Handle review comments.

Abandoned in favor of: D102055.

Thanks I will abandon my review. This fixes the issue for me.

I use: Apple clang version 11.0.3.

I might have misunderstood something. Let me know if I should abandon this in case this a local issue on my machine or there is a better solution.

Handle review comments.

Add more detailed docs.

Handle review comments:

• Fold the 2 phases of the cost-mode into one.

Handle review comments.

Add more tests.

Handle Sean's comments:

• Refactor control-flow model.
• Dedeuce branch op detensoring from block argument detensoring.
• Add more docs.

Handle rridle's comments:

• Use ArrayRef for argument type.
• Use function_ref instead of DenseMap.
• Remove unneeded includes.
• Avoid dynamic memory allocation.

• Rearrange code a bit.

Should be valid with sexti as well right?

Your folder does not do that actually. It is just the framework that does not perform such replacement for all folders.
(side note, you could write simply : If the type of trunci's destination does not match zexti's source types, no folding takes place. ; the intermediate type does not matter here)

• Handle review comments.
• Rebase.

I came across this pattern while working on detensoring in linalg-on-tensors for IREE. Let me know if there are any reasons why this might be a bad idea.

We will halt working on this until a discussion on a good assembly scheme is discussed with a wider audience.

@antiagainst, 2 points:

• I adopted the set of rules you suggested, we can refine them as we discuss on the review.
• I tried to define GlobalVariable using the declarative format, but it seems that is not possible due to the optional directives: bind and built_in. Let me know if I misunderstood something.

Rebase.

In D98017#2608850, @antiagainst wrote:

Actually we can probably keep spv.module as-is. Similarly to spv.func, they maintain consistency with other module/func op names. What do you think?

Fix doc string.

Fix comments.

Rename to spv.mlir.loop.

• Rebase.
• Add a TODO.

Rename function.

Clean-up.

Handle review comments.

Handle some review comments.

Handle review comments.

Update docs.

Properly handle non-entry function blocks.

Assert on 0 inputs.

Simplify converted code by merging some blocks.

Handle review comments:

@silvas Your last comments all seem to be on an old diff :). FuncBufferize changes were already reverted and the detensoring code was moved to its own file.

Thanks @nicolasvasilache and @silvas again for your help in calibrating this to better fit the intended needs. I keep adding and deleting code but that's what happens when you are a newbie :).

• Don't detensorize control flow.
• Don't detensorize across function boundaries.
• Support only rank0 for now.