Add SSAShort, shift, and, or, and fix unaligned store

src/CMakeLists.txt

@@ -1434,6 +1434,7 @@ set (PCH_SOURCES
 	r_compiler/ssa/ssa_if_block.cpp
 	r_compiler/ssa/ssa_int.cpp
 	r_compiler/ssa/ssa_int_ptr.cpp
+	r_compiler/ssa/ssa_short.cpp
 	r_compiler/ssa/ssa_scope.cpp
 	r_compiler/ssa/ssa_struct_type.cpp
 	r_compiler/ssa/ssa_ubyte.cpp

src/r_compiler/ssa/ssa_float_ptr.cpp

@@ -38,7 +38,6 @@ SSAVec4f SSAFloatPtr::load_unaligned_vec4f() const
 {
 	llvm::PointerType *m4xfloattypeptr = llvm::VectorType::get(llvm::Type::getFloatTy(SSAScope::context()), 4)->getPointerTo();
 	return SSAVec4f::from_llvm(SSAScope::builder().Insert(new llvm::LoadInst(SSAScope::builder().CreateBitCast(v, m4xfloattypeptr, SSAScope::hint()), SSAScope::hint(), false, 4), SSAScope::hint()));
-	// return SSAVec4f::from_llvm(SSAScope::builder().CreateCall(get_intrinsic(llvm::Intrinsic::x86_sse2_loadu_dq), SSAScope::builder().CreateBitCast(v, llvm::PointerType::getUnqual(llvm::IntegerType::get(SSAScope::context(), 8)))));
 }
 
 void SSAFloatPtr::store(const SSAFloat &new_value)
@@ -49,17 +48,11 @@ void SSAFloatPtr::store(const SSAFloat &new_value)
 void SSAFloatPtr::store_vec4f(const SSAVec4f &new_value)
 {
 	llvm::PointerType *m4xfloattypeptr = llvm::VectorType::get(llvm::Type::getFloatTy(SSAScope::context()), 4)->getPointerTo();
-	SSAScope::builder().CreateAlignedStore(new_value.v, SSAScope::builder().CreateBitCast(v, m4xfloattypeptr, SSAScope::hint()), 16);
+	SSAScope::builder().CreateStore(new_value.v, SSAScope::builder().CreateBitCast(v, m4xfloattypeptr, SSAScope::hint()));
 }
 
 void SSAFloatPtr::store_unaligned_vec4f(const SSAVec4f &new_value)
 {
-	/*llvm::Value *values[2] =
-	{
-		SSAScope::builder().CreateBitCast(v, llvm::Type::getFloatPtrTy(SSAScope::context())),
-		new_value.v
-	};
-	SSAScope::builder().CreateCall(SSAScope::intrinsic(llvm::Intrinsic::x86_sse_storeu_ps), values);*/
 	llvm::PointerType *m4xfloattypeptr = llvm::VectorType::get(llvm::Type::getFloatTy(SSAScope::context()), 4)->getPointerTo();
-	SSAScope::builder().CreateStore(new_value.v, SSAScope::builder().CreateBitCast(v, m4xfloattypeptr, SSAScope::hint()));
+	SSAScope::builder().CreateAlignedStore(new_value.v, SSAScope::builder().CreateBitCast(v, m4xfloattypeptr, SSAScope::hint()), 4);
 }

src/r_compiler/ssa/ssa_int.cpp

@@ -115,3 +115,33 @@ SSAInt operator>>(const SSAInt &a, int bits)
 {
 	return SSAInt::from_llvm(SSAScope::builder().CreateLShr(a.v, bits, SSAScope::hint()));
 }
+
+SSAInt operator<<(const SSAInt &a, const SSAInt &bits)
+{
+	return SSAInt::from_llvm(SSAScope::builder().CreateShl(a.v, bits.v, SSAScope::hint()));
+}
+
+SSAInt operator>>(const SSAInt &a, const SSAInt &bits)
+{
+	return SSAInt::from_llvm(SSAScope::builder().CreateLShr(a.v, bits.v, SSAScope::hint()));
+}
+
+SSAInt operator&(const SSAInt &a, int b)
+{
+	return SSAInt::from_llvm(SSAScope::builder().CreateAnd(a.v, b, SSAScope::hint()));
+}
+
+SSAInt operator&(const SSAInt &a, const SSAInt &b)
+{
+	return SSAInt::from_llvm(SSAScope::builder().CreateAnd(a.v, b.v, SSAScope::hint()));
+}
+
+SSAInt operator|(const SSAInt &a, int b)
+{
+	return SSAInt::from_llvm(SSAScope::builder().CreateOr(a.v, b, SSAScope::hint()));
+}
+
+SSAInt operator|(const SSAInt &a, const SSAInt &b)
+{
+	return SSAInt::from_llvm(SSAScope::builder().CreateOr(a.v, b.v, SSAScope::hint()));
+}

