Add edge likelihood dumping; fix one edge likelihood update case #99740

BruceForstall · 2024-03-14T02:41:40Z

Add dumping of edge likelihood numbers to the block table and other block dumping. Examples: BB17(0.143), BB18(1).

Edge likelihood dumping is parameterized in the code, but is currently always enabled.

Update setLikelihood and updateLikelihood to print the previous value of the likelihood that is being updated (if there is a previous value).
Fix a case of likelihood updating in fgReplaceJumpTarget(): in a switch block, when the new target block is already a target of the switch, then there already exists an edge from the switch to the new target. In that case, we were updating the edge dup count, but fgAddRefPred wasn't updating the edge likelihood to add the removed edge likelihood to the existing edge likelihood.

This was found as part of work to fix JitOptRepeat: #94250.

Add natvis debugger support for FlowEdge.
In 'fgDebugCheckOutgoingProfileData()`, if outgoing likelihoods are invalid, print out all the likelihoods of the outgoing edges.

1. Add dumping of edge likelihood numbers to the block table and other block dumping. Examples: `BB17(0.143)`, `BB18(1)`. Edge likelihood dumping is parameterized in the code, but is currently always enabled. 2. Update `setLikelihood` and `updateLikelihood` to print the previous value of the likelihood that is being updated (if there is a previous value). 3. Fix a case of likelihood updating in `fgReplaceJumpTarget()`: in a switch block, when the new target block is already a target of the switch, then there already exists an edge from the switch to the new target. In that case, we were updating the edge dup count, but `fgAddRefPred` wasn't updating the edge likelihood to add the removed edge likelihood to the existing edge likelihood. This was found as part of work to fix JitOptRepeat: dotnet#94250. 4. Add natvis debugger support for FlowEdge. 5. In 'fgDebugCheckOutgoingProfileData()`, if outgoing likelihoods are invalid, print out all the likelihoods of the outgoing edges.

dotnet-policy-service · 2024-03-14T02:42:13Z

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

BruceForstall · 2024-03-14T02:42:19Z

@amanasifkhalid @AndyAyersMS PTAL
cc @dotnet/jit-contrib

BruceForstall · 2024-03-14T02:52:33Z

Here's a bigger example:

---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight   IBC [IL range]   [jump]                            [EH region]        [flags]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BB01 [0000]  1                             1    250 [000..00E)-> BB06(0),*(1)            ( cond )                     i LIR IBC
BB02 [0001]  1       BB01                  1    250 [00E..01B)-> BB05(0),*(1)            ( cond )                     i LIR IBC
BB03 [0002]  1       BB02                  1    250 [01B..022)-> BB04(0),*(1)            ( cond )                     i LIR IBC hascall gcsafe
BB06 [0006]  2       BB01,BB03             0      0 [02F..030)-> *(1)                    (always)                     i LIR IBC rare q
BB07 [0007]  3       BB04,BB05,BB06        0      0 [030..034)-> BB09(0),*(1)            ( cond )                     i LIR IBC rare
BB08 [0008]  1       BB07                  0      0 [034..04C)                           (return)                     i LIR IBC rare
BB09 [0009]  1       BB07                  0      0 [04C..04C)-> BB17(0.143),*(1)        ( cond )                     i LIR IBC rare q
BB19 [0044]  1       BB09                  0      0 [04C..06A)-> *(0.286),*(0.286),BB13(0.143),BB14(0.143),BB15(0.143),BB16(0.143) (switch)                     i LIR IBC rare
BB11 [0043]  2       BB19(2)               0      0 [07D..080)-> BB18(1)                 (always)                     i LIR IBC rare
BB13 [0013]  1       BB19                  0      0 [094..0A8)-> BB18(1)                 (always)                     i LIR IBC rare
BB14 [0014]  1       BB19                  0      0 [0A8..0BD)-> BB18(1)                 (always)                     i LIR IBC rare
BB15 [0015]  1       BB19                  0      0 [0BD..0D1)-> BB18(1)                 (always)                     i LIR IBC rare
BB16 [0016]  1       BB19                  0      0 [0D1..0E6)-> BB18(1)                 (always)                     i LIR IBC rare
BB17 [0017]  1       BB09                  0      0 [0E6..0FE)-> *(1)                    (always)                     i LIR IBC rare q
BB18 [0018]  6       BB11,BB13,BB14,BB15,BB16,BB17   0      0 [0FE..100)                           (return)                     i LIR IBC rare
BB04 [0003]  1       BB03                  0      0 [022..029)-> BB07(1)                 (always)                     i LIR IBC rare hascall gcsafe
BB05 [0005]  1       BB02                  0      0 [02C..02F)-> BB07(1)                 (always)                     i LIR IBC rare
---------------------------------------------------------------------------------------------------------------------------------------------------------------------

in the full IR dump, you'll see things like:

------------ BB09 [0009] [04C..04C) -> BB17(0.1428571),BB19(1) (cond), preds={BB07} succs={BB19,BB17}

------------ BB19 [0044] [04C..06A) -> BB11(0.2857143),BB11(0.2857143),BB13(0.1428571),BB14(0.1428571),BB15(0.1428571),BB16(0.1428571) (switch), preds={BB09} succs={BB11,BB13,BB14,BB15,BB16}

This includes various fixes that are being separately PR'ed: 1. dotnet#99744: Introduce HandleKindDataIsInvariant helper 2. dotnet#99743: Add basic support for TYP_MASK constants 3. dotnet#99742: Fix problem with scaling general loop blocks; add dumpers 4. dotnet#99740: Add edge likelihood dumping; fix one edge likelihood update case Also: 1. Add support for running JitOptRepeat under JitStress. This is still over-written by JitOptRepeat being forced on at 4 iterations.

