1. Hoist minus sign as high as possible in an attempt to reveal

   some optimization opportunities (in the enclosing supper-expressions).

   rule 1. (-0.0 - X ) * Y => -0.0 - (X * Y)
     if expression "-0.0 - X" has only one reference.

   rule 2. (0.0 - X ) * Y => -0.0 - (X * Y)
     if expression "0.0 - X" has only one reference, and
        the instruction is marked "noSignedZero".

2. Eliminate negation (The compiler was already able to handle these
    opt if the 0.0s are replaced with -0.0.)

   rule 3: (0.0 - X) * (0.0 - Y) => X * Y
   rule 4: (0.0 - X) * C => X * -C
   if the expr is flagged "noSignedZero".

3. 
  Rule 5: (X*Y) * X => (X*X) * Y
   if X!=Y and the expression is flagged with "UnsafeAlgebra".

   The purpose of this transformation is two-fold:
    a) to form a power expression (of X).
    b) potentially shorten the critical path: After transformation, the
       latency of the instruction Y is amortized by the expression of X*X,
       and therefore Y is in a "less critical" position compared to what it
      was before the transformation. 

4. Remove the InstCombine code about simplifiying "X * select".
   
   The reasons are following:
    a) The "select" is somewhat architecture-dependent, therefore the
       higher level optimizers are not able to precisely predict if
       the simplification really yields any performance improvement
       or not.

    b) The "select" operator is bit complicate, and tends to obscure
       optimization opportunities. It is btter to keep it as low as
       possible in expr tree, and let CodeGen to tackle the optimization.

llvm-svn: 172551

3 files changed, 152 insertions, 64 deletions

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
index 29846c156c4..8e4267f8986 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -377,6 +377,8 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
   if (Value *V = SimplifyFMulInst(Op0, Op1, I.getFastMathFlags(), TD))
     return ReplaceInstUsesWith(I, V);
 
+  bool AllowReassociate = I.hasUnsafeAlgebra();
+
   // Simplify mul instructions with a constant RHS.
   if (isa<Constant>(Op1)) {
     // Try to fold constant mul into select arguments.
@@ -389,7 +391,7 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
         return NV;
 
     ConstantFP *C = dyn_cast<ConstantFP>(Op1);
-    if (C && I.hasUnsafeAlgebra() && C->getValueAPF().isNormal()) {
+    if (C && AllowReassociate && C->getValueAPF().isNormal()) {
       // Let MDC denote an expression in one of these forms:
       // X * C, C/X, X/C, where C is a constant.
       //
@@ -430,7 +432,7 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
                         BinaryOperator::CreateFAdd(M0, M1) :
                         BinaryOperator::CreateFSub(M0, M1);
             Instruction *RI = cast<Instruction>(R);
-            RI->setHasUnsafeAlgebra(true);
+            RI->copyFastMathFlags(&I);
             return RI;
           }
         }
@@ -438,9 +440,6 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
     }
   }
 
-  if (Value *Op0v = dyn_castFNegVal(Op0))     // -X * -Y = X*Y
-    if (Value *Op1v = dyn_castFNegVal(Op1))
-      return BinaryOperator::CreateFMul(Op0v, Op1v);
 
   // Under unsafe algebra do:
   // X * log2(0.5*Y) = X*log2(Y) - X
@@ -469,36 +468,66 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
     }
   }
 
-  // X * cond ? 1.0 : 0.0 => cond ? X : 0.0
-  if (I.hasNoNaNs() && I.hasNoSignedZeros()) {
-    Value *V0 = I.getOperand(0);
-    Value *V1 = I.getOperand(1);
-    Value *Cond, *SLHS, *SRHS;
-    bool Match = false;
-
-    if (match(V0, m_Select(m_Value(Cond), m_Value(SLHS), m_Value(SRHS)))) {
-      Match = true;
-    } else if (match(V1, m_Select(m_Value(Cond), m_Value(SLHS), 
-                     m_Value(SRHS)))) {
-      Match = true;
-      std::swap(V0, V1);
+  // Handle symmetric situation in a 2-iteration loop
+  Value *Opnd0 = Op0;
+  Value *Opnd1 = Op1;
+  for (int i = 0; i < 2; i++) {
+    bool IgnoreZeroSign = I.hasNoSignedZeros();
+    if (BinaryOperator::isFNeg(Opnd0, IgnoreZeroSign)) {
+      Value *N0 = dyn_castFNegVal(Opnd0, IgnoreZeroSign);
+      Value *N1 = dyn_castFNegVal(Opnd1, IgnoreZeroSign);
+
+      // -X * -Y => X*Y
+      if (N1)
+        return BinaryOperator::CreateFMul(N0, N1);
+
+      if (Opnd0->hasOneUse()) {
+        // -X * Y => -(X*Y) (Promote negation as high as possible)
+        Value *T = Builder->CreateFMul(N0, Opnd1);
+        cast<Instruction>(T)->setDebugLoc(I.getDebugLoc());
+        Instruction *Neg = BinaryOperator::CreateFNeg(T);
+        if (I.getFastMathFlags().any()) {
+          cast<Instruction>(T)->copyFastMathFlags(&I);
+          Neg->copyFastMathFlags(&I);
+        }
+        return Neg;
+      }
     }
 
