This patch closes PR28513: an optimization of multiplication by different constants.

It's implemented on DAG combiner level.

llvm-svn: 304209

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -80,6 +80,12 @@ static cl::opt<int> ExperimentalPrefLoopAlignment(
              " of the loop header PC will be 0)."),
     cl::Hidden);
 
+static cl::opt<bool> MulConstantOptimization(
+    "mul-constant-optimization", cl::init(true),
+    cl::desc("Replace 'mul x, Const' with more effective instructions like "
+             "SHIFT, LEA, etc."),
+    cl::Hidden);
+
 /// Call this when the user attempts to do something unsupported, like
 /// returning a double without SSE2 enabled on x86_64. This is not fatal, unlike
 /// report_fatal_error, so calling code should attempt to recover without
@@ -30928,6 +30934,75 @@ static SDValue reduceVMULWidth(SDNode *N, SelectionDAG &DAG,
   }
 }
 
+static SDValue combineMulSpecial(uint64_t MulAmt, SDNode *N, SelectionDAG &DAG,
+                                 EVT VT, SDLoc DL) {
+
+  auto combineMulShlAddOrSub = [&](int Mult, int Shift, bool isAdd) {
+    SDValue Result = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
+                                 DAG.getConstant(Mult, DL, VT));
+    Result = DAG.getNode(ISD::SHL, DL, VT, Result,
+                         DAG.getConstant(Shift, DL, MVT::i8));
+    Result = DAG.getNode(isAdd ? ISD::ADD : ISD::SUB, DL, VT, N->getOperand(0),
+                         Result);
+    return Result;
+  };
+
+  auto combineMulMulAddOrSub = [&](bool isAdd) {
+    SDValue Result = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
+                                 DAG.getConstant(9, DL, VT));
+    Result = DAG.getNode(ISD::MUL, DL, VT, Result, DAG.getConstant(3, DL, VT));
+    Result = DAG.getNode(isAdd ? ISD::ADD : ISD::SUB, DL, VT, N->getOperand(0),
+                         Result);
+    return Result;
+  };
+
+  switch (MulAmt) {
+  default:
+    break;
+  case 11:
+    // mul x, 11 => add ((shl (mul x, 5), 1), x)
+    return combineMulShlAddOrSub(5, 1, /*isAdd*/ true);
+  case 21:
+    // mul x, 21 => add ((shl (mul x, 5), 2), x)
+    return combineMulShlAddOrSub(5, 2, /*isAdd*/ true);
+  case 22:
+    // mul x, 22 => add (add ((shl (mul x, 5), 2), x), x)
+    return DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0),
+                       combineMulShlAddOrSub(5, 2, /*isAdd*/ true));
+  case 19:
+    // mul x, 19 => sub ((shl (mul x, 5), 2), x)
+    return combineMulShlAddOrSub(5, 2, /*isAdd*/ false);
+  case 13:
+    // mul x, 13 => add ((shl (mul x, 3), 2), x)
+    return combineMulShlAddOrSub(3, 2, /*isAdd*/ true);
+  case 23:
+    // mul x, 13 => sub ((shl (mul x, 3), 3), x)
+    return combineMulShlAddOrSub(3, 3, /*isAdd*/ false);
+  case 14:
+    // mul x, 14 => add (add ((shl (mul x, 3), 2), x), x)
+    return DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0),
+                       combineMulShlAddOrSub(3, 2, /*isAdd*/ true));
+  case 26:
+    // mul x, 26 => sub ((mul (mul x, 9), 3), x)
+    return combineMulMulAddOrSub(/*isAdd*/ false);
+  case 28:
+    // mul x, 28 => add ((mul (mul x, 9), 3), x)
+    return combineMulMulAddOrSub(/*isAdd*/ true);
+  case 29:
+    // mul x, 29 => add (add ((mul (mul x, 9), 3), x), x)
+    return DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0),
+                       combineMulMulAddOrSub(/*isAdd*/ true));
+  case 30:
+    // mul x, 30 => sub (sub ((shl x, 5), x), x)
+    return DAG.getNode(
+        ISD::SUB, DL, VT, N->getOperand(0),
+        DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0),
+                    DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
+                                DAG.getConstant(5, DL, MVT::i8))));
+  }
+  return SDValue();
+}
+
 /// Optimize a single multiply with constant into two operations in order to
 /// implement it with two cheaper instructions, e.g. LEA + SHL, LEA + LEA.
 static SDValue combineMul(SDNode *N, SelectionDAG &DAG,
@@ -30937,6 +31012,8 @@ static SDValue combineMul(SDNode *N, SelectionDAG &DAG,
   if (DCI.isBeforeLegalize() && VT.isVector())
     return reduceVMULWidth(N, DAG, Subtarget);
 
