[DAG] Don't try to create an INSERT_SUBVECTOR with an illegal source

[mkuper]

Summary

The vector op legalizer currently can't legalize INSERT_SUBVECTOR with an illegal source, and I'm not sure I'd like to teach it - it seems the general case would have to be similar to the CONCAT legalization (basically turn the INSERT_SUBVECTOR back into a BUILD_VECTOR).
Instead, make the - apparently only? - place that creates such INSERTs stop doing that.

This fixes PR31956.

[andreadb]

LGTM.

Thanks for the quick fix.

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if ((VT.getSizeInBits() % InVT1.getSizeInBits() == 0) && InVT1 == InVT2) {
         // VecIn1 is wider than the output, and we have another, possibly
         // smaller input. Pad the smaller input with undefs, shuffle at the
         // input vector width, and extract the output.
         // The shuffle type is different than VT, so check legality again.
         if (LegalOperations &&
             !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, InVT1))
           return SDValue();
 
-        if (InVT1 != InVT2)
+        // Legalizing INSERT_SUBVECTOR is tricky - you basically have to
+        // lower it back into a BUILD_VECTOR. So if the inserted type is
+        // illegal, don't even try.
+        if (InVT1 != InVT2) {
+          if (!TLI.isTypeLegal(InVT2))
+            return SDValue();
           VecIn2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, InVT1,
                                DAG.getUNDEF(InVT1), VecIn2, ZeroIdx);
+        }
         ShuffleNumElems = NumElems * 2;
       } else {
         // Both VecIn1 and VecIn2 are wider than the output, and VecIn2 is wider
         // than VecIn1. We can't handle this for now - this case will disappear
         // when we start sorting the vectors by type.
         return SDValue();
       }
     } else {

llvm/trunk/test/CodeGen/X86/pr31956.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mattr=+avx < %s | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-scei-ps4"
+
+@G1 = common global <2 x float> zeroinitializer, align 8
+@G2 = common global <8 x float> zeroinitializer, align 32
+
+define <4 x float> @foo() {
+; CHECK-LABEL: foo:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    vblendpd {{.*#+}} ymm0 = ymm0[0],mem[1,2,3]
+; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; CHECK-NEXT:    vshufps {{.*#+}} xmm0 = xmm1[0,2],xmm0[2,0]
+; CHECK-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,2,1,3]
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+entry:
+  %V = load <2 x float>, <2 x float>* @G1, align 8
+  %shuffle = shufflevector <2 x float> %V, <2 x float> undef, <8 x i32> <i32 undef, i32 undef, i32 undef, i32 undef, i32 0, i32 undef, i32 undef, i32 undef>
+  %L = load <8 x float>, <8 x float>* @G2, align 32
+  %shuffle1 = shufflevector <8 x float> %shuffle, <8 x float> %L, <4 x i32> <i32 12, i32 10, i32 14, i32 4>
+  ret <4 x float> %shuffle1
+}