Tentative implementation of HasIRContext.

[math-fehr]

Currently, we have IRContext, WfIRContext, Rewriter, and WfRewriter that all "contains" an IRContext.
What we would like to have is a coercion from all of these to IRContext. However, as IRContext has a parameter that we always implicitely deduce, the coercion almost never works.

One of the solution would be to reimplement all getters so that they can use these different classes, but I wanted to try if there was another solution before trying this.

The implementation here uses a new typeclass HasIRContext, which is essentially just the coercion from the types to IRContext.

Please let me know what you think of the design

[math-fehr]

!radar

[ineol]

Benchmark results for 3a90a7a against c2650f2 are in! @math-fehr

Large changes (13🟥)

• 🟥 build/module/Veir.IR.Basic//instructions: +1.3G (+3.87%)
• 🟥 build/module/Veir.IR.Basic//lines: +19.0 (+0.91%)
• 🟥 build/module/Veir.IR.Fields//instructions: +3.7G (+4.07%)
• 🟥 build/module/Veir.IR.InBounds//instructions: +13.6G (+4.44%)
• 🟥 build/module/Veir.Rewriter.InlineBlock//instructions: +2.3G (+2.12%)
• 🟥 build/module/Veir.Rewriter.LinkedList.Basic//instructions: +534.3M (+2.69%)
• 🟥 build/module/Veir.Rewriter.LinkedList.GetSet//instructions: +3.4G (+2.81%)
• 🟥 build/module/Veir.Rewriter.WellFormed.BlockOperands//instructions: +2.3G (+1.57%)
• 🟥 build/module/Veir.Rewriter.WellFormed.OpResults//instructions: +397.3M (+3.38%)
• 🟥 build/module/Veir.Rewriter.WellFormed.Operation//instructions: +1.9G (+1.12%)
• 🟥 build/module/Veir.Rewriter.WellFormed.ReplaceOp//instructions: +377.8M (+2.91%)
• 🟥 build/module/Veir.Rewriter.WellFormed.Value//instructions: +334.4M (+0.69%)
• 🟥 build/module/Veir.Rewriter.WfRewriter//instructions: +888.7M (+11.50%)

Medium changes (1🟥)

• 🟥 build/module/Veir.Rewriter.WellFormed.OpRegion//instructions: +266.5M (+2.77%)

Small changes (7🟥)

• 🟥 build/module/Veir.IR.GetSet//instructions: +3.5G (+1.27%)
• 🟥 build/module/Veir.IR.Grind//instructions: +8.1M (+0.60%)
• 🟥 build/module/Veir.PatternRewriter.Basic//instructions: +157.6M (+1.19%)
• 🟥 build/module/Veir.Rewriter.LinkedList.WellFormed//instructions: +3.0G (+0.67%)
• 🟥 build/module/Veir.Rewriter.WellFormed.Block//instructions: +114.3M (+1.10%)
• 🟥 build/module/Veir.Rewriter.WellFormed.OpOperands//instructions: +777.0M (+2.01%)
• 🟥 build/module/Veir.Rewriter.WellFormed.Region//instructions: +94.8M (+1.44%)

[github-actions]

VeIR Benchmarks

Details

Benchmark suite	Current: 205d7c1	Previous: 964db92	Ratio
add-fold-worklist/create	2611000 ns	2240000 ns	1.17
add-fold-worklist/rewrite	3806000 ns	3814000 ns	1.00
add-fold-worklist-local/create	2895000 ns	2265000 ns	1.28
add-fold-worklist-local/rewrite	3190000 ns	3205000 ns	1.00
add-zero-worklist/create	3281000 ns	2441000 ns	1.34
add-zero-worklist/rewrite	2550000 ns	2605000 ns	0.98
add-zero-reuse-worklist/create	2584000 ns	2025000 ns	1.28
add-zero-reuse-worklist/rewrite	2098000 ns	2000000 ns	1.05
mul-two-worklist/create	2855000 ns	2439000 ns	1.17
mul-two-worklist/rewrite	5324000 ns	5524000 ns	0.96
add-fold-forwards/create	3034000 ns	2262000 ns	1.34
add-fold-forwards/rewrite	3083000 ns	3141000 ns	0.98
add-zero-forwards/create	3050000 ns	2239000 ns	1.36
add-zero-forwards/rewrite	2062000 ns	2026000 ns	1.02
add-zero-reuse-forwards/create	2593000 ns	1850000 ns	1.40
add-zero-reuse-forwards/rewrite	1711000 ns	1580000 ns	1.08
mul-two-forwards/create	3001000 ns	2285000 ns	1.31
mul-two-forwards/rewrite	3717000 ns	3771000 ns	0.99
add-zero-reuse-first/create	2742000 ns	1865000 ns	1.47
add-zero-reuse-first/rewrite	8000 ns	8000 ns	1
add-zero-lots-of-reuse-first/create	2669000 ns	1834000 ns	1.46
add-zero-lots-of-reuse-first/rewrite	955000 ns	788000 ns	1.21

This comment was automatically generated by workflow using github-action-benchmark.

[math-fehr]

!radar

[ineol]

Benchmark results for 876c9b5 against c2650f2 are in! @math-fehr

Large changes (23🟥)

Too many entries to display here. View the full report on radar instead.

Medium changes (1🟥)

• 🟥 build/profile/grind mark subsingleton//wall-clock: +2s (+14.10%)

Small changes (5🟥)

• 🟥 build/module/Veir.Benchmarks//instructions: +90.1M (+0.97%)
• 🟥 build/module/Veir.Printer//instructions: +43.0M (+0.89%)
• 🟥 build/module/Veir.Rewriter.Basic//instructions: +4.0G (+7.91%)
• 🟥 build/module/Veir.Rewriter.InsertPoint//instructions: +1.4G (+3.09%)
• 🟥 build/module/Veir.Rewriter.WellFormed.IRContext//instructions: +28.9M (+1.37%)

[ineol]

Why did you replace all the grind call in IR/Basic.lean?

[math-fehr]

I'll update it again good point. The issue was originally that InBounds cannot be used by grind anymore, so I changed them.

But now I realized that you can use inBounds_def if the typeclass is already in grind, and it seems I missed some, I'll fix that.

[tobiasgrosser]

Nice. This seems to go in a great direction.

[math-fehr]

Just so you know, I'm first testing this on PatternRewriter before opening this for review.
I want to be sure that if we merge this, this will not cause us problems in the future. Right now this seems to work well for IRContext at least. I also ran the benchmarks and had no performance loss.