src/r_compiler/ssa/ssa_int.h

@@ -39,3 +39,10 @@ SSAInt operator%(const SSAInt &a, int b);
 
 SSAInt operator<<(const SSAInt &a, int bits);
 SSAInt operator>>(const SSAInt &a, int bits);
+SSAInt operator<<(const SSAInt &a, const SSAInt &bits);
+SSAInt operator>>(const SSAInt &a, const SSAInt &bits);
+
+SSAInt operator&(const SSAInt &a, int b);
+SSAInt operator&(const SSAInt &a, const SSAInt &b);
+SSAInt operator|(const SSAInt &a, int b);
+SSAInt operator|(const SSAInt &a, const SSAInt &b);

src/r_compiler/ssa/ssa_int_ptr.cpp

@@ -48,11 +48,11 @@ void SSAIntPtr::store(const SSAInt &new_value)
 void SSAIntPtr::store_vec4i(const SSAVec4i &new_value)
 {
 	llvm::PointerType *m4xint32typeptr = llvm::VectorType::get(llvm::Type::getInt32Ty(SSAScope::context()), 4)->getPointerTo();
-	SSAScope::builder().CreateAlignedStore(new_value.v, SSAScope::builder().CreateBitCast(v, m4xint32typeptr, SSAScope::hint()), 16);
+	SSAScope::builder().CreateStore(new_value.v, SSAScope::builder().CreateBitCast(v, m4xint32typeptr, SSAScope::hint()));
 }
 
 void SSAIntPtr::store_unaligned_vec4i(const SSAVec4i &new_value)
 {
 	llvm::PointerType *m4xint32typeptr = llvm::VectorType::get(llvm::Type::getInt32Ty(SSAScope::context()), 4)->getPointerTo();
-	SSAScope::builder().CreateStore(new_value.v, SSAScope::builder().CreateBitCast(v, m4xint32typeptr, SSAScope::hint()));
+	SSAScope::builder().CreateAlignedStore(new_value.v, SSAScope::builder().CreateBitCast(v, m4xint32typeptr, SSAScope::hint()), 4);
 }

