riscv: Initial vertexjit

CMakeLists.txt

@@ -1485,6 +1485,10 @@ list(APPEND CoreExtra
 	Core/MIPS/MIPS/MipsJit.h
 )
 
+list(APPEND CoreExtra
+	GPU/Common/VertexDecoderRiscV.cpp
+)
+
 if(NOT MOBILE_DEVICE)
 	set(CoreExtra ${CoreExtra}
 		Core/AVIDump.cpp

Common/CommonFuncs.h

@@ -36,6 +36,8 @@
 #define Crash() {asm ("bkpt #0");}
 #elif PPSSPP_ARCH(ARM64)
 #define Crash() {asm ("brk #0");}
+#elif PPSSPP_ARCH(RISCV64)
+#define Crash() {asm ("ebreak");}
 #else
 #include <signal.h>
 #define Crash() {kill(getpid(), SIGINT);}

Common/RiscVEmitter.cpp

@@ -1021,8 +1021,12 @@ void RiscVEmitter::ReserveCodeSpace(u32 bytes) {
 	_assert_msg_((bytes & 3) == 0 || SupportsCompressed(), "Code space should be aligned (no compressed)");
 	for (u32 i = 0; i < bytes / 4; i++)
 		EBREAK();
-	if (bytes & 2)
-		Write16(0);
+	if (bytes & 2) {
+		if (SupportsCompressed())
+			C_EBREAK();
+		else
+			Write16(0);
+	}
 }
 
 const u8 *RiscVEmitter::AlignCode16() {
@@ -4278,4 +4282,18 @@ void RiscVEmitter::C_SDSP(RiscVReg rs2, u32 uimm9) {
 	Write16(EncodeCSS(Opcode16::C2, rs2, imm5_4_3_8_7_6, Funct3::C_SDSP));
 }
 
+void RiscVCodeBlock::PoisonMemory(int offset) {
+	// So we can adjust region to writable space.  Might be zero.
+	ptrdiff_t writable = writable_ - code_;
+
+	u32 *ptr = (u32 *)(region + offset + writable);
+	u32 *maxptr = (u32 *)(region + region_size - offset + writable);
+	// This will only write an even multiple of u32, but not much else to do.
+	// RiscV: 0x00100073 = EBREAK, 0x9002 = C.EBREAK
+	while (ptr + 1 <= maxptr)
+		*ptr++ = 0x00100073;
+	if (SupportsCompressed() && ptr < maxptr && (intptr_t)maxptr - (intptr_t)ptr >= 2)
+		*(u16 *)ptr = 0x9002;
+}
+
 };

Common/RiscVEmitter.h

@@ -47,6 +47,8 @@ enum RiscVReg {
 	V8, V9, V10, V11, V12, V13, V14, V15,
 	V16, V17, V18, V19, V20, V21, V22, V23,
 	V24, V25, V26, V27, V28, V29, V30, V31,
+
+	INVALID_REG = 0xFFFFFFFF,
 };
 
 enum class FixupBranchType {
@@ -302,7 +304,7 @@ class RiscVEmitter {
 	void ECALL();
 	void EBREAK();
 
-	// 64-bit instructions - oens ending in W sign extend result to 32 bits.
+	// 64-bit instructions - ones ending in W sign extend result to 32 bits.
 	void LWU(RiscVReg rd, RiscVReg addr, s32 simm12);
 	void LD(RiscVReg rd, RiscVReg addr, s32 simm12);
 	void SD(RiscVReg rs2, RiscVReg addr, s32 simm12);
@@ -1024,13 +1026,14 @@ class RiscVEmitter {
 		writable_ += 2;
 	}
 
+protected:
 	const u8 *code_ = nullptr;
 	u8 *writable_ = nullptr;
 	const u8 *lastCacheFlushEnd_ = nullptr;
 	bool autoCompress_ = false;
 };
 
-class MIPSCodeBlock : public CodeBlock<RiscVEmitter> {
+class RiscVCodeBlock : public CodeBlock<RiscVEmitter> {
 private:
 	void PoisonMemory(int offset) override;
 };

