Erik de Castro Lopo via llvm-dev
2017-Jan-14 00:57 UTC
[llvm-dev] Problems with tbaa in llvm 4.0
Hi all,
I do a little work on the GHC haskell compiler. One of the things I
do is compile current GHC git HEAD against the next release of LLVM
to find issues and adapt the GHC code as early as possible.
A recent modification to LLVM is now rejecting GHC generated LLVM
IR that was previously accepted. I discussed this on IRC with
sanjoyd who explained what had changed and gave me some hints about
trying to fix it. Basically the issue (as I understood it) was that
no statements should reference the root tbaa node directly.
With that in mind, I came up with a fix to GHC llvm codegen so the
the tbaa metadata is now:
!0 = !{!"root", null}
!1 = !{!"top", !0}
!2 = !{!"stack", !1}
!3 = !{!"heap", !1}
!4 = !{!"rx", !3}
!5 = !{!"base", !1}
However, with this fix I get basically the same set of error messages
A couple of the errors are:
Access type node must be a valid scalar type
%ln59 = load i64, i64* %ln58, !tbaa !6
!6 = !{!2, !2, i64 0}
!2 = !{!"top", !3}
Access type node must be a valid scalar type
%ln5d = load i64, i64* %ln5c, !tbaa !4
!4 = !{!5, !5, i64 0}
!5 = !{!"base", !2}
Access type node must be a valid scalar type
%ln5i = load i64, i64* %ln5h, !tbaa !0
!0 = !{!1, !1, i64 0}
!1 = !{!"stack", !2}
I would appreciate any clues that might help me solve this.
Cheers,
Erik
--
----------------------------------------------------------------------
Erik de Castro Lopo
http://www.mega-nerd.com/
[+Sanjoy] On 01/13/2017 06:57 PM, Erik de Castro Lopo via llvm-dev wrote:> Hi all, > > I do a little work on the GHC haskell compiler. One of the things I > do is compile current GHC git HEAD against the next release of LLVM > to find issues and adapt the GHC code as early as possible. > > A recent modification to LLVM is now rejecting GHC generated LLVM > IR that was previously accepted. I discussed this on IRC with > sanjoyd who explained what had changed and gave me some hints about > trying to fix it. Basically the issue (as I understood it) was that > no statements should reference the root tbaa node directly. > > With that in mind, I came up with a fix to GHC llvm codegen so the > the tbaa metadata is now: > > !0 = !{!"root", null} > !1 = !{!"top", !0} > !2 = !{!"stack", !1} > !3 = !{!"heap", !1} > !4 = !{!"rx", !3} > !5 = !{!"base", !1} > > However, with this fix I get basically the same set of error messages > A couple of the errors are: > > Access type node must be a valid scalar type > %ln59 = load i64, i64* %ln58, !tbaa !6 > !6 = !{!2, !2, i64 0} > !2 = !{!"top", !3} > > Access type node must be a valid scalar type > %ln5d = load i64, i64* %ln5c, !tbaa !4 > !4 = !{!5, !5, i64 0} > !5 = !{!"base", !2} > > Access type node must be a valid scalar type > %ln5i = load i64, i64* %ln5h, !tbaa !0 > !0 = !{!1, !1, i64 0} > !1 = !{!"stack", !2} > > I would appreciate any clues that might help me solve this. > > Cheers, > Erik-- Hal Finkel Lead, Compiler Technology and Programming Languages Leadership Computing Facility Argonne National Laboratory
Hi Erik, (I've already replied on IRC, but echoing that here for others) Hal Finkel wrote:> [+Sanjoy] > > > On 01/13/2017 06:57 PM, Erik de Castro Lopo via llvm-dev wrote: >> Hi all, >> >> I do a little work on the GHC haskell compiler. One of the things I >> do is compile current GHC git HEAD against the next release of LLVM >> to find issues and adapt the GHC code as early as possible. >> >> A recent modification to LLVM is now rejecting GHC generated LLVM >> IR that was previously accepted. I discussed this on IRC with >> sanjoyd who explained what had changed and gave me some hints about >> trying to fix it. Basically the issue (as I understood it) was that >> no statements should reference the root tbaa node directly. >> >> With that in mind, I came up with a fix to GHC llvm codegen so the >> the tbaa metadata is now: >> >> !0 = !{!"root", null}You need to change !0 to !{!"root"}. Root nodes must have a single operand. -- Sanjoy>> !1 = !{!"top", !0} >> !2 = !{!"stack", !1} >> !3 = !{!"heap", !1} >> !4 = !{!"rx", !3} >> !5 = !{!"base", !1} >> >> However, with this fix I get basically the same set of error messages >> A couple of the errors are: >> >> Access type node must be a valid scalar type >> %ln59 = load i64, i64* %ln58, !tbaa !6 >> !6 = !{!2, !2, i64 0} >> !2 = !{!"top", !3} >> >> Access type node must be a valid scalar type >> %ln5d = load i64, i64* %ln5c, !tbaa !4 >> !4 = !{!5, !5, i64 0} >> !5 = !{!"base", !2} >> >> Access type node must be a valid scalar type >> %ln5i = load i64, i64* %ln5h, !tbaa !0 >> !0 = !{!1, !1, i64 0} >> !1 = !{!"stack", !2} >> >> I would appreciate any clues that might help me solve this. >> >> Cheers, >> Erik >
Seemingly Similar Threads
- [LLVMdev] Please review my patch to make GHC calling convention work on ARM
- Make LLD output COFF relocatable object file (like ELF's -r does). How much work is required to implement this?
- [LLVMdev] [PATCH] New calling convention for use by GHC
- Segfault in llc 3.8.0 building GHC
- POssible bug in the Arm code generator