src/r_compiler/ssa/ssa_short.cpp

@@ -0,0 +1,148 @@
+
+#include "ssa_short.h"
+#include "ssa_float.h"
+#include "ssa_int.h"
+#include "ssa_scope.h"
+#include "r_compiler/llvm_include.h"
+
+SSAShort::SSAShort()
+: v(0)
+{
+}
+
+SSAShort::SSAShort(int constant)
+: v(0)
+{
+	v = llvm::ConstantInt::get(SSAScope::context(), llvm::APInt(16, constant, true));
+}
+
+SSAShort::SSAShort(SSAFloat f)
+: v(0)
+{
+	v = SSAScope::builder().CreateFPToSI(f.v, llvm::Type::getInt16Ty(SSAScope::context()), SSAScope::hint());
+}
+
+SSAShort::SSAShort(llvm::Value *v)
+: v(v)
+{
+}
+
+llvm::Type *SSAShort::llvm_type()
+{
+	return llvm::Type::getInt16Ty(SSAScope::context());
+}
+
+SSAShort operator+(const SSAShort &a, const SSAShort &b)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateAdd(a.v, b.v, SSAScope::hint()));
+}
+
+SSAShort operator-(const SSAShort &a, const SSAShort &b)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateSub(a.v, b.v, SSAScope::hint()));
+}
+
+SSAShort operator*(const SSAShort &a, const SSAShort &b)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateMul(a.v, b.v, SSAScope::hint()));
+}
+
+SSAShort operator/(const SSAShort &a, const SSAShort &b)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateSDiv(a.v, b.v, SSAScope::hint()));
+}
+
+SSAShort operator%(const SSAShort &a, const SSAShort &b)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateSRem(a.v, b.v, SSAScope::hint()));
+}
+
+SSAShort operator+(int a, const SSAShort &b)
+{
+	return SSAShort(a) + b;
+}
+
+SSAShort operator-(int a, const SSAShort &b)
+{
+	return SSAShort(a) - b;
+}
+
+SSAShort operator*(int a, const SSAShort &b)
+{
+	return SSAShort(a) * b;
+}
+
+SSAShort operator/(int a, const SSAShort &b)
+{
+	return SSAShort(a) / b;
+}
+
+SSAShort operator%(int a, const SSAShort &b)
+{
+	return SSAShort(a) % b;
+}
+
+SSAShort operator+(const SSAShort &a, int b)
+{
+	return a + SSAShort(b);
+}
+
+SSAShort operator-(const SSAShort &a, int b)
+{
+	return a - SSAShort(b);
+}
+
+SSAShort operator*(const SSAShort &a, int b)
+{
+	return a * SSAShort(b);
+}
+
+SSAShort operator/(const SSAShort &a, int b)
+{
+	return a / SSAShort(b);
+}
+
+SSAShort operator%(const SSAShort &a, int b)
+{
+	return a % SSAShort(b);
+}
+
+SSAShort operator<<(const SSAShort &a, int bits)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateShl(a.v, bits, SSAScope::hint()));
+}
+
+SSAShort operator>>(const SSAShort &a, int bits)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateLShr(a.v, bits, SSAScope::hint()));
+}
+
+SSAShort operator<<(const SSAShort &a, const SSAInt &bits)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateShl(a.v, bits.v, SSAScope::hint()));
+}
+
+SSAShort operator>>(const SSAShort &a, const SSAInt &bits)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateLShr(a.v, bits.v, SSAScope::hint()));
+}
+
+SSAShort operator&(const SSAShort &a, int b)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateAnd(a.v, b, SSAScope::hint()));
+}
+
+SSAShort operator&(const SSAShort &a, const SSAShort &b)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateAnd(a.v, b.v, SSAScope::hint()));
+}
+
+SSAShort operator|(const SSAShort &a, int b)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateOr(a.v, b, SSAScope::hint()));
+}
+
+SSAShort operator|(const SSAShort &a, const SSAShort &b)
+{
+	return SSAShort::from_llvm(SSAScope::builder().CreateOr(a.v, b.v, SSAScope::hint()));
+}

src/r_compiler/ssa/ssa_short.h

@@ -0,0 +1,49 @@
+
+#pragma once
+
+namespace llvm { class Value; }
+namespace llvm { class Type; }
+
+class SSAFloat;
+class SSAInt;
+
+class SSAShort
+{
+public:
+	SSAShort();
+	SSAShort(int constant);
+	SSAShort(SSAFloat f);
+	explicit SSAShort(llvm::Value *v);
+	static SSAShort from_llvm(llvm::Value *v) { return SSAShort(v); }
+	static llvm::Type *llvm_type();
+
+	llvm::Value *v;
+};
+
+SSAShort operator+(const SSAShort &a, const SSAShort &b);
+SSAShort operator-(const SSAShort &a, const SSAShort &b);
+SSAShort operator*(const SSAShort &a, const SSAShort &b);
+SSAShort operator/(const SSAShort &a, const SSAShort &b);
+SSAShort operator%(const SSAShort &a, const SSAShort &b);
+
+SSAShort operator+(int a, const SSAShort &b);
+SSAShort operator-(int a, const SSAShort &b);
+SSAShort operator*(int a, const SSAShort &b);
+SSAShort operator/(int a, const SSAShort &b);
+SSAShort operator%(int a, const SSAShort &b);
+
+SSAShort operator+(const SSAShort &a, int b);
+SSAShort operator-(const SSAShort &a, int b);
+SSAShort operator*(const SSAShort &a, int b);
+SSAShort operator/(const SSAShort &a, int b);
+SSAShort operator%(const SSAShort &a, int b);
+
+SSAShort operator<<(const SSAShort &a, int bits);
+SSAShort operator>>(const SSAShort &a, int bits);
+SSAShort operator<<(const SSAShort &a, const SSAInt &bits);
+SSAShort operator>>(const SSAShort &a, const SSAInt &bits);
+
+SSAShort operator&(const SSAShort &a, int b);
+SSAShort operator&(const SSAShort &a, const SSAShort &b);
+SSAShort operator|(const SSAShort &a, int b);
+SSAShort operator|(const SSAShort &a, const SSAShort &b);

