[llvm] r203311 - ISel: Make VSELECT selection terminate in cases where the condition type has to

[Arnold Schwaighofer]

Author: arnolds
Date: Fri Mar  7 17:25:55 2014
New Revision: 203311

URL: http://llvm.org/viewvc/llvm-project?rev=203311&view=rev
Log:
ISel: Make VSELECT selection terminate in cases where the condition type has to
be split and the result type widened.

When the condition of a vselect has to be split it makes no sense widening the
vselect and thereby widening the condition. We end up in an endless loop of
widening (vselect result type) and splitting (condition mask type) doing this.
Instead, split both the condition and the vselect and widen the result.

I ran this over the test suite with i686 and mattr=+sse and saw no regressions.

Fixes PR18036.

Modified:
    llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
    llvm/trunk/test/CodeGen/X86/sse1.ll

Modified: llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp?rev=203311&r1=203310&r2=203311&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp Fri Mar  7 17:25:55 2014
@@ -2192,6 +2192,17 @@ SDValue DAGTypeLegalizer::WidenVecRes_SE
     if (getTypeAction(CondVT) == TargetLowering::TypeWidenVector)
       Cond1 = GetWidenedVector(Cond1);
 
+    // If we have to split the condition there is no point in widening the
+    // select. This would result in an cycle of widening the select ->
+    // widening the condition operand -> splitting the condition operand ->
+    // splitting the select -> widening the select. Instead split this select
+    // further and widen the resulting type.
+    if (getTypeAction(CondVT) == TargetLowering::TypeSplitVector) {
+      SDValue SplitSelect = SplitVecOp_VSELECT(N, 0);
+      SDValue Res = ModifyToType(SplitSelect, WidenVT);
+      return Res;
+    }
+
     if (Cond1.getValueType() != CondWidenVT)
       Cond1 = ModifyToType(Cond1, CondWidenVT);
   }

Modified: llvm/trunk/test/CodeGen/X86/sse1.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/sse1.ll?rev=203311&r1=203310&r2=203311&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/sse1.ll (original)
+++ llvm/trunk/test/CodeGen/X86/sse1.ll Fri Mar  7 17:25:55 2014
@@ -43,3 +43,17 @@ entry:
 ; CHECK-NOT: shufps	$16
 ; CHECK: ret
 }
+
+; We used to get stuck in type legalization for this example when lowering the
+; vselect. With SSE1 v4f32 is a legal type but v4i1 (or any vector integer type)
+; is not. We used to ping pong between splitting the vselect for the v4i
+; condition operand and widening the resulting vselect for the v4f32 result.
+; PR18036
+
+; CHECK-LABEL: vselect
+define <4 x float> @vselect(<4 x float>*%p, <4 x i32> %q) {
+entry:
+  %a1 = icmp eq <4 x i32> %q, zeroinitializer
+  %a14 = select <4 x i1> %a1, <4 x float> <float 1.000000e+00, float 2.000000e+00, float 3.000000e+00, float 4.000000e+0> , <4 x float> zeroinitializer
+  ret <4 x float> %a14
+}