LLVM 23: Adapt codegen test to moved assume

I wasn't completely sure what the important part of this test is - based on the "Make sure we don't create `br` or `select` instructions" comment I assume it's that the `load` is followed ~immediately by the `ret`, and that the presence or absence of the `assume` doesn't really matter (it's been moved up a few lines to before the `load`):

```ll
; Function Attrs
define noundef i32 @foo(ptr noalias nofree noundef align 8 captures(none) dereferenceable(16) %x) unnamed_addr #0 {
start
 %_6 = load ptr, ptr %x, align 8, !nonnull !4, !noundef !4
 %0 = getelementptr inbounds nuw i8, ptr %x, i64 8
 %_7 = load ptr, ptr %0, align 8, !nonnull !4, !noundef !4
 %_8 = icmp ne ptr %_6, %_7
 tail call void @llvm.assume(i1 %_8)
 %_14 = getelementptr inbounds nuw i8, ptr %_6, i64 4
 store ptr %_14, ptr %x, align 8
 %_4 = load i32, ptr %_6, align 4, !noundef !4
 ret i32 %_4
}
 ```

@rustbot label llvm-main

[original CI failure](https://buildkite.com/llvm-project/rust-llvm-integrate-prototype/builds/46608/list?sid=019ed950-c81f-466d-a3bc-2064e6a26719&tab=output)

Rollup of 10 pull requests

Successful merges:

 - #158410 (Update LLVM for Mach-O __LINKEDIT alignment fix.)
 - #157397 (cmse: clear padding when crossing the secure boundary)
 - #158036 (Add -Zinstrument-mcount=fentry to -Zinstrument-mcount)
 - #158330 (llvm: use intrinsics for f16, f32 minimum/maximum)
 - #158359 (fix(tests): allow either branch direction in ilog_known_base)
 - #158067 (LLVM 23: Adapt codegen test to moved assume)
 - #158261 (Move part of the target checking for `#[may_dangle]` to the parser)
 - #158358 (Fix invalid E0609 raw pointer deref suggestion inside macros)
 - #158392 (delegation: add tests for defaults and infers in generics)
 - #158394 (Generate synthetic generic args only for delegation's child segment)

Rollup merge of #158067 - TimNN:moved-assume, r=dianqk

LLVM 23: Adapt codegen test to moved assume

I wasn't completely sure what the important part of this test is - based on the "Make sure we don't create `br` or `select` instructions" comment I assume it's that the `load` is followed ~immediately by the `ret`, and that the presence or absence of the `assume` doesn't really matter (it's been moved up a few lines to before the `load`):

```ll
; Function Attrs
define noundef i32 @foo(ptr noalias nofree noundef align 8 captures(none) dereferenceable(16) %x) unnamed_addr #0 {
start
 %_6 = load ptr, ptr %x, align 8, !nonnull !4, !noundef !4
 %0 = getelementptr inbounds nuw i8, ptr %x, i64 8
 %_7 = load ptr, ptr %0, align 8, !nonnull !4, !noundef !4
 %_8 = icmp ne ptr %_6, %_7
 tail call void @llvm.assume(i1 %_8)
 %_14 = getelementptr inbounds nuw i8, ptr %_6, i64 4
 store ptr %_14, ptr %x, align 8
 %_4 = load i32, ptr %_6, align 4, !noundef !4
 ret i32 %_4
}
 ```

@rustbot label llvm-main

[original CI failure](https://buildkite.com/llvm-project/rust-llvm-integrate-prototype/builds/46608/list?sid=019ed950-c81f-466d-a3bc-2064e6a26719&tab=output)

…athanBrouwer

Revert "LLVM 23: Adapt codegen test to moved assume"

This reverts #158067.
Fixes #158500

The fix that was applied here is not robust against layout randomization, specifically the IR to load through the slice head pointer needs to come after a gepi, and not just a `load ptr, ptr %x`.

Based on the PR discussion, I think it's possible that the `min-llvm-version: 22` was mistakenly added when CI happened to choose the bad layout during the initial merge.

I'm not yet sure what the right fix is, but this is causing about half of all merges to fail right now so a revert seems prudent.

I don't think we have a way to skip tests when randomize-layout is set. The problem here is that we run codegen tests against the globally-configured sysroot, so setting `-Zrandomize-layout` in the test doesn't do anything.

…-fix, r=JonathanBrouwer

Revert "LLVM 23: Adapt codegen test to moved assume"

This reverts rust-lang#158067.
Fixes rust-lang#158500

The fix that was applied here is not robust against layout randomization, specifically the IR to load through the slice head pointer needs to come after a gepi, and not just a `load ptr, ptr %x`.

Based on the PR discussion, I think it's possible that the `min-llvm-version: 22` was mistakenly added when CI happened to choose the bad layout during the initial merge.

I'm not yet sure what the right fix is, but this is causing about half of all merges to fail right now so a revert seems prudent.

I don't think we have a way to skip tests when randomize-layout is set. The problem here is that we run codegen tests against the globally-configured sysroot, so setting `-Zrandomize-layout` in the test doesn't do anything.

…athanBrouwer

Revert "LLVM 23: Adapt codegen test to moved assume"

This reverts #158067.
Fixes #158500

The fix that was applied here is not robust against layout randomization, specifically the IR to load through the slice head pointer needs to come after a gepi, and not just a `load ptr, ptr %x`.

Based on the PR discussion, I think it's possible that the `min-llvm-version: 22` was mistakenly added when CI happened to choose the bad layout during the initial merge.

I'm not yet sure what the right fix is, but this is causing about half of all merges to fail right now so a revert seems prudent.

I don't think we have a way to skip tests when randomize-layout is set. The problem here is that we run codegen tests against the globally-configured sysroot, so setting `-Zrandomize-layout` in the test doesn't do anything.