[AggressiveInstCombine] Combine consecutive loads which are being merged to form a wider load.

[bipmis]

Summary

The patch simplifies some of the patterns as below

1. (zExt(L1) << shift1) | (zExt(L2) << shift2) -> zExt(L3) << shift1
2. (? | (zExt(L1) << shift1)) | (zExt(L2) << shift2) -> ? | (zExt(L3) << shift1)

The pattern is indicative of the fact that the loads are being merged to a wider load and the only use of this pattern is with a wider load. In this case for a non-atomic/non-volatile loads reduce the pattern to a combined load which would improve the cost of inlining, unrolling, vectorization etc.

[spatel]

I'd like to see some form of load combining in IR, but we need to be very careful about how that is implemented.

Note that LLVM had a dedicated pass for load combining, but it caused problems and was removed:
https://lists.llvm.org/pipermail/llvm-dev/2016-September/105291.html

So any proposal to do it again should be aware of and avoid those problems. Putting it in InstCombine is controversial because it could interfere with other analysis/passes. We recently added a cost-aware transform to AggressiveInstCombine, so that might be a better fit.

I didn't check all of the tests, but at least the simplest cases should already be handled in DAGCombiner. Can you show a larger, motivating example where that backend transform is too late to get to the ideal code?

[bipmis]

@spatel Thanks for reviewing it. The common scenarios in llvm where we see this as an issue is with the cost involved in the inliner, unrolller and vectorizer which results in a sub-optimal code. One such simple example can be with dot product where vectorizer fails to generate a vectorized code as below:
https://godbolt.org/z/Ee4cbf1PG

Similar situation can be seen with the unroller as well as in the example below:
https://godbolt.org/z/dWo8n3Yz8

[spatel]

In D127392#3581102, @bipmis wrote:

https://godbolt.org/z/Ee4cbf1PG
Similar situation can be seen with the unroller as well as in the example below:
https://godbolt.org/z/dWo8n3Yz8

Thanks - those look like good motivating examples. I recommend moving this patch over to AggressiveInstCombine or recreating a limited form of the old LoadCombine pass. Either way, we need to be careful to make sure that the transform does not combine loads illegally and that it doesn't harm other analysis. You might want to start an RFC thread on discourse for more feedback and tag people who commented on earlier proposals to do this kind of transform.

[tschuett]

I am in favour of moving the discussion to discourse.

AggressiveInstCombine and target-dependent InstCombine are two independent concepts for me.

[RKSimon]

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

17 ↗

(On Diff #435486)

clang-format this?

[bipmis]

This is being discussed in the discourse -https://discourse.llvm.org/t/load-widening-in-ir/61952

[nikic]

The main question I see here is whether this needs to be TTI based or not -- if yes, then it also shouldn't be in InstCombine. I think there's two reason why we might want TTI:

• Do we want to create unaligned loads? Creating them is legal, but if the target does not support fast unaligned loads, the backend will break them up again. Should we only perform this optimization if TTI.allowsMisalignedMemoryAccesses reports a fast unaligned access?
• Do we want to create loads with illegal sizes? For example, if we have a 64-bit target, so we want to combine two i64 loads into an i128 load, even though it will later be broken up again? (At least for the current implementation where both loads must have the same size, the question of whether to allow i24 loads or similar does not come up.)

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

3004 ↗

(On Diff #435486)

nullptr

3017 ↗

(On Diff #435486)

Move dyn_cast into initialization, having it in the if is hard to read.

3029 ↗

(On Diff #435486)

GEPOperator

Also directly dyn_cast in the condition.

3035 ↗

(On Diff #435486)

You are probably looking for GEP2->accumulateConstantOffsets() here -- your current code will treat GEPs with multiple indices incorrectly.

3077 ↗

(On Diff #435486)

I don't get this check -- you're checking that there is an available value, but not that this value is the previous load. There might still be a clobber in between.

What you're probably looking for is canInstructionRangeModRef() between the two instructions.

3108 ↗

(On Diff #435486)

Unnecessary cast

3114 ↗

(On Diff #435486)

Why does this not use the builder to create the instruction, only to insert it?

3119 ↗

(On Diff #435486)

Uh, what's the point of oring something with zero?

[RKSimon]

llvm/test/Transforms/InstCombine/or-load.ll

2 ↗

(On Diff #435486)

add little and big endian test coverage

[bipmis]

In D127392#3607415, @nikic wrote:

The main question I see here is whether this needs to be TTI based or not -- if yes, then it also shouldn't be in InstCombine. I think there's two reason why we might want TTI:

• Do we want to create unaligned loads? Creating them is legal, but if the target does not support fast unaligned loads, the backend will break them up again. Should we only perform this optimization if TTI.allowsMisalignedMemoryAccesses reports a fast unaligned access?
• Do we want to create loads with illegal sizes? For example, if we have a 64-bit target, so we want to combine two i64 loads into an i128 load, even though it will later be broken up again? (At least for the current implementation where both loads must have the same size, the question of whether to allow i24 loads or similar does not come up.)

Thanks for the review. I have made most of the other suggested changes and can post a patch for the same. It would also handle some more test cases.

Right I think we need to be sure on this particular path. If we want a TTI based should Aggressive Instcombine be a better choice. We also see some use scenarios of TTI in InstCombine but is our case a vaild use case scenario.
Next patch would be in InstCombine for this reason.
I will also test the patch up with extended tests to see if we are seeing any performance issues and if the backend breaking up happens ok.

[xbolva00]

We also see some use scenarios of TTI in InstCombine

I believe there is a strong consensus to say no to TTI in InstCombine.

[bipmis]

llvm/test/Transforms/InstCombine/or-load.ll

2 ↗

(On Diff #435486)

In IR I believe the checks would be the same for LE and BE. The differences should pop up in the assembly.
https://godbolt.org/z/xboKx47cc

[bipmis]

Handle some of the review comments.
Added support for additional scenarios like reverse order loads and big-endian support.

Moving the implementation to AgressiveInstCombine still open for discussion in addition with the TTI requirement.

[spatel]

In D127392#3625158, @bipmis wrote:

Handle some of the review comments.
Added support for additional scenarios like reverse order loads and big-endian support.

Moving the implementation to AgressiveInstCombine still open for discussion in addition with the TTI requirement.

IMO, we should move this patch to AggressiveInstCombine with conservative TTI limitations to start.

Now that we've heard from several others, I'd summarize as:

1. There's good motivation to have load combining in IR as a canonicalization.
2. But it should be limited with TTI (avoid unaligned loads, illegal operations/types).
3. Adding TTI to InstCombine is a non-starter; we don't want to have that level of target-dependence in IR canonicalization.
4. AggressiveInstCombine is lower-weight canonicalization pass that recently gained access to TTI.
5. Potential follow-ups could relax the TTI constraints and/or allow bswap formation to deal with endian diffs seen in the current tests (we do that type of transform in the backend already).

[bipmis]

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

3035 ↗

(On Diff #435486)

Currently we are handling GEP with single indice. With the multiple indices we exit at the check "isa<GetElementPtrInst>(Op2)" . Will keep a look on this if it can be improved.

3077 ↗

(On Diff #435486)

This is basically to check if there are stores or memory accesses b/w the 2 loads which we should consider for before the load merges. The check starts at L2 in reverse and does the memory check till L1.
There are alias tests loadCombine_4consecutive_with_alias* which should give you the idea.

3119 ↗

(On Diff #435486)

Yeah this is basically cause the visitOr() does an Insert of the returned OR. This may not be needed if we move this to AggressiveInstCombine.

[nikic]

When switching this to AggressiveInstCombine, I would strongly recommend to start with a much more minimal patch. Handle only a single simple case, without any of the possible variants. We can build on that base later.

[spatel]

In D127392#3625257, @nikic wrote:

When switching this to AggressiveInstCombine, I would strongly recommend to start with a much more minimal patch. Handle only a single simple case, without any of the possible variants. We can build on that base later.

Strongly agree - there's a lot of potential for this to go wrong both in correctness and perf regressions, so we need to build up in steps.
AFAIK, the load combine pass did not have correctness problems when it died, so that source code would be a good reference.

[bipmis]

A simple implementation in AggressiveInstCombine which handles the forward consecutive load sequences as provided in the tests.
The implementation is limited to a specific consecutive load pattern which reduces to a combined load only(One and only use of individual loads is to generate a wider load). This is not a generic LoadCombine as combining loads with other uses can result in poison propagation.

[bipmis]

In D127392#3625567, @spatel wrote:

In D127392#3625257, @nikic wrote:

When switching this to AggressiveInstCombine, I would strongly recommend to start with a much more minimal patch. Handle only a single simple case, without any of the possible variants. We can build on that base later.

Strongly agree - there's a lot of potential for this to go wrong both in correctness and perf regressions, so we need to build up in steps.
AFAIK, the load combine pass did not have correctness problems when it died, so that source code would be a good reference.

The load combine pass is different in that it implements load combine in a more generic way and not based on a specific pattern. So if it find 2 loads can be combined it generates a wider load and subsequent usage are derived from this wider load using CreateExtractInteger. The current implementation will be more like a pattern match with offset and shift verify to confirm the loads are consecutive and with the only use to create a wider load.

[spatel]

How does this code account for potential memory accesses between the loads that are getting combined?

define i16 @loadCombine_2consecutive_store_between(ptr %p) {
  %p1 = getelementptr i8, ptr %p, i32 1
  %l1 = load i8, ptr %p, align 2
  store i8 42, ptr %p  ; this must happen after a combined load?
  %l2 = load i8, ptr %p1

  %e1 = zext i8 %l1 to i16
  %e2 = zext i8 %l2 to i16
  %s2 = shl i16 %e2, 8
  %o1 = or i16 %e1, %s2
  ret i16 %o1
}

llvm/test/Transforms/AggressiveInstCombine/or-load.ll

5 ↗

(On Diff #443893)

We need to have this test duplicated on a target (x86?) where the fold actually happens. Target-specific tests will need to go inside target subdirectories, otherwise we'll break testing bots. We can pre-commit those once we have the right set of tests.

24 ↗

(On Diff #443893)

Why do the tests have non-canonical code (shift-by-0)? I don't think we'd ever see this pattern given where AIC is in the opt pipeline.

[bipmis]

In D127392#3653014, @spatel wrote:

How does this code account for potential memory accesses between the loads that are getting combined?

define i16 @loadCombine_2consecutive_store_between(ptr %p) {
  %p1 = getelementptr i8, ptr %p, i32 1
  %l1 = load i8, ptr %p, align 2
  store i8 42, ptr %p  ; this must happen after a combined load?
  %l2 = load i8, ptr %p1

  %e1 = zext i8 %l1 to i16
  %e2 = zext i8 %l2 to i16
  %s2 = shl i16 %e2, 8
  %o1 = or i16 %e1, %s2
  ret i16 %o1
}

I should have mentioned. This being the base version have not enabled the same. Just targeting simple load scenarios in this patch. This was enabled in the the InstCombine patch. The same will be enabled in the subsequent patches to handle memory access b/w loads.

[spatel]

In D127392#3654507, @bipmis wrote:

I should have mentioned. This being the base version have not enabled the same. Just targeting simple load scenarios in this patch. This was enabled in the the InstCombine patch. The same will be enabled in the subsequent patches to handle memory access b/w loads.

I don't understand the comment. Does the posted patch miscompile the example with a store? If so, then the patch can't be committed in this form.
If there are planned changes to the posted patch before it is ready for review, please note that in the patch status (or add "WIP" to the title).

[bipmis]

Updating the patch with AliasAnalysis. Test split up for respective Targets and new tests updated.
Currently handling only simple forward load patterns.

[bipmis]

In D127392#3654734, @spatel wrote:

In D127392#3654507, @bipmis wrote:

I should have mentioned. This being the base version have not enabled the same. Just targeting simple load scenarios in this patch. This was enabled in the the InstCombine patch. The same will be enabled in the subsequent patches to handle memory access b/w loads.

I don't understand the comment. Does the posted patch miscompile the example with a store? If so, then the patch can't be committed in this form.
If there are planned changes to the posted patch before it is ready for review, please note that in the patch status (or add "WIP" to the title).

Yes it should have been enabled to handle the store scenarios. I have updates the tests and added few more tests. Can add more as needed.
Also to keep it simple have handled forward load sequences.

[dmgreen]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

448–449

The comment could flow better:

// 2. (? | (zExt(L1) << shift1)) | (zExt(L2) << shift2)
//  -> ? | (zExt(L3) << shift1)

479

Drop extra brackets from (LI1 == LI2)

489

I think this can use getPointersDiff to check the two pointers are the right distance apart.

531

Do we check anywhere that LI1 and LI2 are in the same block?

533–536

Does std::distance work?

583

This doesn't need to cast to a Value*

592

Allows -> Allowed

593

This could also be DL.isLegalInteger, which would avoid the need to create the Type.

601

Should this check the Fast flag too?

if (!Allowed || !Fast)

605–612

I think this is unneeded, and this can always just create:

NewLoad =
        new LoadInst(IntegerType::get(Load1Ptr->getContext(), LOps.LoadSize),
                     LI1->getPointerOperand(), "", LI1->isVolatile(), LI1->getAlign(),
                     LI1->getOrdering(), LI1->getSyncScopeID());

Or possibly use Builder.CreateLoad to avoid the separate Insert.

625–628

This could avoid the NewOp variable and just do:

if (LOps.zext) {
  NewLoad = Builder.CreateZExt(NewLoad, LOps.zext);

llvm/test/Transforms/AggressiveInstCombine/AArch64/or-load.ll

25–26

Can you run these tests through opt -O1 (without this patch) and use the result as the tests (maybe with a little cleanup). LLVM-IR will almost never include shl 0 nodes, and we should make sure we are testing what will appear in reality.
https://godbolt.org/z/raxKnEE9a

[bipmis]

Handle Review Comments from David.

[bipmis]

Handle GEP in a more generic way as as requested earlier by @nikic
Add check for loads belonging to same BB.

[bipmis]

@spatel I have updated the patch and it should handle the forward load scenarios. Have incorporated most of the review comments received.
The patch has been tested and passing all tests. Do review and suggest if you have more comments. Thanks.

[nikic]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

512

This whole code can be replaced with stripAndAccumulateConstantOffsets().

542

FindAvailableLoadedValue() is not the correct way to check for clobbers. In particular, it will return an "available value" for a direct clobber (store to the same address).

What you want it to loop over the instructions and call getModRefInfo() on AliasAnalysis, together with a small limit (e.g. 16) when you will abort the walk and bail out of the transform.

llvm/test/Transforms/AggressiveInstCombine/AArch64/or-load.ll

203

Try a variant storing to %p3 rather than %pstr here. I believe your current implementation will incorrectly accept this.

[bipmis]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

512

Sure. Will look into this.

llvm/test/Transforms/AggressiveInstCombine/AArch64/or-load.ll

203

It does not return an "available value" for a direct clobber. For example a change to the test

%l1 = load i8, ptr %p
%l2 = load i8, ptr %p1
%l3 = load i8, ptr %p2
store i8 10, i8* %p3
%l4 = load i8, ptr %p3

still returns
; LE-NEXT: [[TMP1:%.*]] = load i16, ptr [[P]], align 1
; LE-NEXT: [[TMP2:%.*]] = zext i16 [[TMP1]] to i32
; LE-NEXT: [[L3:%.*]] = load i8, ptr [[P2]], align 1
; LE-NEXT: store i8 10, ptr [[P3]], align 1
; LE-NEXT: [[L4:%.*]] = load i8, ptr [[P3]], align 1

Can add more tests if you suggest.

[nikic]

llvm/test/Transforms/AggressiveInstCombine/AArch64/or-load.ll

203

Okay, I had to test this patch locally to find a case where it fails. Try this variant:

store i8 0, ptr %p3
store i8 1, ptr %p

We are looking for an available value of %p, so we find the store i8 1, ptr %p and are happy. But before that, there is a clobber of %p3 that makes the transform invalid (the load is moved before the store).

[bipmis]

Handle Review comments from nikic. Made changes to Alias Analysis.

[bipmis]

Gentle ping for review on this! Thanks.

[dmgreen]

I feel like this being recursive makes it more difficult to reason about than it could be. I have added some mostly nitpicks/cleanups below.

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

443

zext -> Zext (Or ZExt? Capitalized at least)

448

Are there any tests for case 2?

467–468

The variables can be defined where they are first used.

482

Should it check that the address space is the same?

502

Does it matter if it is a bitcast or a gep?

520

Capitalize variable names.

549

We checked that loadSize1 == loadSize2 above.

552

Demorgan this.

568

Maybe replace the name of foldLoadsIterative with foldConsecutiveLoads and vice-versa. foldLoadsIterative doesn't really explain what this function is folding, and it's not really iterative.

[bipmis]

Thanks David for reviewing. Have handled most of the nits.

[bipmis]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

448

The case2 is not required as we are handling the pattern recursively. This also keeps the implementation simple.

502

I think this may be needed, so that we fall through and evaluate further if instructions are only of these types.

549

Right. I have added additional comments on why this is needed.

[bipmis]

@dmgreen For pattern matching a chain of or(or,load), recursion seemed to a good choice to go to the root node and evaluate if the entire chain can be reduced. Also there are other instances of pattern match in AggressiveCombine() like matchAndOrChain() which implements it similarly.
@nikic Do you have any further comments on the Alias Analysis used. I have used the limit as the instruction difference between 2 loads for alias analysis b/w which we need to look out for a store.
@spatel Please do suggest if you have any other review comments. Thanks.

[dmgreen]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

502

What is it you mean by fall though? If stripAndAccumulateConstantOffsets could give us an Offset, it seems like we should just always call it and have it do what it can. It will return the original pointer if it couldn't do anything useful.
It may be better to keep Offset1/Offset2 as APInt too. It would help if the pointers were > 64bits.

564

Drop these extra brackets

569

Should this have a limit on the number of instructions?

579

Is there a test for an i128 version of the fold?

646

Is all the metadata on the old instruction (like MD_range) always valid on the new load?

[nikic]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

523

Shouldn't this be !LI1->isSimple() || !LI2->isSimple()`? We want to bail if either load isn't simple, not if both are.

Also, it looks like there are no (negative) tests for volatile/atomic loads.

