Skip to content

V
V8
Switch branch/tag
  1. 18 Aug, 2015 1 commit
    • danno's avatar
      [turbofan] Unify referencing of stack slots · cbbaf9ea
      danno authored 
      Previously, it was not possible to specify StackSlotOperands for all
slots in both the caller and callee stacks. Specifically, the region
of the callee's stack including the saved return address, frame
pointer, function pointer and context pointer could not be addressed
by the register allocator/gap resolver.

In preparation for better tail call support, which will use the gap
resolver to reconcile outgoing parameters, this change makes it
possible to address all slots on the stack, because slots in the
previously inaccessible dead zone may become parameter slots for
outgoing tail calls. All caller stack slots are accessible as they
were before, with slot -1 corresponding to the last stack
parameter. Stack slot indices >= 0 access the callee stack, with slot
0 corresponding to the callee's saved return address, 1 corresponding
to the saved frame pointer, 2 corresponding to the current function
context, 3 corresponding to the frame marker/JSFunction, and slots 4
and above corresponding to spill slots.

The following changes were specifically	needed:

* Frame	has been changed to explicitly manage three areas of the
  callee frame, the fixed header, the spill slot area, and the
  callee-saved register area.
* Conversions from stack slot indices to fp offsets all now go through
  a common bottleneck: OptimizedFrame::StackSlotOffsetRelativeToFp
* The generation of deoptimization translation tables has been changed
  to support the new stack slot indexing scheme. Crankshaft, which
  doesn't support the new slot numbering in its register allocator,
  must adapt the indexes when creating translation tables.
* Callee-saved parameters are now kept below spill slots, not above,
  to support saving only the optimal set of used registers, which is
  only known after register allocation is finished and spill slots
  have been allocated.

Review URL: https://codereview.chromium.org/1261923007

Cr-Commit-Position: refs/heads/master@{#30224}
      cbbaf9ea
  3. 06 Jul, 2015 3 commits
    • bmeurer's avatar
      [turbofan] Reland "Add new JSFrameSpecialization reducer." and "Perform OSR... · ef661b08
      bmeurer authored 
      [turbofan] Reland "Add new JSFrameSpecialization reducer." and "Perform OSR deconstruction early and remove type propagation.".

We have to reland these two commits at once, because the first breaks
some asm.js benchmarks without the second. The change was reverted
because of bogus checks in the verifier, which will not work in the
presence of OSR (and where hidden because of the type back propagation
hack in OSR so far). Original messages are below:

[turbofan] Add new JSFrameSpecialization reducer.

The JSFrameSpecialization specializes an OSR graph to the current
unoptimized frame on which we will perform the on-stack replacement.
This is used for asm.js functions, where we cannot reuse the OSR
code object anyway because of context specialization, and so we could as
well specialize to the max instead.

It works by replacing all OsrValues in the graph with their values
in the JavaScriptFrame.

The idea is that using this trick we get better performance without
doing the unsound backpropagation of types to OsrValues later. This
is the first step towards fixing OSR for TurboFan.

[turbofan] Perform OSR deconstruction early and remove type propagation.

This way we don't have to deal with dead pre-OSR code in the graph
and risk optimizing the wrong code, especially we don't make
optimistic assumptions in the dead code that leaks into the OSR code
(i.e. deopt guards are in dead code, but the types propagate to OSR
code via the OsrValue type back propagation).

BUG=v8:4273
LOG=n
R=jarin@chromium.org

Review URL: https://codereview.chromium.org/1226673005

Cr-Commit-Position: refs/heads/master@{#29486}
      ef661b08
    • machenbach's avatar
      Revert "[turbofan] Add new JSFrameSpecialization reducer." · 9e71cdba
      machenbach authored 
      Also revert "[turbofan] Perform OSR deconstruction early and remove type propagation."

This reverts commit b0a852e8.

This reverts commit cdbb6c48.

