Faster unsigned division by constants (dotnet#49585)

* Faster unsigned division by constants

* Fix Arm build and add some tests.

* Improve register allocation

* Fix ARM64 codegen

* Fix MULHI flags

* Remove ARM32 codegen

* Widen 32bit UDIV to 64bit MULHI when possible. Improve register allocation.

* Always widen 32bit UDIV to 64bit MUL/MULHI

* Cleanup

* Final optimization

* Fix typo

THIRD-PARTY-NOTICES.TXT

@@ -942,3 +942,13 @@ OF SUCH DAMAGES.
 You acknowledge that this software is not designed, licensed or
 intended  for  use  in  the  design, construction, operation or
 maintenance of any nuclear facility.
+
+
+License notice for "Faster Unsigned Division by Constants"
+------------------------------
+
+Reference implementations of computing and using the "magic number" approach to dividing
+by constants, including codegen instructions. The unsigned division incorporates the
+"round down" optimization per ridiculous_fish.
+
+This is free and unencumbered software. Any copyright is dedicated to the Public Domain.

src/coreclr/jit/assertionprop.cpp

@@ -5152,8 +5152,9 @@ Compiler::fgWalkResult Compiler::optVNConstantPropCurStmt(BasicBlock* block, Sta
         case GT_INTRINSIC:
             break;
 
+        case GT_INC_SATURATE:
         case GT_MULHI:
-            assert(false && "Unexpected GT_MULHI node encountered before lowering");
+            assert(false && "Unexpected GT_INC_SATURATE/GT_MULHI node encountered before lowering");
             break;
 
         case GT_JTRUE:

src/coreclr/jit/codegen.h

@@ -843,6 +843,7 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
 
     void genCodeForDivMod(GenTreeOp* treeNode);
     void genCodeForMul(GenTreeOp* treeNode);
+    void genCodeForIncSaturate(GenTree* treeNode);
     void genCodeForMulHi(GenTreeOp* treeNode);
     void genLeaInstruction(GenTreeAddrMode* lea);
     void genSetRegToCond(regNumber dstReg, GenTree* tree);

src/coreclr/jit/codegenarm64.cpp

@@ -1753,6 +1753,28 @@ void CodeGen::genSetRegToConst(regNumber targetReg, var_types targetType, GenTre
     }
 }
 
+// Produce code for a GT_INC_SATURATE node.
+void CodeGen::genCodeForIncSaturate(GenTree* tree)
+{
+    regNumber targetReg  = tree->GetRegNum();
+    var_types targetType = tree->TypeGet();
+
+    // The arithmetic node must be sitting in a register (since it's not contained)
+    assert(!tree->isContained());
+    // The dst can only be a register.
+    assert(targetReg != REG_NA);
+
+    GenTree* operand = tree->gtGetOp1();
+    assert(!operand->isContained());
+    // The src must be a register.
+    regNumber operandReg = genConsumeReg(operand);
+
+    GetEmitter()->emitIns_R_R_I(INS_adds, emitActualTypeSize(tree), targetReg, operandReg, 1);
+    GetEmitter()->emitIns_R_R_COND(INS_cinv, emitActualTypeSize(tree), targetReg, targetReg, INS_COND_HS);
+
+    genProduceReg(tree);
+}
+
 // Generate code to get the high N bits of a N*N=2N bit multiplication result
 void CodeGen::genCodeForMulHi(GenTreeOp* treeNode)
 {

src/coreclr/jit/codegenarmarch.cpp

@@ -303,6 +303,10 @@ void CodeGen::genCodeForTreeNode(GenTree* treeNode)
 
 #ifdef TARGET_ARM64
 
+        case GT_INC_SATURATE:
+            genCodeForIncSaturate(treeNode);
+            break;
+
         case GT_MULHI:
             genCodeForMulHi(treeNode->AsOp());
             break;

src/coreclr/jit/codegenxarch.cpp

@@ -605,6 +605,27 @@ void CodeGen::genCodeForBswap(GenTree* tree)
     genProduceReg(tree);
 }
 
