custom lowering with multiple results

Project Infrastructure LLVM Dev List Archives
1

Here's a concrete example of what I'm having problems with:
trying to use LowerOperation to replace one node with two nodes.

Imagine I have a new instruction "read_world" that sets the status flags
according to some kind of global state. In order to generate this instruction I
have a built-in function with the prototype "int __builtin_read_world(void)".

I also have a new instruction "read_flags" that sets a register according to the
status flags so I can translate a call to __builtin_read_world into a
"read_world" followed by a "read_flags" (at least as a last resort; I'm
expecting the "read_flags" instruction to be optimised away in most cases).

Attached is the patch I have attempted to use to implement this. It's mostly
boring (though I may have made a stupid error in the boring bits). The most
interesting bit is in ARMISelLowering.[ch] where I add the DAG nodes READ_WORLD
and READ_FLAGS and translate Intrinsic::arm_read_world into a READ_WORLD
followed by a READ_FLAGS. This is the bit I'm most worried about as I'm getting
ARMTargetLowering::LowerOperation to do something slightly more complicated than
usual. I may have broken some rule about what LowerOperation is allowed to do.

When I feed the patched Clang/LLVM with the source code

// clang -cc1 -triple thumbv7-eabi -target-cpu cortex-a8 t.c -O3 -S -o t.s

int f1()
{
  return __builtin_read_world();
}

I get:

clang: InstrEmitter.cpp:667: ...
"#operands for dag node doesn't match .td file!"' failed.

What have I done wrong?

Any clues gratefully received.

Edmund

read_world.patch (5.71 KB)

2

Hi Edmund,

We did some custom lowering for the reminder operation (move the
remainder from R1 to R0), maybe that's more or less what you need?

See in the patch I sent last week (rt-abi), or you can ask Evzen about
it directly.

cheers,
--renato

3

Hello, Edmund

Here's a concrete example of what I'm having problems with:
trying to use LowerOperation to replace one node with two nodes.

I don't see any problem with such usage. There are a lot of examples
in the current codebase for this.

clang: InstrEmitter.cpp:667: ...
"#operands for dag node doesn't match .td file!"' failed.

Well... You created a node with wrong number of operands. Check how
you created stuff and declared in .td.

In your case:
- for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) {
+ for (unsigned i = 0, e = Tmp1->getNumValues(); i != e; ++i) {
This hunks looks potentially invalid

And how do you handle chain for your node? You can just gdb at the
place of the assertion and check the expected and provided # of
operands.