[turbofan] Implement tail calls with differing stack parameter counts

[turbofan] Implement tail calls with differing stack parameter counts

WORK IN PROGRESS, I'D APPRECIATE FEEDBACK BEFORE DOING THE OTHER PLATFORM PORTS

This implementation loads as many parameters as possible before the tail call
into registers and pushes the rest on the stack before deconstructing the
frame. As the frame is torn down, callee parameters are replaced for the
tail call by a mini parameter-handling interpreter, either by storing values
that are already in registers or by popping and storing the parameter values
from the stack.

The return address is handled specially. If the number of parameters increases
in tail call tall, the parameter value that will replace the return address
must be in a register before the frame is deconstructed and frame desconstruction
does a special value-swapping dance to make sure that the return address
ends up at the top of all parameters before finishing the tail call.

[danno, 2015-07-29 08:38:30]

danno@chromium.org changed reviewers:
+ bmeurer@chromium.org, titzer@chromium.org

[danno, 2015-07-29 08:49:50]

Not for landing yet, but I'd appreciate feedback before doing the rest of the
platform ports.

[Benedikt Meurer, 2015-07-30 06:47:24]

https://codereview.chromium.org/1259203002/diff/140001/src/compiler/ast-graph...
File src/compiler/ast-graph-builder.cc (right):

https://codereview.chromium.org/1259203002/diff/140001/src/compiler/ast-graph...
src/compiler/ast-graph-builder.cc:2606: TailCallMode mode =
IMHO this will be highly confusing and might be a source of weird bugs in the
future. By just looking at a tail call site in a .js file you will be mostly
unable to tell whether it is actually a tailcall to the runtime or a regular
call. I'd really prefer to have this explicit at the tail call site so people
are aware of what they are doing, just like we do in the native code stubs,
where we have __ TailCallRuntime and __ CallRuntime.

https://codereview.chromium.org/1259203002/diff/140001/src/compiler/ia32/code...
File src/compiler/ia32/code-generator-ia32.cc (right):

https://codereview.chromium.org/1259203002/diff/140001/src/compiler/ia32/code...
src/compiler/ia32/code-generator-ia32.cc:287: void
CodeGenerator::AssembleTailCallSetup(Instruction* instr) {
This is quite a lot of (complex) per-architecture code. I have a bad feeling
about this. Can we make this into explicit instruction opcodes at least and do
this magic as part of the instruction selection? This code within code approach
frightens me.

https://codereview.chromium.org/1259203002/diff/140001/src/compiler/ia32/inst...
File src/compiler/ia32/instruction-selector-ia32.cc (right):

https://codereview.chromium.org/1259203002/diff/140001/src/compiler/ia32/inst...
src/compiler/ia32/instruction-selector-ia32.cc:912: bool
InstructionSelector::EmitTailCallSetup(Node* node, CallBuffer* buffer,
See my comment in code-generator-ia32.cc.
Can we somehow make the tail call setup into regular InstructionCodes? We
already have push and pop.