Merge pull request #124 from SChernykh/master

Optimized loading from scratchpad

src/asm/program_loop_load.inc

@@ -1,5 +1,3 @@
-	mov rdx, rax
-	and eax, RANDOMX_SCRATCHPAD_MASK
 	lea rcx, [rsi+rax]
 	push rcx
 	xor r8,  qword ptr [rcx+0]
@@ -10,8 +8,6 @@
 	xor r13, qword ptr [rcx+40]
 	xor r14, qword ptr [rcx+48]
 	xor r15, qword ptr [rcx+56]
-	ror rdx, 32
-	and edx, RANDOMX_SCRATCHPAD_MASK
 	lea rcx, [rsi+rdx]
 	push rcx
 	cvtdq2pd xmm0, qword ptr [rcx+0]

src/asm/program_loop_store.inc

@@ -1,4 +1,3 @@
-	xor eax, eax
 	pop rcx
 	mov qword ptr [rcx+0], r8
 	mov qword ptr [rcx+8], r9

src/jit_compiler_x86.cpp

@@ -243,7 +243,7 @@ namespace randomx {
 		generateProgramPrologue(prog, pcfg);
 		memcpy(code + codePos, codeReadDataset, readDatasetSize);
 		codePos += readDatasetSize;
-		generateProgramEpilogue(prog);
+		generateProgramEpilogue(prog, pcfg);
 	}
 
 	void JitCompilerX86::generateProgramLight(Program& prog, ProgramConfiguration& pcfg, uint32_t datasetOffset) {
@@ -254,7 +254,7 @@ namespace randomx {
 		emitByte(CALL);
 		emit32(superScalarHashOffset - (codePos + 4));
 		emit(codeReadDatasetLightSshFin, readDatasetLightFinSize);
-		generateProgramEpilogue(prog);
+		generateProgramEpilogue(prog, pcfg);
 	}
 
 	template<size_t N>
@@ -298,12 +298,13 @@ namespace randomx {
 		for (unsigned i = 0; i < 8; ++i) {
 			registerUsage[i] = -1;
 		}
+
+		codePos = ((uint8_t*)randomx_program_prologue_first_load) - ((uint8_t*)randomx_program_prologue);
+		code[codePos + sizeof(REX_XOR_RAX_R64)] = 0xc0 + pcfg.readReg0;
+		code[codePos + sizeof(REX_XOR_RAX_R64) * 2 + 1] = 0xc0 + pcfg.readReg1;
+
 		codePos = prologueSize;
 		memcpy(code + codePos - 48, &pcfg.eMask, sizeof(pcfg.eMask));
-		emit(REX_XOR_RAX_R64);
-		emitByte(0xc0 + pcfg.readReg0);
-		emit(REX_XOR_RAX_R64);
-		emitByte(0xc0 + pcfg.readReg1);
 		memcpy(code + codePos, codeLoopLoad, loopLoadSize);
 		codePos += loopLoadSize;
 		for (unsigned i = 0; i < prog.getSize(); ++i) {
@@ -318,7 +319,12 @@ namespace randomx {
 		emitByte(0xc0 + pcfg.readReg3);
 	}
 
-	void JitCompilerX86::generateProgramEpilogue(Program& prog) {
+	void JitCompilerX86::generateProgramEpilogue(Program& prog, ProgramConfiguration& pcfg) {
+		emit(REX_MOV_RR64);
+		emitByte(0xc0 + pcfg.readReg0);
+		emit(REX_XOR_RAX_R64);
+		emitByte(0xc0 + pcfg.readReg1);
+		emit((const uint8_t*)&randomx_prefetch_scratchpad, ((uint8_t*)&randomx_prefetch_scratchpad_end) - ((uint8_t*)&randomx_prefetch_scratchpad));
 		memcpy(code + codePos, codeLoopStore, loopStoreSize);
 		codePos += loopStoreSize;
 		emit(SUB_EBX);

src/jit_compiler_x86.hpp

@@ -73,7 +73,7 @@ namespace randomx {
 		int32_t codePos;
 
 		void generateProgramPrologue(Program&, ProgramConfiguration&);
-		void generateProgramEpilogue(Program&);
+		void generateProgramEpilogue(Program&, ProgramConfiguration&);
 		void genAddressReg(Instruction&, bool);
 		void genAddressRegDst(Instruction&);
 		void genAddressImm(Instruction&);

src/jit_compiler_x86_static.S

@@ -37,7 +37,10 @@
 #define WINABI
 #endif
 
+.global DECL(randomx_prefetch_scratchpad)
+.global DECL(randomx_prefetch_scratchpad_end)
 .global DECL(randomx_program_prologue)
+.global DECL(randomx_program_prologue_first_load)
 .global DECL(randomx_program_loop_begin)
 .global DECL(randomx_program_loop_load)
 .global DECL(randomx_program_start)
@@ -65,6 +68,16 @@
 
