Merged load/store of integer and FP registers

src/JitCompilerX86-static.S

@@ -27,13 +27,11 @@
 #define DECL(x) x
 #endif
 .global DECL(randomx_program_prologue)
-.global DECL(randomx_loop_begin)
-.global DECL(randomx_program_load_int)
-.global DECL(randomx_program_load_flt)
+.global DECL(randomx_program_loop_begin)
+.global DECL(randomx_program_loop_load)
 .global DECL(randomx_program_start)
 .global DECL(randomx_program_read_dataset)
-.global DECL(randomx_program_store_int)
-.global DECL(randomx_program_store_flt)
+.global DECL(randomx_program_loop_store)
 .global DECL(randomx_program_loop_end)
 .global DECL(randomx_program_epilogue)
 .global DECL(randomx_program_end)
@@ -48,14 +46,11 @@ DECL(randomx_program_prologue):
 	#include "asm/program_xmm_constants.inc"
 
 .align 64
-DECL(randomx_loop_begin):
+DECL(randomx_program_loop_begin):
 	nop
 
-DECL(randomx_program_load_int):
-	#include "asm/program_load_int.inc"
-
-DECL(randomx_program_load_flt):
-	#include "asm/program_load_flt.inc"
+DECL(randomx_program_loop_load):
+	#include "asm/program_loop_load.inc"
 
 DECL(randomx_program_start):
 	nop
@@ -63,11 +58,8 @@ DECL(randomx_program_start):
 DECL(randomx_program_read_dataset):
 	#include "asm/program_read_dataset.inc"
 
-DECL(randomx_program_store_int):
-	#include "asm/program_store_int.inc"
-
-DECL(randomx_program_store_flt):
-	#include "asm/program_store_flt.inc"
+DECL(randomx_program_loop_store):
+	#include "asm/program_loop_store.inc"
 
 DECL(randomx_program_loop_end):
 	nop

src/JitCompilerX86-static.asm

@@ -20,13 +20,11 @@ IFDEF RAX
 _RANDOMX_JITX86_STATIC SEGMENT PAGE READ EXECUTE
 
 PUBLIC randomx_program_prologue
-PUBLIC randomx_loop_begin
-PUBLIC randomx_program_load_int
-PUBLIC randomx_program_load_flt
+PUBLIC randomx_program_loop_begin
+PUBLIC randomx_program_loop_load
 PUBLIC randomx_program_start
 PUBLIC randomx_program_read_dataset
-PUBLIC randomx_program_store_int
-PUBLIC randomx_program_store_flt
+PUBLIC randomx_program_loop_store
 PUBLIC randomx_program_loop_end
 PUBLIC randomx_program_epilogue
 PUBLIC randomx_program_end
@@ -40,17 +38,13 @@ ALIGN 64
 	include asm/program_xmm_constants.inc
 
 ALIGN 64
-randomx_loop_begin PROC
+randomx_program_loop_begin PROC
 	nop
-randomx_loop_begin ENDP
+randomx_program_loop_begin ENDP
 
-randomx_program_load_int PROC
-	include asm/program_load_int.inc
-randomx_program_load_int ENDP
-
-randomx_program_load_flt PROC
-	include asm/program_load_flt.inc
-randomx_program_load_flt ENDP
+randomx_program_loop_load PROC
+	include asm/program_loop_load.inc
+randomx_program_loop_load ENDP
 
 randomx_program_start PROC
 	nop
@@ -60,13 +54,9 @@ randomx_program_read_dataset PROC
 	include asm/program_read_dataset.inc
 randomx_program_read_dataset ENDP
 
-randomx_program_store_int PROC
-	include asm/program_store_int.inc
-randomx_program_store_int ENDP
-
-randomx_program_store_flt PROC
-	include asm/program_store_flt.inc
-randomx_program_store_flt ENDP
+randomx_program_loop_store PROC
+	include asm/program_loop_store.inc
+randomx_program_loop_store ENDP
 
 randomx_program_loop_end PROC
 	nop

src/JitCompilerX86-static.hpp

@@ -19,13 +19,11 @@ along with RandomX.  If not, see<http://www.gnu.org/licenses/>.
 