+// Produce code for a GT_INC_SATURATE node.
+void CodeGen::genCodeForIncSaturate(GenTree* tree)
+{
+    regNumber targetReg  = tree->GetRegNum();
+    var_types targetType = tree->TypeGet();
+
+    GenTree* operand = tree->gtGetOp1();
+    assert(operand->isUsedFromReg());
+    regNumber operandReg = genConsumeReg(operand);
+
+    if (operandReg != targetReg)
+    {
+        inst_RV_RV(INS_mov, targetReg, operandReg, targetType);
+    }
+
+    inst_RV_IV(INS_add, targetReg, 1, emitActualTypeSize(targetType));
+    inst_RV_IV(INS_sbb, targetReg, 0, emitActualTypeSize(targetType));
+
+    genProduceReg(tree);
+}
+
 // Generate code to get the high N bits of a N*N=2N bit multiplication result
 void CodeGen::genCodeForMulHi(GenTreeOp* treeNode)
 {
@@ -1608,6 +1629,10 @@ void CodeGen::genCodeForTreeNode(GenTree* treeNode)
             genCodeForIndir(treeNode->AsIndir());
             break;
 
+        case GT_INC_SATURATE:
+            genCodeForIncSaturate(treeNode);
+            break;
+
         case GT_MULHI:
 #ifdef TARGET_X86
         case GT_MUL_LONG:

src/coreclr/jit/compiler.h

@@ -10793,6 +10793,7 @@ class GenTreeVisitor
             case GT_RETFILT:
             case GT_RUNTIMELOOKUP:
             case GT_KEEPALIVE:
+            case GT_INC_SATURATE:
             {
                 GenTreeUnOp* const unOp = node->AsUnOp();
                 if (unOp->gtOp1 != nullptr)

src/coreclr/jit/compiler.hpp

@@ -4328,6 +4328,7 @@ void GenTree::VisitOperands(TVisitor visitor)
 #endif // FEATURE_ARG_SPLIT
         case GT_RETURNTRAP:
         case GT_KEEPALIVE:
+        case GT_INC_SATURATE:
             visitor(this->AsUnOp()->gtOp1);
             return;
 

src/coreclr/jit/gentree.cpp

@@ -5217,6 +5217,7 @@ bool GenTree::TryGetUse(GenTree* def, GenTree*** use)
         case GT_BSWAP:
         case GT_BSWAP16:
         case GT_KEEPALIVE:
+        case GT_INC_SATURATE:
             if (def == this->AsUnOp()->gtOp1)
             {
                 *use = &this->AsUnOp()->gtOp1;
@@ -9315,6 +9316,7 @@ GenTreeUseEdgeIterator::GenTreeUseEdgeIterator(GenTree* node)
         case GT_BSWAP:
         case GT_BSWAP16:
         case GT_KEEPALIVE:
+        case GT_INC_SATURATE:
 #if FEATURE_ARG_SPLIT
         case GT_PUTARG_SPLIT:
 #endif // FEATURE_ARG_SPLIT

src/coreclr/jit/gtlist.h

@@ -132,6 +132,7 @@ GTNODE(RSH              , GenTreeOp          ,0,GTK_BINOP)
 GTNODE(RSZ              , GenTreeOp          ,0,GTK_BINOP)
 GTNODE(ROL              , GenTreeOp          ,0,GTK_BINOP)
 GTNODE(ROR              , GenTreeOp          ,0,GTK_BINOP)
+GTNODE(INC_SATURATE     , GenTreeOp          ,0,GTK_UNOP) // saturating increment, used in division by a constant (LowerUnsignedDivOrMod)
 GTNODE(MULHI            , GenTreeOp          ,1,GTK_BINOP) // returns high bits (top N bits of the 2N bit result of an NxN multiply)
                                                            // GT_MULHI is used in division by a constant (fgMorphDivByConst). We turn
                                                            // the div into a MULHI + some adjustments. In codegen, we only use the

src/coreclr/jit/lower.cpp

@@ -5171,31 +5171,48 @@ bool Lowering::LowerUnsignedDivOrMod(GenTreeOp* divMod)
     if (!comp->opts.MinOpts() && (divisorValue >= 3))
     {
         size_t magic;
-        bool   add;
-        int    shift;
+        bool   increment;
+        int    preShift;
+        int    postShift;
+        bool   simpleMul = false;
 
