[PATCH] Fix IAT (Import Address Table) alignment on AArch64

[Evgeny Karpov]

On Thu, Apr 09, 2026, Jan Beulich wrote:
> On 09.04.2026 10:10, Evgeny Karpov wrote:
> > On Wed, Apr 08, 2026, Jan Beulich wrote:
> >> On 08.04.2026 12:30, Evgeny Karpov wrote:
> >>> On Wed, Apr 08, 2026, Jan Beulich wrote:
> >>>> On 08.04.2026 11:27, Evgeny Karpov wrote:
> >>>>> On Wed, Apr 08, 2026, Jan Beulich wrote:
> >>>>>> On 08.04.2026 11:05, Evgeny Karpov wrote:
> >>>>>>> On Wed, Apr 08, 2026, Jan Beulich wrote:
> >>>>>>>> On 07.04.2026 19:12, Evgeny Karpov wrote:
> >>>>>>>>> On Tue, Apr 07, 2026, Jan Beulich wrote:
> >>>>>>>>>> On 07.04.2026 16:47, Evgeny Karpov wrote:
> >>>>>>>>>>> On Tue, Apr 07, 2026, Jan Beulich wrote:
> >>>>>>>>>>>> On 07.04.2026 11:48, Evgeny Karpov wrote:
> >>>>>>>>>>>>> The IAT should be aligned to 8 bytes in aarch64-w64-mingw32, otherwise, it might
> >>>>>>>>>>>>> result in relocation issues.
> >>>>>>>>>>>>
> >>>>>>>>>>>> What is it that makes this different on aarch64? IOW I wonder if this is
> >>>>>>>>>>>> going too far or not far enough.
> >>>>>>>>>>>>
> >>>>>>>>>>>>> binutils/ChangeLog:
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> 	* dlltool.c (make_head): Update.
> >>>>>>>>>>>>
> >>>>>>>>>>>> Please either omit the ChangeLog entry, or have it say something meaningful.
> >>>>>>>>>>>
> >>>>>>>>>>> The updated decription:
> >>>>>>>>>>> ---
> >>>>>>>>>>> The IAT should be aligned to 8 bytes in aarch64-w64-mingw32, otherwise, it might
> >>>>>>>>>>> result in relocation issues.
> >>>>>>>>>>>
> >>>>>>>>>>> When a function is imported from DLL, it generates the following code:
> >>>>>>>>>>>
> >>>>>>>>>>> adrp    x19, __imp_fn
> >>>>>>>>>>> ldr     x19, [x19, #:lo12:__imp_fn]
> >>>>>>>>>>>
> >>>>>>>>>>> 8 byte alignment is required for ldr relocation. Addresses are placed in IAT.
> >>>>>>>>>>> The size of the chunk on AArch64 is 8 bytes.
> >>>>>>>>>>> If IAT is not aligned to 8 bytes, the relocation issue appears.
> >>>>>>>>>>> This patch fixes this issue.
> >>>>>>>>>>
> >>>>>>>>>> Much better, thanks. This addresses part of my earlier remark then as well,
> >>>>>>>>>> clarifying it's not "too much" that you do. The "too little" aspect remains,
> >>>>>>>>>> though: Looking at a random ia64 archive, I see .idata$5 to have 8-byte
> >>>>>>>>>> alignment there as well. Shouldn't .idata$5 always have machine-word
> >>>>>>>>>> alignment?
> >>>>>>>>>
> >>>>>>>>> Based on PE Format documentation
> >>>>>>>>> https://learn.microsoft.com/en-us/windows/win32/debug/pe-format#import-address-table
> >>>>>>>>>
> >>>>>>>>> 4 byte addresses are used for PE32 and 8 byte addresses are used for PE32+.
> >>>>>>>>> There is no definition about the alignment, however it might be optional.
> >>>>>>>>> 8 byte alignment appears as a requirement for an architecture.
> >>>>>>>>> For example, it requires 8 byte alignment for ldr relocation on AArch64.
> >>>>>>>>>
> >>>>>>>>> Potentially, 8 byte alignment might be applied to all architectures,
> >>>>>>>>> however this might not be necessary for other relocations and could
> >>>>>>>>> slightly unnecessarily increase the size.
> >>>>>>>>
> >>>>>>>> Applying 8-byte alignment uniformly might add padding between .idata$4 and