 extern "C" {
 	void randomx_program_prologue();
-	void randomx_loop_begin();
-	void randomx_program_load_int();
-	void randomx_program_load_flt();
+	void randomx_program_loop_begin();
+	void randomx_program_loop_load();
 	void randomx_program_start();
 	void randomx_program_read_dataset();
-	void randomx_program_store_int();
-	void randomx_program_store_flt();
+	void randomx_program_loop_store();
 	void randomx_program_loop_end();
 	void randomx_program_epilogue();
 	void randomx_program_end();

src/JitCompilerX86.cpp

@@ -94,13 +94,11 @@ namespace RandomX {
 #include "JitCompilerX86-static.hpp"
 
 	const uint8_t* codePrologue = (uint8_t*)&randomx_program_prologue;
-	const uint8_t* codeLoopBegin = (uint8_t*)&randomx_loop_begin;
-	const uint8_t* codeLoadInt = (uint8_t*)&randomx_program_load_int;
-	const uint8_t* codeLoadFlt = (uint8_t*)&randomx_program_load_flt;
+	const uint8_t* codeLoopBegin = (uint8_t*)&randomx_program_loop_begin;
+	const uint8_t* codeLoopLoad = (uint8_t*)&randomx_program_loop_load;
 	const uint8_t* codeProgamStart = (uint8_t*)&randomx_program_start;
 	const uint8_t* codeReadDataset = (uint8_t*)&randomx_program_read_dataset;
-	const uint8_t* codeStoreInt = (uint8_t*)&randomx_program_store_int;
-	const uint8_t* codeStoreFlt = (uint8_t*)&randomx_program_store_flt;
+	const uint8_t* codeLoopStore = (uint8_t*)&randomx_program_loop_store;
 	const uint8_t* codeLoopEnd = (uint8_t*)&randomx_program_loop_end;
 	const uint8_t* codeEpilogue = (uint8_t*)&randomx_program_epilogue;
 	const uint8_t* codeProgramEnd = (uint8_t*)&randomx_program_end;
@@ -108,11 +106,9 @@ namespace RandomX {
 	const int32_t prologueSize = codeLoopBegin - codePrologue;
 	const int32_t epilogueSize = codeProgramEnd - codeEpilogue;
 
-	const int32_t loadIntSize = codeLoadFlt - codeLoadInt;
-	const int32_t loadFltSize = codeProgamStart - codeLoadFlt;
-	const int32_t readDatasetSize = codeStoreInt - codeReadDataset;
-	const int32_t storeIntSize = codeStoreFlt - codeStoreInt;
-	const int32_t storeFltSize = codeLoopEnd - codeStoreFlt;
+	const int32_t loopLoadSize = codeProgamStart - codeLoopLoad;
+	const int32_t readDatasetSize = codeLoopStore - codeReadDataset;
+	const int32_t loopStoreSize = codeLoopEnd - codeLoopStore;
 
 	const int32_t epilogueOffset = CodeSize - epilogueSize;
 
@@ -179,6 +175,7 @@ namespace RandomX {
 	static const uint8_t SUB_EBX[] = { 0x83, 0xEB, 0x01 };
 	static const uint8_t JNZ[] = { 0x0f, 0x85 };
 	static const uint8_t JMP = 0xe9;
+	static const uint8_t REX_XOR_RAX_R64[] = { 0x49, 0x33 };
 
 	size_t JitCompilerX86::getCodeSize() {
 		return codePos - prologueSize;
@@ -204,18 +201,16 @@ namespace RandomX {
 		addressRegisters >>= 1;
 		int readReg2 = 2 + (addressRegisters & 1);
 		addressRegisters >>= 1;
-		int writeReg1 = 4 + (addressRegisters & 1);
+		int readReg3 = 4 + (addressRegisters & 1);
 		addressRegisters >>= 1;
-		int writeReg2 = 6 + (addressRegisters & 1);
+		int readReg4 = 6 + (addressRegisters & 1);
 		codePos = prologueSize;
-		emit(REX_XOR_EAX);
+		emit(REX_XOR_RAX_R64);
 		emitByte(0xc0 + readReg1);
-		memcpy(code + codePos, codeLoadInt, loadIntSize);
-		codePos += loadIntSize;
-		emit(REX_XOR_EAX);
+		emit(REX_XOR_RAX_R64);
 		emitByte(0xc0 + readReg2);
-		memcpy(code + codePos, codeLoadFlt, loadFltSize);
-		codePos += loadFltSize;
+		memcpy(code + codePos, codeLoopLoad, loopLoadSize);
+		codePos += loopLoadSize;
 		Instruction instr;
 		for (unsigned i = 0; i < ProgramLength; ++i) {
 			for (unsigned j = 0; j < sizeof(instr) / sizeof(Pcg32::result_type); ++j) {
@@ -226,19 +221,13 @@ namespace RandomX {
 			generateCode(instr);
 		}
 		emit(REX_MOV_RR);
-		emitByte(0xc0 + readReg1);
+		emitByte(0xc0 + readReg3);
 		emit(REX_XOR_EAX);
-		emitByte(0xc0 + readReg2);
+		emitByte(0xc0 + readReg4);
 		memcpy(code + codePos, codeReadDataset, readDatasetSize);
 		codePos += readDatasetSize;
-		emit(REX_MOV_RR);
-		emitByte(0xc0 + writeReg1);
-		memcpy(code + codePos, codeStoreInt, storeIntSize);
-		codePos += storeIntSize;
-		emit(REX_XOR_EAX);
-		emitByte(0xc0 + writeReg2);
-		memcpy(code + codePos, codeStoreFlt, storeFltSize);
-		codePos += storeFltSize;
+		memcpy(code + codePos, codeLoopStore, loopStoreSize);
+		codePos += loopStoreSize;
 		emit(SUB_EBX);
 		emit(JNZ);
 		emit32(prologueSize - codePos - 4);

