Mix dataset cacheline with registers r0-r7

src/AssemblyGeneratorX86.cpp

@@ -67,20 +67,16 @@ namespace RandomX {
 	void AssemblyGeneratorX86::gena(Instruction& instr, int i) {
 		asmCode << "\txor " << regR[instr.rega % RegistersCount] << ", 0" << std::hex << instr.addra << "h" << std::dec << std::endl;
 		asmCode << "\tmov eax, " << regR32[instr.rega % RegistersCount] << std::endl;
+		if ((instr.loca & 192) == 0)
+			asmCode << "\txor " << regMx << ", rax" << std::endl;
 		asmCode << "\ttest " << regIc8 << ", 63" << std::endl;
 		asmCode << "\tjnz short rx_body_" << i << std::endl;
+		asmCode << "\tcall rx_read" << std::endl;
+		asmCode << "rx_body_" << i << ":" << std::endl;
 		if (instr.loca & 3) {
-			asmCode << "\tcall rx_read_l1" << std::endl;
-			asmCode << "rx_body_" << i << ":" << std::endl;
-			if ((instr.loca & 192) == 0)
-				asmCode << "\txor " << regMx << ", rax" << std::endl;
 			asmCode << "\tand eax, " << (ScratchpadL1 - 1) << std::endl;
 		}
 		else {
-			asmCode << "\tcall rx_read_l2" << std::endl;
-			asmCode << "rx_body_" << i << ":" << std::endl;
-			if ((instr.loca & 192) == 0)
-				asmCode << "\txor " << regMx << ", rax" << std::endl;
 			asmCode << "\tand eax, " << (ScratchpadL2 - 1) << std::endl;
 		}
 	}

src/JitCompilerX86-static.S

@@ -29,8 +29,7 @@
 .global DECL(randomx_program_prologue)
 .global DECL(randomx_program_begin)
 .global DECL(randomx_program_epilogue)
-.global DECL(randomx_program_read_l1)
+.global DECL(randomx_program_read)
-.global DECL(randomx_program_read_l2)
 .global DECL(randomx_program_end)
 .global DECL(randomx_program_transform)
 
@@ -48,22 +47,10 @@ DECL(randomx_program_begin):
 DECL(randomx_program_epilogue):
 	#include "asm/program_epilogue_linux.inc"
 
-#define scratchpad_mask and eax, 2040
-
 .align 64
-DECL(randomx_program_read_l1):
+DECL(randomx_program_read):
 	#include "asm/program_read.inc"
 
-#undef scratchpad_mask
-
-#define scratchpad_mask and eax, 32760
-
-.align 64
-DECL(randomx_program_read_l2):
-	#include "asm/program_read.inc"
-
-#undef scratchpad_mask
-
 .align 64
 DECL(randomx_program_end):
 	nop

src/JitCompilerX86-static.asm

@@ -20,8 +20,7 @@ _RANDOMX_JITX86_STATIC SEGMENT PAGE READ EXECUTE
 PUBLIC randomx_program_prologue
 PUBLIC randomx_program_begin
 PUBLIC randomx_program_epilogue
-PUBLIC randomx_program_read_l1
+PUBLIC randomx_program_read
-PUBLIC randomx_program_read_l2
 PUBLIC randomx_program_end
 PUBLIC randomx_program_transform
 
@@ -41,23 +40,10 @@ randomx_program_epilogue PROC
 	include asm/program_epilogue_win64.inc
 randomx_program_epilogue ENDP
 
-scratchpad_mask MACRO
-	and eax, 2040
-ENDM
-
 ALIGN 64
-randomx_program_read_l1 PROC
+randomx_program_read PROC
 	include asm/program_read.inc
-randomx_program_read_l1 ENDP
+randomx_program_read ENDP
-
-scratchpad_mask MACRO
-	and eax, 32760
-ENDM
-
-ALIGN 64
-randomx_program_read_l2 PROC
-	include asm/program_read.inc
-randomx_program_read_l2 ENDP
 
 ALIGN 64
 randomx_program_end PROC

src/JitCompilerX86-static.hpp

@@ -22,7 +22,6 @@ extern "C" {
   void randomx_program_begin();
   void randomx_program_epilogue();
   void randomx_program_transform();
-  void randomx_program_read_l1();
+  void randomx_program_read();
-  void randomx_program_read_l2();
   void randomx_program_end();
 }

src/JitCompilerX86.cpp

@@ -102,22 +102,19 @@ namespace RandomX {
 	const uint8_t* codePrologue = (uint8_t*)&randomx_program_prologue;
 	const uint8_t* codeProgramBegin = (uint8_t*)&randomx_program_begin;
 	const uint8_t* codeEpilogue = (uint8_t*)&randomx_program_epilogue;
-	const uint8_t* codeReadDatasetL1 = (uint8_t*)&randomx_program_read_l1;
+	const uint8_t* codeReadDataset = (uint8_t*)&randomx_program_read;
-	const uint8_t* codeReadDatasetL2 = (uint8_t*)&randomx_program_read_l2;
 	const uint8_t* codeProgramEnd = (uint8_t*)&randomx_program_end;
 	const uint32_t* addressTransformations = (uint32_t*)&randomx_program_transform;
 