Core/Dialog/PSPOskDialog.cpp

@@ -350,6 +350,7 @@ int PSPOskDialog::Init(u32 oskPtr) {
 
 	// Eat any keys pressed before the dialog inited.
 	UpdateButtons();
+	InitCommon();
 
 	std::lock_guard<std::mutex> guard(nativeMutex_);
 	nativeStatus_ = PSPOskNativeStatus::IDLE;
@@ -954,6 +955,7 @@ int PSPOskDialog::Update(int animSpeed) {
 	const int framesHeldRepeatRate = 5;
 
 	UpdateButtons();
+	UpdateCommon();
 	int selectedRow = selectedChar / numKeyCols[currentKeyboard];
 	int selectedExtra = selectedChar % numKeyCols[currentKeyboard];
 

Core/MIPS/JitCommon/JitBlockCache.cpp

@@ -698,7 +698,7 @@ JitBlockDebugInfo JitBlockCache::GetBlockDebugInfo(int blockNum) const {
 	debugInfo.targetDisasm = DisassembleArm64(block->normalEntry, block->codeSize);
 #elif PPSSPP_ARCH(X86) || PPSSPP_ARCH(AMD64)
 	debugInfo.targetDisasm = DisassembleX86(block->normalEntry, block->codeSize);
-#elif PPSSPP_ARCH(ARM64)
+#elif PPSSPP_ARCH(RISCV64)
 	debugInfo.targetDisasm = DisassembleRV64(block->normalEntry, block->codeSize);
 #endif
 

Core/MIPS/JitCommon/JitState.cpp

@@ -60,7 +60,7 @@ namespace MIPSComp {
 
 		useStaticAlloc = false;
 		enablePointerify = false;
-#if PPSSPP_ARCH(ARM64)
+#if PPSSPP_ARCH(ARM64) || PPSSPP_ARCH(RISCV64)
 		useStaticAlloc = !Disabled(JitDisable::STATIC_ALLOC);
 		// iOS/etc. may disable at runtime if Memory::base is not nicely aligned.
 		enablePointerify = !Disabled(JitDisable::POINTERIFY);

GPU/Common/VertexDecoderArm64.cpp

@@ -722,7 +722,7 @@ void VertexDecoderJitCache::Jit_NormalS16() {
 
 void VertexDecoderJitCache::Jit_NormalFloat() {
 	// Only need to copy 12 bytes, but copying 16 should be okay (and is faster.)
-	if ((dec_->posoff & 7) == 0 && (dec_->decFmt.posoff & 7) == 0) {
+	if ((dec_->nrmoff & 7) == 0 && (dec_->decFmt.nrmoff & 7) == 0) {
 		LDP(INDEX_SIGNED, EncodeRegTo64(tempReg1), EncodeRegTo64(tempReg2), srcReg, dec_->nrmoff);
 		STP(INDEX_SIGNED, EncodeRegTo64(tempReg1), EncodeRegTo64(tempReg2), dstReg, dec_->decFmt.nrmoff);
 	} else {

GPU/Common/VertexDecoderCommon.h

@@ -37,6 +37,8 @@
 #include "Common/Arm64Emitter.h"
 #elif PPSSPP_ARCH(X86) || PPSSPP_ARCH(AMD64)
 #include "Common/x64Emitter.h"
+#elif PPSSPP_ARCH(RISCV64)
+#include "Common/RiscVEmitter.h"
 #else
 #include "Common/FakeEmitter.h"
 #endif
@@ -489,6 +491,8 @@ class VertexDecoder {
 #define VERTEXDECODER_JIT_BACKEND Arm64Gen::ARM64CodeBlock
 #elif PPSSPP_ARCH(X86) || PPSSPP_ARCH(AMD64)
 #define VERTEXDECODER_JIT_BACKEND Gen::XCodeBlock
+#elif PPSSPP_ARCH(RISCV64)
+#define VERTEXDECODER_JIT_BACKEND RiscVGen::RiscVCodeBlock
 #endif