-    if (Match) {
-      ConstantFP *C0 = dyn_cast<ConstantFP>(SLHS);
-      ConstantFP *C1 = dyn_cast<ConstantFP>(SRHS);
-
-      if (C0 && C1 &&
-          ((C0->isZero() && C1->isExactlyValue(1.0)) ||
-           (C1->isZero() && C0->isExactlyValue(1.0)))) {
-        Value *T;
-        if (C0->isZero())
-          T = Builder->CreateSelect(Cond, SLHS, V1);
-        else
-          T = Builder->CreateSelect(Cond, V1, SRHS);
-        return ReplaceInstUsesWith(I, T);
+    // (X*Y) * X => (X*X) * Y where Y != X
+    //  The purpose is two-fold: 
+    //   1) to form a power expression (of X).
+    //   2) potentially shorten the critical path: After transformation, the
+    //  latency of the instruction Y is amortized by the expression of X*X,
+    //  and therefore Y is in a "less critical" position compared to what it
+    //  was before the transformation.
+    //
+    if (AllowReassociate) {
+      Value *Opnd0_0, *Opnd0_1;
+      if (Opnd0->hasOneUse() &&
+          match(Opnd0, m_FMul(m_Value(Opnd0_0), m_Value(Opnd0_1)))) {
+        Value *Y = 0;
+        if (Opnd0_0 == Opnd1 && Opnd0_1 != Opnd1)
+          Y = Opnd0_1;
+        else if (Opnd0_1 == Opnd1 && Opnd0_0 != Opnd1)
+          Y = Opnd0_0;
+
+        if (Y) {
+          Instruction *T = cast<Instruction>(Builder->CreateFMul(Opnd1, Opnd1));
+          T->copyFastMathFlags(&I);
+          T->setDebugLoc(I.getDebugLoc());
+
+          Instruction *R = BinaryOperator::CreateFMul(T, Y);
+          R->copyFastMathFlags(&I);
+          return R;
+        }
       }
     }
+
+    if (!isa<Constant>(Op1))
+      std::swap(Opnd0, Opnd1);
+    else
+      break;
   }
 
   return Changed ? &I : 0;
diff --git a/llvm/test/Transforms/InstCombine/fast-math.ll b/llvm/test/Transforms/InstCombine/fast-math.ll
index 5a1ad5e6dbb..88e19e343a1 100644
--- a/llvm/test/Transforms/InstCombine/fast-math.ll
+++ b/llvm/test/Transforms/InstCombine/fast-math.ll
@@ -130,37 +130,6 @@ define double @fail2(double %f1, double %f2) {
 ; CHECK: ret
 }
 
-; rdar://12753946:  x * cond ? 1.0 : 0.0 => cond ? x : 0.0
-define double @select1(i32 %cond, double %x, double %y) {
-  %tobool = icmp ne i32 %cond, 0
-  %cond1 = select i1 %tobool, double 1.000000e+00, double 0.000000e+00
-  %mul = fmul nnan nsz double %cond1, %x
-  %add = fadd double %mul, %y
-  ret double %add
-; CHECK: @select1
-; CHECK: select i1 %tobool, double %x, double 0.000000e+00
-}
-
-define double @select2(i32 %cond, double %x, double %y) {
-  %tobool = icmp ne i32 %cond, 0
-  %cond1 = select i1 %tobool, double 0.000000e+00, double 1.000000e+00
-  %mul = fmul nnan nsz double %cond1, %x
-  %add = fadd double %mul, %y
-  ret double %add
-; CHECK: @select2
-; CHECK: select i1 %tobool, double 0.000000e+00, double %x
-}
-
-define double @select3(i32 %cond, double %x, double %y) {
-  %tobool = icmp ne i32 %cond, 0
-  %cond1 = select i1 %tobool, double 0.000000e+00, double 2.000000e+00
-  %mul = fmul nnan nsz double %cond1, %x
-  %add = fadd double %mul, %y
-  ret double %add
-; CHECK: @select3
-; CHECK: fmul nnan nsz double %cond1, %x
-}
-
 ; =========================================================================
 ;
 ;   Testing-cases about fmul begin
@@ -243,6 +212,25 @@ define float @fmul5(float %f1, float %f2) {
 ; CHECK: fdiv fast float %f1, 0x47E8000000000000
 }
 
