why not change the single return 0 in getNewAlignment() with return Align(1), and modify that type ?

And thanks for the contribution.

This isn't a new mode, it's still for measuring latency.

I'd rather have every user forced to think whether they want to use the value or its internal (log) representation, e.g. .p2align vs MIR format. If only for debug, then:

Please rephrase the commit header to something more informative.

Thanks.

Thanks for the explanation.

In D76625#1944389, @gchatelet wrote:

• BaseAlign is never None

In D76550#1935732, @lebedev.ri wrote:

@courbet To be honest

explicit MaybeAlign(uint64_t Value) {
  assert((Value == 0 || llvm::isPowerOf2_64(Value)) &&
         "Alignment is neither 0 nor a power of 2");
  if (Value)
    emplace(Value);
}

that if-check looks like a major rake.

LGTM when the comments are addressed.

Can you please add a test with an explicit parameter alignment ?

Actually I'm actually now convinced that the whole approach is incorrect when there are several blocks. I'm going to take some time to rethink this.

Thanks, PTAL.

address comments

clang-format

Use the right unsigned equivalent for comparing loads instead of the signed predicate from the zero relational compare.

Benchmark results from test-suite. TBH I'm not exactly sure why only the "greater" case gets better, but overall the results are promising.

clang-format

In D75519#1909956, @spatel wrote:

In D75519#1909488, @courbet wrote:

Simplify instructions after expandmemcmp.

LGTM, but as a matter of risk reduction, please make that an independent commit.

Simplify instructions after expandmemcmp.

Revert scheduling changes.

In D75519#1903306, @efriedma wrote:

Given memcmp expansion is running late in the pipeline, do you think we're possibly missing other relevant optimizations on the loads? LICM, maybe?

Update powerpc tests

Note that most test changes are NFC (just interleaving computations). I've highlighted the actual changes.

regenerate pgso tests

Thanks for the review !