[bipmis]

Handle review comments and add additional tests.

[bipmis]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

569

Some tests like load64_farLoads() which have wider instruction gap b/w loads may result in partial combine(when tried with 16). I can possibly go for a bigger limit or can keep the limit on the actual instructions b/w 2 loads.

[dmgreen]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

569

Using what limit?

646

What happens if new metadata gets added in the future, that isn't valid?

Is it better to just drop all the metadata? Or is that too likely to be worse for performance?

[bipmis]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

569

In the current implementation worst case limit could be all instructions in a BB. What is the issue with this?
For the test case load64_farLoads() it does fine with a limit of 35.

646

This being a specific scenario of the pattern match and looking for an or-load chain, I dont think performance should be a big concern. Depends on the end use of the merged load.

What I am seeing in most cases is that they try to retain atleast the AATags.

if (AATags)
      NewVal->setAAMetadata(AATags);

[nikic]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

646

Note that you can't simply take the AATags from one load, they have to be merged appropriately. I believe for this specific case you need the AAMDNodes::concat() method, because you are merging loads from different non-overlapping locations.

[dmgreen]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

569

It protects us where there are thousands of instructions in the block, just to be safe for degenerate cases. If we expect the maximum pattern to be (load+zext+shift+or) * i8->i64, so 4*8, then a limit of 64 instructions sounds fine.

[bipmis]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

569

Sounds right. Will implement the same with the 64 instruction limit. Thanks.

646

Agreed. Thanks for this.
I think we are better off dropping the metadata at this point.

[bipmis]

Add a limit of 64 instructions for Alias Analysis.
Concat AATags Metadata for merged loads.

[bipmis]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

646

The concat method leaves the tbaa blank so maybe we may want to drop the Metadata altogether?
Currently the ‘noalias’ and ‘alias.scope’ Metadata will be concatenated from
AAMDNodes.

[dmgreen]

Thanks for the the changes - as far as I understand, this LGTM now. Thanks for working through the details.

Any other comments?

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

646

I'm a little surprised that if two the tbaa info are the same, we can't use the same on the result node. I think using concat sounds sensible though. I suspect in practice we will often be combining char in any case.

[nikic]

There are a number of test failures in pre-merge checks, probably pipeline tests.

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

909–910

Just like all the other analyses, this should use a reference, not a pointer (the analysis is not optional).

[nikic]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

488

Zext -> ZextType, maybe?

495

Iterative -> Recursive

567

This check should come first, otherwise you don't count store instructions.

591

I think this handles non-byte-sized loads incorrectly. Let's say you do i4 loads with 1 byte offset. Then PrevSize will be 1 and match the offset, even though the loads are not actually consecutive.

Please add some tests for non-byte-sized loads.

624

Combine these declarations with initialization.

645

Why does this not use the IRBuilder?

867

The DataLayout fetch can be moved outside the loop.

[spatel]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

866–867

The transform should be placed ahead of foldSqrt(), or we we may hit a use-after-free bug (because foldSqrt can delete I).
There was a code comment about this in:
https://github.com/llvm/llvm-project/commit/df868edee561eb973edd85ec9df41c67aa0bff6b
...but that patch got reverted. We should probably add that code comment independently (or fix the bug some other way).

[bipmis]

Update the patch with review comments.
-> Support power of 2 loads and minimum load size 8bits.
-> IR builder for new load.
-> New test for a 4 bit load.
-> Nits.

[bipmis]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

567

So it would count stores which do not alias. For the case it aliases we terminate anyways and the count update wont matter?

591

Right. The shift condition will prevent it from merging.
But we do not want to combine loads smaller than a byte. Updated checks.

[bipmis]

Ping.
@nikic Any further comments on the patch.

[nikic]

From a cursory look it looks fine, I assume @dmgreen has reviewed in detail :)

I have one note (that can be addressed later): There's currently a check that the loads have the same size. Does that mean that if you have a partially folded pattern (e.g. one i16 load plus two i8 loads) that it will not get folded? If so, this seems like something we should relax (later).

[bipmis]

In D127392#3807970, @nikic wrote:

From a cursory look it looks fine, I assume @dmgreen has reviewed in detail :)

I have one note (that can be addressed later): There's currently a check that the loads have the same size. Does that mean that if you have a partially folded pattern (e.g. one i16 load plus two i8 loads) that it will not get folded? If so, this seems like something we should relax (later).

Thats correct. Currently we are folding loads of same size only in a chain. So for above case if two i8 loads belong to single lower chain they can be folded.
Agreed. This is more to do with a stricter basic implementation first. We can relax this later.

[RKSimon]

Please ensure you precommit the new tests to trunk and then rebase this patch on it - so we see the effect of the patch on the IR

[gribozavr2]

When I do a 3-stage bootstrap at this commit, the second-stage compiler crashes. The issue does not appear at the previous commit, so I'm reverting this commit.

[gribozavr2]

Reverted: https://github.com/llvm/llvm-project/commit/954d3cd2c6e9a5cdee23fcef60f2c3a21c88be60

[RKSimon]

In D127392#3812152, @gribozavr2 wrote:

Reverted: https://github.com/llvm/llvm-project/commit/954d3cd2c6e9a5cdee23fcef60f2c3a21c88be60

@gribozavr2 Do you have any useful build log that you can provide to speed up triage please?

[bipmis]

In D127392#3812152, @gribozavr2 wrote:

Reverted: https://github.com/llvm/llvm-project/commit/954d3cd2c6e9a5cdee23fcef60f2c3a21c88be60

@gribozavr2 Do let me know how to reproduce this issue. I am trying a 3 stage in "clang>/cmake/caches/3-stage.cmake". Linking the executable in second stage takes quite some time for a clean HEAD code.
Builds fine on AArch64 and x86 with the patch. Stage2 and stage3 binary are bit exact.

[aeubanks]

I'm looking into coming up with a reduced repro.
We're seeing a crash in re2. The weird thing is that the compiler I've repro'd is not bootstrapped, but the crash in re2 happens even when compiling it with -O0. My guess is that something else we compile with the just-built clang like libc++ is getting miscompiled, but more investigation required.

[aeubanks]

it ended up being a function that we always compile with -O3...

anyway, reduced test case

$ cat /tmp/a.ll
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"                                                                                                       
target triple = "x86_64-grtev4-linux-gnu"                                                                                                                                                          
                                                                                                                                                                                                   
define i64 @eggs(ptr noundef readonly %arg) {                                                                                                                                                      
  %tmp3 = load i8, ptr %arg, align 1                                                             
  %tmp4 = getelementptr inbounds i8, ptr %arg, i64 1                                             
  %tmp5 = load i8, ptr %tmp4, align 1                                                            
  %tmp6 = getelementptr inbounds i8, ptr %arg, i64 2                                             
  %tmp7 = load i8, ptr %tmp6, align 1                                                            
  %tmp8 = getelementptr inbounds i8, ptr %arg, i64 3                                             
  %tmp9 = load i8, ptr %tmp8, align 1                                                            
  %tmp10 = getelementptr inbounds i8, ptr %arg, i64 4                                            
  %tmp11 = load i8, ptr %tmp10, align 1                                                          
  %tmp12 = getelementptr inbounds i8, ptr %arg, i64 5                                            
  %tmp13 = load i8, ptr %tmp12, align 1                                                          
  %tmp14 = getelementptr inbounds i8, ptr %arg, i64 6                                            
  %tmp15 = load i8, ptr %tmp14, align 1                                                          
  %tmp16 = getelementptr inbounds i8, ptr %arg, i64 7                                            
  %tmp17 = load i8, ptr %tmp16, align 1
  %tmp18 = zext i8 %tmp17 to i64     
  %tmp19 = shl nuw i64 %tmp18, 56    
  %tmp20 = zext i8 %tmp15 to i64     
  %tmp21 = shl nuw nsw i64 %tmp20, 48
  %tmp22 = or i64 %tmp19, %tmp21     
  %tmp23 = zext i8 %tmp13 to i64     
  %tmp24 = shl nuw nsw i64 %tmp23, 40
  %tmp25 = or i64 %tmp22, %tmp24     
  %tmp26 = zext i8 %tmp11 to i64     
  %tmp27 = shl nuw nsw i64 %tmp26, 32
  %tmp28 = or i64 %tmp25, %tmp27     
  %tmp29 = zext i8 %tmp9 to i64      
  %tmp30 = shl nuw nsw i64 %tmp29, 24
  %tmp31 = or i64 %tmp28, %tmp30     
  %tmp32 = zext i8 %tmp7 to i64      
  %tmp33 = shl nuw nsw i64 %tmp32, 16
  %tmp34 = zext i8 %tmp5 to i64     
  %tmp35 = shl nuw nsw i64 %tmp34, 8
  %tmp36 = or i64 %tmp31, %tmp33
  %tmp37 = zext i8 %tmp3 to i64 
  %tmp38 = or i64 %tmp36, %tmp35                                                                                                                                                                   
  %tmp39 = or i64 %tmp38, %tmp37                                                                                                                                                                   
  ret i64 %tmp39                                                                                                                                                                                   
}
$ ./build/rel/bin/opt -passes=aggressive-instcombine -S /tmp/a.ll
; ModuleID = '/tmp/b.ll'
source_filename = "/tmp/a.ll"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-grtev4-linux-gnu"

define i64 @eggs(ptr noundef readonly %arg) {
  %tmp3 = load i8, ptr %arg, align 1
  %tmp4 = getelementptr inbounds i8, ptr %arg, i64 1
  %tmp5 = load i32, ptr %tmp4, align 1
  %1 = zext i32 %tmp5 to i64
  %2 = shl i64 %1, 8
  %tmp37 = zext i8 %tmp3 to i64
  %tmp39 = or i64 %2, %tmp37
  ret i64 %tmp39
}

that does look wrong, it looks like it should be optimized to load i64 rather than zext(load i32 (%a + 1)) | zext(load i8 %a)

[bipmis]

In D127392#3815730, @aeubanks wrote:

it ended up being a function that we always compile with -O3...

anyway, reduced test case

$ cat /tmp/a.ll
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"                                                                                                       
target triple = "x86_64-grtev4-linux-gnu"                                                                                                                                                          
                                                                                                                                                                                                   
define i64 @eggs(ptr noundef readonly %arg) {                                                                                                                                                      
  %tmp3 = load i8, ptr %arg, align 1                                                             
  %tmp4 = getelementptr inbounds i8, ptr %arg, i64 1                                             
  %tmp5 = load i8, ptr %tmp4, align 1                                                            
  %tmp6 = getelementptr inbounds i8, ptr %arg, i64 2                                             
  %tmp7 = load i8, ptr %tmp6, align 1                                                            
  %tmp8 = getelementptr inbounds i8, ptr %arg, i64 3                                             
  %tmp9 = load i8, ptr %tmp8, align 1                                                            
  %tmp10 = getelementptr inbounds i8, ptr %arg, i64 4                                            
  %tmp11 = load i8, ptr %tmp10, align 1                                                          
  %tmp12 = getelementptr inbounds i8, ptr %arg, i64 5                                            
  %tmp13 = load i8, ptr %tmp12, align 1                                                          
  %tmp14 = getelementptr inbounds i8, ptr %arg, i64 6                                            
  %tmp15 = load i8, ptr %tmp14, align 1                                                          
  %tmp16 = getelementptr inbounds i8, ptr %arg, i64 7                                            
  %tmp17 = load i8, ptr %tmp16, align 1
  %tmp18 = zext i8 %tmp17 to i64     
  %tmp19 = shl nuw i64 %tmp18, 56    
  %tmp20 = zext i8 %tmp15 to i64     
  %tmp21 = shl nuw nsw i64 %tmp20, 48
  %tmp22 = or i64 %tmp19, %tmp21     
  %tmp23 = zext i8 %tmp13 to i64     
  %tmp24 = shl nuw nsw i64 %tmp23, 40
  %tmp25 = or i64 %tmp22, %tmp24     
  %tmp26 = zext i8 %tmp11 to i64     
  %tmp27 = shl nuw nsw i64 %tmp26, 32
  %tmp28 = or i64 %tmp25, %tmp27     
  %tmp29 = zext i8 %tmp9 to i64      
  %tmp30 = shl nuw nsw i64 %tmp29, 24
  %tmp31 = or i64 %tmp28, %tmp30     
  %tmp32 = zext i8 %tmp7 to i64      
  %tmp33 = shl nuw nsw i64 %tmp32, 16
  %tmp34 = zext i8 %tmp5 to i64     
  %tmp35 = shl nuw nsw i64 %tmp34, 8
  %tmp36 = or i64 %tmp31, %tmp33
  %tmp37 = zext i8 %tmp3 to i64 
  %tmp38 = or i64 %tmp36, %tmp35                                                                                                                                                                   
  %tmp39 = or i64 %tmp38, %tmp37                                                                                                                                                                   
  ret i64 %tmp39                                                                                                                                                                                   
}
$ ./build/rel/bin/opt -passes=aggressive-instcombine -S /tmp/a.ll
; ModuleID = '/tmp/b.ll'
source_filename = "/tmp/a.ll"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-grtev4-linux-gnu"

define i64 @eggs(ptr noundef readonly %arg) {
  %tmp3 = load i8, ptr %arg, align 1
  %tmp4 = getelementptr inbounds i8, ptr %arg, i64 1
  %tmp5 = load i32, ptr %tmp4, align 1
  %1 = zext i32 %tmp5 to i64
  %2 = shl i64 %1, 8
  %tmp37 = zext i8 %tmp3 to i64
  %tmp39 = or i64 %2, %tmp37
  ret i64 %tmp39
}

that does look wrong, it looks like it should be optimized to load i64 rather than zext(load i32 (%a + 1)) | zext(load i8 %a)

@aeubanks Thanks for the test case.
So in this patch we have implemented forward load merge as can be seen by the test example. However I had to handle the child node "or(load,load)" as this is being reversed by InstCombine currently as seen below. This needed a corrected check.

define i32 @Load32(ptr noundef %ptr) local_unnamed_addr #0 {
entry:
  %0 = load i8, ptr %ptr, align 1, !tbaa !5
  %conv = zext i8 %0 to i32
  %arrayidx1 = getelementptr inbounds i8, ptr %ptr, i64 1
  %1 = load i8, ptr %arrayidx1, align 1, !tbaa !5
  %conv2 = zext i8 %1 to i32
  %shl = shl i32 %conv2, 8
 ** %or = or i32 %conv, %shl**
  %arrayidx3 = getelementptr inbounds i8, ptr %ptr, i64 2
  %2 = load i8, ptr %arrayidx3, align 1, !tbaa !5
  %conv4 = zext i8 %2 to i32
  %shl5 = shl i32 %conv4, 16
  %or6 = or i32 %or, %shl5
  %arrayidx7 = getelementptr inbounds i8, ptr %ptr, i64 3
  %3 = load i8, ptr %arrayidx7, align 1, !tbaa !5
  %conv8 = zext i8 %3 to i32
  %shl9 = shl i32 %conv8, 24
  %or10 = or i32 %or6, %shl9
  ret i32 %or10
}

*** IR Dump After InstCombinePass on Load32 ***
; Function Attrs: nounwind uwtable
define i32 @Load32(ptr noundef %ptr) local_unnamed_addr #0 {
entry:
  %0 = load i8, ptr %ptr, align 1, !tbaa !5
  %conv = zext i8 %0 to i32
  %arrayidx1 = getelementptr inbounds i8, ptr %ptr, i64 1
  %1 = load i8, ptr %arrayidx1, align 1, !tbaa !5
  %conv2 = zext i8 %1 to i32
  %shl = shl nuw nsw i32 %conv2, 8
**  %or = or i32 %shl, %conv**
  %arrayidx3 = getelementptr inbounds i8, ptr %ptr, i64 2
  %2 = load i8, ptr %arrayidx3, align 1, !tbaa !5
  %conv4 = zext i8 %2 to i32
  %shl5 = shl nuw nsw i32 %conv4, 16
  %or6 = or i32 %or, %shl5
  %arrayidx7 = getelementptr inbounds i8, ptr %ptr, i64 3
  %3 = load i8, ptr %arrayidx7, align 1, !tbaa !5
  %conv8 = zext i8 %3 to i32
  %shl9 = shl nuw i32 %conv8, 24
  %or10 = or i32 %or6, %shl9
  ret i32 %or10

The full implementation of reverse load merge pattern is planned subsequently. So now you will see the lowest node loads being merged but not the other ones. I am updating the patch. Would be great if you can test the same.

[bipmis]

Handle the reverse load pattern checks correctly.
Currently we need to handle the leaf node reverse loads as InstCombine pass folds the pattern from a forward to reverse one.
Full reverse load patterns planned to be implemented subsequently.

[spatel]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

694–697

I didn't notice this limitation before. "Forward" and "reverse" are referring to the order that the or instructions use the loaded values?

I agree that we don't want to complicate the initial implementation any more than necessary, but it might help to see how the backend handles that in DAGCombiner::MatchLoadCombine() (see D133584 for a proposed enhancement to that code).

[bipmis]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

694–697

Right it is the order
Forward - or(or(or(or(0,1),2),3)
Reverse - or(or(or(or(3,2),1),0)
Considering 0,1...as zext and shl of loads with index 0,1 etc.

For simplicity we wanted to implement the forward first and if this looks good we can do the reverse+mixed size loads. There should be minimal changes on top of this. So that is in plan next.

[bipmis]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

694–697

I didn't notice this limitation before. "Forward" and "reverse" are referring to the order that the or instructions use the loaded values?

I agree that we don't want to complicate the initial implementation any more than necessary, but it might help to see how the backend handles that in DAGCombiner::MatchLoadCombine() (see D133584 for a proposed enhancement to that code).

@spatel Verified DAGCombiner::MatchLoadCombine() handles the reverse load pattern fine and a single load is generated.

[spatel]

Please pre-commit the new tests with the baseline CHECK lines.
Also, it would be good to add a test that matches the reported failure case (the @eggs test from the earlier comment).
After that, I think it's fine to re-commit.

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

694–697

Thanks for checking. It doesn't directly impact the initial/immediate patch here, but it might provide a template for the planned enhancements.

[Kai]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

816

FYI: The missing bitcast for the Load1Ptr argument means that this change only works with opaque pointers.

