[llvm-dev] How to debug instruction selection
Krzysztof Parzyszek via llvm-dev
llvm-dev at lists.llvm.org
Tue Aug 15 05:44:50 PDT 2017
The line
0x28ec830: ch,glue = X86ISD::CALL 0x28ec7c0, 0x28ef900, Register:i32
%EDI, Register:i8 %AL, RegisterMask:Untyped, 0x28ec7c0:1
shows a DAG node in the typical debugging output. The first part
(0x28ec830) is the address of that node. Each node can produce multiple
values, and the "ch,glue" is the list of types of the produced value: in
this case it's a chain and a glue. Then there is the opcode
(X86ISD::CALL), and then the list of arguments. Most arguments are
printed as addresses of the nodes, but some are printed inline, like
"Register:i32 %EDI". In a typical case, like "0x28ec7c0" the address by
itself indicates the 0-th value produced by that node. If some other
value is used, the index of that value will be given as :N, as in
"0x28ec7c0:1". The node 0x28ec7c0 also produces a chain and a glue: the
address by itself refers to the 0-th value, i.e. the chain, and the :1
refers to the glue.
The other lines show the nodes used as arguments, and their arguments, etc.
If you want to debug the actual instruction selection, add [-mllvm]
-debug-only=isel to the command line. I'm not sure how to enable it with
JIT (maybe hardcode the debug flag to "true" by hand). It will print the
initial DAG as well as the DAG after various steps of
legalization/combining. Most importantly (from your perspective), it
will print the steps of instruction selection. The instruction selection
is done "backwards", i.e. from the last instruction in the basic block
upwards. More precisely, it happens from the root node, through the
intermediate nodes up to the leaves.
The debug output will have lines like "match failed at 12345, continuing
at 23456". The numbers correspond to the steps in the DAG matcher, which
is auto-generated from the .td files by TableGen. In the build
directory, TableGen has created a file <Target>GenDAGISel.inc. It
implements the function "SelectCode", that has a long array
(MatcherTable) of matcher commands like OPC_CheckType. The comments at
the beginning of each line contain the number of the step implemented by
that line.
When you see "match failed at NNNNN", the step "NNNNN" is usually some
kind of a check, like OPC_CheckOpcode, or OPC_CheckPredicate. The .inc
file will have the expected opcode, or the name of the predicate being
checked, so you can see what check exactly was performed at that step.
-Krzysztof
On 8/15/2017 2:58 AM, Alex Denisov via llvm-dev wrote:
> Hi there,
>
> I try to JIT compile some bitcode and seeing the following error:
>
> LLVM ERROR: Cannot select: 0x28ec830: ch,glue = X86ISD::CALL 0x28ec7c0, 0x28ef900, Register:i32 %EDI, Register:i8 %AL, RegisterMask:Untyped, 0x28ec7c0:1
> 0x28ef900: i32 = X86ISD::Wrapper TargetGlobalAddress:i32<void (i8*, ...)* @_ZN5FooBr7xprintfEPKcz> 0
> 0x28ec520: i32 = TargetGlobalAddress<void (i8*, ...)* @_ZN5FooBr7xprintfEPKcz> 0
> 0x28ec670: i32 = Register %EDI
> 0x28ec750: i8 = Register %AL
> 0x28ec360: Untyped = RegisterMask
> 0x28ec7c0: ch,glue = CopyToReg 0x28ec6e0, Register:i8 %AL, Constant:i8<0>, 0x28ec6e0:1
> 0x28ec750: i8 = Register %AL
> 0x28ec600: i8 = Constant<0>
> 0x28ec6e0: ch,glue = CopyToReg 0x28ec590, Register:i32 %EDI, 0x28ef820
> 0x28ec670: i32 = Register %EDI
> 0x28ef820: i32 = X86ISD::Wrapper TargetGlobalAddress:i32<[47 x i8]* @.str> 0
> 0x28ec910: i32 = TargetGlobalAddress<[47 x i8]* @.str> 0
> In function: _ZN5FooBr10initSystemEv
>
> I have some assumptions about this particular error.
> What is not clear to me is: how am I supposed to read this error message? I.e. what to look at to make some sense out of it?
>
> P.S. I have zero knowledge about instruction selection, so any suggestions on where I can get this knowledge are more than welcome.
> --
> AlexDenisov
> Software Engineer, https://lowlevelbits.org
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation
More information about the llvm-dev
mailing list