Adventures in gdb

I picked up from yesterday wanting to see what code was being generated for record initialization. A colleague pointed me to a page of SpiderMonkey debugging tips. This was helpful, but required being able to run the JS interpreter inside GDB and type some code into the REPL. The problem is that before it got to that point, the interpreter was trying to compile all the self-hosted code; I knew that this wasn’t going to succeed since I’ve only implemented one of the record/tuple opcodes. I wanted to be able to just do:

> x = #{}

(binding the variable x to an empty record literal) and see the generated code. But because the much-more-complicated self-hosted code has to get compiled first, I never get to that point.

Another colleague suggested looking at the IONFLAGS environment variable. This, in turn, seems to only have an effect if you build the compiler with the --enable-jitspew option. Once I did that, I was able to find out more:

$ IONFLAGS=zzzz mach run
found tag: zzzz
Unknown flag.

usage: IONFLAGS=option,option,option,... where options can be:

  aborts        Compilation abort messages
  scripts       Compiled scripts
  mir           MIR information

And so on.

I found that IONFLAGS=codegen mach run would cause the interpreter to print out all the generated assembly code, including all the code for self-hosted methods. This wasn’t entirely helpful, since it was hard to see where the boundaries were between different methods.

I decided to try a different strategy and see what I could do inside gdb. I’ve avoided using debuggers as much as possible throughout my programming career. I’m a fan of printf-style debugging. So much so that I created the printf-style debugging page on Facebook. (This made more sense back when Facebook pages were “fan pages”, so you could be a “fan of” printf-style debugging.) I’ve always had the feeling that any more sophisticated debugging technology wasn’t worth the difficulty of use. Working on a compiler implemented in C++, though, it seems I’m finally having to suck it up and learn.

The first question was how to set a breakpoint on a templated function. I found the rbreak command in gdb, which takes a regular expression. I realized I could also just do:

(gdb) info functions .*emit_InitR.*
All functions matching regular expression ".*emit_InitR.*":

File js/src/jit/BaselineCodeGen.cpp:
2590:   bool js::jit::BaselineCodeGen::emit_InitRecord();
2590:   bool js::jit::BaselineCodeGen::emit_InitRecord();

File js/src/jit/BaselineIC.cpp:
2454:   bool js::jit::FallbackICCodeCompiler::emit_InitRecord();

So I set a breakpoint on the method I wrote to generate code for the InitRecord opcode:

(gdb) b js::jit::BaselineCodeGen::emit_InitRecord
Breakpoint 1 at 0x555558093884: file /home/tjc/gecko-fork/js/src/jit/BaselineCodeGen.cpp, line 2591.
(gdb) b js::jit::FallbackICCodeCompiler::emit_InitRecord
Breakpoint 2 at 0x5555580807b1: file /home/tjc/gecko-fork/js/src/jit/BaselineIC.cpp, line 2455.
(gdb) run
The program being debugged has been started already.
Start it from the beginning? (y or n) y
Starting program: /home/tjc/gecko-fork/obj-x64-debug/dist/bin/js 

Thread 1 "js" hit Breakpoint 2, js::jit::FallbackICCodeCompiler::emit_InitRecord (this=0x7fffffffd1b0)
    at /home/tjc/gecko-fork/js/src/jit/BaselineIC.cpp:2455
2455      EmitRestoreTailCallReg(masm);

Finally! At this point, I was hoping to be able to view the code that was being generated for the empty record literal. Stepping through the code from here gave me what I was looking for:

(gdb) s
js::jit::FallbackICCodeCompiler::tailCallVMInternal (
    this=0x7fffffffd1b0, masm=..., 
    at /home/tjc/gecko-fork/js/src/jit/BaselineIC.cpp:510
510   TrampolinePtr code = cx->runtime()->jitRuntime()->getVMWrapper(id);
(gdb) n
511   const VMFunctionData& fun = GetVMFunction(id);
(gdb) n
512   MOZ_ASSERT(fun.expectTailCall == TailCall);
(gdb) n
513   uint32_t argSize = fun.explicitStackSlots() * sizeof(void*);
(gdb) n
514   EmitBaselineTailCallVM(code, masm, argSize);
(gdb) n
515   return true;
(gdb) p code
$18 = {value = 0x1e4412b875e0 "H\277"}
(gdb) p code.value
$19 = (uint8_t *) 0x1e4412b875e0 "H\277"
(gdb) x/64i code.value
   0x1e4412b875e0:  movabs $0x7ffff4219000,%rdi
   0x1e4412b875ea:  mov    0x1c0(%rdi),%rax
   0x1e4412b875f1:  mov    %rsp,0x70(%rax)
   0x1e4412b875f5:  movabs $0x55555903de60,%r11
   0x1e4412b875ff:  push   %r11
   0x1e4412b87601:  lea    0x18(%rsp),%r10
   0x1e4412b87606:  movabs $0xfff9800000000000,%r11

