[TypePromotion] Don't insert Truncate for a no-op ZExt

[avieira]

Summary

Hi,

As the title says, I don't believe there's a point in inserting Truncates before ZExt's that have no-effect, i.e. ZExts where the argument's type width is the same as the result's type width.

This is a prerequisite for the test in D111237 to pass.

[samparker]

Thanks.

[avieira]

I'm just gonna rebase this to see if I also get the failure on TypePromotion/ARM/calls.ll that the bot is reporting.

[avieira]

I had forgotten to build for ARM too, which had similar testisms as AArch64 where this patch removes unnecessary Truncates. Updated tests.

[samparker]

Good catch.

[barannikov88]

llvm/lib/CodeGen/TypePromotion.cpp

554

The comment is kind of misleading. According to LangRef, the size of the input operand of zext must be less than the size of the output operand. It is also asserted in the constructor of the operation (see CastInst::castIsValid).

[avieira]

Is that somewhat clearer? I suspect the reason we never see a failure is because we 'replace' the input, rather than create a new ZExt. But like I said in the comment, it get's deleted. And as it is the trunc would just stay there dangling and get in the way of things.

[barannikov88]

In D131487#3713019, @avieira wrote:

Is that somewhat clearer? I suspect the reason we never see a failure is because we 'replace' the input, rather than create a new ZExt. But like I said in the comment, it get's deleted. And as it is the trunc would just stay there dangling and get in the way of things.

With that comment is is pretty much clear now, thanks.