[HVX] Fix EliminateInterleaves

[rootjalex]

Minor fix for an edge case that introduces more shuffling.

Fixes #7100

[steven-johnson]

Huh, the win32 failure is one I've never seen before:

FAILED: bin/li2018_demo_gradient_autoscheduler.exe 
li2018_demo_gradient.runtime.lib(li2018_demo_gradient.runtime.obj) : error LNK2019: unresolved external symbol ___floatundisf referenced in function _halide_profiler_report_unlocked

[rootjalex]

Huh, the win32 failure is one I've never seen before:

FAILED: bin/li2018_demo_gradient_autoscheduler.exe 
li2018_demo_gradient.runtime.lib(li2018_demo_gradient.runtime.obj) : error LNK2019: unresolved external symbol ___floatundisf referenced in function _halide_profiler_report_unlocked

That is super weird, and entirely unrelated to this PR... Strange.

[steven-johnson]

Huh, the win32 failure is one I've never seen before:

FAILED: bin/li2018_demo_gradient_autoscheduler.exe 
li2018_demo_gradient.runtime.lib(li2018_demo_gradient.runtime.obj) : error LNK2019: unresolved external symbol ___floatundisf referenced in function _halide_profiler_report_unlocked

That is super weird, and entirely unrelated to this PR... Strange.

Maybe related to something new/different in LLVM codegen?

[steven-johnson]

Runtime Function: float __floatundisf (unsigned long i)
Runtime Function: double __floatundidf (unsigned long i)
Runtime Function: long double __floatunditf (unsigned long i)
Runtime Function: long double __floatundixf (unsigned long i)
These functions convert i, an unsigned long, to floating point.

[rootjalex]

That test is this file: https://github.com/halide/Halide/blob/main/test/autoschedulers/li2018/demo_generator.cpp , right? I don't even see an int -> float conversion in that code.

[rootjalex]

Oh it's in the profiler. Hmm

[steven-johnson]

So I experimentally crosscompiled _halide_profiler_report_unlocked to x86-32-windows-profile -- I don't see any calls to floatundisf there, but looking at the .ll file, there are a number of uitofp instructions, which could be the culprit. That said... that would be used all over the place, so why is the failure being reported only there?

EDIT: aha, objdump -t shows ___floatundisf in the syms. Now let me see if I can bisect where it first appeared.

[steven-johnson]

Anyway, this is clearly unrelated, feel free to land.

[steven-johnson]

Still finishing the bisect but the culprit is almost certainly https://reviews.llvm.org/D141074