[llvm] [VPlan] Don't use the legacy cost model for loop conditions (PR #156864)

[John Brawn via llvm-commits]

================
@@ -386,7 +386,7 @@ define i32 @diff_exit_block_needs_scev_check(i32 %end) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = trunc i32 [[END]] to i10
 ; CHECK-NEXT:    [[TMP1:%.*]] = zext i10 [[TMP0]] to i64
 ; CHECK-NEXT:    [[UMAX1:%.*]] = call i64 @llvm.umax.i64(i64 [[TMP1]], i64 1)
-; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[UMAX1]], 8
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[UMAX1]], 12
----------------
john-brawn-arm wrote:

Looking at a diff of the ``-debug`` output before and after the last marge it looks like there's two things going on:

```
Cost of 0 for VF 4: EMIT vp<%index.next> = add nuw vp<%4>, vp<%1>                                       Cost of 0 for VF 4: EMIT vp<%index.next> = add nuw vp<%4>, vp<%1>
Cost of 2 for VF 4: EMIT vp<%9> = any-of ir<%cmp3>                                                      Cost of 2 for VF 4: EMIT vp<%9> = any-of ir<%cmp3>
Cost of 1 for VF 4: EMIT vp<%10> = icmp eq vp<%index.next>, vp<%2>                                 |    Cost of 2 for VF 4: EMIT vp<%10> = icmp eq vp<%index.next>, vp<%2>
Cost of 0 for VF 4: EMIT vp<%11> = or vp<%9>, vp<%10>                                              |    Cost of 0 for VF 4: EMIT branch-on-two-conds vp<%9>, vp<%10>
Cost of 0 for VF 4: EMIT branch-on-cond vp<%11>                                                    <
```
Cost of the icmp is now 2. It looks like this is because of 524b1788c4f5a5fd4fdda5e51dd1e484b4124870: BranchOnTwoConds isn't listed in VPInstruction::usesFirstLaneOnly which causes vputils::onlyFirstLaneUsed to return false, so VPInstruction::computeCost calculates the cost of a vector compare. I think this can be fixed by adding BranchOnTwoConds to VPInstruction::usesFirstLaneOnly.


```
Calculating cost of work in exit block vector.early.exit:                                               Calculating cost of work in exit block vector.early.exit:
Cost of 12 for VF 4: EMIT vp<%17> = first-active-lane ir<%cmp>                                     |    Cost of 12 for VF 4: EMIT vp<%15> = first-active-lane ir<%cmp>
Cost of 0 for VF 4: EMIT vp<%18> = add vp<%4>, vp<%17>                                             |    Cost of 0 for VF 4: EMIT vp<%16> = add vp<%4>, vp<%15>
Cost of 0 for VF 4: vp<%19> = DERIVED-IV ir<3> + vp<%18> * ir<1>                                   |    Cost of 0 for VF 4: vp<%17> = DERIVED-IV ir<3> + vp<%16> * ir<1>
                                                                                                   >    Cost of 0 for VF 4: EMIT vp<%12> = extract-last-part ir<%ld2>
                                                                                                   >    Cost of 2 for VF 4: EMIT vp<%13> = extract-last-lane vp<%12>
Cost of 1 for VF 4: EMIT vp<%cmp.n> = icmp eq ir<64>, vp<%2>                                            Cost of 1 for VF 4: EMIT vp<%cmp.n> = icmp eq ir<64>, vp<%2>
Cost of 0 for VF 4: EMIT branch-on-cond vp<%cmp.n>                                                      Cost of 0 for VF 4: EMIT branch-on-cond vp<%cmp.n>
LV: Minimum required TC for runtime checks to be profitable:20                                     |    LV: Minimum required TC for runtime checks to be profitable:24
```
I haven't figured out what's going on here yet, will continue looking at it.

https://github.com/llvm/llvm-project/pull/156864