[PATCH][DAGCombine] Fix bug in DAGCombine

Shuxin Yang shuxin.llvm at gmail.com
Fri Feb 1 16:09:34 PST 2013


Hi,

     rdar://13126763.

     The attached patch fixes an inadvertent bug in DAG-combine. This 
bug  was
  introduced by r162956 back to Aug 2012.  The symptom is to mistakenly
optimize "x + x*x" into "x * 3.0".   Following pseudo-code illustrate 
the mistaken
logic. The fix is quite straightforward.

    -----------------------
DAGCombiner::visitFADD(SDNode *N) {

    Let N0, N1 be the 1st and 2nd operand of N;
    if (N0 is fmul) {
         ...
         !!!! The sub-expr must be fmul, Yuck !!!!!
        // (fadd x, (fadd x, x)) -> (fmul 3.0, x)
    }
   -----------------------------

Thanks
Shuxin


-------------- next part --------------
Index: test/CodeGen/X86/dagcombine_unsafe_math.ll
===================================================================
--- test/CodeGen/X86/dagcombine_unsafe_math.ll	(revision 0)
+++ test/CodeGen/X86/dagcombine_unsafe_math.ll	(revision 0)
@@ -0,0 +1,42 @@
+; RUN: llc < %s -enable-unsafe-fp-math -mtriple=x86_64-apple-darwin -mcpu=corei7-avx | FileCheck %s 
+
+
+; rdar://13126763
+; Expression "x + x*x" was mistakenly transformed into "x * 3.0f".
+
+define float @test1(float %x) {
+  %t1 = fmul fast float %x, %x
+  %t2 = fadd fast float %t1, %x
+  ret float %t2
+; CHECK: test1
+; CHECK: vaddss
+}
+
+; (x + x) + x => x * 3.0
+define float @test2(float %x) {
+  %t1 = fadd fast float %x, %x
+  %t2 = fadd fast float %t1, %x
+  ret float %t2
+; CHECK: .long  1077936128
+; CHECK: test2
+; CHECK: vmulss LCPI1_0(%rip), %xmm0, %xmm0
+}
+
+; x + (x + x) => x * 3.0
+define float @test3(float %x) {
+  %t1 = fadd fast float %x, %x
+  %t2 = fadd fast float %t1, %x
+  ret float %t2
+; CHECK: .long  1077936128
+; CHECK: test3
+; CHECK: vmulss LCPI2_0(%rip), %xmm0, %xmm0
+}
+
+; (y + x) + x != x * 3.0
+define float @test4(float %x, float %y) {
+  %t1 = fadd fast float %x, %y
+  %t2 = fadd fast float %t1, %x
+  ret float %t2
+; CHECK: test4
+; CHECK: vaddss
+}
Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/DAGCombiner.cpp	(revision 174228)
+++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp	(working copy)
@@ -5832,13 +5832,6 @@
                            N1, NewCFP);
       }
 
-      // (fadd (fadd x, x), x) -> (fmul 3.0, x)
-      if (!CFP00 && !CFP01 && N0.getOperand(0) == N0.getOperand(1) &&
-          N0.getOperand(0) == N1) {
-        return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
-                           N1, DAG.getConstantFP(3.0, VT));
-      }
-
       // (fadd (fmul c, x), (fadd x, x)) -> (fmul c+2, x)
       if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
           N1.getOperand(0) == N1.getOperand(1) &&
@@ -5884,12 +5877,6 @@
                            N0, NewCFP);
       }
 
-      // (fadd x, (fadd x, x)) -> (fmul 3.0, x)
-      if (!CFP10 && !CFP11 && N1.getOperand(0) == N1.getOperand(1) &&
-          N1.getOperand(0) == N0) {
-        return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
-                           N0, DAG.getConstantFP(3.0, VT));
-      }
 
       // (fadd (fadd x, x), (fmul c, x)) -> (fmul c+2, x)
       if (CFP10 && !CFP11 && N1.getOpcode() == ISD::FADD &&
@@ -5914,6 +5901,26 @@
       }
     }
 
+    if (N0.getOpcode() == ISD::FADD) {
+      ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
+      // (fadd (fadd x, x), x) -> (fmul 3.0, x)
+      if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
+          (N0.getOperand(0) == N1)) {
+        return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
+                           N1, DAG.getConstantFP(3.0, VT));
+      }
+    }
+
+    if (N1.getOpcode() == ISD::FADD) {
+      ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
+      // (fadd x, (fadd x, x)) -> (fmul 3.0, x)
+      if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
+          N1.getOperand(0) == N0) {
+        return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
+                           N0, DAG.getConstantFP(3.0, VT));
+      }
+    }
+
     // (fadd (fadd x, x), (fadd x, x)) -> (fmul 4.0, x)
     if (N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
         N0.getOperand(0) == N0.getOperand(1) &&


More information about the llvm-commits mailing list