When the visitSub method was split into visitSub and visitFSub, this xform was

added to the FSub version. However, the original version of this xform guarded
against doing this for floating point (!Op0->getType()->isFPOrFPVector()).

This is causing LLVM to perform incorrect xforms for code like:

void func(double *rhi, double *rlo, double xh, double xl, double yh, double yl){
  double mh, ml;
  double c = 134217729.0;
  double up, u1, u2, vp, v1, v2;

  up = xh*c;
  u1 = (xh - up) + up;
  u2 = xh - u1;

  vp = yh*c;
  v1 = (yh - vp) + vp;
  v2 = yh - v1;

  mh = xh*yh;
  ml = (((u1*v1 - mh) + (u1*v2)) + (u2*v1)) + (u2*v2);
  ml += xh*yl + xl*yh;

  *rhi = mh + ml;
  *rlo = (mh - (*rhi)) + ml;
}

The last line was optimized away, but rl is intended to be the difference
between the infinitely precise result of mh + ml and after it has been rounded
to double precision.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@93369 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index c8556ea01d8f8843e06333d473e46d9c29b94226..4891ff00e7b1ff7e0e1b1bf72f280aa9ca4c3472 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -736,16 +736,5 @@ Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
   if (Value *V = dyn_castFNegVal(Op1))
     return BinaryOperator::CreateFAdd(Op0, V);
 
-  if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
-    if (Op1I->getOpcode() == Instruction::FAdd) {
-      if (Op1I->getOperand(0) == Op0)              // X-(X+Y) == -Y
-        return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
-                                          I.getName());
-      else if (Op1I->getOperand(1) == Op0)         // X-(Y+X) == -Y
-        return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
-                                          I.getName());
-    }
-  }
-
   return 0;
 }

diff --git a/test/Transforms/InstCombine/fsub-fadd.ll b/test/Transforms/InstCombine/fsub-fadd.ll
new file mode 100644 (file)
index 0000000..f4cff88
--- /dev/null
+++ b/test/Transforms/InstCombine/fsub-fadd.ll
@@ -0,0 +1,39 @@
+; RUN: opt < %s -instcombine -S | FileCheck %s
+; <rdar://problem/7530098>
+
+define void @func(double* %rhi, double* %rlo, double %xh, double %xl, double %yh, double %yl) nounwind ssp {
+entry:
+  %"alloca point" = bitcast i32 0 to i32          ; <i32> [#uses=0]
+  %tmp = fmul double %xh, 0x41A0000002000000      ; <double> [#uses=2]
+  %tmp1 = fsub double %xh, %tmp                   ; <double> [#uses=1]
+  %tmp2 = fadd double %tmp1, %tmp                 ; <double> [#uses=3]
+  %tmp3 = fsub double %xh, %tmp2                  ; <double> [#uses=2]
+  %tmp4 = fmul double %yh, 0x41A0000002000000     ; <double> [#uses=2]
+  %tmp5 = fsub double %yh, %tmp4                  ; <double> [#uses=1]
+  %tmp6 = fadd double %tmp5, %tmp4                ; <double> [#uses=3]
+  %tmp7 = fsub double %yh, %tmp6                  ; <double> [#uses=2]
+  %tmp8 = fmul double %xh, %yh                    ; <double> [#uses=3]
+  %tmp9 = fmul double %tmp2, %tmp6                ; <double> [#uses=1]
+  %tmp10 = fsub double %tmp9, %tmp8               ; <double> [#uses=1]
+  %tmp11 = fmul double %tmp2, %tmp7               ; <double> [#uses=1]
+  %tmp12 = fadd double %tmp10, %tmp11             ; <double> [#uses=1]
+  %tmp13 = fmul double %tmp3, %tmp6               ; <double> [#uses=1]
+  %tmp14 = fadd double %tmp12, %tmp13             ; <double> [#uses=1]
+  %tmp15 = fmul double %tmp3, %tmp7               ; <double> [#uses=1]
+  %tmp16 = fadd double %tmp14, %tmp15             ; <double> [#uses=1]
+  %tmp17 = fmul double %xh, %yl                   ; <double> [#uses=1]
+  %tmp18 = fmul double %xl, %yh                   ; <double> [#uses=1]
+  %tmp19 = fadd double %tmp17, %tmp18             ; <double> [#uses=1]
+  %tmp20 = fadd double %tmp19, %tmp16             ; <double> [#uses=2]
+  %tmp21 = fadd double %tmp8, %tmp20              ; <double> [#uses=1]
+  store double %tmp21, double* %rhi, align 8
+  %tmp22 = load double* %rhi, align 8             ; <double> [#uses=1]
+  %tmp23 = fsub double %tmp8, %tmp22              ; <double> [#uses=1]
+  %tmp24 = fadd double %tmp23, %tmp20             ; <double> [#uses=1]
+
+; CHECK: %tmp23 = fsub double %tmp8, %tmp21
+; CHECK: %tmp24 = fadd double %tmp23, %tmp20
+
+  store double %tmp24, double* %rlo, align 8
+  ret void
+}