[LLVMdev] Assertion failure in tablegen: rationale ?

[Amaury Pouly]

Hello,
I was fiddling with TableGen (for a use that has nothing to do with a
compiler but it's doesn't matter) and TableGen triggers an assertion failure
on this code (I reduced the case to the minimum, it's a parsing bug):

class Bli<string _t>
{
    string t = _t;
}

class Bla<list<Bli> _bli>
: Bli<!car(_bli).t>
{
}

#0  0x00007ffff6ebda75 in *__GI_raise (sig=<value optimized out>) at
../nptl/sysdeps/unix/sysv/linux/raise.c:64
#1  0x00007ffff6ec15c0 in *__GI_abort () at abort.c:92
#2  0x00007ffff6eb6941 in *__GI___assert_fail (assertion=0x5aa663 "0 &&
\"Unknown unop\"", file=<value optimized out>, line=633,
    function=0x5abb20 "virtual llvm::RecTy*
llvm::UnOpInit::getFieldType(const std::string&) const") at assert.c:81
#3  0x000000000050ddc8 in llvm::UnOpInit::getFieldType (this=0x850f00,
FieldName=...) at Record.cpp:633
#4  0x0000000000543f64 in llvm::TGParser::ParseValue (this=0x7fffffffd480,
CurRec=0x8511d0, ItemType=0x851540) at TGParser.cpp:1342
#5  0x0000000000544574 in llvm::TGParser::ParseValueList
(this=0x7fffffffd480, CurRec=0x8511d0, ArgsRec=0x8510d0, EltTy=0x0) at
TGParser.cpp:1405
#6  0x000000000053e520 in llvm::TGParser::ParseSubClassReference
(this=0x7fffffffd480, CurRec=0x8511d0, isDefm=false) at TGParser.cpp:397
#7  0x00000000005458e5 in llvm::TGParser::ParseObjectBody
(this=0x7fffffffd480, CurRec=0x8511d0) at TGParser.cpp:1613
#8  0x00000000005463ad in llvm::TGParser::ParseClass (this=0x7fffffffd480)
at TGParser.cpp:1727
#9  0x0000000000548b86 in llvm::TGParser::ParseObject (this=0x7fffffffd480,
MC=0x0) at TGParser.cpp:2098
#10 0x0000000000548bbb in llvm::TGParser::ParseObjectList
(this=0x7fffffffd480, MC=0x0) at TGParser.cpp:2107
#11 0x0000000000548c10 in llvm::TGParser::ParseFile (this=0x7fffffffd480) at
TGParser.cpp:2115
#12 0x0000000000550035 in ParseFile (Filename=..., IncludeDirs=...,
SrcMgr=...) at TableGen.cpp:196
#13 0x0000000000550116 in main (argc=3, argv=0x7fffffffe208) at
TableGen.cpp:206

Looking at the code, it's triggered because UnOpInit::getFieldType only
handles the cast<> operator. I don't see a good reason for this given that
!car can give a value of which we can access a subfield. Of course, a
trivial solution to the problem is to use !cast with !car but this seems
ugly to me:


class Bli<string _t>
{
    string t = _t;
}

class Bla<list<Bli> _bli>
: Bli<!cast<Bli>(!car(_bli)).t>
{
}

Is this a bug or is it the wanted behaviour that a cast must be introduced ?
Regards


Amaury Pouly
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20100901/b043a6a5/attachment.html>