1710025 - Make Ion's regalloc debug output easier to read

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, enhancement, P3)

Description

[Julian Seward [:jseward]]

This is in some sense a followup to bug 1686658.

The backtracking allocator's debug printing of live ranges and bundles is hard
to read and occasionally misleading. This patch makes it somewhat less "busy"
whilst not taking any more space than it did before. There is no functional
change. Details/examples to follow.

Comment 1

[Julian Seward [:jseward]]

Changes:

• Instruction-point ranges were previously printed as [first_index, last_index+1),
  except when printing LIR, where the format first_index,last_index is used.
  Both are replaced with, simply, first_index-last_index.

• For live range and live bundle printing, the set of use points is enclosed
  in { and } so as to make it easier to see where the range ends.

• For live range use/def points, the index is printed first, followed by _.

• LUse policy printing has been changed to be less mysterious:

  • "register": r changed to R
  • "fixed": r<reg> changed to F:reg
  • "any": r? (which is misleading, since it includes stack) changed to A
  • "keepalive": * changed to KA
  • "recovered_input": ** changed to RI

Hence

[RegAlloc] Instructions:
[RegAlloc]   Block 0 [successor 1] [successor 2]
[RegAlloc]     [2,3 WasmParameter] [def v1<g>:x23]
[RegAlloc]     [4,5 WasmLoadSlot] [def v2<i>] [use v1:r]
...
[RegAlloc] Live ranges by bundle (after grouping/queueing regs):
[RegAlloc]   v1 [3,32) (def) v1:r@4 v1:x23@15 v1:x23@25 v1:x23@31
[RegAlloc]   v2 [5,12) (def) v2:r@6 v2:r@6 v2:r@8 ## v2 [16,24) v2:r@23

becomes

[RegAlloc] Instructions:
[RegAlloc]   Block 0 [successor 1] [successor 2]
[RegAlloc]     2-3 WasmParameter [def v1<g>:x23]
[RegAlloc]     4-5 WasmLoadSlot [def v2<i>] [use v1:R]
...
[RegAlloc] Live ranges by bundle (after grouping/queueing regs):
[RegAlloc]   v1 3-31 { 3_def 4_v1:R 15_v1:F:x23 25_v1:F:x23 31_v1:F:x23 }
[RegAlloc]   v2 5-11 { 5_def 6_v2:R 6_v2:R 8_v2:R } ## v2 16-23 { 23_v2:R }

Finally, in AllocationIntegrityState::checkIntegrity(), there is an
assertion which checks that no definition has multiple conflicting
assignments. This patch also adds printing to assist diagnosis of such
failures. It is guarded by both #ifdef JS_JITSPEW and IONFLAGS=regalloc.

Comment 2

[Julian Seward [:jseward]]

Attachment: Bug 1710025 - Make Ion's regalloc debug output easier to read. r=jandem.

The backtracking allocator's debug printing of live ranges, bundles and passes
is hard to read and occasionally misleading. This patch has the following
goals:

• make printing of instruction-point ranges, live ranges and bundles less
  cluttered

• enable consistent printing of fixed-register constraints in LUse and
  LDefinition

• give LiveBundles identities of the form LB<number> in what is printed

• be clearer and generally more elegant about when regalloc passes begin and
  end

There are no functional changes, and for non-JS_JITSPEW builds, no storage
changes.

Changes in detail (may not be exhaustive):

• Instruction-point ranges were previously printed as
  [first_index, last_index+1), except when printing LIR, where the format
  first_index,last_index is used. Both are replaced with
  first_index-last_index.

• For live range and live bundle printing, the set of use points is enclosed
  in { and } so as to make it easier to see where the range ends.

• For live range use/def points, the index is printed first, followed by _.
  If a def has an associated fixed assignment, that is now printed too.

• For LIR MoveGroups, the destinations are now printed on the left. That
  fixes an inconsistency in LIR printing wherein "normal" nodes had
  destinations on the left but MoveGroups had destinations on the right.

• LUse policy printing has been changed to be less mysterious:

  • "register": r changed to R
  • "fixed": r<reg> changed to F:reg
  • "any": r? (which is misleading, since it includes stack)
    changed to A
  • "keepalive": * changed to KA
  • "recovered_input": ** changed to RI

• To make LiveBundle creation/splitting easier to follow, LiveBundles now have
  a zero-based integer index associated with them. This gives them identities
  in the debug output. Also, the spillParent of each bundle is shown, so as to
  expose the spill-parent/child relationships:

  LB4(parent=LB2 v1 3-3 { 3_def:F:rdi })
  LB5(parent=LB2 v1 6-7 { 7_v1:F:rax })
  LB2(parent=none v1 4-7 { })
  

• There is some new debug printing in fn ResolveControlFlow.

• The initial live ranges (prior to grouping) and initial live bundles (after
  grouping) are printed differently, to make it obvious that these things are
  different C++ types. Function dumpVregs has accordingly been split into
  dumpLiveRangesByVReg and dumpLiveRangesByBundle.

• Some finessing of blank lines to make the boundaries between passes clear.

Finally, in AllocationIntegrityState::checkIntegrity, there is an existing
assertion which checks that no definition has multiple conflicting
assignments. This patch also adds printing to assist diagnosis of such
failures. It is guarded by both JS_JITSPEW and IONFLAGS=regalloc.

Comment 3

[Julian Seward [:jseward]]

Attachment: Example output before comment 2 patch (D137509)

Comment 4

[Julian Seward [:jseward]]

Attachment: Example output after comment 2 patch (D137509)

Comment 5

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/4e8e89273ddb
Make Ion's regalloc debug output easier to read.  r=jandem.

Comment 6

[Andreea Pavel [:apavel]]

https://hg.mozilla.org/mozilla-central/rev/4e8e89273ddb