"Unexpected illegal type!" assertion

[llvmbot]

Bugzilla Link	9190
Resolution	FIXED
Resolved on	Feb 11, 2011 13:29
Version	trunk
OS	Windows NT
Reporter	LLVM Bugzilla Contributor

Extended Description

I get "Unexpected illegal type!" assertion when trying to compile the following kernel:

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-f80:128:128-v64:64:64-v128:128:128-a0:0:64-f80:32:32-n8:16:32"
target triple = "i686-pc-win32"

define void @​test() nounwind {
%1 = fdiv <3 x double> zeroinitializer, undef
%2 = fdiv <2 x double> zeroinitializer, undef
%3 = shufflevector <2 x double> %2, <2 x double> undef, <3 x i32> <i32 0, i32 1, i32 undef>
%4 = insertelement <3 x double> %3, double undef, i32 2
%5 = bitcast <3 x double> %1 to <3 x i64>
%6 = bitcast <3 x double> %4 to <3 x i64>
%7 = sub <3 x i64> %5, %6
%8 = shufflevector <3 x i64> %7, <3 x i64> undef, <2 x i32> <i32 0, i32 1>
%9 = xor <2 x i64> %8, zeroinitializer
%10 = add nsw <2 x i64> %9, zeroinitializer
%11 = shufflevector <2 x i64> %10, <2 x i64> undef, <3 x i32> <i32 0, i32 1, i32 undef>
%12 = insertelement <3 x i64> %11, i64 0, i32 2
%13 = shufflevector <3 x i64> %12, <3 x i64> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%14 = shufflevector <4 x i64> %13, <4 x i64> undef, <2 x i32> <i32 0, i32 1>
%15 = bitcast <2 x i64> %14 to <4 x i32>
%16 = shufflevector <4 x i32> %15, <4 x i32> undef, <4 x i32> <i32 0, i32 2, i32 0, i32 2>
%17 = bitcast <4 x i32> %16 to <2 x i64>
%18 = shufflevector <2 x i64> %17, <2 x i64> undef, <2 x i32> <i32 0, i32 2>
%19 = bitcast <2 x i64> %18 to <4 x i32>
%20 = shufflevector <4 x i32> %19, <4 x i32> undef, <3 x i32> <i32 0, i32 1, i32 2>
%21 = or <3 x i32> %20, zeroinitializer
store <3 x i32> %21, <3 x i32> addrspace(1)* undef, align 16
ret void
}

running "llc test.ll" produces:

llc: /llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp:800: llvm::SDValue::SelectionDAGLegalize::LegalizeOp(llvm::SDValue): Assertion `getTypeAction(Node->getValueType(i)) == Legal && "Unexpected illegal type!"' failed.
Stack dump:
0. Program arguments: llc test.ll

1.  Running pass 'Function Pass Manager' on module 'test.ll'.
   

2.  Running pass 'X86 DAG->DAG Instruction Selection' on function '@test'
   

Abort

[llvmbot]

Here is a patch to the problem.

The bug happens when the DAGCombiner attempts to optimize one of the patterns of the SUB opcode. It tries to create a zero of type v2i64. This type is legal on 32bit machines, but the initializer is target dependent. Currently, the initializer attempts to create an i64 zero constant, which fails.
Added a flag to tell the DagCombiner to create a legal zero, if we require that the pass would generate legal types.

Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp

--- lib/CodeGen/SelectionDAG/DAGCombiner.cpp (revision 124985)
+++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp (working copy)
@@ -1533,7 +1533,7 @@

// fold (sub x, x) -> 0
if (N0 == N1)

• return DAG.getConstant(0, N->getValueType(0));

• return DAG.getConstant(0, N->getValueType(0), LegalTypes);
  // fold (sub c1, c2) -> c1-c2
  if (N0C && N1C)
  return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);

[llvmbot]

Fixed in #​125391