         if (type == TYP_INT)
         {
-            magic = MagicDivide::GetUnsigned32Magic(static_cast<uint32_t>(divisorValue), &add, &shift);
+            magic =
+                MagicDivide::GetUnsigned32Magic(static_cast<uint32_t>(divisorValue), &increment, &preShift, &postShift);
+
+#ifdef TARGET_64BIT
+            // avoid inc_saturate/multiple shifts by widening to 32x64 MULHI
+            if (increment || (preShift
+#ifdef TARGET_XARCH
+                              // IMUL reg,reg,imm32 can't be used if magic<0 because of sign-extension
+                              && static_cast<int32_t>(magic) < 0
+#endif
+                              ))
+            {
+                magic = MagicDivide::GetUnsigned64Magic(static_cast<uint64_t>(divisorValue), &increment, &preShift,
+                                                        &postShift, 32);
+            }
+            // otherwise just widen to regular multiplication
+            else
+            {
+                postShift += 32;
+                simpleMul = true;
+            }
+#endif
         }
         else
         {
 #ifdef TARGET_64BIT
-            magic = MagicDivide::GetUnsigned64Magic(static_cast<uint64_t>(divisorValue), &add, &shift);
+            magic =
+                MagicDivide::GetUnsigned64Magic(static_cast<uint64_t>(divisorValue), &increment, &preShift, &postShift);
 #else
             unreached();
 #endif
         }
         assert(divMod->MarkedDivideByConstOptimized());
 
-        // Depending on the "add" flag returned by GetUnsignedMagicNumberForDivide we need to generate:
-        // add == false (when divisor == 3 for example):
-        //     div = (dividend MULHI magic) RSZ shift
-        // add == true (when divisor == 7 for example):
-        //     mulhi = dividend MULHI magic
-        //     div   = (((dividend SUB mulhi) RSZ 1) ADD mulhi)) RSZ (shift - 1)
-        const bool                 requiresAdjustment       = add;
-        const bool                 requiresDividendMultiuse = requiresAdjustment || !isDiv;
+        const bool                 requiresDividendMultiuse = !isDiv;
         const BasicBlock::weight_t curBBWeight              = m_block->getBBWeight(comp);
 
         if (requiresDividendMultiuse)
@@ -5204,62 +5221,107 @@ bool Lowering::LowerUnsignedDivOrMod(GenTreeOp* divMod)
             dividend = ReplaceWithLclVar(dividendUse);
         }
 
-        // Insert a new GT_MULHI node before the existing GT_UDIV/GT_UMOD node.
-        // The existing node will later be transformed into a GT_RSZ/GT_SUB that
-        // computes the final result. This way don't need to find and change the use
-        // of the existing node.
-        GenTree* mulhi = comp->gtNewOperNode(GT_MULHI, type, dividend, divisor);
-        mulhi->gtFlags |= GTF_UNSIGNED;
-        divisor->AsIntCon()->SetIconValue(magic);
-        BlockRange().InsertBefore(divMod, mulhi);
-        GenTree* firstNode = mulhi;
+        GenTree* firstNode        = nullptr;
+        GenTree* adjustedDividend = dividend;
 
-        if (requiresAdjustment)
+        // If "increment" flag is returned by GetUnsignedMagic we need to do Saturating Increment first
+        if (increment)
         {
-            dividend     = comp->gtNewLclvNode(dividend->AsLclVar()->GetLclNum(), dividend->TypeGet());
-            GenTree* sub = comp->gtNewOperNode(GT_SUB, type, dividend, mulhi);
-            BlockRange().InsertBefore(divMod, dividend, sub);
-
-            GenTree* one = comp->gtNewIconNode(1, TYP_INT);
-            GenTree* rsz = comp->gtNewOperNode(GT_RSZ, type, sub, one);
-            BlockRange().InsertBefore(divMod, one, rsz);
-
-            LIR::Use mulhiUse(BlockRange(), &sub->AsOp()->gtOp2, sub);
-            mulhi = ReplaceWithLclVar(mulhiUse);
-
-            mulhi        = comp->gtNewLclvNode(mulhi->AsLclVar()->GetLclNum(), mulhi->TypeGet());
-            GenTree* add = comp->gtNewOperNode(GT_ADD, type, rsz, mulhi);
-            BlockRange().InsertBefore(divMod, mulhi, add);
-
-            mulhi = add;
-            shift -= 1;
+            adjustedDividend = comp->gtNewOperNode(GT_INC_SATURATE, type, adjustedDividend);
+            BlockRange().InsertBefore(divMod, adjustedDividend);
+            firstNode = adjustedDividend;
+            assert(!preShift);
         }
+        // if "preShift" is required, then do a right shift before
+        else if (preShift)
+        {
+            GenTree* preShiftBy = comp->gtNewIconNode(preShift, TYP_INT);
+            adjustedDividend    = comp->gtNewOperNode(GT_RSZ, type, adjustedDividend, preShiftBy);
+            BlockRange().InsertBefore(divMod, preShiftBy, adjustedDividend);
+            firstNode = preShiftBy;
+        }
+        else if (type != TYP_I_IMPL)
+        {
+            adjustedDividend = comp->gtNewCastNode(TYP_I_IMPL, adjustedDividend, true, TYP_U_IMPL);
+            BlockRange().InsertBefore(divMod, adjustedDividend);
+            firstNode = adjustedDividend;
+        }
+
+#ifdef TARGET_XARCH
+        // force input transformation to RAX because the following MULHI will kill RDX:RAX anyway and LSRA often causes
+        // reduntant copies otherwise
+        if (firstNode && !simpleMul)
+            adjustedDividend->SetRegNum(REG_RAX);
+#endif
 