 	const int32_t prologueSize = codeProgramBegin - codePrologue;
-	const int32_t epilogueSize = codeReadDatasetL1 - codeEpilogue;
+	const int32_t epilogueSize = codeReadDataset - codeEpilogue;
-	const int32_t readDatasetL1Size = codeReadDatasetL2 - codeReadDatasetL1;
+	const int32_t readDatasetSize = codeProgramEnd - codeReadDataset;
-	const int32_t readDatasetL2Size = codeProgramEnd - codeReadDatasetL2;
 
-	const int32_t readDatasetL2Offset = CodeSize - readDatasetL2Size;
+	const int32_t readDatasetOffset = CodeSize - readDatasetSize;
-	const int32_t readDatasetL1Offset = readDatasetL2Offset - readDatasetL1Size;
+	const int32_t epilogueOffset = readDatasetOffset - epilogueSize;
-	const int32_t epilogueOffset = readDatasetL1Offset - epilogueSize;
 
 	size_t JitCompilerX86::getCodeSize() {
-		return codePos - prologueSize + readDatasetL1Size + readDatasetL2Size;
+		return codePos - prologueSize + readDatasetSize;
 	}
 
 	JitCompilerX86::JitCompilerX86() {
@@ -131,9 +128,8 @@ namespace RandomX {
 			throw std::runtime_error("mmap failed");
 #endif
 		memcpy(code, codePrologue, prologueSize);
-		memcpy(code + CodeSize - epilogueSize - readDatasetL1Size - readDatasetL2Size, codeEpilogue, epilogueSize);
+		memcpy(code + CodeSize - epilogueSize - readDatasetSize, codeEpilogue, epilogueSize);
-		memcpy(code + CodeSize - readDatasetL1Size - readDatasetL2Size, codeReadDatasetL1, readDatasetL1Size);
+		memcpy(code + CodeSize - readDatasetSize, codeReadDataset, readDatasetSize);
-		memcpy(code + CodeSize - readDatasetL2Size, codeReadDatasetL2, readDatasetL2Size);
 	}
 
 	void JitCompilerX86::generateProgram(Pcg32& gen) {
@@ -150,10 +146,8 @@ namespace RandomX {
 		emitByte(0xe9);
 		emit(instructionOffsets[0] - (codePos + 4));
 		fixCallOffsets();
-		uint32_t transformL1 = addressTransformations[gen.getUniform(0, TransformationCount - 1)];
+		uint32_t transform = addressTransformations[gen.getUniform(0, TransformationCount - 1)];
-		uint32_t transformL2 = addressTransformations[gen.getUniform(0, TransformationCount - 1)];
+		*reinterpret_cast<uint32_t*>(code + readDatasetOffset) = transform;
-		*reinterpret_cast<uint32_t*>(code + readDatasetL1Offset + 1) = transformL1;
-		*reinterpret_cast<uint32_t*>(code + readDatasetL2Offset + 1) = transformL2;
 	}
 
 	void JitCompilerX86::generateCode(Instruction& instr, int i) {
@@ -176,18 +170,13 @@ namespace RandomX {
 		emit(instr.addra);
 		emit(uint16_t(0x8b41)); //mov
 		emitByte(0xc0 + (instr.rega % RegistersCount)); //eax, rega
-		emit(0x753fc3f6); //test bl,0x3f; jne
-		emit(uint16_t(0xe805));
-		if (instr.loca & 3) { //A.LOC.W
-			emit(readDatasetL1Offset - (codePos + 4));
-		}
-		else {
-			emit(readDatasetL2Offset - (codePos + 4));
-		}
 		if ((instr.loca & 192) == 0) { //A.LOC.X
 			emit(uint16_t(0x3348));
 			emitByte(0xe8); //xor rbp, rax
 		}
+		emit(0x753fc3f6); //test bl,0x3f; jne
+		emit(uint16_t(0xe805));
+		emit(readDatasetOffset - (codePos + 4));
 		emitByte(0x25); //and eax,
 		if (instr.loca & 3) {
 			emit(ScratchpadL1 - 1); //first 16 KiB of scratchpad