[dstuttard]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

816

We had the same issue - I've just uploaded a patch. See D135249

[uabelho]

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

738

Since the NumScanned counter includes debug instructions, this code makes the result depend on the presence on debug instructions.
I wrote a ticket about that: https://github.com/llvm/llvm-project/issues/69925

llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp

 
 STATISTIC(NumAnyOrAllBitsSet, "Number of any/all-bits-set patterns folded");
 STATISTIC(NumGuardedRotates,
           "Number of guarded rotates transformed into funnel shifts");
 STATISTIC(NumGuardedFunnelShifts,
           "Number of guarded funnel shifts transformed into funnel shifts");
 STATISTIC(NumPopCountRecognized, "Number of popcount idioms recognized");
 
+static cl::opt<unsigned> MaxInstrsToScan(
+    "aggressive-instcombine-max-scan-instrs", cl::init(64), cl::Hidden,
+    cl::desc("Max number of instructions to scan for aggressive instcombine."));
+
 namespace {
 /// Contains expression pattern combiner logic.
 /// This class provides both the logic to combine expression patterns and
 /// combine them. It differs from InstCombiner class in that each pattern
 /// combiner runs only once as opposed to InstCombine's multi-iteration,
 /// which allows pattern combiner to have higher complexity than the O(1)
 /// required by the instruction combiner.
 class AggressiveInstCombinerLegacyPass : public FunctionPass {

 /// Try to replace a mathlib call to sqrt with the LLVM intrinsic. This avoids
 /// pessimistic codegen that has to account for setting errno and can enable
 /// vectorization.
 static bool
 foldSqrt(Instruction &I, TargetTransformInfo &TTI, TargetLibraryInfo &TLI) {
   // Match a call to sqrt mathlib function.
   auto *Call = dyn_cast<CallInst>(&I);
   if (!Call)
     return false;
 
   Module *M = Call->getModule();
   LibFunc Func;
   if (!TLI.getLibFunc(*Call, Func) || !isLibFuncEmittable(M, &TLI, Func))
     return false;
 
   if (Func != LibFunc_sqrt && Func != LibFunc_sqrtf && Func != LibFunc_sqrtl)
     return false;
 
   // If (1) this is a sqrt libcall, (2) we can assume that NAN is not created
   // (because NNAN or the operand arg must not be less than -0.0) and (2) we
   // would not end up lowering to a libcall anyway (which could change the value
   // of errno), then:
   // (1) errno won't be set.
   // (2) it is safe to convert this to an intrinsic call.
   Type *Ty = Call->getType();
   Value *Arg = Call->getArgOperand(0);
   if (TTI.haveFastSqrt(Ty) &&
       (Call->hasNoNaNs() || CannotBeOrderedLessThanZero(Arg, &TLI))) {
     IRBuilder<> Builder(&I);
     IRBuilderBase::FastMathFlagGuard Guard(Builder);
     Builder.setFastMathFlags(Call->getFastMathFlags());
 
     Function *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, Ty);
     Value *NewSqrt = Builder.CreateCall(Sqrt, Arg, "sqrt");
     I.replaceAllUsesWith(NewSqrt);
 
     // Explicitly erase the old call because a call with side effects is not
     // trivially dead.
     I.eraseFromParent();
     return true;
   }
 
   return false;
 }
 
 // Check if this array of constants represents a cttz table.
 // Iterate over the elements from \p Table by trying to find/match all
 // the numbers from 0 to \p InputBits that should represent cttz results.
 static bool isCTTZTable(const ConstantDataArray &Table, uint64_t Mul,
                         uint64_t Shift, uint64_t InputBits) {
   unsigned Length = Table.getNumElements();
   if (Length < InputBits || Length > InputBits * 2)
     return false;
 
   APInt Mask = APInt::getBitsSetFrom(InputBits, Shift);
   unsigned Matched = 0;
 
   for (unsigned i = 0; i < Length; i++) {
     uint64_t Element = Table.getElementAsInteger(i);
     if (Element >= InputBits)
       continue;
 
     // Check if \p Element matches a concrete answer. It could fail for some
     // elements that are never accessed, so we keep iterating over each element
     // from the table. The number of matched elements should be equal to the
     // number of potential right answers which is \p InputBits actually.
     if ((((Mul << Element) & Mask.getZExtValue()) >> Shift) == i)
       Matched++;
   }
 
   return Matched == InputBits;
 }
 
 // Try to recognize table-based ctz implementation.
 // E.g., an example in C (for more cases please see the llvm/tests):
 // int f(unsigned x) {
 //    static const char table[32] =
 //      {0, 1, 28, 2, 29, 14, 24, 3, 30,
 //       22, 20, 15, 25, 17, 4, 8, 31, 27,
 //       13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9};
 //    return table[((unsigned)((x & -x) * 0x077CB531U)) >> 27];
 // }
 // this can be lowered to `cttz` instruction.
 // There is also a special case when the element is 0.
 //
 // Here are some examples or LLVM IR for a 64-bit target:
 //
 // CASE 1:
 // %sub = sub i32 0, %x
 // %and = and i32 %sub, %x
 // %mul = mul i32 %and, 125613361
 // %shr = lshr i32 %mul, 27
 // %idxprom = zext i32 %shr to i64
 // %arrayidx = getelementptr inbounds [32 x i8], [32 x i8]* @ctz1.table, i64 0,
 // i64 %idxprom %0 = load i8, i8* %arrayidx, align 1, !tbaa !8
 //
 // CASE 2:
 // %sub = sub i32 0, %x
 // %and = and i32 %sub, %x
 // %mul = mul i32 %and, 72416175
 // %shr = lshr i32 %mul, 26
 // %idxprom = zext i32 %shr to i64
 // %arrayidx = getelementptr inbounds [64 x i16], [64 x i16]* @ctz2.table, i64
 // 0, i64 %idxprom %0 = load i16, i16* %arrayidx, align 2, !tbaa !8
 //
 // CASE 3:
 // %sub = sub i32 0, %x
 // %and = and i32 %sub, %x
 // %mul = mul i32 %and, 81224991
 // %shr = lshr i32 %mul, 27
 // %idxprom = zext i32 %shr to i64
 // %arrayidx = getelementptr inbounds [32 x i32], [32 x i32]* @ctz3.table, i64
 // 0, i64 %idxprom %0 = load i32, i32* %arrayidx, align 4, !tbaa !8
 //
 // CASE 4:
 // %sub = sub i64 0, %x
 // %and = and i64 %sub, %x
 // %mul = mul i64 %and, 283881067100198605
 // %shr = lshr i64 %mul, 58
 // %arrayidx = getelementptr inbounds [64 x i8], [64 x i8]* @table, i64 0, i64
 // %shr %0 = load i8, i8* %arrayidx, align 1, !tbaa !8
 //
 // All this can be lowered to @llvm.cttz.i32/64 intrinsic.
 static bool tryToRecognizeTableBasedCttz(Instruction &I) {
   LoadInst *LI = dyn_cast<LoadInst>(&I);
   if (!LI)
     return false;
 
   Type *AccessType = LI->getType();
   if (!AccessType->isIntegerTy())
     return false;
 
   GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getPointerOperand());
   if (!GEP || !GEP->isInBounds() || GEP->getNumIndices() != 2)
     return false;
 
   if (!GEP->getSourceElementType()->isArrayTy())
     return false;
 
   uint64_t ArraySize = GEP->getSourceElementType()->getArrayNumElements();
   if (ArraySize != 32 && ArraySize != 64)
     return false;
 
   GlobalVariable *GVTable = dyn_cast<GlobalVariable>(GEP->getPointerOperand());
   if (!GVTable || !GVTable->hasInitializer() || !GVTable->isConstant())
     return false;
 
   ConstantDataArray *ConstData =
       dyn_cast<ConstantDataArray>(GVTable->getInitializer());
   if (!ConstData)
     return false;
 
   if (!match(GEP->idx_begin()->get(), m_ZeroInt()))
     return false;
 
   Value *Idx2 = std::next(GEP->idx_begin())->get();
   Value *X1;
   uint64_t MulConst, ShiftConst;
   // FIXME: 64-bit targets have `i64` type for the GEP index, so this match will
   // probably fail for other (e.g. 32-bit) targets.
   if (!match(Idx2, m_ZExtOrSelf(
                        m_LShr(m_Mul(m_c_And(m_Neg(m_Value(X1)), m_Deferred(X1)),
                                     m_ConstantInt(MulConst)),
                               m_ConstantInt(ShiftConst)))))
     return false;
 
   unsigned InputBits = X1->getType()->getScalarSizeInBits();
   if (InputBits != 32 && InputBits != 64)
     return false;
 
   // Shift should extract top 5..7 bits.
   if (InputBits - Log2_32(InputBits) != ShiftConst &&
       InputBits - Log2_32(InputBits) - 1 != ShiftConst)
     return false;
 
   if (!isCTTZTable(*ConstData, MulConst, ShiftConst, InputBits))
     return false;
 
   auto ZeroTableElem = ConstData->getElementAsInteger(0);
   bool DefinedForZero = ZeroTableElem == InputBits;
 
   IRBuilder<> B(LI);
   ConstantInt *BoolConst = B.getInt1(!DefinedForZero);
   Type *XType = X1->getType();
   auto Cttz = B.CreateIntrinsic(Intrinsic::cttz, {XType}, {X1, BoolConst});
   Value *ZExtOrTrunc = nullptr;
 
   if (DefinedForZero) {
     ZExtOrTrunc = B.CreateZExtOrTrunc(Cttz, AccessType);
   } else {
     // If the value in elem 0 isn't the same as InputBits, we still want to
     // produce the value from the table.
     auto Cmp = B.CreateICmpEQ(X1, ConstantInt::get(XType, 0));
     auto Select =
         B.CreateSelect(Cmp, ConstantInt::get(XType, ZeroTableElem), Cttz);
 
     // NOTE: If the table[0] is 0, but the cttz(0) is defined by the Target
     // it should be handled as: `cttz(x) & (typeSize - 1)`.
 
     ZExtOrTrunc = B.CreateZExtOrTrunc(Select, AccessType);
   }
 
   LI->replaceAllUsesWith(ZExtOrTrunc);
 
   return true;
 }
 
+/// This is used by foldLoadsRecursive() to capture a Root Load node which is
+/// of type or(load, load) and recursively build the wide load. Also capture the
+/// shift amount, zero extend type and loadSize.
+struct LoadOps {
+  LoadInst *Root = nullptr;
+  bool FoundRoot = false;
+  uint64_t LoadSize = 0;
+  Value *Shift = nullptr;
+  Type *ZextType;
+  AAMDNodes AATags;
+};
+
+// Identify and Merge consecutive loads recursively which is of the form
+// (ZExt(L1) << shift1) | (ZExt(L2) << shift2) -> ZExt(L3) << shift1
+// (ZExt(L1) << shift1) | ZExt(L2) -> ZExt(L3)
+static bool foldLoadsRecursive(Value *V, LoadOps &LOps, const DataLayout &DL,
+                               AliasAnalysis &AA) {
+  Value *ShAmt2 = nullptr;
+  Value *X;
+  Instruction *L1, *L2;
+
+  // Go to the last node with loads.
+  if (match(V, m_OneUse(m_c_Or(
+                   m_Value(X),
+                   m_OneUse(m_Shl(m_OneUse(m_ZExt(m_OneUse(m_Instruction(L2)))),
+                                  m_Value(ShAmt2)))))) ||
+      match(V, m_OneUse(m_Or(m_Value(X),
+                             m_OneUse(m_ZExt(m_OneUse(m_Instruction(L2))))))))
+    foldLoadsRecursive(X, LOps, DL, AA);
+  else
+    return false;
+
+  // Check if the pattern has loads
+  LoadInst *LI1 = LOps.Root;
+  Value *ShAmt1 = LOps.Shift;
+  if (LOps.FoundRoot == false &&
+      (match(X, m_OneUse(m_ZExt(m_Instruction(L1)))) ||
+       match(X, m_OneUse(m_Shl(m_OneUse(m_ZExt(m_OneUse(m_Instruction(L1)))),
+                               m_Value(ShAmt1)))))) {
+    LI1 = dyn_cast<LoadInst>(L1);
+  }
+  LoadInst *LI2 = dyn_cast<LoadInst>(L2);
+
+  // Check if loads are same, atomic, volatile and having same address space.
+  if (LI1 == LI2 || !LI1 || !LI2 || !LI1->isSimple() || !LI2->isSimple() ||
+      LI1->getPointerAddressSpace() != LI2->getPointerAddressSpace())
+    return false;
+
+  // Check if Loads come from same BB.
+  if (LI1->getParent() != LI2->getParent())
+    return false;
+
+  // Swap loads if LI1 comes later as we handle only forward loads.
+  // This is done as InstCombine folds lowest node forward loads to reverse.
+  // The implementation will be subsequently extended to handle all reverse
+  // loads.
+  if (!LI1->comesBefore(LI2)) {
+    if (LOps.FoundRoot == false) {
+      std::swap(LI1, LI2);
+      std::swap(ShAmt1, ShAmt2);
+    } else
+      return false;
+  }
+
+  // Find the data layout
+  bool IsBigEndian = DL.isBigEndian();
+
+  // Check if loads are consecutive and same size.
+  Value *Load1Ptr = LI1->getPointerOperand();
+  APInt Offset1(DL.getIndexTypeSizeInBits(Load1Ptr->getType()), 0);
+  Load1Ptr =
+      Load1Ptr->stripAndAccumulateConstantOffsets(DL, Offset1,
+                                                  /* AllowNonInbounds */ true);
+
+  Value *Load2Ptr = LI2->getPointerOperand();
+  APInt Offset2(DL.getIndexTypeSizeInBits(Load2Ptr->getType()), 0);
+  Load2Ptr =
+      Load2Ptr->stripAndAccumulateConstantOffsets(DL, Offset2,
+                                                  /* AllowNonInbounds */ true);
+
+  // Verify if both loads have same base pointers and load sizes are same.
+  uint64_t LoadSize1 = LI1->getType()->getPrimitiveSizeInBits();
+  uint64_t LoadSize2 = LI2->getType()->getPrimitiveSizeInBits();
+  if (Load1Ptr != Load2Ptr || LoadSize1 != LoadSize2)
+    return false;
+
+  // Support Loadsizes greater or equal to 8bits and only power of 2.
+  if (LoadSize1 < 8 || !isPowerOf2_64(LoadSize1))
+    return false;
+
+  // Alias Analysis to check for store b/w the loads.
+  MemoryLocation Loc = MemoryLocation::get(LI2);
+  unsigned NumScanned = 0;
+  for (Instruction &Inst : make_range(LI1->getIterator(), LI2->getIterator())) {
+    if (Inst.mayWriteToMemory() && isModSet(AA.getModRefInfo(&Inst, Loc)))
+      return false;
+    if (++NumScanned > MaxInstrsToScan)
+      return false;
+  }
+
+  // Big endian swap the shifts
+  if (IsBigEndian)
+    std::swap(ShAmt1, ShAmt2);
+
+  // Find Shifts values.
+  const APInt *Temp;
+  uint64_t Shift1 = 0, Shift2 = 0;
+  if (ShAmt1 && match(ShAmt1, m_APInt(Temp)))
+    Shift1 = Temp->getZExtValue();
+  if (ShAmt2 && match(ShAmt2, m_APInt(Temp)))
+    Shift2 = Temp->getZExtValue();
+
+  // First load is always LI1. This is where we put the new load.
+  // Use the merged load size available from LI1, if we already combined loads.
+  if (LOps.FoundRoot)
+    LoadSize1 = LOps.LoadSize;
+
+  // Verify if shift amount and load index aligns and verifies that loads
+  // are consecutive.
+  uint64_t ShiftDiff = IsBigEndian ? LoadSize2 : LoadSize1;
+  uint64_t PrevSize =
+      DL.getTypeStoreSize(IntegerType::get(LI1->getContext(), LoadSize1));
+  if ((Shift2 - Shift1) != ShiftDiff || (Offset2 - Offset1) != PrevSize)
+    return false;
+
+  // Update LOps
+  AAMDNodes AATags1 = LOps.AATags;
+  AAMDNodes AATags2 = LI2->getAAMetadata();
+  if (LOps.FoundRoot == false) {
+    LOps.FoundRoot = true;
+    LOps.LoadSize = LoadSize1 + LoadSize2;
+    AATags1 = LI1->getAAMetadata();
+  } else
+    LOps.LoadSize = LOps.LoadSize + LoadSize2;
+
+  // Concatenate the AATags of the Merged Loads.
+  LOps.AATags = AATags1.concat(AATags2);
+
+  LOps.Root = LI1;
+  LOps.Shift = ShAmt1;
+  LOps.ZextType = X->getType();
+  return true;
+}
+
+// For a given BB instruction, evaluate all loads in the chain that form a
+// pattern which suggests that the loads can be combined. The one and only use
+// of the loads is to form a wider load.
+static bool foldConsecutiveLoads(Instruction &I, const DataLayout &DL,
+                                 TargetTransformInfo &TTI, AliasAnalysis &AA) {
+  LoadOps LOps;
+  if (!foldLoadsRecursive(&I, LOps, DL, AA) || !LOps.FoundRoot)
+    return false;
+
+  IRBuilder<> Builder(&I);
+  LoadInst *NewLoad = nullptr, *LI1 = LOps.Root;
+
+  // TTI based checks if we want to proceed with wider load
+  bool Allowed =
+      TTI.isTypeLegal(IntegerType::get(I.getContext(), LOps.LoadSize));
+  if (!Allowed)
+    return false;
+
+  unsigned AS = LI1->getPointerAddressSpace();
+  bool Fast = false;
+  Allowed = TTI.allowsMisalignedMemoryAccesses(I.getContext(), LOps.LoadSize,
+                                               AS, LI1->getAlign(), &Fast);
+  if (!Allowed || !Fast)
+    return false;
+
+  // New load can be generated
+  Value *Load1Ptr = LI1->getPointerOperand();
+  Builder.SetInsertPoint(LI1);
+  NewLoad = Builder.CreateAlignedLoad(
+      IntegerType::get(Load1Ptr->getContext(), LOps.LoadSize), Load1Ptr,
+      LI1->getAlign(), LI1->isVolatile(), "");
+  NewLoad->takeName(LI1);
+  // Set the New Load AATags Metadata.
+  if (LOps.AATags)
+    NewLoad->setAAMetadata(LOps.AATags);
+
+  Value *NewOp = NewLoad;
+  // Check if zero extend needed.
+  if (LOps.ZextType)
+    NewOp = Builder.CreateZExt(NewOp, LOps.ZextType);
+
+  // Check if shift needed. We need to shift with the amount of load1
+  // shift if not zero.
+  if (LOps.Shift)
+    NewOp = Builder.CreateShl(NewOp, LOps.Shift);
+  I.replaceAllUsesWith(NewOp);
+
+  return true;
+}
+
 /// This is the entry point for folds that could be implemented in regular
 /// InstCombine, but they are separated because they are not expected to
 /// occur frequently and/or have more than a constant-length pattern match.
 static bool foldUnusualPatterns(Function &F, DominatorTree &DT,
                                 TargetTransformInfo &TTI,
-                                TargetLibraryInfo &TLI) {
+                                TargetLibraryInfo &TLI, AliasAnalysis &AA) {
   bool MadeChange = false;
   for (BasicBlock &BB : F) {
     // Ignore unreachable basic blocks.
     if (!DT.isReachableFromEntry(&BB))
       continue;
 