AndyAyersMS · 2024-03-14T06:08:24Z

Fix a case of likelihood updating in fgReplaceJumpTarget(): in a switch block, when the new target block is already a target of the switch, then there already exists an edge from the switch to the new target. In that case, we were updating the edge dup count, but fgAddRefPred wasn't updating the edge likelihood to add the removed edge likelihood to the existing edge likelihood.

Thanks for fixing this—it sounds exactly like the bug I hit in libraries jitstress in #99628.

jakobbotsch · 2024-03-14T09:16:09Z

src/coreclr/jit/optimizer.cpp

+    JITDUMP("Created new exit " FMT_BB " to replace " FMT_BB " exit for " FMT_LP "\n", newExit->bbNum, exit->bbNum,
+            loop->GetIndex());


optSetWeightForPreheaderOrExit references the block created, so moving its dumping here makes the jitdump look a bit strange I think.

I added the "Canonicalize exit" line above so it's obvious what process is starting when new blocks start getting created -- before I was confused at why new blocks were getting created. So moving this line to the end "brackets" the process by telling you what happened when it's done. E.g.,

Canonicalize exit BB15 for L00 to have only loop predecessors New Basic Block BB18 [0116] created. Setting edge weights for BB18 -> BB15 to [0 .. 3.402823e+38] setting likelihood of BB18 -> BB15 to 1 Estimated likelihood BB04 -> BB18 to be 0.6585233 (contribution: 35.62623) Setting edge weights for BB04 -> BB18 to [35.62623 .. 35.62623] Estimated likelihood BB05 -> BB18 to be 0.3993941 (contribution: 10.80365) Setting edge weights for BB05 -> BB18 to [10.80365 .. 10.80365] Setting edge weights for BB18 -> BB15 to [46.42988 .. 46.42988] Created new exit BB18 to replace BB15 exit for L00

Ok, seems fine -- I think you should change it in optCreatePreheader as well then.

amanasifkhalid

LGTM, thanks!

amanasifkhalid · 2024-03-14T11:40:36Z

src/coreclr/jit/fgbasic.cpp

+                assert(newEdge->getSourceBlock() == block);
+                assert(newEdge->getDestinationBlock() == newTarget);
+
+                if (newEdge->hasLikelihood() && oldEdge->hasLikelihood())


I wonder if we can start skipping these checks and asserting that the edges have likelihoods already. Once Andy's last PR for propagating likelihoods is merged in, I'll take a look at removing these.

BruceForstall · 2024-03-14T16:48:09Z

No diffs

BruceForstall · 2024-03-14T16:55:42Z

Failures are known or infra.

dotnet-issue-labeler bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Mar 14, 2024

BruceForstall requested review from amanasifkhalid and AndyAyersMS March 14, 2024 02:41

dotnet-policy-service bot assigned BruceForstall Mar 14, 2024

build-analysis bot mentioned this pull request Mar 14, 2024

System.DirectoryServices.Protocols.Tests.BerConverterTests.Decode_Bytes_ReturnsExpected failing on Windows x86 #99725

Closed

AndyAyersMS approved these changes Mar 14, 2024

View reviewed changes

jakobbotsch reviewed Mar 14, 2024

View reviewed changes

amanasifkhalid approved these changes Mar 14, 2024

View reviewed changes

BruceForstall merged commit b4bc0aa into dotnet:main Mar 14, 2024
124 of 129 checks passed

BruceForstall deleted the FixSwitchLikelihoods branch March 14, 2024 16:55

AndyAyersMS mentioned this pull request Mar 14, 2024

JIT: enable edge checks throughout #99628

Merged

github-actions bot locked and limited conversation to collaborators Apr 14, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add edge likelihood dumping; fix one edge likelihood update case #99740

Add edge likelihood dumping; fix one edge likelihood update case #99740

BruceForstall commented Mar 14, 2024

dotnet-policy-service bot commented Mar 14, 2024

BruceForstall commented Mar 14, 2024

BruceForstall commented Mar 14, 2024

AndyAyersMS commented Mar 14, 2024

jakobbotsch Mar 14, 2024

BruceForstall Mar 14, 2024

jakobbotsch Mar 14, 2024

amanasifkhalid left a comment

amanasifkhalid Mar 14, 2024

BruceForstall commented Mar 14, 2024

BruceForstall commented Mar 14, 2024

		JITDUMP("Created new exit " FMT_BB " to replace " FMT_BB " exit for " FMT_LP "\n", newExit->bbNum, exit->bbNum,
		loop->GetIndex());

Add edge likelihood dumping; fix one edge likelihood update case #99740

Add edge likelihood dumping; fix one edge likelihood update case #99740

Conversation

BruceForstall commented Mar 14, 2024

dotnet-policy-service bot commented Mar 14, 2024

BruceForstall commented Mar 14, 2024

BruceForstall commented Mar 14, 2024

AndyAyersMS commented Mar 14, 2024

jakobbotsch Mar 14, 2024

Choose a reason for hiding this comment

BruceForstall Mar 14, 2024

Choose a reason for hiding this comment

jakobbotsch Mar 14, 2024

Choose a reason for hiding this comment

amanasifkhalid left a comment

Choose a reason for hiding this comment

amanasifkhalid Mar 14, 2024

Choose a reason for hiding this comment

BruceForstall commented Mar 14, 2024

BruceForstall commented Mar 14, 2024