Mikael Holmén via llvm-dev
2017-May-10 06:37 UTC
[llvm-dev] -speculative-execution moving load before store
Hi, A few days ago I stumbled upon a problem where SpeculativeExecution changed the order of a load and a store to the same address. I wrote https://bugs.llvm.org//show_bug.cgi?id=32964 about it but no response there so far. In the input we have store i8 0, i8* @i %.pre = load i8, i8* @i and then in the output the load is moved so it's before the store which clearly makes it load the wrong value. isSafeToSpeculativelyExecute happily says "true" for the load, and there is a comment on isSafeToSpeculativelyExecute saying /// This method can return true for instructions that read memory; /// for such instructions, moving them may change the resulting value. And that is indeed the case in PR32964. How is this supposed to work? I started seeing this after the commit SpeculativeExecution: Stop using whitelist for costs Before that, e.g. a load got the cost UINT_MAX which I suppose disabled hoisting, but since the commit TTI.getUserCost(I) is used for all instructions. Regards, Mikael
Friedman, Eli via llvm-dev
2017-May-10 17:51 UTC
[llvm-dev] -speculative-execution moving load before store
On 5/9/2017 11:37 PM, Mikael Holmén via llvm-dev wrote:> Hi, > > A few days ago I stumbled upon a problem where SpeculativeExecution > changed the order of a load and a store to the same address. > > I wrote > > https://bugs.llvm.org//show_bug.cgi?id=32964 > > about it but no response there so far. > > In the input we have > > store i8 0, i8* @i > %.pre = load i8, i8* @i > > and then in the output the load is moved so it's before the store > which clearly makes it load the wrong value. > > isSafeToSpeculativelyExecute happily says "true" for the load, and > there is a comment on isSafeToSpeculativelyExecute saying > > /// This method can return true for instructions that read memory; > /// for such instructions, moving them may change the resulting value. > > And that is indeed the case in PR32964. > > How is this supposed to work?Passes are supposed to check whether the moved instruction actually produces the right result. As a short-term fix, you can just add a check for mayReadFromMemory(). If you want to do something fancier, MemorySSA provides a convenient interface to check aliasing. -Eli -- Employee of Qualcomm Innovation Center, Inc. Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
Amara Emerson via llvm-dev
2017-May-10 19:00 UTC
[llvm-dev] -speculative-execution moving load before store
This looks like it's due to r301950, which has today been reverted in r302640. Can you try with the latest trunk? Regardless, there seems to be an underlying bug exposed. On 10 May 2017 at 07:37, Mikael Holmén via llvm-dev <llvm-dev at lists.llvm.org> wrote:> Hi, > > A few days ago I stumbled upon a problem where SpeculativeExecution > changed the order of a load and a store to the same address. > > I wrote > > https://bugs.llvm.org//show_bug.cgi?id=32964 > > about it but no response there so far. > > In the input we have > > store i8 0, i8* @i > %.pre = load i8, i8* @i > > and then in the output the load is moved so it's before the store which > clearly makes it load the wrong value. > > isSafeToSpeculativelyExecute happily says "true" for the load, and there is > a comment on isSafeToSpeculativelyExecute saying > > /// This method can return true for instructions that read memory; > /// for such instructions, moving them may change the resulting value. > > And that is indeed the case in PR32964. > > How is this supposed to work? > > I started seeing this after the commit > > SpeculativeExecution: Stop using whitelist for costs > > Before that, e.g. a load got the cost UINT_MAX which I suppose disabled > hoisting, but since the commit TTI.getUserCost(I) is used for all > instructions. > > Regards, > Mikael > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
Daniel Berlin via llvm-dev
2017-May-10 19:53 UTC
[llvm-dev] -speculative-execution moving load before store
(staring at this pass, i'm not sure it's supposed to work in all cases on all processors. If it is, it ... uh, needs help :P) On Wed, May 10, 2017 at 12:00 PM, Amara Emerson via llvm-dev < llvm-dev at lists.llvm.org> wrote:> This looks like it's due to r301950, which has today been reverted in > r302640. Can you try with the latest trunk? > > Regardless, there seems to be an underlying bug exposed. > > On 10 May 2017 at 07:37, Mikael Holmén via llvm-dev > <llvm-dev at lists.llvm.org> wrote: > > Hi, > > > > A few days ago I stumbled upon a problem where SpeculativeExecution > > changed the order of a load and a store to the same address. > > > > I wrote > > > > https://bugs.llvm.org//show_bug.cgi?id=32964 > > > > about it but no response there so far. > > > > In the input we have > > > > store i8 0, i8* @i > > %.pre = load i8, i8* @i > > > > and then in the output the load is moved so it's before the store which > > clearly makes it load the wrong value. > > > > isSafeToSpeculativelyExecute happily says "true" for the load, and there > is > > a comment on isSafeToSpeculativelyExecute saying > > > > /// This method can return true for instructions that read memory; > > /// for such instructions, moving them may change the resulting value. > > > > And that is indeed the case in PR32964. > > > > How is this supposed to work? > > > > I started seeing this after the commit > > > > SpeculativeExecution: Stop using whitelist for costs > > > > Before that, e.g. a load got the cost UINT_MAX which I suppose disabled > > hoisting, but since the commit TTI.getUserCost(I) is used for all > > instructions. > > > > Regards, > > Mikael > > _______________________________________________ > > LLVM Developers mailing list > > llvm-dev at lists.llvm.org > > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170510/9895bcc5/attachment.html>
Mikael Holmén via llvm-dev
2017-May-11 05:20 UTC
[llvm-dev] -speculative-execution moving load before store
On 05/10/2017 09:00 PM, Amara Emerson wrote:> This looks like it's due to r301950, which has today been reverted in > r302640. Can you try with the latest trunk? >Yep after the revert it works again. Chandler (who reverted r301950) updated the original review for r302640 with the findings: https://reviews.llvm.org/D24544 /Mikael> Regardless, there seems to be an underlying bug exposed. > > On 10 May 2017 at 07:37, Mikael Holmén via llvm-dev > <llvm-dev at lists.llvm.org> wrote: >> Hi, >> >> A few days ago I stumbled upon a problem where SpeculativeExecution >> changed the order of a load and a store to the same address. >> >> I wrote >> >> https://bugs.llvm.org//show_bug.cgi?id=32964 >> >> about it but no response there so far. >> >> In the input we have >> >> store i8 0, i8* @i >> %.pre = load i8, i8* @i >> >> and then in the output the load is moved so it's before the store which >> clearly makes it load the wrong value. >> >> isSafeToSpeculativelyExecute happily says "true" for the load, and there is >> a comment on isSafeToSpeculativelyExecute saying >> >> /// This method can return true for instructions that read memory; >> /// for such instructions, moving them may change the resulting value. >> >> And that is indeed the case in PR32964. >> >> How is this supposed to work? >> >> I started seeing this after the commit >> >> SpeculativeExecution: Stop using whitelist for costs >> >> Before that, e.g. a load got the cost UINT_MAX which I suppose disabled >> hoisting, but since the commit TTI.getUserCost(I) is used for all >> instructions. >> >> Regards, >> Mikael >> _______________________________________________ >> LLVM Developers mailing list >> llvm-dev at lists.llvm.org >> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
Reasonably Related Threads
- Instruction selection pattern for intrinsic returning llvm_any_ty
- [LLVMdev] Physical register definition removed by MachineCSE
- Instruction selection pattern for intrinsic returning llvm_any_ty
- Instruction selection pattern for intrinsic returning llvm_any_ty
- DirectoryWatcher causing build failures on Redhat linux with kernel version 2.6.32