src/r_compiler/ssa/ssa_ubyte_ptr.cpp

@@ -86,7 +86,7 @@ void SSAUBytePtr::store_vec4ub(const SSAVec4i &new_value)
 void SSAUBytePtr::store_vec16ub(const SSAVec16ub &new_value)
 {
 	llvm::PointerType *m16xint8typeptr = llvm::VectorType::get(llvm::Type::getInt8Ty(SSAScope::context()), 16)->getPointerTo();
-	llvm::StoreInst *inst = SSAScope::builder().CreateAlignedStore(new_value.v, SSAScope::builder().CreateBitCast(v, m16xint8typeptr, SSAScope::hint()), 16);
+	llvm::StoreInst *inst = SSAScope::builder().CreateStore(new_value.v, SSAScope::builder().CreateBitCast(v, m16xint8typeptr, SSAScope::hint()));
 
 	// The following generates _mm_stream_si128, maybe!
 	// llvm::MDNode *node = llvm::MDNode::get(SSAScope::context(), SSAScope::builder().getInt32(1));
@@ -95,12 +95,6 @@ void SSAUBytePtr::store_vec16ub(const SSAVec16ub &new_value)
 
 void SSAUBytePtr::store_unaligned_vec16ub(const SSAVec16ub &new_value)
 {
-	/*llvm::Value *values[2] =
-	{
-		SSAScope::builder().CreateBitCast(v, llvm::Type::getInt8PtrTy(SSAScope::context())),
-		new_value.v
-	};
-	SSAScope::builder().CreateCall(SSAScope::intrinsic(llvm::Intrinsic::x86_sse2_storeu_dq), values);*/
 	llvm::PointerType *m16xint8typeptr = llvm::VectorType::get(llvm::Type::getInt8Ty(SSAScope::context()), 16)->getPointerTo();
-	llvm::StoreInst *inst = SSAScope::builder().CreateStore(new_value.v, SSAScope::builder().CreateBitCast(v, m16xint8typeptr, SSAScope::hint()));
+	llvm::StoreInst *inst = SSAScope::builder().CreateAlignedStore(new_value.v, SSAScope::builder().CreateBitCast(v, m16xint8typeptr, SSAScope::hint()), 4);
 }

src/r_compiler/ssa/ssa_vec4f_ptr.cpp

@@ -35,16 +35,10 @@ SSAVec4f SSAVec4fPtr::load_unaligned() const
 
 void SSAVec4fPtr::store(const SSAVec4f &new_value)
 {
-	SSAScope::builder().CreateAlignedStore(new_value.v, v, 16, false);
+	SSAScope::builder().CreateStore(new_value.v, v, false);
 }
 
 void SSAVec4fPtr::store_unaligned(const SSAVec4f &new_value)
 {
-	/*llvm::Value *values[2] =
+	SSAScope::builder().CreateAlignedStore(new_value.v, v, 4, false);
-	{
-		SSAScope::builder().CreateBitCast(v, llvm::Type::getFloatPtrTy(SSAScope::context())),
-		new_value.v
-	};
-	SSAScope::builder().CreateCall(SSAScope::intrinsic(llvm::Intrinsic::x86_sse_storeu_ps), values);*/
-	SSAScope::builder().CreateStore(new_value.v, v, false);
 }

src/r_compiler/ssa/ssa_vec4i_ptr.cpp

@@ -35,16 +35,10 @@ SSAVec4i SSAVec4iPtr::load_unaligned() const
 
 void SSAVec4iPtr::store(const SSAVec4i &new_value)
 {
-	SSAScope::builder().CreateAlignedStore(new_value.v, v, 16, false);
+	SSAScope::builder().CreateStore(new_value.v, v, false);
 }
 
 void SSAVec4iPtr::store_unaligned(const SSAVec4i &new_value)
 {
-	/*llvm::Value *values[2] =
+	SSAScope::builder().CreateAlignedStore(new_value.v, v, 4, false);
-	{
-		v,
-		new_value.v
-	};
-	SSAScope::builder().CreateCall(SSAScope::intrinsic(llvm::Intrinsic::x86_sse2_storeu_pd), values);*/
-	SSAScope::builder().CreateStore(new_value.v, v, false);
 }