assertion failure handling 64-bit umulh sequence

[sunfishcode]

Bugzilla Link	1198
Resolution	FIXED
Resolved on	Feb 22, 2010 12:53
Version	trunk
OS	All

Extended Description

This test case:

define i64 @​foo(i64 %x, i64 %y) {
%tmp0 = zext i64 %x to i128
%tmp1 = zext i64 %y to i128
%tmp2 = mul i128 %tmp0, %tmp1
%tmp7 = zext i32 64 to i128
%tmp3 = lshr i128 %tmp2, %tmp7
%tmp4 = trunc i128 %tmp3 to i64
ret i64 %tmp4
}

compiled with -march=x86-64 triggers the following assertion failure:

lib/CodeGen/SelectionDAG/TargetLowering.cpp:2197: llvm::MVT::ValueType
llvm::TargetLowering::getValueType(const llvm::Type*) const: Assertion `0 &&
"Invalid width for value type"' failed.

The following patch fixes it.

Index: TargetLowering.cpp

RCS file: /var/cvs/llvm/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp,v
retrieving revision 1.89
diff -u -r1.89 TargetLowering.cpp
--- TargetLowering.cpp
+++ TargetLowering.cpp
@@ -2200,6 +2200,7 @@
case 16: return MVT::i16;
case 32: return MVT::i32;
case 64: return MVT::i64;

•  case 128:  return MVT::i128;
  

  }
  break;
  case Type::FloatTyID: return MVT::f32;

[lattner]

Very nice catch. I applied your patch:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070212/044465.html

and checked in your testcase here: test/CodeGen/X86/i128-mul.ll

Note that 128-bit integer support in the LLVM level is desired, but not really ready yet. Things like 128-
bit immediates are not handled right. Fortunately, the APInt stuff is great steps in the right direction, so
we should have it for LLVM 2.0.

-Chris