+    const DataLayout &DL = F.getParent()->getDataLayout();
+
     // Walk the block backwards for efficiency. We're matching a chain of
     // use->defs, so we're more likely to succeed by starting from the bottom.
     // Also, we want to avoid matching partial patterns.
     // TODO: It would be more efficient if we removed dead instructions
     // iteratively in this loop rather than waiting until the end.
     for (Instruction &I : make_early_inc_range(llvm::reverse(BB))) {
       MadeChange |= foldAnyOrAllBitsSet(I);
       MadeChange |= foldGuardedFunnelShift(I, DT);
       MadeChange |= tryToRecognizePopCount(I);
       MadeChange |= tryToFPToSat(I, TTI);
       MadeChange |= tryToRecognizeTableBasedCttz(I);
+      MadeChange |= foldConsecutiveLoads(I, DL, TTI, AA);
       // NOTE: This function introduces erasing of the instruction `I`, so it
       // needs to be called at the end of this sequence, otherwise we may make
       // bugs.
       MadeChange |= foldSqrt(I, TTI, TLI);
     }
   }
 
   // We're done with transforms, so remove dead instructions.
   if (MadeChange)
     for (BasicBlock &BB : F)
       SimplifyInstructionsInBlock(&BB);
 
   return MadeChange;
 }
 
 /// This is the entry point for all transforms. Pass manager differences are
 /// handled in the callers of this function.
 static bool runImpl(Function &F, AssumptionCache &AC, TargetTransformInfo &TTI,
-                    TargetLibraryInfo &TLI, DominatorTree &DT) {
+                    TargetLibraryInfo &TLI, DominatorTree &DT,
+                    AliasAnalysis &AA) {
   bool MadeChange = false;
   const DataLayout &DL = F.getParent()->getDataLayout();
   TruncInstCombine TIC(AC, TLI, DL, DT);
   MadeChange |= TIC.run(F);
-  MadeChange |= foldUnusualPatterns(F, DT, TTI, TLI);
+  MadeChange |= foldUnusualPatterns(F, DT, TTI, TLI, AA);
   return MadeChange;
 }
 
 void AggressiveInstCombinerLegacyPass::getAnalysisUsage(
     AnalysisUsage &AU) const {
   AU.setPreservesCFG();
   AU.addRequired<AssumptionCacheTracker>();
   AU.addRequired<DominatorTreeWrapperPass>();
   AU.addRequired<TargetLibraryInfoWrapperPass>();
   AU.addRequired<TargetTransformInfoWrapperPass>();
   AU.addPreserved<AAResultsWrapperPass>();
   AU.addPreserved<BasicAAWrapperPass>();
   AU.addPreserved<DominatorTreeWrapperPass>();
   AU.addPreserved<GlobalsAAWrapperPass>();
+  AU.addRequired<AAResultsWrapperPass>();
 }
 
 bool AggressiveInstCombinerLegacyPass::runOnFunction(Function &F) {
   auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
   auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
   auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
-  return runImpl(F, AC, TTI, TLI, DT);
+  auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
+  return runImpl(F, AC, TTI, TLI, DT, AA);
 }
 
 PreservedAnalyses AggressiveInstCombinePass::run(Function &F,
                                                  FunctionAnalysisManager &AM) {
   auto &AC = AM.getResult<AssumptionAnalysis>(F);
   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
-  if (!runImpl(F, AC, TTI, TLI, DT)) {
+  auto &AA = AM.getResult<AAManager>(F);
+  if (!runImpl(F, AC, TTI, TLI, DT, AA)) {
     // No changes, all analyses are preserved.
     return PreservedAnalyses::all();
   }
   // Mark all the analyses that instcombine updates as preserved.
   PreservedAnalyses PA;
   PA.preserveSet<CFGAnalyses>();
   return PA;
 }

llvm/test/Transforms/AggressiveInstCombine/AArch64/or-load.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt < %s -passes=aggressive-instcombine -S -mtriple aarch64 -data-layout="e-n64" | FileCheck %s --check-prefixes=ALL
-; RUN: opt < %s -passes=aggressive-instcombine -S -mtriple aarch64 -data-layout="E-n64" | FileCheck %s --check-prefixes=ALL
+; RUN: opt < %s -passes=aggressive-instcombine -S -mtriple aarch64 -data-layout="e-n64" | FileCheck %s --check-prefixes=ALL,LE
+; RUN: opt < %s -passes=aggressive-instcombine -S -mtriple aarch64 -data-layout="E-n64" | FileCheck %s --check-prefixes=ALL,BE
 
 define i16 @loadCombine_2consecutive(ptr %p) {
 ;
 ; ALL-LABEL: @loadCombine_2consecutive(
 ; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
 ; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
 ; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
 ; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i16
 ; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i16
 ; ALL-NEXT:    [[S2:%.*]] = shl i16 [[E2]], 8
 ; ALL-NEXT:    [[O1:%.*]] = or i16 [[E1]], [[S2]]
 ; ALL-NEXT:    ret i16 [[O1]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %e1 = zext i8 %l1 to i16
   %e2 = zext i8 %l2 to i16
   %s2 = shl i16 %e2, 8
   %o1 = or i16 %e1, %s2
   ret i16 %o1
 }
 
 define i16 @loadCombine_2consecutive_BE(ptr %p) {
 ; ALL-LABEL: @loadCombine_2consecutive_BE(
 ; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
 ; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
 ; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
 ; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i16
 ; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i16
 ; ALL-NEXT:    [[S1:%.*]] = shl i16 [[E1]], 8
 ; ALL-NEXT:    [[O1:%.*]] = or i16 [[S1]], [[E2]]
 ; ALL-NEXT:    ret i16 [[O1]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %e1 = zext i8 %l1 to i16
   %e2 = zext i8 %l2 to i16
   %s1 = shl i16 %e1, 8
   %o1 = or i16 %s1, %e2
   ret i16 %o1
 }
 
 define i32 @loadCombine_4consecutive(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
-; ALL-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive(
+; LE-NEXT:    [[L1:%.*]] = load i32, ptr [[P:%.*]], align 1
+; LE-NEXT:    ret i32 [[L1]]
+;
+; BE-LABEL: @loadCombine_4consecutive(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; BE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; BE-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
   %l4 = load i8, ptr %p3
;
 
   %o1 = or i32 %e1, %s2
   %o2 = or i32 %o1, %s3
   %o3 = or i32 %o2, %s4
   ret i32 %o3
 }
 
 define i32 @loadCombine_4consecutive_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; LE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; LE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; LE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; LE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; LE-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
+; LE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
+; LE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
+; LE-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
+; LE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
+; LE-NEXT:    ret i32 [[O3]]
+;
+; BE-LABEL: @loadCombine_4consecutive_BE(
+; BE-NEXT:    [[L1:%.*]] = load i32, ptr [[P:%.*]], align 1
+; BE-NEXT:    ret i32 [[L1]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
   %l4 = load i8, ptr %p3
;
 
   %o1 = or i32 %s1, %s2
   %o2 = or i32 %o1, %s3
   %o3 = or i32 %o2, %e4
   ret i32 %o3
 }
 
 define i32 @loadCombine_4consecutive_alias(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive_alias(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    store i8 10, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
-; ALL-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_alias(
+; LE-NEXT:    [[L1:%.*]] = load i32, ptr [[P:%.*]], align 1
+; LE-NEXT:    store i8 10, ptr [[P]], align 1
+; LE-NEXT:    ret i32 [[L1]]
+;
+; BE-LABEL: @loadCombine_4consecutive_alias(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    store i8 10, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; BE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; BE-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   store i8 10, i8* %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
;
 
   %o1 = or i32 %e1, %s2
   %o2 = or i32 %o1, %s3
   %o3 = or i32 %o2, %s4
   ret i32 %o3
 }
 
 define i32 @loadCombine_4consecutive_alias_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive_alias_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    store i8 10, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_alias_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; LE-NEXT:    store i8 10, ptr [[P]], align 1
+; LE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; LE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; LE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; LE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; LE-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
+; LE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
+; LE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
+; LE-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
+; LE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
+; LE-NEXT:    ret i32 [[O3]]
+;
+; BE-LABEL: @loadCombine_4consecutive_alias_BE(
+; BE-NEXT:    [[L1:%.*]] = load i32, ptr [[P:%.*]], align 1
+; BE-NEXT:    store i8 10, ptr [[P]], align 1
+; BE-NEXT:    ret i32 [[L1]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   store i8 10, i8* %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
;
   %s3 = shl i32 %e3, 16
 
   %o1 = or i32 %e1, %s2
   %o2 = or i32 %o1, %s3
   ret i32 %o2
 }
 
 define i128 @loadCombine_i128(ptr %p) {
-; ALL-LABEL: @loadCombine_i128(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i32, ptr [[P]], align 4
-; ALL-NEXT:    [[L2:%.*]] = load i32, ptr [[P1]], align 4
-; ALL-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
-; ALL-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
-; ALL-NEXT:    [[E1:%.*]] = zext i32 [[L1]] to i128
-; ALL-NEXT:    [[E2:%.*]] = zext i32 [[L2]] to i128
-; ALL-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
-; ALL-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
-; ALL-NEXT:    [[S2:%.*]] = shl i128 [[E2]], 32
-; ALL-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 64
-; ALL-NEXT:    [[S4:%.*]] = shl i128 [[E4]], 96
-; ALL-NEXT:    [[O1:%.*]] = or i128 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i128 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[S4]]
-; ALL-NEXT:    ret i128 [[O3]]
+; LE-LABEL: @loadCombine_i128(
+; LE-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i64, ptr [[P]], align 4
+; LE-NEXT:    [[TMP1:%.*]] = zext i64 [[L1]] to i128
+; LE-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
+; LE-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
+; LE-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
+; LE-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
+; LE-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 64
+; LE-NEXT:    [[S4:%.*]] = shl i128 [[E4]], 96
+; LE-NEXT:    [[O2:%.*]] = or i128 [[TMP1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[S4]]
+; LE-NEXT:    ret i128 [[O3]]
+;
+; BE-LABEL: @loadCombine_i128(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i32, ptr [[P]], align 4
+; BE-NEXT:    [[L2:%.*]] = load i32, ptr [[P1]], align 4
+; BE-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
+; BE-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
+; BE-NEXT:    [[E1:%.*]] = zext i32 [[L1]] to i128
+; BE-NEXT:    [[E2:%.*]] = zext i32 [[L2]] to i128
+; BE-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
+; BE-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
+; BE-NEXT:    [[S2:%.*]] = shl i128 [[E2]], 32
+; BE-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 64
+; BE-NEXT:    [[S4:%.*]] = shl i128 [[E4]], 96
+; BE-NEXT:    [[O1:%.*]] = or i128 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i128 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[S4]]
+; BE-NEXT:    ret i128 [[O3]]
 ;
   %p1 = getelementptr i32, ptr %p, i32 1
   %p2 = getelementptr i32, ptr %p, i32 2
   %p3 = getelementptr i32, ptr %p, i32 3
   %l1 = load i32, ptr %p
   %l2 = load i32, ptr %p1
   %l3 = load i32, ptr %p2
   %l4 = load i32, ptr %p3
;
 
   %o1 = or i128 %e1, %s2
   %o2 = or i128 %o1, %s3
   %o3 = or i128 %o2, %s4
   ret i128 %o3
 }
 
 define i128 @loadCombine_i128_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_i128_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i32, ptr [[P]], align 4
-; ALL-NEXT:    [[L2:%.*]] = load i32, ptr [[P1]], align 4
-; ALL-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
-; ALL-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
-; ALL-NEXT:    [[E1:%.*]] = zext i32 [[L1]] to i128
-; ALL-NEXT:    [[E2:%.*]] = zext i32 [[L2]] to i128
-; ALL-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
-; ALL-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
-; ALL-NEXT:    [[S1:%.*]] = shl i128 [[E1]], 96
-; ALL-NEXT:    [[S2:%.*]] = shl i128 [[E2]], 64
-; ALL-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 32
-; ALL-NEXT:    [[O1:%.*]] = or i128 [[S1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i128 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[E4]]
-; ALL-NEXT:    ret i128 [[O3]]
+; LE-LABEL: @loadCombine_i128_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i32, ptr [[P]], align 4
+; LE-NEXT:    [[L2:%.*]] = load i32, ptr [[P1]], align 4
+; LE-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
+; LE-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
+; LE-NEXT:    [[E1:%.*]] = zext i32 [[L1]] to i128
+; LE-NEXT:    [[E2:%.*]] = zext i32 [[L2]] to i128
+; LE-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
+; LE-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
+; LE-NEXT:    [[S1:%.*]] = shl i128 [[E1]], 96
+; LE-NEXT:    [[S2:%.*]] = shl i128 [[E2]], 64
+; LE-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 32
+; LE-NEXT:    [[O1:%.*]] = or i128 [[S1]], [[S2]]
+; LE-NEXT:    [[O2:%.*]] = or i128 [[O1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[E4]]
+; LE-NEXT:    ret i128 [[O3]]
+;
+; BE-LABEL: @loadCombine_i128_BE(
+; BE-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i64, ptr [[P]], align 4
+; BE-NEXT:    [[TMP1:%.*]] = zext i64 [[L1]] to i128
+; BE-NEXT:    [[TMP2:%.*]] = shl i128 [[TMP1]], 64
+; BE-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
+; BE-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
+; BE-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
+; BE-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
+; BE-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 32
+; BE-NEXT:    [[O2:%.*]] = or i128 [[TMP2]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[E4]]
+; BE-NEXT:    ret i128 [[O3]]
 ;
   %p1 = getelementptr i32, ptr %p, i32 1
   %p2 = getelementptr i32, ptr %p, i32 2
   %p3 = getelementptr i32, ptr %p, i32 3
   %l1 = load i32, ptr %p
   %l2 = load i32, ptr %p1
   %l3 = load i32, ptr %p2
   %l4 = load i32, ptr %p3
;
 
   %o1 = or i128 %s1, %s2
   %o2 = or i128 %o1, %s3
   %o3 = or i128 %o2, %e4
   ret i128 %o3
 }
 
 define i64 @loadCombine_i64(ptr %p) {
-; ALL-LABEL: @loadCombine_i64(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i16, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i16, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i16, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 2
-; ALL-NEXT:    [[L2:%.*]] = load i16, ptr [[P1]], align 2
-; ALL-NEXT:    [[L3:%.*]] = load i16, ptr [[P2]], align 2
-; ALL-NEXT:    [[L4:%.*]] = load i16, ptr [[P3]], align 2
-; ALL-NEXT:    [[E1:%.*]] = zext i16 [[L1]] to i64
-; ALL-NEXT:    [[E2:%.*]] = zext i16 [[L2]] to i64
-; ALL-NEXT:    [[E3:%.*]] = zext i16 [[L3]] to i64
-; ALL-NEXT:    [[E4:%.*]] = zext i16 [[L4]] to i64
-; ALL-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 16
-; ALL-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 32
-; ALL-NEXT:    [[S4:%.*]] = shl i64 [[E4]], 48
-; ALL-NEXT:    [[O1:%.*]] = or i64 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i64 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i64 [[O2]], [[S4]]
-; ALL-NEXT:    ret i64 [[O3]]
+; LE-LABEL: @loadCombine_i64(
+; LE-NEXT:    [[L1:%.*]] = load i64, ptr [[P:%.*]], align 2
+; LE-NEXT:    ret i64 [[L1]]
+;
+; BE-LABEL: @loadCombine_i64(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i16, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i16, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i16, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 2
+; BE-NEXT:    [[L2:%.*]] = load i16, ptr [[P1]], align 2
+; BE-NEXT:    [[L3:%.*]] = load i16, ptr [[P2]], align 2
+; BE-NEXT:    [[L4:%.*]] = load i16, ptr [[P3]], align 2
+; BE-NEXT:    [[E1:%.*]] = zext i16 [[L1]] to i64
+; BE-NEXT:    [[E2:%.*]] = zext i16 [[L2]] to i64
+; BE-NEXT:    [[E3:%.*]] = zext i16 [[L3]] to i64
+; BE-NEXT:    [[E4:%.*]] = zext i16 [[L4]] to i64
+; BE-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 16
+; BE-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 32
+; BE-NEXT:    [[S4:%.*]] = shl i64 [[E4]], 48
+; BE-NEXT:    [[O1:%.*]] = or i64 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i64 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i64 [[O2]], [[S4]]
+; BE-NEXT:    ret i64 [[O3]]
 ;
   %p1 = getelementptr i16, ptr %p, i32 1
   %p2 = getelementptr i16, ptr %p, i32 2
   %p3 = getelementptr i16, ptr %p, i32 3
   %l1 = load i16, ptr %p
   %l2 = load i16, ptr %p1
   %l3 = load i16, ptr %p2
   %l4 = load i16, ptr %p3
;
 
