v0.0.0-7603-g0686143f0

Integrate LLVM at llvm/llvm-project@2d287f51eff2

Updates LLVM usage to match
[2d287f51eff2](https://github.com/llvm/llvm-project/commit/2d287f51eff2)

PiperOrigin-RevId: 756508479

v0.0.0-7593-g727265954

Improve const-ness of module in run routines, and some of its methods.

PiperOrigin-RevId: 756032997

v0.0.0-7576-ge200ad5ab

Automated rollback of commit 97a4b6efdb652a9ae06092712a9a785558a49f40.

PiperOrigin-RevId: 755680460

v0.0.0-7566-gb94271454

Improve metrics reporting of the optimization pipeline.

This includes a number of fixes/changes:
* Gather metrics in un-aggregated form, one proto for each time a pass is run. The data is organized hierarchically mirroring the hiearchical structure of the pass pipeline.
* Add a utility which dumps a table from the proto metrics data. The utility produces two tables: one aggregated by pass name, and one hierarchically organized.
* Fix bug where the invariant checkers were not being run for some passes, specifically those within CapOptLevel or IfOptLevelAtLeast compound passes. As part of this fix clean up the pass_base code a bit. Make RunNested a method on PassBase which makes CompoundPassBase and PassBase a bit more interchangeable.
* Add numerous test for various bits of PassBase functionality which were previously untested.

PiperOrigin-RevId: 755096116

v0.0.0-7559-g54d29b76e

[XLS] Simplify apfloat using abs_diff

APFloat previously used a custom sign-magnitude difference function; this is now redundant with the new absolute difference library, so we switch to use that instead.

To enable this, we also add `abs_diff` to the list of built-in libraries, so it will be imported from the DSLX stdlib location.

PiperOrigin-RevId: 754159246

v0.0.0-7544-g2698b115a

Refactored block_conversion and added accessor functions for proc-scoped

channels to nodes.

PiperOrigin-RevId: 753768577

v0.0.0-7538-g8e0a48b8c

Update more typecheck module tests to run in both versions. Document …

v0.0.0-7533-g12b2fa99c

[scheduling] Add optional schedule optimization for dynamic throughput

Adds a new codegen/scheduling flag `--dynamic_throughput_objective_weight`. This controls how strongly the SDC scheduler will attempt to optimize for dynamic throughput as opposed to area. If set to `K`, XLS will roughly consider decreasing the dynamic throughput of one state element by one cycle to be worth adding up to `K` flops, under the assumption that all data-dependent state-feedback paths are equally likely.

This was motivated by the discovery of a real-world case where it's possible to improve throughput substantially while sacrificing approximately one flop's worth of area.

PiperOrigin-RevId: 752962519

v0.0.0-7525-g6fd20bd2b

Integrate LLVM at llvm/llvm-project@c60f24dca96d

Updates LLVM usage to match
[c60f24dca96d](https://github.com/llvm/llvm-project/commit/c60f24dca96d)

PiperOrigin-RevId: 751669474

v0.0.0-7517-ga4c74e757

Disable equivalence verification in tests when MSAN is enabled

Z3 is enough slower when MSAN is enabled that this substantially increases timeouts - and does not improve coverage, since we also run these tests with MSAN disabled, and Z3's proof's validity should not depend on whether the sanitizer is enabled.

PiperOrigin-RevId: 751071676