Print indices of assertions instead of raw bitsets #54928

SingleAccretion · 2021-06-29T23:01:08Z

This is to improve the experience of reading assertion propagation dumps.

Before:

GenTreeNode creates assertion:
N004 (  5,  5) [000017] ------------              *  JTRUE     void  
In BB04 New Global Constant Assertion: (337, 64) ($151,$40) Oper_Bnd { {ADD($46, $1c1)} LT  {ARR_LENGTH($80)}ADD {IntCns -1}} is  {IntCns 0} index=#08, mask=0000000000000080
BB01 valueGen = 0000000000000005 => BB03 valueGen = 0000000000000003,
BB02 valueGen = 0000000000000000
BB03 valueGen = 0000000000000010 => BB05 valueGen = 0000000000000008,
BB06 valueGen = 0000000000000000
BB04 valueGen = 00000000000000A0 => BB04 valueGen = 0000000000000060,
BB05 valueGen = 0000000000000000
AssertionPropCallback::StartMerge: BB01 in -> 0000000000000000
AssertionPropCallback::EndMerge  : BB01 in -> 0000000000000000

After:

GenTreeNode creates assertion:
N004 (  5,  5) [000017] ------------              *  JTRUE     void  
In BB04 New Global Constant Assertion: ($151,$40) Oper_Bnd { {ADD($46, $1c1)} LT  {ARR_LENGTH($80)}ADD {IntCns -1}} is  {IntCns 0}, index = #08

BB01 valueGen = #01 #03 => BB03 valueGen = #01 #02
BB02 valueGen = #NA
BB03 valueGen = #05 => BB05 valueGen = #04
BB06 valueGen = #NA
BB04 valueGen = #06 #08 => BB04 valueGen = #06 #07
BB05 valueGen = #NA

AssertionPropFlowCallback:

StartMerge: BB01 in -> #NA
EndMerge  : BB01 in -> #NA

Before:

AssertionPropCallback::StartMerge: BB02 in -> 00000000000000FF
AssertionPropCallback::Merge     : BB02 in -> 00000000000000FF, predBlock BB01 out -> 0000000000000005
AssertionPropCallback::EndMerge  : BB02 in -> 0000000000000005

AssertionPropCallback::Changed   : BB02 before out -> 00000000000000FF; after out -> 0000000000000005;
		jumpDest before out -> 00000000000000FF; jumpDest after out -> 0000000000000005;

After

StartMerge: BB02 in -> #01 #02 #03 #04 #05 #06 #07 #08
Merge     : BB02 in -> #01 #02 #03 #04 #05 #06 #07 #08; pred BB01 out -> #01 #03
EndMerge  : BB02 in -> #01 #03

Changed   : BB02 before out -> #01 #02 #03 #04 #05 #06 #07 #08; after out -> #01 #03;
        jumpDest before out -> #01 #02 #03 #04 #05 #06 #07 #08; jumpDest after out -> #01 #03;

Before:

BB01 valueIn  = 0000000000000000 valueOut = 0000000000000005 => BB03 valueOut= 0000000000000003
BB02 valueIn  = 0000000000000005 valueOut = 0000000000000005
BB03 valueIn  = 0000000000000003 valueOut = 0000000000000013 => BB05 valueOut= 000000000000000B
BB06 valueIn  = 0000000000000013 valueOut = 0000000000000013
BB04 valueIn  = 0000000000000013 valueOut = 00000000000000B3 => BB04 valueOut= 0000000000000073
BB05 valueIn  = 0000000000000003 valueOut = 0000000000000003

After:

BB01 in = #NA; out = #01 #03 => BB03 out = #01 #02
BB02 in = #01 #03; out = #01 #03
BB03 in = #01 #02; out = #01 #02 #05 => BB05 out = #01 #02 #04
BB06 in = #01 #02 #05; out = #01 #02 #05
BB04 in = #01 #02 #05; out = #01 #02 #05 #06 #08 => BB04 out = #01 #02 #05 #06 #07
BB05 in = #01 #02; out = #01 #02

(This part I am unsure about - would it be worth it to increase the vertical size of the dump to align ins with outs?)

Before:

Propagating 0000000000000013 assertions for BB04, stmt STMT00004, tree [000020], tree -> 0
Propagating 0000000000000013 assertions for BB04, stmt STMT00004, tree [000082], tree -> 0
Propagating 0000000000000013 assertions for BB04, stmt STMT00004, tree [000037], tree -> 6
Propagating 0000000000000033 assertions for BB04, stmt STMT00004, tree [000034], tree -> 0
Propagating 0000000000000033 assertions for BB04, stmt STMT00004, tree [000035], tree -> 0

After:

Propagating #01 #02 #05 for BB04, stmt STMT00004, tree [000020], tree -> #NA
Propagating #01 #02 #05 for BB04, stmt STMT00004, tree [000082], tree -> #NA
Propagating #01 #02 #05 for BB04, stmt STMT00004, tree [000037], tree -> #06
Propagating #01 #02 #05 #06 for BB04, stmt STMT00004, tree [000034], tree -> #NA
Propagating #01 #02 #05 #06 for BB04, stmt STMT00004, tree [000035], tree -> #NA

Before:

Merging assertions from pred edges of BB04 for op [000066] $ffffffff
Merge assertions from pred BB04 JTrue edge: 0000000000000073
Constant Assertion: (337, 64) ($151,$40) Oper_Bnd { {ADD($46, $1c1)} LT  {ARR_LENGTH($80)}ADD {IntCns -1}} is not  {IntCns 0} index=#07, mask=0000000000000040
The range after edge merging:<Dependent, $c0 + -2>

After:

Merging assertions from pred edges of BB04 for op [000066] $ffffffff
Merge assertions from pred BB04 JTrue edge: #01 #02 #05 #06 #07
Constant Assertion: ($151,$40) Oper_Bnd { {ADD($46, $1c1)} LT  {ARR_LENGTH($80)}ADD {IntCns -1}} is not  {IntCns 0}, index = #07
The range after edge merging:<Dependent, $c0 + -2>

The downside to this style of printing is the horizontal size of the dump increases when working with methods that have a large number of assertions. I think it is worth it.