So that’s the generated code for DoInitRecordFallback (the fallback method implemented in the inline cache module of the baseline compiler), but I realized this wasn’t really what I was hoping to find. I wanted to see the intermediate representation first.

From there, I realized I was barking up the wrong tree, since the baseline compiler just goes straight from JS to assembly; only the more sophisticated compilers (which weren’t being invoked at this point) use MIR and LIR. (A blog post from Matthew Gaudet, “A Beginners Guide To SpiderMonkey’s MacroAssembler”], explains some of the pipeline.)

So at least I knew one way to get to the generated assembly code for one opcode, but it wasn’t particularly helpful. My co-worker suggested putting in no-op implementations for the other opcodes so that it would be able to compile all the self-hosted code (even if the generated code wouldn’t work). This seemed like the fastest way to get to a functioning REPL so I could experiment with simpler code snippets, and it worked. After just adding a no-op emit_ method in BaselineCodeGen.cpp for each opcode, the interpreter was able to start up.

When I typed code into the REPL, I could tell it was only being interpreted, not compiled, since everything still worked, and I would expect anything that used records/tuples except for an empty record literal to fail. I found the --baseline-eager flag with a little bit of digging, and:

obj-x64-debug/dist/bin/js --baseline-eager
js> function f() { return #{}; }
function f() { return #{}; }
js> f()
Assertion failure: !BytecodeOpHasIC(op) (Missing entry in OpToFallbackKindTable for JOF_IC op), at js/src/jit/BaselineIC.cpp:353
Segmentation fault

Excellent! This pointed to something I didn’t change yesterday (since the compiler didn’t make me) — I had to update the OpToFallbackKindTable in BaselineIC.cpp.

Once I did that, I realized that I couldn’t get very far with just InitRecord, since I wouldn’t expect even the empty record to compile without being able to compile the FinishRecord opcode. (Since records are immutable, Nicolò’s implementation adds three opcodes for creating records: one to initialize the empty record, one to add a new record field, and one to finish initialization, the last of which marks the record as immutable so that no more fields can be added.)

So I implemented FinishRecord, similarly to the work from yesterday. Now what? I was able to type in an empty record literal without errors:

> x = #{}

But how do I know that x is bound to a well-formed record that satisfies its interface? There’s not too much you can do with an empty record. I decided to check that typeof(x) worked (it should return “record”), and got an assertion failure in the emitGuardNonDoubleType() method in CacheIRCompiler.cpp). It took me some time to make sense of various calls through generated code, but the issue was the TypeOfIRGenerator::tryAttachStub() method in CacheIR.cpp:

AttachDecision TypeOfIRGenerator::tryAttachStub() {

  MOZ_ASSERT_UNREACHABLE("Failed to attach TypeOf");
  return AttachDecision::NoAction;

This code decides, based on the type of the operand (valId) whether to use the typeOf code for primitives or for objects. The record/tuple implementation adds “object primitives”, which share some qualities with objects but aren’t objects (since, among other things, objects are mutable). The tryAttachPrimitive() call was successfully selecting the typeOf code for primitives, since the isPrimitive() method on the Value type returns true for object primitives. Because there was no explicit case in the code for records, the code for double values was getting called as a fallback and that’s where the assertion failure was coming from. Tracking this down took much more time than actually implementing typeOf for records, which I proceeded to do. And now I can get the type of a record-valued variable in compiled code:

js> x = #{}
js> typeof(x)

This provides at least some evidence that the code I’m generating is laying out records properly. Next up, I’ll try implementing the opcode that adds record properties, so that I can test out non-empty records!

Tags: ,

Leave a Reply