Transform div to mul with reciprocal only when fp imm is legal.

This fixes PR12516 and uncovers one weird problem in legalize (workarounded)


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154394 91177308-0d34-0410-b5e6-96231b3b80d8

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -5769,8 +5769,15 @@ SDValue DAGCombiner::visitFDIV(SDNode *N) {
     APFloat N1APF = N1CFP->getValueAPF();
     APFloat Recip(N1APF.getSemantics(), 1); // 1.0
     APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
-    // Only do the transform if the reciprocal is not too horrible (eg not NaN).
-    if (st == APFloat::opOK || st == APFloat::opInexact)
+    // Only do the transform if the reciprocal is not too horrible (eg not NaN)
+    // and the reciprocal is a legal fp imm.
+    if ((st == APFloat::opOK || st == APFloat::opInexact) &&
+        (!LegalOperations ||
+         // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
+         // backend)... we should handle this gracefully after Legalize.
+         // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
+         TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
+         TLI.isFPImmLegal(Recip, VT)))
       return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0,
                          DAG.getConstantFP(Recip, VT));
   }

test/CodeGen/ARM/2012-04-10-DAGCombine.ll

@@ -0,0 +1,31 @@
+; RUN: llc < %s -march=arm -mcpu=cortex-a9 -enable-unsafe-fp-math
+;target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:64:128-a0:0:64-n32-S64"
+;target triple = "armv7-none-linux-gnueabi"
+
+define arm_aapcs_vfpcc void @foo(<4 x float> %arg) nounwind align 2 {
+bb4:
+  %tmp = extractelement <2 x float> undef, i32 0
+  br i1 undef, label %bb18, label %bb5
+
+bb5:                                              ; preds = %bb4
+  %tmp6 = fadd float %tmp, -1.500000e+01
+  %tmp7 = fdiv float %tmp6, 2.000000e+01
+  %tmp8 = fadd float %tmp7, 1.000000e+00
+  %tmp9 = fdiv float 1.000000e+00, %tmp8
+  %tmp10 = fsub float 1.000000e+00, %tmp9
+  %tmp11 = fmul float %tmp10, 1.000000e+01
+  %tmp12 = fadd float %tmp11, 1.500000e+01
+  %tmp13 = fdiv float %tmp12, %tmp
+  %tmp14 = insertelement <2 x float> undef, float %tmp13, i32 0
+  %tmp15 = shufflevector <2 x float> %tmp14, <2 x float> undef, <4 x i32> zeroinitializer
+  %tmp16 = fmul <4 x float> zeroinitializer, %tmp15
+  %tmp17 = fadd <4 x float> %tmp16, %arg
+  store <4 x float> %tmp17, <4 x float>* undef, align 8, !tbaa !0
+  br label %bb18
+
+bb18:                                             ; preds = %bb5, %bb4
+  ret void
+}
+
+!0 = metadata !{metadata !"omnipotent char", metadata !1}
+!1 = metadata !{metadata !"Simple C/C++ TBAA", null}