> >>>>>>>> .idata$5 for PE32, which likely is unwanted (and possibly is wrong).
> >>>>>>>
> >>>>>>> Most likely, it does not break the PE format as it has an RVA to IAT, however it is 
> >>>>>>> definitely might bring unwanted padding to PE32.
> >>>>>>>
> >>>>>>>>>>> binutils/ChangeLog:
> >>>>>>>>>>>
> >>>>>>>>>>>         * dlltool.c (make_head): Add 8 byte alignment.
> >>>>>>>>>>
> >>>>>>>>>> This sadly still gives the impression of wider a change (all targets) than
> >>>>>>>>>> there is.
> >>>>>>>>>
> >>>>>>>>> It will be changed to "Add 8 byte alignment on AArch64.".
> >>>>>>>>
> >>>>>>>> If you insist on enforcing the alignment only for Arm64, then the code
> >>>>>>>> comment itself also wants extending to clarify why that target is special.
> >>>>>>>> As said, I think we should follow what MS tools do and arrange for 4- / 8-
> >>>>>>>> byte alignment for all targets.
> >>>>>>>
> >>>>>>> Based on this suggestion, the patch can be changed to:
> >>>>>>>
> >>>>>>> +      /* Align IAT (Import Address Table) to 8 bytes for PE32+.
> >>>>>>> +        https://learn.microsoft.com/en-us/windows/win32/debug/pe-format#import-address-table.  */
> >>>>>>> +      if (create_for_pep)
> >>>>>>> +       fprintf (f, "\t.align 3\n");
> >>>>>>
> >>>>>> Any reason to then not also go the final small step and use
> >>>>>>
> >>>>>>        fprintf (f, "\t.p2align %d\n", create_for_pep ? 3 : 2);
> >>>>>
> >>>>> It looks like the alignment to 4 is not necessary because it is already aligned to
> >>>>> 4 by INIT_SEC_DATA (IDATA5, ".idata$5", SEC_HAS_CONTENTS, 2).
> >>>>
> >>>> Hmm, good point, yet then wouldn't we better correct things there? And then
> >>>> also for .didat? secdata_{plain,delay}[] aren't even const, so a "rude"
> >>>> approach could be to edit those tables at runtime. (Preferably we'd solve
> >>>> this differently though, e.g. by setting .align to -1 to indicate 2 or 3
> >>>> want using depending on the target.)
> >>>
> >>> It might be better to keep default 4 byte alignment for readability.
> >>> It turns the change into following:
> >>>
> >>>    create_for_pep = (strcmp (mname, "i386:x86-64") == 0
> >>>                     || strcmp (mname, "arm64") == 0);
> >>>  
> >>> +  if (create_for_pep)
> >>> +    {
> >>> +      /* Update default 4 byte IAT (Import Address Table) alignemnt to 8 bytes
> >>> +        for PE32+.
> >>> +        https://learn.microsoft.com/en-us/windows/win32/debug/pe-format#import-address-table.  */
> >>> +      secdata[IDATA5].align = 3;
> >>
> >> I guess you mean secdata_plain[] here? Other than that I think all that is
> >> now needed is a clean v2 submission.
> > 
> > Actually, applying alignment in secdata_plain is already too late, and it 
> > should be done in the header for .idata$5 before the first chunk symbol,
> > as it was done initially.
> 
> I fear I don't understand: When modifying secdata_{plain,delay}[] right
> after setting create_for_pep, how can that be too late? gen_lib_file()
> (and hence make_one_lib_file()) is only called later.

There is a sequence of .idata$5 sections that starts with a section created
by using make_head.

Regards,
Evgeny