src/asm/program_load_int.inc

@@ -1,10 +0,0 @@
-	and eax, 1048512
-	lea rcx, [rsi+rax]
-	xor r8,  qword ptr [rcx+0]
-	xor r9,  qword ptr [rcx+8]
-	xor r10, qword ptr [rcx+16]
-	xor r11, qword ptr [rcx+24]
-	xor r12, qword ptr [rcx+32]
-	xor r13, qword ptr [rcx+40]
-	xor r14, qword ptr [rcx+48]
-	xor r15, qword ptr [rcx+56]

src/asm/program_load_flt.inc → src/asm/program_loop_load.inc

@@ -1,5 +1,19 @@
+	mov rdx, rax
 	and eax, 1048512
 	lea rcx, [rsi+rax]
+	push rcx
+	xor r8,  qword ptr [rcx+0]
+	xor r9,  qword ptr [rcx+8]
+	xor r10, qword ptr [rcx+16]
+	xor r11, qword ptr [rcx+24]
+	xor r12, qword ptr [rcx+32]
+	xor r13, qword ptr [rcx+40]
+	xor r14, qword ptr [rcx+48]
+	xor r15, qword ptr [rcx+56]
+	ror rdx, 32
+	and edx, 1048512
+	lea rcx, [rsi+rdx]
+	push rcx
 	cvtdq2pd xmm0, qword ptr [rcx+0]
 	cvtdq2pd xmm1, qword ptr [rcx+8]
 	cvtdq2pd xmm2, qword ptr [rcx+16]

src/asm/program_loop_store.inc

@@ -0,0 +1,18 @@
+	pop rcx
+	mov qword ptr [rcx+0], r8
+	mov qword ptr [rcx+8], r9
+	mov qword ptr [rcx+16], r10
+	mov qword ptr [rcx+24], r11
+	mov qword ptr [rcx+32], r12
+	mov qword ptr [rcx+40], r13
+	mov qword ptr [rcx+48], r14
+	mov qword ptr [rcx+56], r15
+	pop rcx
+	mulpd xmm0, xmm4
+	mulpd xmm1, xmm5
+	mulpd xmm2, xmm6
+	mulpd xmm3, xmm7
+	movapd xmmword ptr [rcx+0], xmm0
+	movapd xmmword ptr [rcx+16], xmm1
+	movapd xmmword ptr [rcx+32], xmm2
+	movapd xmmword ptr [rcx+48], xmm3

src/asm/program_prologue_linux.inc

@@ -11,10 +11,9 @@
 	push rdi                    ;# RegisterFile& registerFile
 	mov rcx, rdi
 	mov rbp, qword ptr [rsi]    ;# "mx", "ma"
-	mov eax, ebp                ;# "mx"
 	mov rdi, qword ptr [rsi+8]  ;# uint8_t* dataset
 	mov rsi, rdx                ;# convertible_t* scratchpad
 
 	#include "program_prologue_load.inc"
 
