JIT: profile synthesis blend and repair modes

[AndyAyersMS]

Implement blend and repair modes for synthesis. Blend merges a bit of synthesized data into an existing PGO data set; repair tries to fix any local inconsistencies (via heuristics). Both run count construction afterwards.

Trust blended data like we trust dynamic data. Probably will want more nuance here (eg trust dynamic blend, but not static blend) but this is sufficent for now.

Also implement random and reverse modes; these will ultimately be used for stress testing (not called anywhere yet).

Parameterize some of the magic constants that have cropped up.

Add blend mode as a new weekend pgo stress mode; fix the other synthesis mode I added recently to pgo stress to set the config properly.

Contributes to #82964.

[ghost]

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

Issue Details

---

Implement blend and repair modes for synthesis. Blend merges a bit of synthesized data into an existing PGO data set; repair tries to fix any local inconsistencies (via heuristics). Both run count construction afterwards.

Trust blended data like we trust dynamic data. Probably will want more nuance here (eg trust dynamic blend, but not static blend) but this is sufficent for now.

Also implement random and reverse modes; these will ultimately be used for stress testing (not called anywhere yet).

Parameterize some of the magic constants that have cropped up.

Add blend mode as a new weekend pgo stress mode; fix the other synthesis mode I added recently to pgo stress to set the config properly.

Contributes to #82964.

Author:	AndyAyersMS
Assignees:	AndyAyersMS
Labels:	area-CodeGen-coreclr
Milestone:	-

[AndyAyersMS]

@jakobbotsch PTAL
cc @dotnet/jit-contrib

Example of blend mode in action:

;; original
-----------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight     IBC  lp [IL range]     [jump]      [EH region]         [flags]
-----------------------------------------------------------------------------------------------------------------------------------------
BB01 [0000]  1                           30532. 30532    [000..004)-> BB03 ( cond )                     IBC 
BB02 [0001]  1       BB01                  0        0    [004..006)        (return)                     rare IBC 
BB03 [0002]  1       BB01                30532. 30532    [006..009)-> BB05 ( cond )                     IBC 
BB04 [0003]  1       BB03                30532. 30532    [009..00C)-> BB06 ( cond )                     IBC 
BB05 [0004]  2       BB03,BB04             0        0    [00C..00E)        (return)                     rare IBC 
BB06 [0005]  1       BB04                30532. 30532    [00E..01C)-> BB08 ( cond )                     IBC 
BB07 [0006]  1       BB06                  0        0    [01C..01E)        (return)                     rare IBC 
BB08 [0007]  1       BB06                30532. 30532    [01E..036)        (return)                     IBC 
-----------------------------------------------------------------------------------------------------------------------------------------

;; blend
-----------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight     IBC  lp [IL range]     [jump]      [EH region]         [flags]
-----------------------------------------------------------------------------------------------------------------------------------------
BB01 [0000]  1                           30532. 30532    [000..004)-> BB03 ( cond )                     IBC 
BB02 [0001]  1       BB01                 61.06    61    [004..006)        (return)                     IBC 
BB03 [0002]  1       BB01                30471. 30471    [006..009)-> BB05 ( cond )                     IBC 
BB04 [0003]  1       BB03                30410. 30410    [009..00C)-> BB06 ( cond )                     IBC 
BB05 [0004]  2       BB03,BB04           121.76   122    [00C..00E)        (return)                     IBC 
BB06 [0005]  1       BB04                30349. 30349    [00E..01C)-> BB08 ( cond )                     IBC 
BB07 [0006]  1       BB06                157.82   158    [01C..01E)        (return)                     IBC 
BB08 [0007]  1       BB06                30191. 30191    [01E..036)        (return)                     IBC 
-----------------------------------------------------------------------------------------------------------------------------------------

[AndyAyersMS]

Mono failure is known: #81123.
Wasm test timed out.

[AndyAyersMS]

3 diffs in asp.net, because fgProfileWeightsEqual comparison changed from < to <=. Seems odd that such a tiny change could have a large impact. The relevant check is here:

runtime/src/coreclr/jit/importer.cpp

Lines 13022 to 13033 in 3c38189

	// If the call site has profile data, report the relative frequency of the site.
	//
	if ((pInlineInfo != nullptr) && rootCompiler->fgHaveSufficientProfileWeights())
	{
	const weight_t callSiteWeight = pInlineInfo->iciBlock->bbWeight;
	const weight_t entryWeight = rootCompiler->fgFirstBB->bbWeight;
	profileFreq = fgProfileWeightsEqual(entryWeight, 0.0) ? 0.0 : callSiteWeight / entryWeight;
	hasProfile = true;
	assert(callSiteWeight >= 0);
	assert(entryWeight >= 0);
	}

We have

-----------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight   IBC  lp [IL range]     [jump]      [EH region]         [flags]
-----------------------------------------------------------------------------------------------------------------------------------------
BB01 [0041]  1                             0.01   0    [???..???)-> BB32 (always)                     i internal IBC 
BB02 [0004]  1       BB37                 15     15    [01E..067)-> BB36 ( cond )                     i idxlen bwd bwd-target IBC 

and so previously we'd determine that the entry weight was different than zero, and now we think it is not.

This is an OSR method and we currently don't have a good mechanism for determining the appropriate entry weight for such methods. We set the weight to 1% of the weight of the target block, guessing that the loop that inspired OSR must have executed at least 100 times.

Since there are just a handful of diffs I'll leave it as is.