 #define db .byte
 
+DECL(randomx_prefetch_scratchpad):
+	mov rdx, rax
+	and eax, RANDOMX_SCRATCHPAD_MASK
+	prefetcht0 [rsi+rax]
+	ror rdx, 32
+	and edx, RANDOMX_SCRATCHPAD_MASK
+	prefetcht0 [rsi+rdx]
+
+DECL(randomx_prefetch_scratchpad_end):
+
 .balign 64
 DECL(randomx_program_prologue):
 #if defined(WINABI)
@@ -75,6 +88,14 @@ DECL(randomx_program_prologue):
 	movapd xmm13, xmmword ptr [mantissaMask+rip]
 	movapd xmm14, xmmword ptr [exp240+rip]
 	movapd xmm15, xmmword ptr [scaleMask+rip]
+
+DECL(randomx_program_prologue_first_load):
+	xor rax, r8
+	xor rax, r8
+	mov rdx, rax
+	and eax, RANDOMX_SCRATCHPAD_MASK
+	ror rdx, 32
+	and edx, RANDOMX_SCRATCHPAD_MASK
 	jmp DECL(randomx_program_loop_begin)
 
 .balign 64

src/jit_compiler_x86_static.asm

@@ -28,7 +28,10 @@ IFDEF RAX
 
 _RANDOMX_JITX86_STATIC SEGMENT PAGE READ EXECUTE
 
+PUBLIC randomx_prefetch_scratchpad
+PUBLIC randomx_prefetch_scratchpad_end
 PUBLIC randomx_program_prologue
+PUBLIC randomx_program_prologue_first_load
 PUBLIC randomx_program_loop_begin
 PUBLIC randomx_program_loop_load
 PUBLIC randomx_program_start
@@ -54,15 +57,36 @@ RANDOMX_CACHE_MASK          EQU (RANDOMX_ARGON_MEMORY*16-1)
 RANDOMX_ALIGN               EQU 4096
 SUPERSCALAR_OFFSET          EQU ((((RANDOMX_ALIGN + 32 * RANDOMX_PROGRAM_SIZE) - 1) / (RANDOMX_ALIGN) + 1) * (RANDOMX_ALIGN))
 
+randomx_prefetch_scratchpad PROC
+	mov rdx, rax
+	and eax, RANDOMX_SCRATCHPAD_MASK
+	prefetcht0 [rsi+rax]
+	ror rdx, 32
+	and edx, RANDOMX_SCRATCHPAD_MASK
+	prefetcht0 [rsi+rdx]
+randomx_prefetch_scratchpad ENDP
+
+randomx_prefetch_scratchpad_end PROC
+randomx_prefetch_scratchpad_end ENDP
+
 ALIGN 64
 randomx_program_prologue PROC
 	include asm/program_prologue_win64.inc
 	movapd xmm13, xmmword ptr [mantissaMask]
 	movapd xmm14, xmmword ptr [exp240]
 	movapd xmm15, xmmword ptr [scaleMask]
-	jmp randomx_program_loop_begin
 randomx_program_prologue ENDP
 
+randomx_program_prologue_first_load PROC
+	xor rax, r8
+	xor rax, r8
+	mov rdx, rax
+	and eax, RANDOMX_SCRATCHPAD_MASK
+	ror rdx, 32
+	and edx, RANDOMX_SCRATCHPAD_MASK
+	jmp randomx_program_loop_begin
+randomx_program_prologue_first_load ENDP
+
 ALIGN 64
 	include asm/program_xmm_constants.inc
 

src/jit_compiler_x86_static.hpp

@@ -29,7 +29,10 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #pragma once
 
 extern "C" {
+	void randomx_prefetch_scratchpad();
+	void randomx_prefetch_scratchpad_end();
 	void randomx_program_prologue();
+	void randomx_program_prologue_first_load();
 	void randomx_program_loop_begin();
 	void randomx_program_loop_load();
 	void randomx_program_start();