   %o1 = or i64 %e1, %s2
   %o2 = or i64 %o1, %s3
   %o3 = or i64 %o2, %s4
   ret i64 %o3
 }
 
 define i64 @loadCombine_i64_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_i64_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i16, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i16, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i16, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 2
-; ALL-NEXT:    [[L2:%.*]] = load i16, ptr [[P1]], align 2
-; ALL-NEXT:    [[L3:%.*]] = load i16, ptr [[P2]], align 2
-; ALL-NEXT:    [[L4:%.*]] = load i16, ptr [[P3]], align 2
-; ALL-NEXT:    [[E1:%.*]] = zext i16 [[L1]] to i64
-; ALL-NEXT:    [[E2:%.*]] = zext i16 [[L2]] to i64
-; ALL-NEXT:    [[E3:%.*]] = zext i16 [[L3]] to i64
-; ALL-NEXT:    [[E4:%.*]] = zext i16 [[L4]] to i64
-; ALL-NEXT:    [[S1:%.*]] = shl i64 [[E1]], 48
-; ALL-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 32
-; ALL-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 16
-; ALL-NEXT:    [[O1:%.*]] = or i64 [[S1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i64 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i64 [[O2]], [[E4]]
-; ALL-NEXT:    ret i64 [[O3]]
+; LE-LABEL: @loadCombine_i64_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i16, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i16, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i16, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 2
+; LE-NEXT:    [[L2:%.*]] = load i16, ptr [[P1]], align 2
+; LE-NEXT:    [[L3:%.*]] = load i16, ptr [[P2]], align 2
+; LE-NEXT:    [[L4:%.*]] = load i16, ptr [[P3]], align 2
+; LE-NEXT:    [[E1:%.*]] = zext i16 [[L1]] to i64
+; LE-NEXT:    [[E2:%.*]] = zext i16 [[L2]] to i64
+; LE-NEXT:    [[E3:%.*]] = zext i16 [[L3]] to i64
+; LE-NEXT:    [[E4:%.*]] = zext i16 [[L4]] to i64
+; LE-NEXT:    [[S1:%.*]] = shl i64 [[E1]], 48
+; LE-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 32
+; LE-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 16
+; LE-NEXT:    [[O1:%.*]] = or i64 [[S1]], [[S2]]
+; LE-NEXT:    [[O2:%.*]] = or i64 [[O1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i64 [[O2]], [[E4]]
+; LE-NEXT:    ret i64 [[O3]]
+;
+; BE-LABEL: @loadCombine_i64_BE(
+; BE-NEXT:    [[L1:%.*]] = load i64, ptr [[P:%.*]], align 2
+; BE-NEXT:    ret i64 [[L1]]
 ;
   %p1 = getelementptr i16, ptr %p, i32 1
   %p2 = getelementptr i16, ptr %p, i32 2
   %p3 = getelementptr i16, ptr %p, i32 3
   %l1 = load i16, ptr %p
   %l2 = load i16, ptr %p1
   %l3 = load i16, ptr %p2
   %l4 = load i16, ptr %p3
;
   %e1 = zext i8 %l1 to i16
   %e2 = zext i8 %l2 to i16
   %s2 = shl i16 %e2, 8
   %o1 = or i16 %e1, %s2
   ret i16 %o1
 }
 
 define i64 @load64_farLoads(ptr %ptr) {
-; ALL-LABEL: @load64_farLoads(
-; ALL-NEXT:  entry:
-; ALL-NEXT:    [[TMP0:%.*]] = load i8, ptr [[PTR:%.*]], align 1
-; ALL-NEXT:    [[CONV:%.*]] = zext i8 [[TMP0]] to i64
-; ALL-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 1
-; ALL-NEXT:    [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX1]], align 1
-; ALL-NEXT:    [[CONV2:%.*]] = zext i8 [[TMP1]] to i64
-; ALL-NEXT:    [[SHL:%.*]] = shl i64 [[CONV2]], 8
-; ALL-NEXT:    [[OR:%.*]] = or i64 [[CONV]], [[SHL]]
-; ALL-NEXT:    [[ARRAYIDX3:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 2
-; ALL-NEXT:    [[TMP2:%.*]] = load i8, ptr [[ARRAYIDX3]], align 1
-; ALL-NEXT:    [[CONV4:%.*]] = zext i8 [[TMP2]] to i64
-; ALL-NEXT:    [[SHL5:%.*]] = shl i64 [[CONV4]], 16
-; ALL-NEXT:    [[OR6:%.*]] = or i64 [[OR]], [[SHL5]]
-; ALL-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 3
-; ALL-NEXT:    [[TMP3:%.*]] = load i8, ptr [[ARRAYIDX7]], align 1
-; ALL-NEXT:    [[CONV8:%.*]] = zext i8 [[TMP3]] to i64
-; ALL-NEXT:    [[SHL9:%.*]] = shl i64 [[CONV8]], 24
-; ALL-NEXT:    [[OR10:%.*]] = or i64 [[OR6]], [[SHL9]]
-; ALL-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 4
-; ALL-NEXT:    [[TMP4:%.*]] = load i8, ptr [[ARRAYIDX11]], align 1
-; ALL-NEXT:    [[CONV12:%.*]] = zext i8 [[TMP4]] to i64
-; ALL-NEXT:    [[SHL13:%.*]] = shl i64 [[CONV12]], 32
-; ALL-NEXT:    [[OR14:%.*]] = or i64 [[OR10]], [[SHL13]]
-; ALL-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 5
-; ALL-NEXT:    [[TMP5:%.*]] = load i8, ptr [[ARRAYIDX15]], align 1
-; ALL-NEXT:    [[CONV16:%.*]] = zext i8 [[TMP5]] to i64
-; ALL-NEXT:    [[SHL17:%.*]] = shl i64 [[CONV16]], 40
-; ALL-NEXT:    [[OR18:%.*]] = or i64 [[OR14]], [[SHL17]]
-; ALL-NEXT:    [[ARRAYIDX19:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 6
-; ALL-NEXT:    [[TMP6:%.*]] = load i8, ptr [[ARRAYIDX19]], align 1
-; ALL-NEXT:    [[CONV20:%.*]] = zext i8 [[TMP6]] to i64
-; ALL-NEXT:    [[SHL21:%.*]] = shl i64 [[CONV20]], 48
-; ALL-NEXT:    [[OR22:%.*]] = or i64 [[OR18]], [[SHL21]]
-; ALL-NEXT:    [[ARRAYIDX23:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 7
-; ALL-NEXT:    [[TMP7:%.*]] = load i8, ptr [[ARRAYIDX23]], align 1
-; ALL-NEXT:    [[CONV24:%.*]] = zext i8 [[TMP7]] to i64
-; ALL-NEXT:    [[SHL25:%.*]] = shl i64 [[CONV24]], 56
-; ALL-NEXT:    [[OR26:%.*]] = or i64 [[OR22]], [[SHL25]]
-; ALL-NEXT:    ret i64 [[OR26]]
+; LE-LABEL: @load64_farLoads(
+; LE-NEXT:  entry:
+; LE-NEXT:    [[TMP0:%.*]] = load i64, ptr [[PTR:%.*]], align 1
+; LE-NEXT:    ret i64 [[TMP0]]
+;
+; BE-LABEL: @load64_farLoads(
+; BE-NEXT:  entry:
+; BE-NEXT:    [[TMP0:%.*]] = load i8, ptr [[PTR:%.*]], align 1
+; BE-NEXT:    [[CONV:%.*]] = zext i8 [[TMP0]] to i64
+; BE-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 1
+; BE-NEXT:    [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX1]], align 1
+; BE-NEXT:    [[CONV2:%.*]] = zext i8 [[TMP1]] to i64
+; BE-NEXT:    [[SHL:%.*]] = shl i64 [[CONV2]], 8
+; BE-NEXT:    [[OR:%.*]] = or i64 [[CONV]], [[SHL]]
+; BE-NEXT:    [[ARRAYIDX3:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 2
+; BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[ARRAYIDX3]], align 1
+; BE-NEXT:    [[CONV4:%.*]] = zext i8 [[TMP2]] to i64
+; BE-NEXT:    [[SHL5:%.*]] = shl i64 [[CONV4]], 16
+; BE-NEXT:    [[OR6:%.*]] = or i64 [[OR]], [[SHL5]]
+; BE-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 3
+; BE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[ARRAYIDX7]], align 1
+; BE-NEXT:    [[CONV8:%.*]] = zext i8 [[TMP3]] to i64
+; BE-NEXT:    [[SHL9:%.*]] = shl i64 [[CONV8]], 24
+; BE-NEXT:    [[OR10:%.*]] = or i64 [[OR6]], [[SHL9]]
+; BE-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 4
+; BE-NEXT:    [[TMP4:%.*]] = load i8, ptr [[ARRAYIDX11]], align 1
+; BE-NEXT:    [[CONV12:%.*]] = zext i8 [[TMP4]] to i64
+; BE-NEXT:    [[SHL13:%.*]] = shl i64 [[CONV12]], 32
+; BE-NEXT:    [[OR14:%.*]] = or i64 [[OR10]], [[SHL13]]
+; BE-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 5
+; BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[ARRAYIDX15]], align 1
+; BE-NEXT:    [[CONV16:%.*]] = zext i8 [[TMP5]] to i64
+; BE-NEXT:    [[SHL17:%.*]] = shl i64 [[CONV16]], 40
+; BE-NEXT:    [[OR18:%.*]] = or i64 [[OR14]], [[SHL17]]
+; BE-NEXT:    [[ARRAYIDX19:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 6
+; BE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[ARRAYIDX19]], align 1
+; BE-NEXT:    [[CONV20:%.*]] = zext i8 [[TMP6]] to i64
+; BE-NEXT:    [[SHL21:%.*]] = shl i64 [[CONV20]], 48
+; BE-NEXT:    [[OR22:%.*]] = or i64 [[OR18]], [[SHL21]]
+; BE-NEXT:    [[ARRAYIDX23:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 7
+; BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[ARRAYIDX23]], align 1
+; BE-NEXT:    [[CONV24:%.*]] = zext i8 [[TMP7]] to i64
+; BE-NEXT:    [[SHL25:%.*]] = shl i64 [[CONV24]], 56
+; BE-NEXT:    [[OR26:%.*]] = or i64 [[OR22]], [[SHL25]]
+; BE-NEXT:    ret i64 [[OR26]]
 ;
 entry:
   %0 = load i8, ptr %ptr, align 1
   %conv = zext i8 %0 to i64
   %arrayidx1 = getelementptr inbounds i8, ptr %ptr, i64 1
   %1 = load i8, ptr %arrayidx1, align 1
   %conv2 = zext i8 %1 to i64
   %shl = shl i64 %conv2, 8
entry:
   %7 = load i8, ptr %arrayidx23, align 1
   %conv24 = zext i8 %7 to i64
   %shl25 = shl i64 %conv24, 56
   %or26 = or i64 %or22, %shl25
   ret i64 %or26
 }
 
 define i32 @loadCombine_4consecutive_metadata(ptr %p, ptr %pstr) {
-; ALL-LABEL: @loadCombine_4consecutive_metadata(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1, !alias.scope !0
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1, !alias.scope !0
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1, !alias.scope !0
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1, !alias.scope !0
-; ALL-NEXT:    store i32 25, ptr [[PSTR:%.*]], align 4, !noalias !0
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
-; ALL-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_metadata(
+; LE-NEXT:    [[L1:%.*]] = load i32, ptr [[P:%.*]], align 1, !alias.scope !0
+; LE-NEXT:    store i32 25, ptr [[PSTR:%.*]], align 4, !noalias !0
+; LE-NEXT:    ret i32 [[L1]]
+;
+; BE-LABEL: @loadCombine_4consecutive_metadata(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1, !alias.scope !0
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1, !alias.scope !0
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1, !alias.scope !0
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1, !alias.scope !0
+; BE-NEXT:    store i32 25, ptr [[PSTR:%.*]], align 4, !noalias !0
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; BE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; BE-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p, !alias.scope !2
   %l2 = load i8, ptr %p1, !alias.scope !2
   %l3 = load i8, ptr %p2, !alias.scope !2
   %l4 = load i8, ptr %p3, !alias.scope !2

llvm/test/Transforms/AggressiveInstCombine/X86/or-load.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt < %s -passes=aggressive-instcombine -mtriple x86_64-none-eabi -mattr=avx2 -data-layout="e-n64" -S | FileCheck %s --check-prefixes=ALL
-; RUN: opt < %s -passes=aggressive-instcombine -mtriple x86_64-none-eabi -mattr=avx2 -data-layout="E-n64" -S | FileCheck %s --check-prefixes=ALL
+; RUN: opt < %s -passes=aggressive-instcombine -mtriple x86_64-none-eabi -mattr=avx2 -data-layout="e-n64" -S | FileCheck %s --check-prefixes=ALL,LE
+; RUN: opt < %s -passes=aggressive-instcombine -mtriple x86_64-none-eabi -mattr=avx2 -data-layout="E-n64" -S | FileCheck %s --check-prefixes=ALL,BE
 
 define i16 @loadCombine_2consecutive(ptr %p) {
 ;
-; ALL-LABEL: @loadCombine_2consecutive(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i16
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i16
-; ALL-NEXT:    [[S2:%.*]] = shl i16 [[E2]], 8
-; ALL-NEXT:    [[O1:%.*]] = or i16 [[E1]], [[S2]]
-; ALL-NEXT:    ret i16 [[O1]]
+; LE-LABEL: @loadCombine_2consecutive(
+; LE-NEXT:    [[L1:%.*]] = load i16, ptr [[P:%.*]], align 1
+; LE-NEXT:    ret i16 [[L1]]
+;
+; BE-LABEL: @loadCombine_2consecutive(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i16
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i16
+; BE-NEXT:    [[S2:%.*]] = shl i16 [[E2]], 8
+; BE-NEXT:    [[O1:%.*]] = or i16 [[E1]], [[S2]]
+; BE-NEXT:    ret i16 [[O1]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %e1 = zext i8 %l1 to i16
   %e2 = zext i8 %l2 to i16
   %s2 = shl i16 %e2, 8
   %o1 = or i16 %e1, %s2
   ret i16 %o1
 }
 
 define i16 @loadCombine_2consecutive_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_2consecutive_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i16
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i16
-; ALL-NEXT:    [[S1:%.*]] = shl i16 [[E1]], 8
-; ALL-NEXT:    [[O1:%.*]] = or i16 [[S1]], [[E2]]
-; ALL-NEXT:    ret i16 [[O1]]
+; LE-LABEL: @loadCombine_2consecutive_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; LE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; LE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i16
+; LE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i16
+; LE-NEXT:    [[S1:%.*]] = shl i16 [[E1]], 8
+; LE-NEXT:    [[O1:%.*]] = or i16 [[S1]], [[E2]]
+; LE-NEXT:    ret i16 [[O1]]
+;
+; BE-LABEL: @loadCombine_2consecutive_BE(
+; BE-NEXT:    [[L1:%.*]] = load i16, ptr [[P:%.*]], align 1
+; BE-NEXT:    ret i16 [[L1]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %e1 = zext i8 %l1 to i16
   %e2 = zext i8 %l2 to i16
   %s1 = shl i16 %e1, 8
   %o1 = or i16 %s1, %e2
   ret i16 %o1
 }
 
 define i32 @loadCombine_4consecutive(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
-; ALL-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive(
+; LE-NEXT:    [[L1:%.*]] = load i32, ptr [[P:%.*]], align 1
+; LE-NEXT:    ret i32 [[L1]]
+;
+; BE-LABEL: @loadCombine_4consecutive(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; BE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; BE-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
   %l4 = load i8, ptr %p3
;
 
   %o1 = or i32 %e1, %s2
   %o2 = or i32 %o1, %s3
   %o3 = or i32 %o2, %s4
   ret i32 %o3
 }
 
 define i32 @loadCombine_4consecutive_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; LE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; LE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; LE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; LE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; LE-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
+; LE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
+; LE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
+; LE-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
+; LE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
+; LE-NEXT:    ret i32 [[O3]]
+;
+; BE-LABEL: @loadCombine_4consecutive_BE(
+; BE-NEXT:    [[L1:%.*]] = load i32, ptr [[P:%.*]], align 1
+; BE-NEXT:    ret i32 [[L1]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
   %l4 = load i8, ptr %p3
;
 
   %o1 = or i32 %s1, %s2
   %o2 = or i32 %o1, %s3
   %o3 = or i32 %o2, %e4
   ret i32 %o3
 }
 
 define i32 @loadCombine_4consecutive_alias(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive_alias(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    store i8 10, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
-; ALL-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_alias(
+; LE-NEXT:    [[L1:%.*]] = load i32, ptr [[P:%.*]], align 1
+; LE-NEXT:    store i8 10, ptr [[P]], align 1
+; LE-NEXT:    ret i32 [[L1]]
+;
+; BE-LABEL: @loadCombine_4consecutive_alias(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    store i8 10, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; BE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; BE-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   store i8 10, i8* %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
;
 
   %o1 = or i32 %e1, %s2
   %o2 = or i32 %o1, %s3
   %o3 = or i32 %o2, %s4
   ret i32 %o3
 }
 
 define i32 @loadCombine_4consecutive_alias_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive_alias_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    store i8 10, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_alias_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; LE-NEXT:    store i8 10, ptr [[P]], align 1
+; LE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; LE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; LE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; LE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; LE-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
+; LE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
+; LE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
+; LE-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
+; LE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
+; LE-NEXT:    ret i32 [[O3]]
+;
+; BE-LABEL: @loadCombine_4consecutive_alias_BE(
+; BE-NEXT:    [[L1:%.*]] = load i32, ptr [[P:%.*]], align 1
+; BE-NEXT:    store i8 10, ptr [[P]], align 1
+; BE-NEXT:    ret i32 [[L1]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   store i8 10, i8* %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
;
 
   %o1 = or i32 %s1, %s2
   %o2 = or i32 %o1, %s3
   %o3 = or i32 %o2, %e4
   ret i32 %o3
 }
 
 define i32 @loadCombine_4consecutive_alias2(ptr %p, ptr %pstr) {
-; ALL-LABEL: @loadCombine_4consecutive_alias2(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    store i8 10, ptr [[PSTR:%.*]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
-; ALL-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_alias2(
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 1
+; LE-NEXT:    [[TMP1:%.*]] = zext i16 [[L1]] to i32
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    store i8 10, ptr [[PSTR:%.*]], align 1
+; LE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; LE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; LE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; LE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; LE-NEXT:    [[O2:%.*]] = or i32 [[TMP1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
+; LE-NEXT:    ret i32 [[O3]]
+;
+; BE-LABEL: @loadCombine_4consecutive_alias2(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    store i8 10, ptr [[PSTR:%.*]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; BE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; BE-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
   store i8 10, i8* %pstr
;
 