+  if (!MulConstantOptimization)
+    return SDValue();
   // An imul is usually smaller than the alternative sequence.
   if (DAG.getMachineFunction().getFunction()->optForMinSize())
     return SDValue();
@@ -30992,7 +31069,8 @@ static SDValue combineMul(SDNode *N, SelectionDAG &DAG,
     else
       NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
                            DAG.getConstant(MulAmt2, DL, VT));
-  }
+  } else if (!Subtarget.slowLEA())
+    NewMul = combineMulSpecial(MulAmt, N, DAG, VT, DL);
 
   if (!NewMul) {
     assert(MulAmt != 0 &&

llvm/test/CodeGen/X86/mul-constant-i16.ll

@@ -188,13 +188,16 @@ define i16 @test_mul_by_11(i16 %x) {
 ; X86-LABEL: test_mul_by_11:
 ; X86:       # BB#0:
 ; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $11, %eax, %eax
+; X86-NEXT:    leal (%eax,%eax,4), %ecx
+; X86-NEXT:    leal (%eax,%ecx,2), %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_11:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $11, %edi, %eax
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal (%rdi,%rdi,4), %eax
+; X64-NEXT:    leal (%rdi,%rax,2), %eax
 ; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 11
@@ -225,13 +228,16 @@ define i16 @test_mul_by_13(i16 %x) {
 ; X86-LABEL: test_mul_by_13:
 ; X86:       # BB#0:
 ; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $13, %eax, %eax
+; X86-NEXT:    leal (%eax,%eax,2), %ecx
+; X86-NEXT:    leal (%eax,%ecx,4), %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_13:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $13, %edi, %eax
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal (%rdi,%rdi,2), %eax
+; X64-NEXT:    leal (%rdi,%rax,4), %eax
 ; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 13
@@ -241,14 +247,19 @@ define i16 @test_mul_by_13(i16 %x) {
 define i16 @test_mul_by_14(i16 %x) {
 ; X86-LABEL: test_mul_by_14:
 ; X86:       # BB#0:
-; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $14, %eax, %eax
+; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    leal (%ecx,%ecx,2), %eax
+; X86-NEXT:    leal (%ecx,%eax,4), %eax
+; X86-NEXT:    addl %ecx, %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_14:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $14, %edi, %eax
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal (%rdi,%rdi,2), %eax
+; X64-NEXT:    leal (%rdi,%rax,4), %eax
+; X64-NEXT:    addl %edi, %eax
 ; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 14
@@ -338,14 +349,19 @@ define i16 @test_mul_by_19(i16 %x) {
 ; X86-LABEL: test_mul_by_19:
 ; X86:       # BB#0:
 ; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $19, %eax, %eax
+; X86-NEXT:    leal (%eax,%eax,4), %ecx
+; X86-NEXT:    shll $2, %ecx
+; X86-NEXT:    subl %ecx, %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_19:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $19, %edi, %eax
-; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal (%rdi,%rdi,4), %eax
+; X64-NEXT:    shll $2, %eax
+; X64-NEXT:    subl %eax, %edi
+; X64-NEXT:    movl %edi, %eax
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 19
   ret i16 %mul
@@ -375,13 +391,16 @@ define i16 @test_mul_by_21(i16 %x) {
 ; X86-LABEL: test_mul_by_21:
 ; X86:       # BB#0:
 ; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $21, %eax, %eax
+; X86-NEXT:    leal (%eax,%eax,4), %ecx
+; X86-NEXT:    leal (%eax,%ecx,4), %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_21:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $21, %edi, %eax
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal (%rdi,%rdi,4), %eax
+; X64-NEXT:    leal (%rdi,%rax,4), %eax
 ; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 21
@@ -391,14 +410,19 @@ define i16 @test_mul_by_21(i16 %x) {
 define i16 @test_mul_by_22(i16 %x) {
 ; X86-LABEL: test_mul_by_22:
 ; X86:       # BB#0:
-; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $22, %eax, %eax
+; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    leal (%ecx,%ecx,4), %eax
+; X86-NEXT:    leal (%ecx,%eax,4), %eax
+; X86-NEXT:    addl %ecx, %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_22:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $22, %edi, %eax
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal (%rdi,%rdi,4), %eax
+; X64-NEXT:    leal (%rdi,%rax,4), %eax
+; X64-NEXT:    addl %edi, %eax
 ; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 22
@@ -409,14 +433,19 @@ define i16 @test_mul_by_23(i16 %x) {