src/asm/program_read.inc

@@ -1,4 +1,3 @@
-	push rax                        ;# preserve eax
 	db 0, 0, 0, 0                   ;# TransformAddress placeholder
 	mov rcx, qword ptr [rdi]        ;# load the dataset address
 	xor rbp, rax                    ;# modify "mx"
@@ -9,24 +8,13 @@
 	;# read cacheline "ma"
 	ror rbp, 32                     ;# swap "ma" and "mx"
 	mov edx, ebp                    ;# edx = ma
-	scratchpad_mask                 ;# limit address to the specified scratchpad size aligned to multiple of 8
-	lea rax, [rsi+rax*8]            ;# scratchpad cache line
 	lea rcx, [rcx+rdx]              ;# dataset cache line
-	mov rdx, qword ptr [rcx+0]      ;# load first dataset quadword (prefetched into the cache by now)
+	xor r8,  qword ptr [rcx+0]
-	xor qword ptr [rax+0], rdx      ;# XOR the dataset item with a scratchpad item, repeat for the rest of the cacheline
+	xor r9,  qword ptr [rcx+8]
-	mov rdx, qword ptr [rcx+8]
+	xor r10, qword ptr [rcx+16]
-	xor qword ptr [rax+8], rdx
+	xor r11, qword ptr [rcx+24]
-	mov rdx, qword ptr [rcx+16]
+	xor r12, qword ptr [rcx+32]
-	xor qword ptr [rax+16], rdx
+	xor r13, qword ptr [rcx+40]
-	mov rdx, qword ptr [rcx+24]
+	xor r14, qword ptr [rcx+48]
-	xor qword ptr [rax+24], rdx
+	xor r15, qword ptr [rcx+56]
-	mov rdx, qword ptr [rcx+32]
-	xor qword ptr [rax+32], rdx
-	mov rdx, qword ptr [rcx+40]
-	xor qword ptr [rax+40], rdx
-	mov rdx, qword ptr [rcx+48]
-	xor qword ptr [rax+48], rdx
-	mov rdx, qword ptr [rcx+56]
-	xor qword ptr [rax+56], rdx
-	pop rax                         ;# restore eax
 	ret

src/executeProgram-win64.asm

@@ -221,54 +221,33 @@ TransformAddress MACRO reg32, reg64
 	;xor reg32, -8                   ;# C = all except 0 to 7
 ENDM
 
-ReadMemoryRandom MACRO spmask
+ALIGN 64
+rx_read:
 ;# IN     eax = random 32-bit address
 ;# GLOBAL rdi = address of the dataset address
 ;# GLOBAL rsi = address of the scratchpad
 ;# GLOBAL rbp = low 32 bits = "mx", high 32 bits = "ma"
 ;# MODIFY rcx, rdx
-	push rax                        ;# preserve eax
 	TransformAddress eax, rax       ;# TransformAddress function
 	mov rcx, qword ptr [rdi]        ;# load the dataset address
 	xor rbp, rax                    ;# modify "mx"
-	; prefetch cacheline "mx"
+	;# prefetch cacheline "mx"
 	and rbp, -64                    ;# align "mx" to the start of a cache line
 	mov edx, ebp                    ;# edx = mx
 	prefetchnta byte ptr [rcx+rdx]
-	; read cacheline "ma"
+	;# read cacheline "ma"
 	ror rbp, 32                     ;# swap "ma" and "mx"
 	mov edx, ebp                    ;# edx = ma
-	and eax, spmask-7               ;# limit address to the specified scratchpad size aligned to multiple of 8
-	lea rax, [rsi+rax*8]            ;# scratchpad cache line
 	lea rcx, [rcx+rdx]              ;# dataset cache line
-	mov rdx, qword ptr [rcx+0]      ;# load first dataset quadword (prefetched into the cache by now)
+	xor r8,  qword ptr [rcx+0]
-	xor qword ptr [rax+0], rdx      ;# XOR the dataset item with a scratchpad item, repeat for the rest of the cacheline
+	xor r9,  qword ptr [rcx+8]
-	mov rdx, qword ptr [rcx+8]
+	xor r10, qword ptr [rcx+16]
-	xor qword ptr [rax+8], rdx
+	xor r11, qword ptr [rcx+24]
-	mov rdx, qword ptr [rcx+16]
+	xor r12, qword ptr [rcx+32]
-	xor qword ptr [rax+16], rdx
+	xor r13, qword ptr [rcx+40]
-	mov rdx, qword ptr [rcx+24]
+	xor r14, qword ptr [rcx+48]
-	xor qword ptr [rax+24], rdx
+	xor r15, qword ptr [rcx+56]
-	mov rdx, qword ptr [rcx+32]
-	xor qword ptr [rax+32], rdx
-	mov rdx, qword ptr [rcx+40]
-	xor qword ptr [rax+40], rdx
-	mov rdx, qword ptr [rcx+48]
-	xor qword ptr [rax+48], rdx
-	mov rdx, qword ptr [rcx+56]
-	xor qword ptr [rax+56], rdx
-	pop rax                         ;# restore eax
 	ret
-ENDM
-
-ALIGN 64
-rx_read_l1:
-ReadMemoryRandom 2047
-
-ALIGN 64
-rx_read_l2:
-ReadMemoryRandom 32767
-
 executeProgram ENDP
 
 _RANDOMX_EXECUTE_PROGRAM ENDS