   %o1 = or i32 %e1, %s2
   %o2 = or i32 %o1, %s3
   %o3 = or i32 %o2, %s4
   ret i32 %o3
 }
 
 define i32 @loadCombine_4consecutive_alias2_BE(ptr %p, ptr %pstr) {
-; ALL-LABEL: @loadCombine_4consecutive_alias2_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    store i8 10, ptr [[PSTR:%.*]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_alias2_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; LE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    store i8 10, ptr [[PSTR:%.*]], align 1
+; LE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; LE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; LE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; LE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; LE-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
+; LE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
+; LE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
+; LE-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
+; LE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
+; LE-NEXT:    ret i32 [[O3]]
+;
+; BE-LABEL: @loadCombine_4consecutive_alias2_BE(
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 1
+; BE-NEXT:    [[TMP1:%.*]] = zext i16 [[L1]] to i32
+; BE-NEXT:    [[TMP2:%.*]] = shl i32 [[TMP1]], 16
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    store i8 10, ptr [[PSTR:%.*]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
+; BE-NEXT:    [[O2:%.*]] = or i32 [[TMP2]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
   store i8 10, i8* %pstr
;
 
   %o1 = or i32 %s1, %s2
   %o2 = or i32 %o1, %s3
   %o3 = or i32 %o2, %e4
   ret i32 %o3
 }
 
 define i32 @loadCombine_4consecutive_alias3(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive_alias3(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    store i8 10, ptr [[P3]], align 1
-; ALL-NEXT:    store i8 5, ptr [[P]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
-; ALL-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_alias3(
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 1
+; LE-NEXT:    [[TMP1:%.*]] = zext i16 [[L1]] to i32
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    store i8 10, ptr [[P3]], align 1
+; LE-NEXT:    store i8 5, ptr [[P]], align 1
+; LE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; LE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; LE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; LE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; LE-NEXT:    [[O2:%.*]] = or i32 [[TMP1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
+; LE-NEXT:    ret i32 [[O3]]
+;
+; BE-LABEL: @loadCombine_4consecutive_alias3(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    store i8 10, ptr [[P3]], align 1
+; BE-NEXT:    store i8 5, ptr [[P]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; BE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; BE-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
   store i8 10, i8* %p3
;
 
   %o1 = or i32 %e1, %s2
   %o2 = or i32 %o1, %s3
   %o3 = or i32 %o2, %s4
   ret i32 %o3
 }
 
 define i32 @loadCombine_4consecutive_alias3_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive_alias3_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    store i8 10, ptr [[P3]], align 1
-; ALL-NEXT:    store i8 5, ptr [[P]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_alias3_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; LE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    store i8 10, ptr [[P3]], align 1
+; LE-NEXT:    store i8 5, ptr [[P]], align 1
+; LE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; LE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; LE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; LE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; LE-NEXT:    [[S1:%.*]] = shl i32 [[E1]], 24
+; LE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 16
+; LE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
+; LE-NEXT:    [[O1:%.*]] = or i32 [[S1]], [[S2]]
+; LE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
+; LE-NEXT:    ret i32 [[O3]]
+;
+; BE-LABEL: @loadCombine_4consecutive_alias3_BE(
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 1
+; BE-NEXT:    [[TMP1:%.*]] = zext i16 [[L1]] to i32
+; BE-NEXT:    [[TMP2:%.*]] = shl i32 [[TMP1]], 16
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    store i8 10, ptr [[P3]], align 1
+; BE-NEXT:    store i8 5, ptr [[P]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 8
+; BE-NEXT:    [[O2:%.*]] = or i32 [[TMP2]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E4]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
   store i8 10, i8* %p3
;
   call void @use(i32 %o1)
   %o2 = or i32 %o1, %s3
   call void @use(i32 %o2)
   %o3 = or i32 %o2, %s4
   ret i32 %o3
 }
 
 define i32 @loadCombine_parLoad1(ptr %p) {
-; ALL-LABEL: @loadCombine_parLoad1(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    ret i32 [[O2]]
+; LE-LABEL: @loadCombine_parLoad1(
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 2
+; LE-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 1
+; LE-NEXT:    [[TMP1:%.*]] = zext i16 [[L1]] to i32
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; LE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; LE-NEXT:    [[O2:%.*]] = or i32 [[TMP1]], [[S3]]
+; LE-NEXT:    ret i32 [[O2]]
+;
+; BE-LABEL: @loadCombine_parLoad1(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; BE-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; BE-NEXT:    ret i32 [[O2]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
 
   %e1 = zext i8 %l1 to i32
   %e2 = zext i8 %l2 to i32
   %e3 = zext i8 %l3 to i32
 
   %s2 = shl i32 %e2, 8
   %s3 = shl i32 %e3, 16
 
   %o1 = or i32 %e1, %s2
   %o2 = or i32 %o1, %s3
   ret i32 %o2
 }
 
 define i128 @loadCombine_i128(ptr %p) {
-; ALL-LABEL: @loadCombine_i128(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i32, ptr [[P]], align 4
-; ALL-NEXT:    [[L2:%.*]] = load i32, ptr [[P1]], align 4
-; ALL-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
-; ALL-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
-; ALL-NEXT:    [[E1:%.*]] = zext i32 [[L1]] to i128
-; ALL-NEXT:    [[E2:%.*]] = zext i32 [[L2]] to i128
-; ALL-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
-; ALL-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
-; ALL-NEXT:    [[S2:%.*]] = shl i128 [[E2]], 32
-; ALL-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 64
-; ALL-NEXT:    [[S4:%.*]] = shl i128 [[E4]], 96
-; ALL-NEXT:    [[O1:%.*]] = or i128 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i128 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[S4]]
-; ALL-NEXT:    ret i128 [[O3]]
+; LE-LABEL: @loadCombine_i128(
+; LE-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i64, ptr [[P]], align 4
+; LE-NEXT:    [[TMP1:%.*]] = zext i64 [[L1]] to i128
+; LE-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
+; LE-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
+; LE-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
+; LE-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
+; LE-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 64
+; LE-NEXT:    [[S4:%.*]] = shl i128 [[E4]], 96
+; LE-NEXT:    [[O2:%.*]] = or i128 [[TMP1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[S4]]
+; LE-NEXT:    ret i128 [[O3]]
+;
+; BE-LABEL: @loadCombine_i128(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i32, ptr [[P]], align 4
+; BE-NEXT:    [[L2:%.*]] = load i32, ptr [[P1]], align 4
+; BE-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
+; BE-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
+; BE-NEXT:    [[E1:%.*]] = zext i32 [[L1]] to i128
+; BE-NEXT:    [[E2:%.*]] = zext i32 [[L2]] to i128
+; BE-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
+; BE-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
+; BE-NEXT:    [[S2:%.*]] = shl i128 [[E2]], 32
+; BE-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 64
+; BE-NEXT:    [[S4:%.*]] = shl i128 [[E4]], 96
+; BE-NEXT:    [[O1:%.*]] = or i128 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i128 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[S4]]
+; BE-NEXT:    ret i128 [[O3]]
 ;
   %p1 = getelementptr i32, ptr %p, i32 1
   %p2 = getelementptr i32, ptr %p, i32 2
   %p3 = getelementptr i32, ptr %p, i32 3
   %l1 = load i32, ptr %p
   %l2 = load i32, ptr %p1
   %l3 = load i32, ptr %p2
   %l4 = load i32, ptr %p3
;
 
   %o1 = or i128 %e1, %s2
   %o2 = or i128 %o1, %s3
   %o3 = or i128 %o2, %s4
   ret i128 %o3
 }
 
 define i128 @loadCombine_i128_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_i128_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i32, ptr [[P]], align 4
-; ALL-NEXT:    [[L2:%.*]] = load i32, ptr [[P1]], align 4
-; ALL-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
-; ALL-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
-; ALL-NEXT:    [[E1:%.*]] = zext i32 [[L1]] to i128
-; ALL-NEXT:    [[E2:%.*]] = zext i32 [[L2]] to i128
-; ALL-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
-; ALL-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
-; ALL-NEXT:    [[S1:%.*]] = shl i128 [[E1]], 96
-; ALL-NEXT:    [[S2:%.*]] = shl i128 [[E2]], 64
-; ALL-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 32
-; ALL-NEXT:    [[O1:%.*]] = or i128 [[S1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i128 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[E4]]
-; ALL-NEXT:    ret i128 [[O3]]
+; LE-LABEL: @loadCombine_i128_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i32, ptr [[P]], align 4
+; LE-NEXT:    [[L2:%.*]] = load i32, ptr [[P1]], align 4
+; LE-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
+; LE-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
+; LE-NEXT:    [[E1:%.*]] = zext i32 [[L1]] to i128
+; LE-NEXT:    [[E2:%.*]] = zext i32 [[L2]] to i128
+; LE-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
+; LE-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
+; LE-NEXT:    [[S1:%.*]] = shl i128 [[E1]], 96
+; LE-NEXT:    [[S2:%.*]] = shl i128 [[E2]], 64
+; LE-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 32
+; LE-NEXT:    [[O1:%.*]] = or i128 [[S1]], [[S2]]
+; LE-NEXT:    [[O2:%.*]] = or i128 [[O1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[E4]]
+; LE-NEXT:    ret i128 [[O3]]
+;
+; BE-LABEL: @loadCombine_i128_BE(
+; BE-NEXT:    [[P2:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i32, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i64, ptr [[P]], align 4
+; BE-NEXT:    [[TMP1:%.*]] = zext i64 [[L1]] to i128
+; BE-NEXT:    [[TMP2:%.*]] = shl i128 [[TMP1]], 64
+; BE-NEXT:    [[L3:%.*]] = load i32, ptr [[P2]], align 4
+; BE-NEXT:    [[L4:%.*]] = load i32, ptr [[P3]], align 4
+; BE-NEXT:    [[E3:%.*]] = zext i32 [[L3]] to i128
+; BE-NEXT:    [[E4:%.*]] = zext i32 [[L4]] to i128
+; BE-NEXT:    [[S3:%.*]] = shl i128 [[E3]], 32
+; BE-NEXT:    [[O2:%.*]] = or i128 [[TMP2]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i128 [[O2]], [[E4]]
+; BE-NEXT:    ret i128 [[O3]]
 ;
   %p1 = getelementptr i32, ptr %p, i32 1
   %p2 = getelementptr i32, ptr %p, i32 2
   %p3 = getelementptr i32, ptr %p, i32 3
   %l1 = load i32, ptr %p
   %l2 = load i32, ptr %p1
   %l3 = load i32, ptr %p2
   %l4 = load i32, ptr %p3
;
 
   %o1 = or i128 %s1, %s2
   %o2 = or i128 %o1, %s3
   %o3 = or i128 %o2, %e4
   ret i128 %o3
 }
 
 define i64 @loadCombine_i64(ptr %p) {
-; ALL-LABEL: @loadCombine_i64(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i16, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i16, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i16, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 2
-; ALL-NEXT:    [[L2:%.*]] = load i16, ptr [[P1]], align 2
-; ALL-NEXT:    [[L3:%.*]] = load i16, ptr [[P2]], align 2
-; ALL-NEXT:    [[L4:%.*]] = load i16, ptr [[P3]], align 2
-; ALL-NEXT:    [[E1:%.*]] = zext i16 [[L1]] to i64
-; ALL-NEXT:    [[E2:%.*]] = zext i16 [[L2]] to i64
-; ALL-NEXT:    [[E3:%.*]] = zext i16 [[L3]] to i64
-; ALL-NEXT:    [[E4:%.*]] = zext i16 [[L4]] to i64
-; ALL-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 16
-; ALL-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 32
-; ALL-NEXT:    [[S4:%.*]] = shl i64 [[E4]], 48
-; ALL-NEXT:    [[O1:%.*]] = or i64 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i64 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i64 [[O2]], [[S4]]
-; ALL-NEXT:    ret i64 [[O3]]
+; LE-LABEL: @loadCombine_i64(
+; LE-NEXT:    [[L1:%.*]] = load i64, ptr [[P:%.*]], align 2
+; LE-NEXT:    ret i64 [[L1]]
+;
+; BE-LABEL: @loadCombine_i64(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i16, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i16, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i16, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 2
+; BE-NEXT:    [[L2:%.*]] = load i16, ptr [[P1]], align 2
+; BE-NEXT:    [[L3:%.*]] = load i16, ptr [[P2]], align 2
+; BE-NEXT:    [[L4:%.*]] = load i16, ptr [[P3]], align 2
+; BE-NEXT:    [[E1:%.*]] = zext i16 [[L1]] to i64
+; BE-NEXT:    [[E2:%.*]] = zext i16 [[L2]] to i64
+; BE-NEXT:    [[E3:%.*]] = zext i16 [[L3]] to i64
+; BE-NEXT:    [[E4:%.*]] = zext i16 [[L4]] to i64
+; BE-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 16
+; BE-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 32
+; BE-NEXT:    [[S4:%.*]] = shl i64 [[E4]], 48
+; BE-NEXT:    [[O1:%.*]] = or i64 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i64 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i64 [[O2]], [[S4]]
+; BE-NEXT:    ret i64 [[O3]]
 ;
   %p1 = getelementptr i16, ptr %p, i32 1
   %p2 = getelementptr i16, ptr %p, i32 2
   %p3 = getelementptr i16, ptr %p, i32 3
   %l1 = load i16, ptr %p
   %l2 = load i16, ptr %p1
   %l3 = load i16, ptr %p2
   %l4 = load i16, ptr %p3
;
 
   %o1 = or i64 %e1, %s2
   %o2 = or i64 %o1, %s3
   %o3 = or i64 %o2, %s4
   ret i64 %o3
 }
 
 define i64 @loadCombine_i64_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_i64_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i16, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i16, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i16, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 2
-; ALL-NEXT:    [[L2:%.*]] = load i16, ptr [[P1]], align 2
-; ALL-NEXT:    [[L3:%.*]] = load i16, ptr [[P2]], align 2
-; ALL-NEXT:    [[L4:%.*]] = load i16, ptr [[P3]], align 2
-; ALL-NEXT:    [[E1:%.*]] = zext i16 [[L1]] to i64
-; ALL-NEXT:    [[E2:%.*]] = zext i16 [[L2]] to i64
-; ALL-NEXT:    [[E3:%.*]] = zext i16 [[L3]] to i64
-; ALL-NEXT:    [[E4:%.*]] = zext i16 [[L4]] to i64
-; ALL-NEXT:    [[S1:%.*]] = shl i64 [[E1]], 48
-; ALL-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 32
-; ALL-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 16
-; ALL-NEXT:    [[O1:%.*]] = or i64 [[S1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i64 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i64 [[O2]], [[E4]]
-; ALL-NEXT:    ret i64 [[O3]]
+; LE-LABEL: @loadCombine_i64_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i16, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i16, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i16, ptr [[P]], i32 3
+; LE-NEXT:    [[L1:%.*]] = load i16, ptr [[P]], align 2
+; LE-NEXT:    [[L2:%.*]] = load i16, ptr [[P1]], align 2
+; LE-NEXT:    [[L3:%.*]] = load i16, ptr [[P2]], align 2
+; LE-NEXT:    [[L4:%.*]] = load i16, ptr [[P3]], align 2
+; LE-NEXT:    [[E1:%.*]] = zext i16 [[L1]] to i64
+; LE-NEXT:    [[E2:%.*]] = zext i16 [[L2]] to i64
+; LE-NEXT:    [[E3:%.*]] = zext i16 [[L3]] to i64
+; LE-NEXT:    [[E4:%.*]] = zext i16 [[L4]] to i64
+; LE-NEXT:    [[S1:%.*]] = shl i64 [[E1]], 48
+; LE-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 32
+; LE-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 16
+; LE-NEXT:    [[O1:%.*]] = or i64 [[S1]], [[S2]]
+; LE-NEXT:    [[O2:%.*]] = or i64 [[O1]], [[S3]]
+; LE-NEXT:    [[O3:%.*]] = or i64 [[O2]], [[E4]]
+; LE-NEXT:    ret i64 [[O3]]
+;
+; BE-LABEL: @loadCombine_i64_BE(
+; BE-NEXT:    [[L1:%.*]] = load i64, ptr [[P:%.*]], align 2
+; BE-NEXT:    ret i64 [[L1]]
 ;
   %p1 = getelementptr i16, ptr %p, i32 1
   %p2 = getelementptr i16, ptr %p, i32 2
   %p3 = getelementptr i16, ptr %p, i32 3
   %l1 = load i16, ptr %p
   %l2 = load i16, ptr %p1
   %l3 = load i16, ptr %p2
   %l4 = load i16, ptr %p3
;
   %e1 = zext i8 %l1 to i16
   %e2 = zext i8 %l2 to i16
   %s2 = shl i16 %e2, 8
   %o1 = or i16 %e1, %s2
   ret i16 %o1
 }
 