-        GenTree* shiftBy = comp->gtNewIconNode(shift, TYP_INT);
-        BlockRange().InsertBefore(divMod, shiftBy);
+        divisor->gtType = TYP_I_IMPL;
+        divisor->AsIntCon()->SetIconValue(magic);
 
-        if (isDiv)
+        if (isDiv && !postShift && type == TYP_I_IMPL)
         {
-            divMod->SetOper(GT_RSZ);
-            divMod->gtOp1 = mulhi;
-            divMod->gtOp2 = shiftBy;
+            divMod->SetOper(GT_MULHI);
+            divMod->gtOp1 = adjustedDividend;
+            divMod->gtFlags |= GTF_UNSIGNED;
         }
         else
         {
-            GenTree* div = comp->gtNewOperNode(GT_RSZ, type, mulhi, shiftBy);
+            // Insert a new GT_MULHI node before the existing GT_UDIV/GT_UMOD node.
+            // The existing node will later be transformed into a GT_RSZ/GT_SUB that
+            // computes the final result. This way don't need to find and change the use
+            // of the existing node.
+            GenTree* mulhi = comp->gtNewOperNode(simpleMul ? GT_MUL : GT_MULHI, TYP_I_IMPL, adjustedDividend, divisor);
+            mulhi->gtFlags |= GTF_UNSIGNED;
+            BlockRange().InsertBefore(divMod, mulhi);
+            if (!firstNode)
+                firstNode = mulhi;
+
+            if (postShift)
+            {
+                GenTree* shiftBy = comp->gtNewIconNode(postShift, TYP_INT);
+                BlockRange().InsertBefore(divMod, shiftBy);
 
-            // divisor UMOD dividend = dividend SUB (div MUL divisor)
-            GenTree* divisor = comp->gtNewIconNode(divisorValue, type);
-            GenTree* mul     = comp->gtNewOperNode(GT_MUL, type, div, divisor);
-            dividend         = comp->gtNewLclvNode(dividend->AsLclVar()->GetLclNum(), dividend->TypeGet());
+                if (isDiv && type == TYP_I_IMPL)
+                {
+                    divMod->SetOper(GT_RSZ);
+                    divMod->gtOp1 = mulhi;
+                    divMod->gtOp2 = shiftBy;
+                }
+                else
+                {
+                    mulhi = comp->gtNewOperNode(GT_RSZ, TYP_I_IMPL, mulhi, shiftBy);
+                    BlockRange().InsertBefore(divMod, mulhi);
+                }
+            }
+
+            if (!isDiv)
+            {
+                // divisor UMOD dividend = dividend SUB (div MUL divisor)
+                GenTree* divisor = comp->gtNewIconNode(divisorValue, type);
+                GenTree* mul     = comp->gtNewOperNode(GT_MUL, type, mulhi, divisor);
+                dividend         = comp->gtNewLclvNode(dividend->AsLclVar()->GetLclNum(), dividend->TypeGet());
 
-            divMod->SetOper(GT_SUB);
-            divMod->gtOp1 = dividend;
-            divMod->gtOp2 = mul;
+                divMod->SetOper(GT_SUB);
+                divMod->gtOp1 = dividend;
+                divMod->gtOp2 = mul;
 
-            BlockRange().InsertBefore(divMod, div, divisor, mul, dividend);
+                BlockRange().InsertBefore(divMod, divisor, mul, dividend);
+            }
+            else if (type != TYP_I_IMPL)
+            {
+#ifdef TARGET_ARMARCH
+                divMod->SetOper(GT_CAST);
+                divMod->gtFlags |= GTF_UNSIGNED;
+                divMod->AsCast()->gtCastType = TYP_UINT;
+#else
+                divMod->SetOper(GT_BITCAST);
+#endif
+                divMod->gtOp1 = mulhi;
+                divMod->gtOp2 = nullptr;
+            }
         }
-        ContainCheckRange(firstNode, divMod);
+
+        if (firstNode)
+            ContainCheckRange(firstNode, divMod);
         return true;
     }
 #endif