[llvm] [VPlan] First step towards VPlan cost modeling. (PR #92555)

[Arthur Eubanks via llvm-commits]

aeubanks wrote:

```
$ cat /tmp/a.ll
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

%"class.llvm::MCOperand" = type { i8, %union.anon.125 }
%union.anon.125 = type { i64 }

define ptr @_ZNK4llvm6MCInst10getOperandEj(i32 %i) {
entry:
  %conv = zext i32 %i to i64
  %call = call ptr @_ZNK4llvm25SmallVectorTemplateCommonINS_9MCOperandEvEixEm(i64 %conv)
  ret ptr %call
}

define i32 @_ZL13getOpIdxForMORKN4llvm6MCInstERKNS_9MCOperandE(ptr %MI, i32 %call) {
entry:
  br label %for.cond

for.cond:                                         ; preds = %for.inc, %entry
  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.inc ]
  %cmp = icmp ult i32 %i.0, %call
  call void @llvm.assume(i1 %cmp)
  %call12 = call ptr @_ZNK4llvm6MCInst10getOperandEj(i32 %i.0)
  %cmp2.not = icmp eq ptr %call12, %MI
  br i1 %cmp2.not, label %cleanup3, label %for.inc

for.inc:                                          ; preds = %for.cond
  %inc = add i32 %i.0, 1
  br label %for.cond

cleanup3:                                         ; preds = %for.cond
  ret i32 0
}

define ptr @_ZNK4llvm25SmallVectorTemplateCommonINS_9MCOperandEvEixEm(i64 %idx) {
entry:
  %arrayidx = getelementptr nusw %"class.llvm::MCOperand", ptr null, i64 %idx
  ret ptr %arrayidx
}

; Function Attrs: nocallback nofree nosync nounwind willreturn memory(inaccessiblemem: write)
declare void @llvm.assume(i1 noundef) #0

attributes #0 = { nocallback nofree nosync nounwind willreturn memory(inaccessiblemem: write) }

$ clang -fprofile-generate -O3  -S -emit-llvm -o /dev/null /tmp/a.ll
clang++: ../../llvm/lib/Transforms/Vectorize/LoopVectorize.cpp:10361: bool llvm::LoopVectorizePass::processLoop(Loop *): Assertion `VF.Width == Width && "VPlan cost model and legacy cost model disagreed"' failed.
```

I couldn't reproduce with `opt` for some reason, even when inserting the instrprof pass. I also couldn't reproduce by dumping the IR right before, so perhaps there's some incorrect analysis invalidation going on

https://github.com/llvm/llvm-project/pull/92555