 define i64 @load64_farLoads(ptr %ptr) {
-; ALL-LABEL: @load64_farLoads(
-; ALL-NEXT:  entry:
-; ALL-NEXT:    [[TMP0:%.*]] = load i8, ptr [[PTR:%.*]], align 1
-; ALL-NEXT:    [[CONV:%.*]] = zext i8 [[TMP0]] to i64
-; ALL-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 1
-; ALL-NEXT:    [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX1]], align 1
-; ALL-NEXT:    [[CONV2:%.*]] = zext i8 [[TMP1]] to i64
-; ALL-NEXT:    [[SHL:%.*]] = shl i64 [[CONV2]], 8
-; ALL-NEXT:    [[OR:%.*]] = or i64 [[CONV]], [[SHL]]
-; ALL-NEXT:    [[ARRAYIDX3:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 2
-; ALL-NEXT:    [[TMP2:%.*]] = load i8, ptr [[ARRAYIDX3]], align 1
-; ALL-NEXT:    [[CONV4:%.*]] = zext i8 [[TMP2]] to i64
-; ALL-NEXT:    [[SHL5:%.*]] = shl i64 [[CONV4]], 16
-; ALL-NEXT:    [[OR6:%.*]] = or i64 [[OR]], [[SHL5]]
-; ALL-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 3
-; ALL-NEXT:    [[TMP3:%.*]] = load i8, ptr [[ARRAYIDX7]], align 1
-; ALL-NEXT:    [[CONV8:%.*]] = zext i8 [[TMP3]] to i64
-; ALL-NEXT:    [[SHL9:%.*]] = shl i64 [[CONV8]], 24
-; ALL-NEXT:    [[OR10:%.*]] = or i64 [[OR6]], [[SHL9]]
-; ALL-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 4
-; ALL-NEXT:    [[TMP4:%.*]] = load i8, ptr [[ARRAYIDX11]], align 1
-; ALL-NEXT:    [[CONV12:%.*]] = zext i8 [[TMP4]] to i64
-; ALL-NEXT:    [[SHL13:%.*]] = shl i64 [[CONV12]], 32
-; ALL-NEXT:    [[OR14:%.*]] = or i64 [[OR10]], [[SHL13]]
-; ALL-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 5
-; ALL-NEXT:    [[TMP5:%.*]] = load i8, ptr [[ARRAYIDX15]], align 1
-; ALL-NEXT:    [[CONV16:%.*]] = zext i8 [[TMP5]] to i64
-; ALL-NEXT:    [[SHL17:%.*]] = shl i64 [[CONV16]], 40
-; ALL-NEXT:    [[OR18:%.*]] = or i64 [[OR14]], [[SHL17]]
-; ALL-NEXT:    [[ARRAYIDX19:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 6
-; ALL-NEXT:    [[TMP6:%.*]] = load i8, ptr [[ARRAYIDX19]], align 1
-; ALL-NEXT:    [[CONV20:%.*]] = zext i8 [[TMP6]] to i64
-; ALL-NEXT:    [[SHL21:%.*]] = shl i64 [[CONV20]], 48
-; ALL-NEXT:    [[OR22:%.*]] = or i64 [[OR18]], [[SHL21]]
-; ALL-NEXT:    [[ARRAYIDX23:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 7
-; ALL-NEXT:    [[TMP7:%.*]] = load i8, ptr [[ARRAYIDX23]], align 1
-; ALL-NEXT:    [[CONV24:%.*]] = zext i8 [[TMP7]] to i64
-; ALL-NEXT:    [[SHL25:%.*]] = shl i64 [[CONV24]], 56
-; ALL-NEXT:    [[OR26:%.*]] = or i64 [[OR22]], [[SHL25]]
-; ALL-NEXT:    ret i64 [[OR26]]
+; LE-LABEL: @load64_farLoads(
+; LE-NEXT:  entry:
+; LE-NEXT:    [[TMP0:%.*]] = load i64, ptr [[PTR:%.*]], align 1
+; LE-NEXT:    ret i64 [[TMP0]]
+;
+; BE-LABEL: @load64_farLoads(
+; BE-NEXT:  entry:
+; BE-NEXT:    [[TMP0:%.*]] = load i8, ptr [[PTR:%.*]], align 1
+; BE-NEXT:    [[CONV:%.*]] = zext i8 [[TMP0]] to i64
+; BE-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 1
+; BE-NEXT:    [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX1]], align 1
+; BE-NEXT:    [[CONV2:%.*]] = zext i8 [[TMP1]] to i64
+; BE-NEXT:    [[SHL:%.*]] = shl i64 [[CONV2]], 8
+; BE-NEXT:    [[OR:%.*]] = or i64 [[CONV]], [[SHL]]
+; BE-NEXT:    [[ARRAYIDX3:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 2
+; BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[ARRAYIDX3]], align 1
+; BE-NEXT:    [[CONV4:%.*]] = zext i8 [[TMP2]] to i64
+; BE-NEXT:    [[SHL5:%.*]] = shl i64 [[CONV4]], 16
+; BE-NEXT:    [[OR6:%.*]] = or i64 [[OR]], [[SHL5]]
+; BE-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 3
+; BE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[ARRAYIDX7]], align 1
+; BE-NEXT:    [[CONV8:%.*]] = zext i8 [[TMP3]] to i64
+; BE-NEXT:    [[SHL9:%.*]] = shl i64 [[CONV8]], 24
+; BE-NEXT:    [[OR10:%.*]] = or i64 [[OR6]], [[SHL9]]
+; BE-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 4
+; BE-NEXT:    [[TMP4:%.*]] = load i8, ptr [[ARRAYIDX11]], align 1
+; BE-NEXT:    [[CONV12:%.*]] = zext i8 [[TMP4]] to i64
+; BE-NEXT:    [[SHL13:%.*]] = shl i64 [[CONV12]], 32
+; BE-NEXT:    [[OR14:%.*]] = or i64 [[OR10]], [[SHL13]]
+; BE-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 5
+; BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[ARRAYIDX15]], align 1
+; BE-NEXT:    [[CONV16:%.*]] = zext i8 [[TMP5]] to i64
+; BE-NEXT:    [[SHL17:%.*]] = shl i64 [[CONV16]], 40
+; BE-NEXT:    [[OR18:%.*]] = or i64 [[OR14]], [[SHL17]]
+; BE-NEXT:    [[ARRAYIDX19:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 6
+; BE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[ARRAYIDX19]], align 1
+; BE-NEXT:    [[CONV20:%.*]] = zext i8 [[TMP6]] to i64
+; BE-NEXT:    [[SHL21:%.*]] = shl i64 [[CONV20]], 48
+; BE-NEXT:    [[OR22:%.*]] = or i64 [[OR18]], [[SHL21]]
+; BE-NEXT:    [[ARRAYIDX23:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 7
+; BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[ARRAYIDX23]], align 1
+; BE-NEXT:    [[CONV24:%.*]] = zext i8 [[TMP7]] to i64
+; BE-NEXT:    [[SHL25:%.*]] = shl i64 [[CONV24]], 56
+; BE-NEXT:    [[OR26:%.*]] = or i64 [[OR22]], [[SHL25]]
+; BE-NEXT:    ret i64 [[OR26]]
 ;
 entry:
   %0 = load i8, ptr %ptr, align 1
   %conv = zext i8 %0 to i64
   %arrayidx1 = getelementptr inbounds i8, ptr %ptr, i64 1
   %1 = load i8, ptr %arrayidx1, align 1
   %conv2 = zext i8 %1 to i64
   %shl = shl i64 %conv2, 8
entry:
   %7 = load i8, ptr %arrayidx23, align 1
   %conv24 = zext i8 %7 to i64
   %shl25 = shl i64 %conv24, 56
   %or26 = or i64 %or22, %shl25
   ret i64 %or26
 }
 
 define i32 @loadCombine_4consecutive_metadata(ptr %p, ptr %pstr) {
-; ALL-LABEL: @loadCombine_4consecutive_metadata(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1, !alias.scope !0
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1, !alias.scope !0
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1, !alias.scope !0
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1, !alias.scope !0
-; ALL-NEXT:    store i32 25, ptr [[PSTR:%.*]], align 4, !noalias !0
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
-; ALL-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_metadata(
+; LE-NEXT:    [[L1:%.*]] = load i32, ptr [[P:%.*]], align 1, !alias.scope !0
+; LE-NEXT:    store i32 25, ptr [[PSTR:%.*]], align 4, !noalias !0
+; LE-NEXT:    ret i32 [[L1]]
+;
+; BE-LABEL: @loadCombine_4consecutive_metadata(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1, !alias.scope !0
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1, !alias.scope !0
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1, !alias.scope !0
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1, !alias.scope !0
+; BE-NEXT:    store i32 25, ptr [[PSTR:%.*]], align 4, !noalias !0
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; BE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; BE-NEXT:    [[O1:%.*]] = or i32 [[E1]], [[S2]]
+; BE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S3]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[S4]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p, !alias.scope !2
   %l2 = load i8, ptr %p1, !alias.scope !2
   %l3 = load i8, ptr %p2, !alias.scope !2
   %l4 = load i8, ptr %p3, !alias.scope !2
;
   %s4 = shl i16 %e4, 12
   %o1 = or i16 %e1, %s2
   %o2 = or i16 %o1, %s3
   %o3 = or i16 %o2, %s4
   ret i16 %o3
 }
 
 define i32 @loadCombine_4consecutive_rev(ptr %p) {
-; ALL-LABEL: @loadCombine_4consecutive_rev(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
-; ALL-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
-; ALL-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
-; ALL-NEXT:    [[O1:%.*]] = or i32 [[S4]], [[S3]]
-; ALL-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S2]]
-; ALL-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E1]]
-; ALL-NEXT:    ret i32 [[O3]]
+; LE-LABEL: @loadCombine_4consecutive_rev(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; LE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; LE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; LE-NEXT:    [[L3:%.*]] = load i16, ptr [[P2]], align 1
+; LE-NEXT:    [[TMP1:%.*]] = zext i16 [[L3]] to i32
+; LE-NEXT:    [[TMP2:%.*]] = shl i32 [[TMP1]], 16
+; LE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; LE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; LE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; LE-NEXT:    [[O2:%.*]] = or i32 [[TMP2]], [[S2]]
+; LE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E1]]
+; LE-NEXT:    ret i32 [[O3]]
+;
+; BE-LABEL: @loadCombine_4consecutive_rev(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i32
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i32
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i32
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i32
+; BE-NEXT:    [[S2:%.*]] = shl i32 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i32 [[E3]], 16
+; BE-NEXT:    [[S4:%.*]] = shl i32 [[E4]], 24
+; BE-NEXT:    [[O1:%.*]] = or i32 [[S4]], [[S3]]
+; BE-NEXT:    [[O2:%.*]] = or i32 [[O1]], [[S2]]
+; BE-NEXT:    [[O3:%.*]] = or i32 [[O2]], [[E1]]
+; BE-NEXT:    ret i32 [[O3]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %l1 = load i8, ptr %p
   %l2 = load i8, ptr %p1
   %l3 = load i8, ptr %p2
   %l4 = load i8, ptr %p3
;
 
   %o1 = or i32 %s4, %s3
   %o2 = or i32 %o1, %s2
   %o3 = or i32 %o2, %e1
   ret i32 %o3
 }
 
 define i64 @loadCombine_8consecutive_rev(ptr %p) {
-; ALL-LABEL: @loadCombine_8consecutive_rev(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[P4:%.*]] = getelementptr i8, ptr [[P]], i32 4
-; ALL-NEXT:    [[P5:%.*]] = getelementptr i8, ptr [[P]], i32 5
-; ALL-NEXT:    [[P6:%.*]] = getelementptr i8, ptr [[P]], i32 6
-; ALL-NEXT:    [[P7:%.*]] = getelementptr i8, ptr [[P]], i32 7
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[L5:%.*]] = load i8, ptr [[P4]], align 1
-; ALL-NEXT:    [[L6:%.*]] = load i8, ptr [[P5]], align 1
-; ALL-NEXT:    [[L7:%.*]] = load i8, ptr [[P6]], align 1
-; ALL-NEXT:    [[L8:%.*]] = load i8, ptr [[P7]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i64
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i64
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i64
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i64
-; ALL-NEXT:    [[E5:%.*]] = zext i8 [[L5]] to i64
-; ALL-NEXT:    [[E6:%.*]] = zext i8 [[L6]] to i64
-; ALL-NEXT:    [[E7:%.*]] = zext i8 [[L7]] to i64
-; ALL-NEXT:    [[E8:%.*]] = zext i8 [[L8]] to i64
-; ALL-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 8
-; ALL-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 16
-; ALL-NEXT:    [[S4:%.*]] = shl i64 [[E4]], 24
-; ALL-NEXT:    [[S5:%.*]] = shl i64 [[E5]], 32
-; ALL-NEXT:    [[S6:%.*]] = shl i64 [[E6]], 40
-; ALL-NEXT:    [[S7:%.*]] = shl i64 [[E7]], 48
-; ALL-NEXT:    [[S8:%.*]] = shl i64 [[E8]], 56
-; ALL-NEXT:    [[O7:%.*]] = or i64 [[S8]], [[S7]]
-; ALL-NEXT:    [[O6:%.*]] = or i64 [[O7]], [[S6]]
-; ALL-NEXT:    [[O5:%.*]] = or i64 [[O6]], [[S5]]
-; ALL-NEXT:    [[O4:%.*]] = or i64 [[O5]], [[S4]]
-; ALL-NEXT:    [[O3:%.*]] = or i64 [[O4]], [[S3]]
-; ALL-NEXT:    [[O2:%.*]] = or i64 [[O3]], [[S2]]
-; ALL-NEXT:    [[O1:%.*]] = or i64 [[O2]], [[E1]]
-; ALL-NEXT:    ret i64 [[O1]]
+; LE-LABEL: @loadCombine_8consecutive_rev(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; LE-NEXT:    [[P4:%.*]] = getelementptr i8, ptr [[P]], i32 4
+; LE-NEXT:    [[P5:%.*]] = getelementptr i8, ptr [[P]], i32 5
+; LE-NEXT:    [[P6:%.*]] = getelementptr i8, ptr [[P]], i32 6
+; LE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; LE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; LE-NEXT:    [[L5:%.*]] = load i8, ptr [[P4]], align 1
+; LE-NEXT:    [[L6:%.*]] = load i8, ptr [[P5]], align 1
+; LE-NEXT:    [[L7:%.*]] = load i16, ptr [[P6]], align 1
+; LE-NEXT:    [[TMP1:%.*]] = zext i16 [[L7]] to i64
+; LE-NEXT:    [[TMP2:%.*]] = shl i64 [[TMP1]], 48
+; LE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i64
+; LE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i64
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i64
+; LE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i64
+; LE-NEXT:    [[E5:%.*]] = zext i8 [[L5]] to i64
+; LE-NEXT:    [[E6:%.*]] = zext i8 [[L6]] to i64
+; LE-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 8
+; LE-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 16
+; LE-NEXT:    [[S4:%.*]] = shl i64 [[E4]], 24
+; LE-NEXT:    [[S5:%.*]] = shl i64 [[E5]], 32
+; LE-NEXT:    [[S6:%.*]] = shl i64 [[E6]], 40
+; LE-NEXT:    [[O6:%.*]] = or i64 [[TMP2]], [[S6]]
+; LE-NEXT:    [[O5:%.*]] = or i64 [[O6]], [[S5]]
+; LE-NEXT:    [[O4:%.*]] = or i64 [[O5]], [[S4]]
+; LE-NEXT:    [[O3:%.*]] = or i64 [[O4]], [[S3]]
+; LE-NEXT:    [[O2:%.*]] = or i64 [[O3]], [[S2]]
+; LE-NEXT:    [[O1:%.*]] = or i64 [[O2]], [[E1]]
+; LE-NEXT:    ret i64 [[O1]]
+;
+; BE-LABEL: @loadCombine_8consecutive_rev(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[P4:%.*]] = getelementptr i8, ptr [[P]], i32 4
+; BE-NEXT:    [[P5:%.*]] = getelementptr i8, ptr [[P]], i32 5
+; BE-NEXT:    [[P6:%.*]] = getelementptr i8, ptr [[P]], i32 6
+; BE-NEXT:    [[P7:%.*]] = getelementptr i8, ptr [[P]], i32 7
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[L5:%.*]] = load i8, ptr [[P4]], align 1
+; BE-NEXT:    [[L6:%.*]] = load i8, ptr [[P5]], align 1
+; BE-NEXT:    [[L7:%.*]] = load i8, ptr [[P6]], align 1
+; BE-NEXT:    [[L8:%.*]] = load i8, ptr [[P7]], align 1
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i64
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i64
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i64
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i64
+; BE-NEXT:    [[E5:%.*]] = zext i8 [[L5]] to i64
+; BE-NEXT:    [[E6:%.*]] = zext i8 [[L6]] to i64
+; BE-NEXT:    [[E7:%.*]] = zext i8 [[L7]] to i64
+; BE-NEXT:    [[E8:%.*]] = zext i8 [[L8]] to i64
+; BE-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 8
+; BE-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 16
+; BE-NEXT:    [[S4:%.*]] = shl i64 [[E4]], 24
+; BE-NEXT:    [[S5:%.*]] = shl i64 [[E5]], 32
+; BE-NEXT:    [[S6:%.*]] = shl i64 [[E6]], 40
+; BE-NEXT:    [[S7:%.*]] = shl i64 [[E7]], 48
+; BE-NEXT:    [[S8:%.*]] = shl i64 [[E8]], 56
+; BE-NEXT:    [[O7:%.*]] = or i64 [[S8]], [[S7]]
+; BE-NEXT:    [[O6:%.*]] = or i64 [[O7]], [[S6]]
+; BE-NEXT:    [[O5:%.*]] = or i64 [[O6]], [[S5]]
+; BE-NEXT:    [[O4:%.*]] = or i64 [[O5]], [[S4]]
+; BE-NEXT:    [[O3:%.*]] = or i64 [[O4]], [[S3]]
+; BE-NEXT:    [[O2:%.*]] = or i64 [[O3]], [[S2]]
+; BE-NEXT:    [[O1:%.*]] = or i64 [[O2]], [[E1]]
+; BE-NEXT:    ret i64 [[O1]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %p4 = getelementptr i8, ptr %p, i32 4
   %p5 = getelementptr i8, ptr %p, i32 5
   %p6 = getelementptr i8, ptr %p, i32 6
   %p7 = getelementptr i8, ptr %p, i32 7
;
   %o4 = or i64 %o5, %s4
   %o3 = or i64 %o4, %s3
   %o2 = or i64 %o3, %s2
   %o1 = or i64 %o2, %e1
   ret i64 %o1
 }
 
