I am trying to understand more deeply the instruction selection process in llvm and for that I am debuging step-by-step the CodeGenAndEmitDAG function. I have printed a small function (see below) just before the combine step - the first step in the above function. In the graph I see blue lines and it seems that they are always pointing at "ch" , which I think means "other" machine value type. What I don't understand is the meaning of the blue lines... what is this dependency ? And, am I right about the meaning of "ch" ? is it "other" ?
Dashed blue arrows represent non-dataflow dependencies between
instructions and enforce specific order between them. For example,
stores and loads which may access the same memory shouldn't be reordered,
though there's no data dependency between them. In such cases blue
arrows are used to represent such hidden dependency. These blue arrows
consume chain values (ch) of type Other
.
Every DAG has a special EntryToken
of type Other
which supplies
the initial chain value for the basic block.
Consider the following example. Notice the control dependency (blue arrow) between load and store because they're allowed to point to the same memory. Also notice the red arrow (Glue) which glues two instructions together.
int foo(int *a, int *b) {
a[0] = 42;
return b[0];
}
来源:https://stackoverflow.com/questions/26661529/what-does-machine-value-type-other-mean-in-llvm-sdnodes