-	jmp DECL(randomx_loop_begin)
+	jmp DECL(randomx_program_loop_begin)

src/asm/program_prologue_load.inc

@@ -1,3 +1,5 @@
+	mov rax, rbp
+
 	;# zero integer registers
 	xor r8, r8
 	xor r9, r9

src/asm/program_prologue_win64.inc

@@ -23,11 +23,10 @@
 	; function arguments
 	push rcx                    ; RegisterFile& registerFile
 	mov rbp, qword ptr [rdx]    ; "mx", "ma"
-	mov eax, ebp                ; "mx"
 	mov rdi, qword ptr [rdx+8]  ; uint8_t* dataset
 	mov rsi, r8                 ; convertible_t* scratchpad
 	mov rbx, r9                 ; loop counter
 
 	include program_prologue_load.inc
 
-	jmp randomx_loop_begin
+	jmp randomx_program_loop_begin

src/asm/program_read_dataset.inc

@@ -1,4 +1,5 @@
 	xor rbp, rax                       ;# modify "mx"
+	xor eax, eax
 	and rbp, -64                       ;# align "mx" to the start of a cache line
 	mov edx, ebp                       ;# edx = mx
 	prefetchnta byte ptr [rdi+rdx]

src/asm/program_store_flt.inc

@@ -1,11 +0,0 @@
-	and eax, 1048512
-	lea rcx, [rsi+rax]
-	mulpd xmm0, xmm4
-	mulpd xmm1, xmm5
-	mulpd xmm2, xmm6
-	mulpd xmm3, xmm7
-	movapd xmmword ptr [rcx+0], xmm0
-	movapd xmmword ptr [rcx+16], xmm1
-	movapd xmmword ptr [rcx+32], xmm2
-	movapd xmmword ptr [rcx+48], xmm3
-

src/asm/program_store_int.inc

@@ -1,10 +0,0 @@
-	and eax, 1048512
-	lea rcx, [rsi+rax]
-	mov qword ptr [rcx+0], r8
-	mov qword ptr [rcx+8], r9
-	mov qword ptr [rcx+16], r10
-	mov qword ptr [rcx+24], r11
-	mov qword ptr [rcx+32], r12
-	mov qword ptr [rcx+40], r13
-	mov qword ptr [rcx+48], r14
-	mov qword ptr [rcx+56], r15

src/executeProgram-win64.asm

@@ -118,8 +118,11 @@ signMask:
 
 ALIGN 64
 program_begin:
-	xor eax, r8d                      ;# read address register 1
+	xor rax, r8                      ;# read address register 1
+	xor rax, r9
+	mov rdx, rax
 	and eax, 1048512
+	push rax
 	lea rcx, [rsi+rax]
 	xor r8,  qword ptr [rcx+0]
 	xor r9,  qword ptr [rcx+8]
@@ -129,9 +132,10 @@ program_begin:
 	xor r13, qword ptr [rcx+40]
 	xor r14, qword ptr [rcx+48]
 	xor r15, qword ptr [rcx+56]
-	xor eax, r9d                      ;# read address register 2
-	and eax, 1048512
-	lea rcx, [rsi+rax]
+	ror rdx, 32
+	and edx, 1048512
+	push rdx
+	lea rcx, [rsi+rdx]
 	cvtdq2pd xmm0, qword ptr [rcx+0]
 	cvtdq2pd xmm1, qword ptr [rcx+8]
 	cvtdq2pd xmm2, qword ptr [rcx+16]
@@ -164,9 +168,8 @@ program_begin:
 	xor r12, qword ptr [rcx+32]
 	xor r13, qword ptr [rcx+40]
 	xor r14, qword ptr [rcx+48]
-	xor r15, qword ptr [rcx+56]                 
-	mov eax, r12d                      ;# write address register 1
-	and eax, 1048512
+	xor r15, qword ptr [rcx+56]
+	pop rax
 	lea rcx, [rsi+rax]
 	mov qword ptr [rcx+0], r8
 	mov qword ptr [rcx+8], r9
@@ -176,8 +179,7 @@ program_begin:
 	mov qword ptr [rcx+40], r13
 	mov qword ptr [rcx+48], r14
 	mov qword ptr [rcx+56], r15
-	xor eax, r13d                      ;# write address register 2
-	and eax, 1048512
+	pop rax
 	lea rcx, [rsi+rax]
 	mulpd xmm0, xmm4
 	mulpd xmm1, xmm5
@@ -187,6 +189,7 @@ program_begin:
 	movapd xmmword ptr [rcx+16], xmm1
 	movapd xmmword ptr [rcx+32], xmm2
 	movapd xmmword ptr [rcx+48], xmm3
+	xor eax, eax
 	dec ebx
 	jnz program_begin