RandomX x86 JIT: remove redundant CFROUND

src/core/Miner.cpp

@@ -254,6 +254,8 @@ public:
             return strcmp(a->m_threadId, b->m_threadId) < 0;
         });
 
+        std::map<std::string, std::pair<uint32_t, double>> averageTime;
+
         for (uint32_t i = 0; i < n;)
         {
             uint32_t n1 = i;
@@ -267,19 +269,27 @@ public:
 
             for (uint32_t j = i; j < n1; ++j) {
                 ProfileScopeData* p = data[j];
+                const double t = p->m_totalCycles / p->m_totalSamples * 1e9 / ProfileScopeData::s_tscSpeed;
                 LOG_INFO("%s Thread %6s | %-30s | %7.3f%% | %9.0f ns",
                     Tags::profiler(),
                     p->m_threadId,
                     p->m_name,
                     p->m_totalCycles * 100.0 / data[i]->m_totalCycles,
-                    p->m_totalCycles / p->m_totalSamples * 1e9 / ProfileScopeData::s_tscSpeed
+                    t
                 );
+                auto& value = averageTime[p->m_name];
+                ++value.first;
+                value.second += t;
             }
 
             LOG_INFO("%s --------------|--------------------------------|----------|-------------", Tags::profiler());
 
             i = n1;
         }
+
+        for (auto& data : averageTime) {
+            LOG_INFO("%s %-30s %9.1f ns", Tags::profiler(), data.first.c_str(), data.second.second / data.second.first);
+        }
 #       endif
     }
 

src/crypto/randomx/jit_compiler_x86.cpp

@@ -164,8 +164,9 @@ namespace randomx {
 	static const uint8_t NOP6[] = { 0x66, 0x0F, 0x1F, 0x44, 0x00, 0x00 };
 	static const uint8_t NOP7[] = { 0x0F, 0x1F, 0x80, 0x00, 0x00, 0x00, 0x00 };
 	static const uint8_t NOP8[] = { 0x0F, 0x1F, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00 };
+	static const uint8_t NOP9[] = { 0x66, 0x0F, 0x1F, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00 };
 
-	static const uint8_t* NOPX[] = { NOP1, NOP2, NOP3, NOP4, NOP5, NOP6, NOP7, NOP8 };
+	static const uint8_t* NOPX[] = { NOP1, NOP2, NOP3, NOP4, NOP5, NOP6, NOP7, NOP8, NOP9 };
 
 	static const uint8_t JMP_ALIGN_PREFIX[14][16] = {
 		{},
@@ -431,6 +432,7 @@ namespace randomx {
 
 		memcpy(code + prologueSize - 48, &pcfg.eMask, sizeof(pcfg.eMask));
 		codePos = codePosFirst;
+		prevCFROUND = 0;
 
 		//mark all registers as used
 		uint64_t* r = (uint64_t*)registerUsage;
@@ -1155,6 +1157,8 @@ namespace randomx {
 		uint8_t* const p = code;
 		uint32_t pos = codePos;
 
+		prevCFROUND = 0;
+
 		const uint64_t dst = instr.dst % RegisterCountFlt;
 		const uint64_t src = instr.src % RegisterCountFlt;
 
@@ -1168,6 +1172,8 @@ namespace randomx {
 		uint8_t* const p = code;
 		uint32_t pos = codePos;
 
+		prevCFROUND = 0;
+
 		const uint32_t src = instr.src % RegistersCount;
 		const uint32_t dst = instr.dst % RegisterCountFlt;
 
@@ -1183,6 +1189,8 @@ namespace randomx {
 		uint8_t* const p = code;
 		uint32_t pos = codePos;
 
+		prevCFROUND = 0;
+
 		const uint64_t dst = instr.dst % RegisterCountFlt;
 		const uint64_t src = instr.src % RegisterCountFlt;
 
@@ -1196,6 +1204,8 @@ namespace randomx {
 		uint8_t* const p = code;
 		uint32_t pos = codePos;
 
+		prevCFROUND = 0;
+
 		const uint32_t src = instr.src % RegistersCount;
 		const uint32_t dst = instr.dst % RegisterCountFlt;
 
@@ -1222,6 +1232,8 @@ namespace randomx {
 		uint8_t* const p = code;
 		uint32_t pos = codePos;
 
+		prevCFROUND = 0;
+
 		const uint64_t dst = instr.dst % RegisterCountFlt;
 		const uint64_t src = instr.src % RegisterCountFlt;
 
@@ -1235,6 +1247,8 @@ namespace randomx {
 		uint8_t* const p = code;
 		uint32_t pos = codePos;
 
+		prevCFROUND = 0;
+
 		const uint32_t src = instr.src % RegistersCount;
 		const uint64_t dst = instr.dst % RegisterCountFlt;
 
@@ -1260,6 +1274,8 @@ namespace randomx {
 		uint8_t* const p = code;
 		uint32_t pos = codePos;
 
+		prevCFROUND = 0;
+
 		const uint32_t dst = instr.dst % RegisterCountFlt;
 
 		emit32(0xe4510f66 + (((dst << 3) + dst) << 24), p, pos);
@@ -1269,7 +1285,22 @@ namespace randomx {
 
 	void JitCompilerX86::h_CFROUND(const Instruction& instr) {
 		uint8_t* const p = code;
-		uint32_t pos = codePos;
+		uint32_t pos = prevCFROUND;
+
+		if (pos) {
+			if (vm_flags & RANDOMX_FLAG_AMD) {
+				memcpy(p + pos + 0, NOP9, 9);
+				memcpy(p + pos + 9, NOP9, 9);
+				memcpy(p + pos + 18, NOP8, 8);
+			}
+			else {
+				memcpy(p + pos + 0, NOP8, 8);
+				memcpy(p + pos + 8, NOP6, 6);
+			}
+		}
+
+		pos = codePos;
+		prevCFROUND = pos;
 
 		const uint32_t src = instr.src % RegistersCount;
 
@@ -1293,7 +1324,22 @@ namespace randomx {
 
 	void JitCompilerX86::h_CFROUND_BMI2(const Instruction& instr) {
 		uint8_t* const p = code;
-		uint32_t pos = codePos;
+		uint32_t pos = prevCFROUND;
+
+		if (pos) {
+			if (vm_flags & RANDOMX_FLAG_AMD) {
+				memcpy(p + pos + 0, NOP9, 9);
+				memcpy(p + pos + 9, NOP9, 9);
+				memcpy(p + pos + 18, NOP7, 7);
+			}
+			else {
+				memcpy(p + pos + 0, NOP8, 8);
+				memcpy(p + pos + 8, NOP5, 5);
+			}
+		}
+
+		pos = codePos;
+		prevCFROUND = pos;
 
 		const uint64_t src = instr.src % RegistersCount;
 
@@ -1319,6 +1365,8 @@ namespace randomx {
 		uint8_t* const p = code;
 		uint32_t pos = codePos;
 
+		prevCFROUND = 0;
+
 		const int reg = instr.dst % RegistersCount;
 		int32_t jmp_offset = registerUsage[reg] - (pos + 16);
 

src/crypto/randomx/jit_compiler_x86.hpp

@@ -89,6 +89,7 @@ namespace randomx {
 		uint32_t codePos = 0;
 		uint32_t codePosFirst = 0;
 		uint32_t vm_flags = 0;
+		uint32_t prevCFROUND = 0;
 
 #		ifdef XMRIG_FIX_RYZEN
 		std::pair<const void*, const void*> mainLoopBounds;