NOTRY=true
NOTREECHECKS=true
BUG=v8:4273
LOG=n
TBR=bmeurer@chromium.org

Review URL: https://codereview.chromium.org/1225743002

Cr-Commit-Position: refs/heads/master@{#29480}
      9e71cdba
    • bmeurer's avatar
      [turbofan] Perform OSR deconstruction early and remove type propagation. · cdbb6c48
      bmeurer authored 
      This way we don't have to deal with dead pre-OSR code in the graph and
risk optimizing the wrong code, especially we don't make optimistic
assumptions in the dead code that leaks into the OSR code (i.e. deopt
guards are in dead code, but the types propagate to OSR code via the
OsrValue type back propagation).

BUG=v8:4273
LOG=n
R=jarin@chromium.org

Review URL: https://codereview.chromium.org/1215333005

Cr-Commit-Position: refs/heads/master@{#29478}
      cdbb6c48
  5. 19 Jun, 2015 1 commit
    • bmeurer's avatar
      [turbofan] Proper dead code elimination as regular reducer. · 733a2463
      bmeurer authored 
      The three different concerns that the ControlReducer used to deal with
are now properly separated into

  a.) DeadCodeElimination, which is a regular AdvancedReducer, that
      propagates Dead via control edges,
  b.) CommonOperatorReducer, which does strength reduction on common
      operators (i.e. Branch, Phi, and friends), and
  c.) GraphTrimming, which removes dead->live edges from the graph.

This will make it possible to run the DeadCodeElimination together with
other passes that actually introduce Dead nodes, i.e. typed lowering;
and it opens the door for general inlining without two stage fix point
iteration.

To make the DeadCodeElimination easier and more uniform, we basically
reverted the introduction of DeadValue and DeadEffect, and changed the
Dead operator to produce control, value and effect. Note however that
this is not a requirement, but merely a way to make dead propagation
easier and more uniform. We could always go back and decide to have
different Dead operators if some other change requires that.

Note that there are several additional opportunities for cleanup now,
i.e. OSR deconstruction could be a regular reducer now, and we don't
need to use TheHole as dead value marker in the GraphReducer. And we can
actually run the dead code elimination together with the other passes
instead of using separate passes over the graph.  We will do this in
follow up CLs.

R=jarin@chromium.org, mstarzinger@chromium.org

Review URL: https://codereview.chromium.org/1193833002

Cr-Commit-Position: refs/heads/master@{#29146}
      733a2463
  7. 17 Jun, 2015 1 commit
  9. 12 Jun, 2015 1 commit
    • bmeurer's avatar
      [turbofan] Use appropriate type for NodeId. · c487aba7
      bmeurer authored 
      Up until now we used int32_t for NodeId, but that was not ideal because
negative values are invalid for NodeId and we use it as an array index
for example in the NodeMarker class, where C++ compilers on x64 have to
generate code that does proper sign extension for the indices, which is
completely unnecessary.

R=svenpanne@chromium.org

Review URL: https://codereview.chromium.org/1178403004

Cr-Commit-Position: refs/heads/master@{#28997}
      c487aba7
  11. 26 May, 2015 1 commit
  13. 06 May, 2015 1 commit
    • bmeurer's avatar
      [turbofan] Add support for advanced reducers. · 7b33409b
      bmeurer authored 
      An AdvancedReducer is basically a regular Reducer with an editor
that can perform graph editing operations beyond changing or
replacing the node that is currently being reduced. The GraphReducer
is the default implementation of the AdvancedReducer::Editor interface.

The ControlReducerImpl is now just an AdvancedReducer, which
temporarily requires a Finish method in the reducer to implement
the dead node trimming until we move that to the GraphReducer
(which in turn requires that all loops are connected to End).

Review URL: https://codereview.chromium.org/1122423003

Cr-Commit-Position: refs/heads/master@{#28251}
      7b33409b
  15. 27 Apr, 2015 1 commit
  17. 17 Mar, 2015 1 commit
  19. 19 Feb, 2015 2 commits
  21. 17 Feb, 2015 2 commits
  23. 16 Feb, 2015 1 commit
  25. 11 Feb, 2015 1 commit
  27. 02 Feb, 2015 1 commit
  29. 30 Jan, 2015 3 commits
  31. 29 Jan, 2015 1 commit
  33. 12 Jan, 2015 1 commit