 define i64 @loadCombine_8consecutive_rev_BE(ptr %p) {
-; ALL-LABEL: @loadCombine_8consecutive_rev_BE(
-; ALL-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
-; ALL-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
-; ALL-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
-; ALL-NEXT:    [[P4:%.*]] = getelementptr i8, ptr [[P]], i32 4
-; ALL-NEXT:    [[P5:%.*]] = getelementptr i8, ptr [[P]], i32 5
-; ALL-NEXT:    [[P6:%.*]] = getelementptr i8, ptr [[P]], i32 6
-; ALL-NEXT:    [[P7:%.*]] = getelementptr i8, ptr [[P]], i32 7
-; ALL-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
-; ALL-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
-; ALL-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
-; ALL-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
-; ALL-NEXT:    [[L5:%.*]] = load i8, ptr [[P4]], align 1
-; ALL-NEXT:    [[L6:%.*]] = load i8, ptr [[P5]], align 1
-; ALL-NEXT:    [[L7:%.*]] = load i8, ptr [[P6]], align 1
-; ALL-NEXT:    [[L8:%.*]] = load i8, ptr [[P7]], align 1
-; ALL-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i64
-; ALL-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i64
-; ALL-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i64
-; ALL-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i64
-; ALL-NEXT:    [[E5:%.*]] = zext i8 [[L5]] to i64
-; ALL-NEXT:    [[E6:%.*]] = zext i8 [[L6]] to i64
-; ALL-NEXT:    [[E7:%.*]] = zext i8 [[L7]] to i64
-; ALL-NEXT:    [[E8:%.*]] = zext i8 [[L8]] to i64
-; ALL-NEXT:    [[S1:%.*]] = shl i64 [[E1]], 56
-; ALL-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 48
-; ALL-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 40
-; ALL-NEXT:    [[S4:%.*]] = shl i64 [[E4]], 32
-; ALL-NEXT:    [[S5:%.*]] = shl i64 [[E5]], 24
-; ALL-NEXT:    [[S6:%.*]] = shl i64 [[E6]], 16
-; ALL-NEXT:    [[S7:%.*]] = shl i64 [[E7]], 8
-; ALL-NEXT:    [[O7:%.*]] = or i64 [[E8]], [[S7]]
-; ALL-NEXT:    [[O6:%.*]] = or i64 [[O7]], [[S6]]
-; ALL-NEXT:    [[O5:%.*]] = or i64 [[O6]], [[S5]]
-; ALL-NEXT:    [[O4:%.*]] = or i64 [[O5]], [[S4]]
-; ALL-NEXT:    [[O3:%.*]] = or i64 [[O4]], [[S3]]
-; ALL-NEXT:    [[O2:%.*]] = or i64 [[O3]], [[S2]]
-; ALL-NEXT:    [[O1:%.*]] = or i64 [[O2]], [[S1]]
-; ALL-NEXT:    ret i64 [[O1]]
+; LE-LABEL: @loadCombine_8consecutive_rev_BE(
+; LE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; LE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; LE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; LE-NEXT:    [[P4:%.*]] = getelementptr i8, ptr [[P]], i32 4
+; LE-NEXT:    [[P5:%.*]] = getelementptr i8, ptr [[P]], i32 5
+; LE-NEXT:    [[P6:%.*]] = getelementptr i8, ptr [[P]], i32 6
+; LE-NEXT:    [[P7:%.*]] = getelementptr i8, ptr [[P]], i32 7
+; LE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; LE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; LE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; LE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; LE-NEXT:    [[L5:%.*]] = load i8, ptr [[P4]], align 1
+; LE-NEXT:    [[L6:%.*]] = load i8, ptr [[P5]], align 1
+; LE-NEXT:    [[L7:%.*]] = load i8, ptr [[P6]], align 1
+; LE-NEXT:    [[L8:%.*]] = load i8, ptr [[P7]], align 1
+; LE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i64
+; LE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i64
+; LE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i64
+; LE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i64
+; LE-NEXT:    [[E5:%.*]] = zext i8 [[L5]] to i64
+; LE-NEXT:    [[E6:%.*]] = zext i8 [[L6]] to i64
+; LE-NEXT:    [[E7:%.*]] = zext i8 [[L7]] to i64
+; LE-NEXT:    [[E8:%.*]] = zext i8 [[L8]] to i64
+; LE-NEXT:    [[S1:%.*]] = shl i64 [[E1]], 56
+; LE-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 48
+; LE-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 40
+; LE-NEXT:    [[S4:%.*]] = shl i64 [[E4]], 32
+; LE-NEXT:    [[S5:%.*]] = shl i64 [[E5]], 24
+; LE-NEXT:    [[S6:%.*]] = shl i64 [[E6]], 16
+; LE-NEXT:    [[S7:%.*]] = shl i64 [[E7]], 8
+; LE-NEXT:    [[O7:%.*]] = or i64 [[E8]], [[S7]]
+; LE-NEXT:    [[O6:%.*]] = or i64 [[O7]], [[S6]]
+; LE-NEXT:    [[O5:%.*]] = or i64 [[O6]], [[S5]]
+; LE-NEXT:    [[O4:%.*]] = or i64 [[O5]], [[S4]]
+; LE-NEXT:    [[O3:%.*]] = or i64 [[O4]], [[S3]]
+; LE-NEXT:    [[O2:%.*]] = or i64 [[O3]], [[S2]]
+; LE-NEXT:    [[O1:%.*]] = or i64 [[O2]], [[S1]]
+; LE-NEXT:    ret i64 [[O1]]
+;
+; BE-LABEL: @loadCombine_8consecutive_rev_BE(
+; BE-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 1
+; BE-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 2
+; BE-NEXT:    [[P3:%.*]] = getelementptr i8, ptr [[P]], i32 3
+; BE-NEXT:    [[P4:%.*]] = getelementptr i8, ptr [[P]], i32 4
+; BE-NEXT:    [[P5:%.*]] = getelementptr i8, ptr [[P]], i32 5
+; BE-NEXT:    [[P6:%.*]] = getelementptr i8, ptr [[P]], i32 6
+; BE-NEXT:    [[L1:%.*]] = load i8, ptr [[P]], align 1
+; BE-NEXT:    [[L2:%.*]] = load i8, ptr [[P1]], align 1
+; BE-NEXT:    [[L3:%.*]] = load i8, ptr [[P2]], align 1
+; BE-NEXT:    [[L4:%.*]] = load i8, ptr [[P3]], align 1
+; BE-NEXT:    [[L5:%.*]] = load i8, ptr [[P4]], align 1
+; BE-NEXT:    [[L6:%.*]] = load i8, ptr [[P5]], align 1
+; BE-NEXT:    [[L7:%.*]] = load i16, ptr [[P6]], align 1
+; BE-NEXT:    [[TMP1:%.*]] = zext i16 [[L7]] to i64
+; BE-NEXT:    [[E1:%.*]] = zext i8 [[L1]] to i64
+; BE-NEXT:    [[E2:%.*]] = zext i8 [[L2]] to i64
+; BE-NEXT:    [[E3:%.*]] = zext i8 [[L3]] to i64
+; BE-NEXT:    [[E4:%.*]] = zext i8 [[L4]] to i64
+; BE-NEXT:    [[E5:%.*]] = zext i8 [[L5]] to i64
+; BE-NEXT:    [[E6:%.*]] = zext i8 [[L6]] to i64
+; BE-NEXT:    [[S1:%.*]] = shl i64 [[E1]], 56
+; BE-NEXT:    [[S2:%.*]] = shl i64 [[E2]], 48
+; BE-NEXT:    [[S3:%.*]] = shl i64 [[E3]], 40
+; BE-NEXT:    [[S4:%.*]] = shl i64 [[E4]], 32
+; BE-NEXT:    [[S5:%.*]] = shl i64 [[E5]], 24
+; BE-NEXT:    [[S6:%.*]] = shl i64 [[E6]], 16
+; BE-NEXT:    [[O6:%.*]] = or i64 [[TMP1]], [[S6]]
+; BE-NEXT:    [[O5:%.*]] = or i64 [[O6]], [[S5]]
+; BE-NEXT:    [[O4:%.*]] = or i64 [[O5]], [[S4]]
+; BE-NEXT:    [[O3:%.*]] = or i64 [[O4]], [[S3]]
+; BE-NEXT:    [[O2:%.*]] = or i64 [[O3]], [[S2]]
+; BE-NEXT:    [[O1:%.*]] = or i64 [[O2]], [[S1]]
+; BE-NEXT:    ret i64 [[O1]]
 ;
   %p1 = getelementptr i8, ptr %p, i32 1
   %p2 = getelementptr i8, ptr %p, i32 2
   %p3 = getelementptr i8, ptr %p, i32 3
   %p4 = getelementptr i8, ptr %p, i32 4
   %p5 = getelementptr i8, ptr %p, i32 5
   %p6 = getelementptr i8, ptr %p, i32 6
   %p7 = getelementptr i8, ptr %p, i32 7
;
   %o4 = or i64 %o5, %s4
   %o3 = or i64 %o4, %s3
   %o2 = or i64 %o3, %s2
   %o1 = or i64 %o2, %s1
   ret i64 %o1
 }
 
 define i64 @eggs(ptr noundef readonly %arg) {
-; ALL-LABEL: @eggs(
-; ALL-NEXT:    [[TMP3:%.*]] = load i8, ptr [[ARG:%.*]], align 1
-; ALL-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 1
-; ALL-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
-; ALL-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 2
-; ALL-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
-; ALL-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 3
-; ALL-NEXT:    [[TMP9:%.*]] = load i8, ptr [[TMP8]], align 1
-; ALL-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 4
-; ALL-NEXT:    [[TMP11:%.*]] = load i8, ptr [[TMP10]], align 1
-; ALL-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 5
-; ALL-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 1
-; ALL-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 6
-; ALL-NEXT:    [[TMP15:%.*]] = load i8, ptr [[TMP14]], align 1
-; ALL-NEXT:    [[TMP16:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 7
-; ALL-NEXT:    [[TMP17:%.*]] = load i8, ptr [[TMP16]], align 1
-; ALL-NEXT:    [[TMP18:%.*]] = zext i8 [[TMP17]] to i64
-; ALL-NEXT:    [[TMP19:%.*]] = shl nuw i64 [[TMP18]], 56
-; ALL-NEXT:    [[TMP20:%.*]] = zext i8 [[TMP15]] to i64
-; ALL-NEXT:    [[TMP21:%.*]] = shl nuw nsw i64 [[TMP20]], 48
-; ALL-NEXT:    [[TMP22:%.*]] = or i64 [[TMP19]], [[TMP21]]
-; ALL-NEXT:    [[TMP23:%.*]] = zext i8 [[TMP13]] to i64
-; ALL-NEXT:    [[TMP24:%.*]] = shl nuw nsw i64 [[TMP23]], 40
-; ALL-NEXT:    [[TMP25:%.*]] = or i64 [[TMP22]], [[TMP24]]
-; ALL-NEXT:    [[TMP26:%.*]] = zext i8 [[TMP11]] to i64
-; ALL-NEXT:    [[TMP27:%.*]] = shl nuw nsw i64 [[TMP26]], 32
-; ALL-NEXT:    [[TMP28:%.*]] = or i64 [[TMP25]], [[TMP27]]
-; ALL-NEXT:    [[TMP29:%.*]] = zext i8 [[TMP9]] to i64
-; ALL-NEXT:    [[TMP30:%.*]] = shl nuw nsw i64 [[TMP29]], 24
-; ALL-NEXT:    [[TMP31:%.*]] = or i64 [[TMP28]], [[TMP30]]
-; ALL-NEXT:    [[TMP32:%.*]] = zext i8 [[TMP7]] to i64
-; ALL-NEXT:    [[TMP33:%.*]] = shl nuw nsw i64 [[TMP32]], 16
-; ALL-NEXT:    [[TMP34:%.*]] = zext i8 [[TMP5]] to i64
-; ALL-NEXT:    [[TMP35:%.*]] = shl nuw nsw i64 [[TMP34]], 8
-; ALL-NEXT:    [[TMP36:%.*]] = or i64 [[TMP31]], [[TMP33]]
-; ALL-NEXT:    [[TMP37:%.*]] = zext i8 [[TMP3]] to i64
-; ALL-NEXT:    [[TMP38:%.*]] = or i64 [[TMP36]], [[TMP35]]
-; ALL-NEXT:    [[TMP39:%.*]] = or i64 [[TMP38]], [[TMP37]]
-; ALL-NEXT:    ret i64 [[TMP39]]
+; LE-LABEL: @eggs(
+; LE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[ARG:%.*]], align 1
+; LE-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 1
+; LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+; LE-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 2
+; LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+; LE-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 3
+; LE-NEXT:    [[TMP9:%.*]] = load i8, ptr [[TMP8]], align 1
+; LE-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 4
+; LE-NEXT:    [[TMP11:%.*]] = load i8, ptr [[TMP10]], align 1
+; LE-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 5
+; LE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 1
+; LE-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 6
+; LE-NEXT:    [[TMP15:%.*]] = load i16, ptr [[TMP14]], align 1
+; LE-NEXT:    [[TMP1:%.*]] = zext i16 [[TMP15]] to i64
+; LE-NEXT:    [[TMP2:%.*]] = shl i64 [[TMP1]], 48
+; LE-NEXT:    [[TMP23:%.*]] = zext i8 [[TMP13]] to i64
+; LE-NEXT:    [[TMP24:%.*]] = shl nuw nsw i64 [[TMP23]], 40
+; LE-NEXT:    [[TMP25:%.*]] = or i64 [[TMP2]], [[TMP24]]
+; LE-NEXT:    [[TMP26:%.*]] = zext i8 [[TMP11]] to i64
+; LE-NEXT:    [[TMP27:%.*]] = shl nuw nsw i64 [[TMP26]], 32
+; LE-NEXT:    [[TMP28:%.*]] = or i64 [[TMP25]], [[TMP27]]
+; LE-NEXT:    [[TMP29:%.*]] = zext i8 [[TMP9]] to i64
+; LE-NEXT:    [[TMP30:%.*]] = shl nuw nsw i64 [[TMP29]], 24
+; LE-NEXT:    [[TMP31:%.*]] = or i64 [[TMP28]], [[TMP30]]
+; LE-NEXT:    [[TMP32:%.*]] = zext i8 [[TMP7]] to i64
+; LE-NEXT:    [[TMP33:%.*]] = shl nuw nsw i64 [[TMP32]], 16
+; LE-NEXT:    [[TMP34:%.*]] = zext i8 [[TMP5]] to i64
+; LE-NEXT:    [[TMP35:%.*]] = shl nuw nsw i64 [[TMP34]], 8
+; LE-NEXT:    [[TMP36:%.*]] = or i64 [[TMP31]], [[TMP33]]
+; LE-NEXT:    [[TMP37:%.*]] = zext i8 [[TMP3]] to i64
+; LE-NEXT:    [[TMP38:%.*]] = or i64 [[TMP36]], [[TMP35]]
+; LE-NEXT:    [[TMP39:%.*]] = or i64 [[TMP38]], [[TMP37]]
+; LE-NEXT:    ret i64 [[TMP39]]
+;
+; BE-LABEL: @eggs(
+; BE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[ARG:%.*]], align 1
+; BE-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 1
+; BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+; BE-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 2
+; BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+; BE-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 3
+; BE-NEXT:    [[TMP9:%.*]] = load i8, ptr [[TMP8]], align 1
+; BE-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 4
+; BE-NEXT:    [[TMP11:%.*]] = load i8, ptr [[TMP10]], align 1
+; BE-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 5
+; BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 1
+; BE-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 6
+; BE-NEXT:    [[TMP15:%.*]] = load i8, ptr [[TMP14]], align 1
+; BE-NEXT:    [[TMP16:%.*]] = getelementptr inbounds i8, ptr [[ARG]], i64 7
+; BE-NEXT:    [[TMP17:%.*]] = load i8, ptr [[TMP16]], align 1
+; BE-NEXT:    [[TMP18:%.*]] = zext i8 [[TMP17]] to i64
+; BE-NEXT:    [[TMP19:%.*]] = shl nuw i64 [[TMP18]], 56
+; BE-NEXT:    [[TMP20:%.*]] = zext i8 [[TMP15]] to i64
+; BE-NEXT:    [[TMP21:%.*]] = shl nuw nsw i64 [[TMP20]], 48
+; BE-NEXT:    [[TMP22:%.*]] = or i64 [[TMP19]], [[TMP21]]
+; BE-NEXT:    [[TMP23:%.*]] = zext i8 [[TMP13]] to i64
+; BE-NEXT:    [[TMP24:%.*]] = shl nuw nsw i64 [[TMP23]], 40
+; BE-NEXT:    [[TMP25:%.*]] = or i64 [[TMP22]], [[TMP24]]
+; BE-NEXT:    [[TMP26:%.*]] = zext i8 [[TMP11]] to i64
+; BE-NEXT:    [[TMP27:%.*]] = shl nuw nsw i64 [[TMP26]], 32
+; BE-NEXT:    [[TMP28:%.*]] = or i64 [[TMP25]], [[TMP27]]
+; BE-NEXT:    [[TMP29:%.*]] = zext i8 [[TMP9]] to i64
+; BE-NEXT:    [[TMP30:%.*]] = shl nuw nsw i64 [[TMP29]], 24
+; BE-NEXT:    [[TMP31:%.*]] = or i64 [[TMP28]], [[TMP30]]
+; BE-NEXT:    [[TMP32:%.*]] = zext i8 [[TMP7]] to i64
+; BE-NEXT:    [[TMP33:%.*]] = shl nuw nsw i64 [[TMP32]], 16
+; BE-NEXT:    [[TMP34:%.*]] = zext i8 [[TMP5]] to i64
+; BE-NEXT:    [[TMP35:%.*]] = shl nuw nsw i64 [[TMP34]], 8
+; BE-NEXT:    [[TMP36:%.*]] = or i64 [[TMP31]], [[TMP33]]
+; BE-NEXT:    [[TMP37:%.*]] = zext i8 [[TMP3]] to i64
+; BE-NEXT:    [[TMP38:%.*]] = or i64 [[TMP36]], [[TMP35]]
+; BE-NEXT:    [[TMP39:%.*]] = or i64 [[TMP38]], [[TMP37]]
+; BE-NEXT:    ret i64 [[TMP39]]
 ;
   %tmp3 = load i8, ptr %arg, align 1
   %tmp4 = getelementptr inbounds i8, ptr %arg, i64 1
   %tmp5 = load i8, ptr %tmp4, align 1
   %tmp6 = getelementptr inbounds i8, ptr %arg, i64 2
   %tmp7 = load i8, ptr %tmp6, align 1
   %tmp8 = getelementptr inbounds i8, ptr %arg, i64 3
   %tmp9 = load i8, ptr %tmp8, align 1