[llvm-bugs] [Bug 48712] New: [coroutines] Merging frame variables drops alignment information

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=48712

            Bug ID: 48712
           Summary: [coroutines] Merging frame variables drops alignment
                    information
           Product: clang
           Version: unspecified
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: C++2a
          Assignee: unassignedclangbugs at nondot.org
          Reporter: jeremy.morse.llvm at gmail.com
                CC: blitzrakete at gmail.com, erik.pilkington at gmail.com,
                    llvm-bugs at lists.llvm.org, richard-llvm at metafoo.co.uk,
                    yedeng.yd at linux.alibaba.com

A recent patch to merge non-overlapping allocas in coroutine frames [0] appears
to drop alignment information from one of the merged allocas. I have a fairly
complicated C++ reproducer that manages to crash as a result of this, however I
think it's best illustrated by one of the frame-reuse tests [1]. In that file,
%a and %b are merged into a single frame element:

  %a.Frame = type { 
     void (%a.Frame*)*, 
     void (%a.Frame*)*, 
     %"struct.task::promise_type", 
     %struct.big_structure, 
     i1 }

Where the big_structure and big_structure.2 allocas have been merged into
big_structure. If I change %a to have 64 byte alignment:

  %a = alloca %struct.big_structure, align 64

Then the frame is reshaped accordingly:

  %a.Frame = type {
    void (%a.Frame*)*, 
    void (%a.Frame*)*, 
    %"struct.task::promise_type", 
    i1,
    [46 x i8], 
    %struct.big_structure }

With padding place before the big_structure type. If I add alignment to %b
instead of %a:

  %b = alloca %struct.big_structure.2, align 64

Then the field in the frame doesn't get aligned:

  %a.Frame = type {
    void (%a.Frame*)*,
    void (%a.Frame*)*,
    %"struct.task::promise_type",
    %struct.big_structure,
    i1 }

The subsequent use of that unaligned field for %b's pointer:

%a.reload.addr = getelementptr inbounds %a.Frame, %a.Frame* %FramePtr, i32 0,
i32 3
%b.reload.addr = bitcast %struct.big_structure* %a.reload.addr to
%struct.big_structure.2*

Produces a pointer for %b that was supposed to be aligned to 64 bytes before
being split into a coroutine; and afterwards will not be aligned to 64 bytes.
I'm still hunting for why this causes my reproducer to crash later on, but I
believe this loss of alignment is a bug.

[0] https://reviews.llvm.org/rGb3a722e66b75
[1]
https://github.com/llvm/llvm-project/blob/a6db7cf1ce7f3523adb132819c1697a572bdcfde/llvm/test/Transforms/Coroutines/coro-frame-reuse-alloca-02.ll

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20210111/78a66717/attachment-0001.html>