From 8f7c5a7f18d0482fdc99c1c02679830fac89bcc6 Mon Sep 17 00:00:00 2001
From: Hal Finkel <hfinkel@anl.gov>
Date: Mon, 23 Mar 2015 08:22:43 +0000
Subject: [PATCH] [SDAG] Don't widen VSETCC during type legalization for split
 operands

Because the operands of a vector SETCC node can be of a different type from the
result (and often are), it can happen that even if we'd prefer to widen the
result type of the SETCC, the operands have been split instead. In this case,
the SETCC result also must be split. This mirrors what is done in
WidenVecRes_SELECT, and should be NFC elsewhere because if the operands are not
widened the following calls to GetWidenedVector will assert (which is what was
happening in the test case).

llvm-svn: 232935
---
 .../CodeGen/SelectionDAG/LegalizeVectorTypes.cpp   | 10 ++++++
 llvm/test/CodeGen/PowerPC/qpx-split-vsetcc.ll      | 40 ++++++++++++++++++++++
 2 files changed, 50 insertions(+)
 create mode 100644 llvm/test/CodeGen/PowerPC/qpx-split-vsetcc.ll

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index 63671f7..f7e4557 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -2553,6 +2553,16 @@ SDValue DAGTypeLegalizer::WidenVecRes_VSETCC(SDNode *N) {
   assert(InVT.isVector() && "can not widen non-vector type");
   EVT WidenInVT = EVT::getVectorVT(*DAG.getContext(),
                                    InVT.getVectorElementType(), WidenNumElts);
+
+  // The input and output types often differ here, and it could be that while
+  // we'd prefer to widen the result type, the input operands have been split.
+  // In this case, we also need to split the result of this node as well.
+  if (getTypeAction(InVT) == TargetLowering::TypeSplitVector) {
+    SDValue SplitVSetCC = SplitVecOp_VSETCC(N);
+    SDValue Res = ModifyToType(SplitVSetCC, WidenVT);
+    return Res;
+  }
+
   InOp1 = GetWidenedVector(InOp1);
   SDValue InOp2 = GetWidenedVector(N->getOperand(1));
 
diff --git a/llvm/test/CodeGen/PowerPC/qpx-split-vsetcc.ll b/llvm/test/CodeGen/PowerPC/qpx-split-vsetcc.ll
new file mode 100644
index 0000000..c8cef0f
--- /dev/null
+++ b/llvm/test/CodeGen/PowerPC/qpx-split-vsetcc.ll
@@ -0,0 +1,40 @@
+; RUN: llc -mcpu=a2q < %s | FileCheck %s
+target datalayout = "E-m:e-i64:64-n32:64"
+target triple = "powerpc64-bgq-linux"
+
+; Function Attrs: nounwind
+define void @gsl_sf_legendre_Pl_deriv_array() #0 {
+entry:
+  br i1 undef, label %do.body.i, label %if.else.i
+
+do.body.i:                                        ; preds = %entry
+  unreachable
+
+if.else.i:                                        ; preds = %entry
+  br i1 undef, label %return, label %for.body46.lr.ph
+
+for.body46.lr.ph:                                 ; preds = %if.else.i
+  br label %vector.body198
+
+vector.body198:                                   ; preds = %vector.body198, %for.body46.lr.ph
+  %0 = icmp ne <4 x i32> undef, zeroinitializer
+  %1 = select <4 x i1> %0, <4 x double> <double 5.000000e-01, double 5.000000e-01, double 5.000000e-01, double 5.000000e-01>, <4 x double> <double -5.000000e-01, double -5.000000e-01, double -5.000000e-01, double -5.000000e-01>
+  %2 = fmul <4 x double> undef, %1
+  %3 = fmul <4 x double> undef, %2
+  %4 = fmul <4 x double> %3, undef
+  store <4 x double> %4, <4 x double>* undef, align 8
+  br label %vector.body198
+
+; CHECK-LABEL: @gsl_sf_legendre_Pl_deriv_array
+; CHECK: qvlfiwzx
+; CHECK: qvfcfidu
+; CHECK: qvfcmpeq
+; CHECK: qvfsel
+; CHECK: qvfmul
+
+return:                                           ; preds = %if.else.i
+  ret void
+}
+
+attributes #0 = { nounwind }
+
-- 
2.7.4