetBlockCache() override { return &blocks; }
 	JitBlockCacheDebugInterface *GetBlockCacheDebugInterface() override { return &blocks; }
@@ -171,6 +171,7 @@ public:
 			blocks.InvalidateICache(em_address, length);
 		}
 	}
+	void UpdateFCR31() override;
 
 	const u8 *GetDispatcher() const override {
 		return dispatcher;
@@ -323,7 +324,6 @@ private:
 
 	const u8 *restoreRoundingMode;
 	const u8 *applyRoundingMode;
-	const u8 *updateRoundingMode;
 
 	const u8 *endOfPregeneratedCode;