From 1d9e50cb4735df46d3de0cee5791e97295eaf588 Mon Sep 17 00:00:00 2001
From: Egor Bogatov <egorbo@gmail.com>
Date: Wed, 30 Sep 2020 02:35:44 +0300
Subject: [PATCH] [RyuJIT] Add "rorx" instruction (BMI2) and emit it instead of
 "rol" when possible (#41772)

* Use rorx instead of rol when possible
---
 src/coreclr/src/jit/codegenxarch.cpp | 22 ++++++++++++++++++----
 src/coreclr/src/jit/emitxarch.cpp    |  4 ++++
 src/coreclr/src/jit/instrsxarch.h    |  1 +
 3 files changed, 23 insertions(+), 4 deletions(-)

diff --git a/src/coreclr/src/jit/codegenxarch.cpp b/src/coreclr/src/jit/codegenxarch.cpp
index 6bca839d10514b..3bcda89f437d15 100644
--- a/src/coreclr/src/jit/codegenxarch.cpp
+++ b/src/coreclr/src/jit/codegenxarch.cpp
@@ -4080,10 +4080,11 @@ void CodeGen::genCodeForShift(GenTree* tree)
 
     if (shiftBy->isContainedIntOrIImmed())
     {
+        emitAttr size = emitTypeSize(tree);
+
         // Optimize "X<<1" to "lea [reg+reg]" or "add reg, reg"
         if (tree->OperIs(GT_LSH) && !tree->gtOverflowEx() && !tree->gtSetFlags() && shiftBy->IsIntegralConst(1))
         {
-            emitAttr size = emitTypeSize(tree);
             if (tree->GetRegNum() == operandReg)
             {
                 GetEmitter()->emitIns_R_R(INS_add, size, tree->GetRegNum(), operandReg);
@@ -4095,6 +4096,21 @@ void CodeGen::genCodeForShift(GenTree* tree)
         }
         else
         {
+            int shiftByValue = (int)shiftBy->AsIntConCommon()->IconValue();
+
+#if defined(TARGET_64BIT)
+            // Try to emit rorx if BMI2 is available instead of mov+rol
+            // it makes sense only for 64bit integers
+            if ((genActualType(targetType) == TYP_LONG) && (tree->GetRegNum() != operandReg) &&
+                compiler->compOpportunisticallyDependsOn(InstructionSet_BMI2) && tree->OperIs(GT_ROL, GT_ROR) &&
+                (shiftByValue > 0) && (shiftByValue < 64))
+            {
+                const int value = tree->OperIs(GT_ROL) ? (64 - shiftByValue) : shiftByValue;
+                GetEmitter()->emitIns_R_R_I(INS_rorx, size, tree->GetRegNum(), operandReg, value);
+                genProduceReg(tree);
+                return;
+            }
+#endif
             // First, move the operand to the destination register and
             // later on perform the shift in-place.
             // (LSRA will try to avoid this situation through preferencing.)
@@ -4102,9 +4118,7 @@ void CodeGen::genCodeForShift(GenTree* tree)
             {
                 inst_RV_RV(INS_mov, tree->GetRegNum(), operandReg, targetType);
             }
-
-            int shiftByValue = (int)shiftBy->AsIntConCommon()->IconValue();
-            inst_RV_SH(ins, emitTypeSize(tree), tree->GetRegNum(), shiftByValue);
+            inst_RV_SH(ins, size, tree->GetRegNum(), shiftByValue);
         }
     }
     else
diff --git a/src/coreclr/src/jit/emitxarch.cpp b/src/coreclr/src/jit/emitxarch.cpp
index 7248dcb106aa6f..58af8ad5306d58 100644
--- a/src/coreclr/src/jit/emitxarch.cpp
+++ b/src/coreclr/src/jit/emitxarch.cpp
@@ -524,6 +524,7 @@ bool TakesRexWPrefix(instruction ins, emitAttr attr)
             case INS_mulx:
             case INS_pdep:
             case INS_pext:
+            case INS_rorx:
                 return true;
             default:
                 return false;
@@ -758,6 +759,7 @@ unsigned emitter::emitOutputRexOrVexPrefixIfNeeded(instruction ins, BYTE* dst, c
                         {
                             switch (ins)
                             {
+                                case INS_rorx:
                                 case INS_pdep:
                                 case INS_mulx:
                                 {
@@ -1242,6 +1244,7 @@ bool emitter::emitInsCanOnlyWriteSSE2OrAVXReg(instrDesc* id)
         case INS_pextrq:
         case INS_pextrw:
         case INS_pextrw_sse41:
+        case INS_rorx:
         {
             // These SSE instructions write to a general purpose integer register.
             return false;
@@ -14944,6 +14947,7 @@ emitter::insExecutionCharacteristics emitter::getInsExecutionCharacteristics(ins
         case INS_tzcnt:
         case INS_popcnt:
         case INS_crc32:
+        case INS_rorx:
         case INS_pdep:
         case INS_pext:
         case INS_addsubps:
diff --git a/src/coreclr/src/jit/instrsxarch.h b/src/coreclr/src/jit/instrsxarch.h
index 60616c5177b58f..c6a50690f70290 100644
--- a/src/coreclr/src/jit/instrsxarch.h
+++ b/src/coreclr/src/jit/instrsxarch.h
@@ -594,6 +594,7 @@ INST3(blsr,             "blsr",             IUM_WR, BAD_CODE,     BAD_CODE,
 INST3(bextr,            "bextr",            IUM_WR, BAD_CODE,     BAD_CODE,     SSE38(0xF7),                             INS_Flags_IsDstDstSrcAVXInstruction)    // Bit Field Extract
 
 // BMI2
+INST3(rorx,             "rorx",             IUM_WR, BAD_CODE,     BAD_CODE,     SSE3A(0xF0),                             INS_FLAGS_None)
 INST3(pdep,             "pdep",             IUM_WR, BAD_CODE,     BAD_CODE,     SSE38(0xF5),                             INS_Flags_IsDstDstSrcAVXInstruction)    // Parallel Bits Deposit
 INST3(pext,             "pext",             IUM_WR, BAD_CODE,     BAD_CODE,     SSE38(0xF5),                             INS_Flags_IsDstDstSrcAVXInstruction)    // Parallel Bits Extract
 INST3(bzhi,             "bzhi",             IUM_WR, BAD_CODE,     BAD_CODE,     SSE38(0xF5),                             INS_Flags_IsDstDstSrcAVXInstruction)    // Zero High Bits Starting with Specified Bit Position