Naturally, no diffs for this change.

SingleAccretion · 2021-07-07T17:40:12Z

cc @AndyAyersMS

sandreenko

LGTM, I like the new dump on small methods.
cc @dotnet/jit-contrib

BruceForstall · 2021-07-08T00:26:17Z

I like the symbolic output.

IMO, we don't need to dump anything for the actual dataflow propagation; that output is only useful if you're debugging that dataflow algorithm, which shouldn't be frequent, whereas debugging which assertions are in/out could be more generally useful. I'd #if 0 that part.

I'd put in/out output (post-dataflow) on separate lines (and aligned).

SingleAccretion · 2021-07-08T10:54:27Z

I've pushed out two changes, as per Bruce's feedback:

1) Changed the format of the final output for the assertions:

BB01:
 in   = #NA
 out  = #NA
BB02:
 in   = #NA
 out  = #02
 BB06 = #01
BB03:
 in   = #02
 out  = #02 #03 #04 #05
BB04:
 in   = #02 #03 #04 #05
 out  = #02 #03 #04 #05
 BB06 = #01 #02 #03 #04 #05
BB05:
 in   = #02 #03 #04 #05
 out  = #02 #03 #04 #05
BB06:
 in   = #NA
 out  = #02
 BB08 = #01
BB07:
 in   = #02
 out  = #02 #06
BB08:
 in   = #NA
 out  = #NA
BB09:
 in   = #02 #03 #04 #05
 out  = #02 #03 #04 #05
BB10:
 in   = #02 #03 #04 #05
 out  = #02 #03 #04 #05 #07
BB11:
 in   = #02 #03 #04 #05 #07
 out  = #02 #03 #04 #05 #07

It is not what I would call "ideal", but it is aligned.

2) Removed the printing of dataflow from the regular dumps. This is intentionally conditioned on an instance method so that it can be easily enabled via the usual environment variable config switch mechanism if needed.

BruceForstall

Looks good!

BruceForstall · 2021-07-08T15:31:56Z

Can you rebase and ensure there are no formatting issues? (rerun jit-format)

For use in contexts where some printing method should only be executed when "verbose" is true.

To aid in understanding what assertions are being propagated and merged when reading the dumps.

It can still be enabled if needed.

SingleAccretion · 2021-07-08T21:04:34Z

CI failures do not appear related:

1) Libraries Test Run checked coreclr OSX x64 Debug - #55313.
2) CoreCLR Pri0 Runtime Tests Run OSX arm64 checked - completely timed out, no tests were run.

dotnet-issue-labeler bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Jun 29, 2021

SingleAccretion force-pushed the Improve-Assertion-Prop-Dump branch from 820ff19 to 11b1a20 Compare June 30, 2021 10:22

sandreenko approved these changes Jul 7, 2021

View reviewed changes

BruceForstall approved these changes Jul 8, 2021

View reviewed changes

SingleAccretion added 7 commits July 8, 2021 18:33

Add JITDUMPEXEC macro

b6bf91c

For use in contexts where some printing method should only be executed when "verbose" is true.

Add helpers for printing assertion indexes

89dabd6

Print assertion indices instead of raw bitsets

b698dde

To aid in understanding what assertions are being propagated and merged when reading the dumps.

Don't print VNs for the same assertion twice

da5e2a4

Also correctly print VNs in CopyProp

8620e64

Align "in"s with "out"s for final assertions

62f1ce1

Don't print the assertion dataflow in usual dumps

157746d

It can still be enabled if needed.

SingleAccretion force-pushed the Improve-Assertion-Prop-Dump branch from 2ad39fa to 157746d Compare July 8, 2021 15:33

BruceForstall merged commit f3b7775 into dotnet:main Jul 8, 2021

SingleAccretion deleted the Improve-Assertion-Prop-Dump branch July 8, 2021 21:22

ManickaP mentioned this pull request Jul 20, 2021

[QUIC] Remove AppContext switch from S.N.Quic #56027

Merged

ghost locked as resolved and limited conversation to collaborators Aug 7, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Print indices of assertions instead of raw bitsets #54928

Print indices of assertions instead of raw bitsets #54928

SingleAccretion commented Jun 29, 2021 •

edited

Loading

SingleAccretion commented Jul 7, 2021

sandreenko left a comment

BruceForstall commented Jul 8, 2021

SingleAccretion commented Jul 8, 2021

BruceForstall left a comment

BruceForstall commented Jul 8, 2021

SingleAccretion commented Jul 8, 2021 •

edited

Loading

Print indices of assertions instead of raw bitsets #54928

Print indices of assertions instead of raw bitsets #54928

Conversation

SingleAccretion commented Jun 29, 2021 • edited Loading

SingleAccretion commented Jul 7, 2021

sandreenko left a comment

Choose a reason for hiding this comment

BruceForstall commented Jul 8, 2021

SingleAccretion commented Jul 8, 2021

BruceForstall left a comment

Choose a reason for hiding this comment

BruceForstall commented Jul 8, 2021

SingleAccretion commented Jul 8, 2021 • edited Loading

SingleAccretion commented Jun 29, 2021 •

edited

Loading

SingleAccretion commented Jul 8, 2021 •

edited

Loading