 ; X86-LABEL: test_mul_by_23:
 ; X86:       # BB#0:
 ; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $23, %eax, %eax
+; X86-NEXT:    leal (%eax,%eax,2), %ecx
+; X86-NEXT:    shll $3, %ecx
+; X86-NEXT:    subl %ecx, %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_23:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $23, %edi, %eax
-; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal (%rdi,%rdi,2), %eax
+; X64-NEXT:    shll $3, %eax
+; X64-NEXT:    subl %eax, %edi
+; X64-NEXT:    movl %edi, %eax
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 23
   ret i16 %mul
@@ -466,14 +495,19 @@ define i16 @test_mul_by_26(i16 %x) {
 ; X86-LABEL: test_mul_by_26:
 ; X86:       # BB#0:
 ; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $26, %eax, %eax
+; X86-NEXT:    leal (%eax,%eax,8), %ecx
+; X86-NEXT:    leal (%ecx,%ecx,2), %ecx
+; X86-NEXT:    subl %ecx, %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_26:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $26, %edi, %eax
-; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal (%rdi,%rdi,8), %eax
+; X64-NEXT:    leal (%rax,%rax,2), %eax
+; X64-NEXT:    subl %eax, %edi
+; X64-NEXT:    movl %edi, %eax
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 26
   ret i16 %mul
@@ -502,14 +536,19 @@ define i16 @test_mul_by_27(i16 %x) {
 define i16 @test_mul_by_28(i16 %x) {
 ; X86-LABEL: test_mul_by_28:
 ; X86:       # BB#0:
-; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $28, %eax, %eax
+; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    leal (%ecx,%ecx,8), %eax
+; X86-NEXT:    leal (%eax,%eax,2), %eax
+; X86-NEXT:    addl %ecx, %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_28:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $28, %edi, %eax
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal (%rdi,%rdi,8), %eax
+; X64-NEXT:    leal (%rax,%rax,2), %eax
+; X64-NEXT:    addl %edi, %eax
 ; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 28
@@ -519,14 +558,21 @@ define i16 @test_mul_by_28(i16 %x) {
 define i16 @test_mul_by_29(i16 %x) {
 ; X86-LABEL: test_mul_by_29:
 ; X86:       # BB#0:
-; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $29, %eax, %eax
+; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    leal (%ecx,%ecx,8), %eax
+; X86-NEXT:    leal (%eax,%eax,2), %eax
+; X86-NEXT:    addl %ecx, %eax
+; X86-NEXT:    addl %ecx, %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_29:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $29, %edi, %eax
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal (%rdi,%rdi,8), %eax
+; X64-NEXT:    leal (%rax,%rax,2), %eax
+; X64-NEXT:    addl %edi, %eax
+; X64-NEXT:    addl %edi, %eax
 ; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 29
@@ -537,14 +583,22 @@ define i16 @test_mul_by_30(i16 %x) {
 ; X86-LABEL: test_mul_by_30:
 ; X86:       # BB#0:
 ; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    imull $30, %eax, %eax
+; X86-NEXT:    movl %eax, %ecx
+; X86-NEXT:    shll $5, %ecx
+; X86-NEXT:    movl %eax, %edx
+; X86-NEXT:    subl %ecx, %edx
+; X86-NEXT:    subl %edx, %eax
 ; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mul_by_30:
 ; X64:       # BB#0:
-; X64-NEXT:    imull $30, %edi, %eax
-; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-NEXT:    movl %edi, %eax
+; X64-NEXT:    shll $5, %eax
+; X64-NEXT:    movl %edi, %ecx
+; X64-NEXT:    subl %eax, %ecx
+; X64-NEXT:    subl %ecx, %edi
+; X64-NEXT:    movl %edi, %eax
 ; X64-NEXT:    retq
   %mul = mul nsw i16 %x, 30
   ret i16 %mul
@@ -587,3 +641,30 @@ define i16 @test_mul_by_32(i16 %x) {
   %mul = mul nsw i16 %x, 32
   ret i16 %mul
 }
+
+; (x*9+42)*(x*5+2)
+define i16 @test_mul_spec(i16 %x) nounwind {
+; X86-LABEL: test_mul_spec:
+; X86:       # BB#0:
+; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    leal 42(%eax,%eax,8), %ecx
+; X86-NEXT:    leal 2(%eax,%eax,4), %eax
+; X86-NEXT:    imull %ecx, %eax
+; X86-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_mul_spec:
+; X64:       # BB#0:
+; X64-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
+; X64-NEXT:    leal 42(%rdi,%rdi,8), %ecx
+; X64-NEXT:    leal 2(%rdi,%rdi,4), %eax
+; X64-NEXT:    imull %ecx, %eax
+; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-NEXT:    retq
+  %mul = mul nsw i16 %x, 9
+  %add = add nsw i16 %mul, 42
+  %mul2 = mul nsw i16 %x, 5
+  %add2 = add nsw i16 %mul2, 2
+  %mul3 = mul nsw i16 %add, %add2
+  ret i16 %mul3
+}