[ARITH] detect iter affine map with predicate

[spectrometerHBH]

Update on previous PR #6667

Enhance the detect_iter_map utility to be able to detect split patterns with predicates.

cc @tqchen

[tqchen]

Thanks @spectrometerHBH . @MasterJH5574 @hzfan @Hzfengsy please also help to take a look when you have time

[tqchen]

some minor comments

[tqchen]

move to src/analysis/expr_complexity.cc analysis.h and expose as a function size_t ExprComplexity(const PrimExpr& expr); document as number of expressions in the child

[FrozenGene]

Yes. In my previous pr #7469 I have done something. We could borrow it. My pr has the left problem of max ops counter, however, it is not an urgent problem IMO, which just a more safer guard for complex expr analysis.

[tqchen]

Thanks @spectrometerHBH There seems to be three places where the ExprComplexity is involved.

• https://github.com/apache/tvm/blob/main/src/arith/solve_linear_inequality.cc#L55
• [tir]Analysis expr complexity support and simplify pass is turned on based on condition #7469

Given your PrimExprSizeCounter counting is simpler, can we create an ExprComplexity function in analysis and call from there?

[tqchen]

cc @FrozenGene @yzhliu

[MasterJH5574]

Thanks @spectrometerHBH! Overall LGTM. Just some nits.

[spectrometerHBH]

Thanks @spectrometerHBH There seems to be three places where the ExprComplexity is involved.

• https://github.com/apache/tvm/blob/main/src/arith/solve_linear_inequality.cc#L55
• [tir]Analysis expr complexity support and simplify pass is turned on based on condition #7469

Given your PrimExprSizeCounter counting is simpler, can we create an ExprComplexity function in analysis and call from there?

@tqchen I'm not sure whther current ExprComplexity is compatible with mine. Looks like current ExprComplexity doesn't count Nodes like Load/BufferLoad/Cast....

[tqchen]

Thanks @spectrometerHBH I think calculating all exprs is the right def. Skipping load do makes sense for index expressions, but my guess is that we can always try the other way if it turns out to be not good

[tqchen]

Thanks @spectrometerHBH @FrozenGene @MasterJH5574 this PR is now merged