+; (X*Y) * X => (X*X) * Y
+define float @fmul6(float %f1, float %f2) {
+  %mul = fmul float %f1, %f2
+  %mul1 = fmul fast float %mul, %f1
+  ret float %mul1
+; CHECK: @fmul6
+; CHECK: fmul fast float %f1, %f1 
+}
+
+; "(X*Y) * X => (X*X) * Y" is disabled if "X*Y" has multiple uses
+define float @fmul7(float %f1, float %f2) {
+  %mul = fmul float %f1, %f2
+  %mul1 = fmul fast float %mul, %f1
+  %add = fadd float %mul1, %mul
+  ret float %add
+; CHECK: @fmul7
+; CHECK: fmul fast float %mul, %f1
+}
+
 ; =========================================================================
 ;
 ;   Testing-cases about negation
@@ -262,8 +250,8 @@ define float @fneg1(float %f1, float %f2) {
 ;   Testing-cases about div
 ;
 ; =========================================================================
-; X/C1 / C2 => X * (1/(C2*C1))
 
+; X/C1 / C2 => X * (1/(C2*C1))
 define float @fdiv1(float %x) {
   %div = fdiv float %x, 0x3FF3333340000000
   %div1 = fdiv fast float %div, 0x4002666660000000
@@ -351,4 +339,3 @@ define float @fdiv9(float %x) {
 ; CHECK: @fdiv9
 ; CHECK: fmul fast float %x, 5.000000e+00
 }
-
diff --git a/llvm/test/Transforms/InstCombine/fmul.ll b/llvm/test/Transforms/InstCombine/fmul.ll
new file mode 100644
index 00000000000..3671b4c6991
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/fmul.ll
@@ -0,0 +1,72 @@
+; RUN: opt -S -instcombine < %s | FileCheck %s
+
+; (-0.0 - X) * C => X * -C
+define float @test1(float %x) {
+  %sub = fsub float -0.000000e+00, %x
+  %mul = fmul float %sub, 2.0e+1
+  ret float %mul
+
+; CHECK: @test1
+; CHECK: fmul float %x, -2.000000e+01
+}
+
+; (0.0 - X) * C => X * -C
+define float @test2(float %x) {
+  %sub = fsub nsz float 0.000000e+00, %x
+  %mul = fmul float %sub, 2.0e+1
+  ret float %mul
+
+; CHECK: @test2
+; CHECK: fmul float %x, -2.000000e+01
+}
+
+; (-0.0 - X) * (-0.0 - Y) => X * Y
+define float @test3(float %x, float %y) {
+  %sub1 = fsub float -0.000000e+00, %x
+  %sub2 = fsub float -0.000000e+00, %y
+  %mul = fmul float %sub1, %sub2
+  ret float %mul
+; CHECK: @test3
+; CHECK: fmul float %x, %y
+}
+
+; (0.0 - X) * (0.0 - Y) => X * Y
+define float @test4(float %x, float %y) {
+  %sub1 = fsub nsz float 0.000000e+00, %x
+  %sub2 = fsub nsz float 0.000000e+00, %y
+  %mul = fmul float %sub1, %sub2
+  ret float %mul
+; CHECK: @test4
+; CHECK: fmul float %x, %y
+}
+
+; (-0.0 - X) * Y => -0.0 - (X * Y)
+define float @test5(float %x, float %y) {
+  %sub1 = fsub float -0.000000e+00, %x
+  %mul = fmul float %sub1, %y
+  ret float %mul
+; CHECK: @test5
+; CHECK: %1 = fmul float %x, %y
+; CHECK: %mul = fsub float -0.000000e+00, %1
+}
+
+; (0.0 - X) * Y => 0.0 - (X * Y)
+define float @test6(float %x, float %y) {
+  %sub1 = fsub nsz float 0.000000e+00, %x
+  %mul = fmul float %sub1, %y
+  ret float %mul
+; CHECK: @test6
+; CHECK: %1 = fmul float %x, %y
+; CHECK: %mul = fsub float -0.000000e+00, %1
+}
+
+; "(-0.0 - X) * Y => -0.0 - (X * Y)" is disabled if expression "-0.0 - X"
+; has multiple uses.
+define float @test7(float %x, float %y) {
+  %sub1 = fsub float -0.000000e+00, %x
+  %mul = fmul float %sub1, %y
+  %mul2 = fmul float %mul, %sub1
+  ret float %mul2
+; CHECK: @test7
+; CHECK: fsub float -0.000000e+00, %x
+}