[llvm-dev] Suboptimal code generated by clang+llc in quite a common scenario (?)

Thu Aug 8 09:07:29 PDT 2019

This might not be the workaround you want because it is only available in
C, but you can use restrict to allow such optimizations.

https://godbolt.org/z/2gQ26f

Alex

On Thu, Aug 8, 2019 at 11:50 AM Michael Kruse via llvm-dev <
llvm-dev at lists.llvm.org> wrote:

> Hi,
>
> char* scscx is an universal pointer and may point to anything,
> including itself. That is, scscx might point to itself:
>
> scscx = (char*)&scscx;
>
> such that
>
> scscx[0] = ...
>
> changes the address scscx point to. A pointer to (int*) in contrast is
> only allowed to point to integers in memory, it is not an universal
> pointer. In particular, when accessing it the compiler can assume that
> it is not aliasing with something that is of type char*.
>
> For more details, see e.g. Wikipedia [1] or Stackoverflow [2]
>
> [1]
> https://en.wikipedia.org/wiki/Pointer_aliasing#Aliasing_and_re-ordering
> [2]
> https://stackoverflow.com/questions/98650/what-is-the-strict-aliasing-rule
>
> Michael
>
>
> Am Do., 8. Aug. 2019 um 10:19 Uhr schrieb Joan Lluch via llvm-dev
> <llvm-dev at lists.llvm.org>:
> >
> > I found a something that I quite not understand when compiling a common
> piece of code using the -Os flags.
> > I found it while testing my own backend but then I got deeper and found
> that at least the x86 is affected as well. This is the referred code:
> >
> > char pp[3];
> > char *scscx = pp;
> > int tst( char i, char j, char k )
> > {
> >   scscx[0] = i;
> >   scscx[1] = j;
> >   scscx[2] = k;
> >   return 0;
> > }
> >
> > The above gets compiled for the x86 architecture like this:
> >
> > ; Function Attrs: nofree norecurse nounwind optsize uwtable
> > define i32 @tst(i8 signext %i, i8 signext %j, i8 signext %k)
> local_unnamed_addr #1 {
> > entry:
> >   %0 = load i8*, i8** @scscx, align 8, !tbaa !11
> >   store i8 %i, i8* %0, align 1, !tbaa !13
> >   %1 = load i8*, i8** @scscx, align 8, !tbaa !11
> >   %arrayidx1 = getelementptr inbounds i8, i8* %1, i64 1
> >   store i8 %j, i8* %arrayidx1, align 1, !tbaa !13
> >   %2 = load i8*, i8** @scscx, align 8, !tbaa !11
> >   %arrayidx2 = getelementptr inbounds i8, i8* %2, i64 2
> >   store i8 %k, i8* %arrayidx2, align 1, !tbaa !13
> >   ret i32 0
> > }
> >
> > According to that, the variable ‘scscx’ is loaded three times despite
> it’s never modified. The resulting assembly code is this:
> >
> > .globl _tst
> > _tst:
> > .cfi_startproc
> > pushl %ebp
> > .cfi_def_cfa_offset 8
> > .cfi_offset %ebp, -8
> > movl %esp, %ebp
> > .cfi_def_cfa_register %ebp
> > pushl %esi
> > .cfi_offset %esi, -12
> > movb 16(%ebp), %al
> > movb 12(%ebp), %cl
> > movb 8(%ebp), %dl
> > movl _scscx, %esi
> > movb %dl, (%esi)
> > movl _scscx, %edx
> > movb %cl, 1(%edx)
> > movl _scscx, %ecx
> > movb %al, 2(%ecx)
> > xorl %eax, %eax
> > popl %esi
> > popl %ebp
> > retl
> > .cfi_endproc
> >
> > .comm _pp,3,0
> > .section __DATA,__data
> > .globl _scscx
> > .p2align 3
> > _scscx:
> > .long _pp
> >
> >
> > Again, the _scscx is loaded three times instead of reusing a register,
> which is suboptimal.
> >
> >
> > NOW, if I replace the original code by this:
> >
> > int pp[3];
> > int *scscx = pp;
> > int tst( int i, int j, int k )
> > {
> >   scscx[0] = i;
> >   scscx[1] = j;
> >   scscx[2] = k;
> >   return 0;
> > }
> >
> > I get the following:
> >
> >
> > ; Function Attrs: nofree norecurse nounwind optsize uwtable
> > define i32 @tst(i32 %i, i32 %j, i32 %k) local_unnamed_addr #1 {
> > entry:
> >   %0 = load i32*, i32** @scscx, align 8, !tbaa !11
> >   store i32 %i, i32* %0, align 4, !tbaa !13
> >   %arrayidx1 = getelementptr inbounds i32, i32* %0, i64 1
> >   store i32 %j, i32* %arrayidx1, align 4, !tbaa !13
> >   %arrayidx2 = getelementptr inbounds i32, i32* %0, i64 2
> >   store i32 %k, i32* %arrayidx2, align 4, !tbaa !13
> >   ret i32 0
> > }
> >
> >
> > .globl _tst
> > _tst:
> > .cfi_startproc
> > pushl %ebp
> > .cfi_def_cfa_offset 8
> > .cfi_offset %ebp, -8
> > movl %esp, %ebp
> > .cfi_def_cfa_register %ebp
> > pushl %esi
> > .cfi_offset %esi, -12
> > movl 16(%ebp), %eax
> > movl 12(%ebp), %ecx
> > movl 8(%ebp), %edx
> > movl _scscx, %esi
> > movl %edx, (%esi)
> > movl %ecx, 4(%esi)
> > movl %eax, 8(%esi)
> > xorl %eax, %eax
> > popl %esi
> > popl %ebp
> > retl
> > .cfi_endproc
> >
> > .comm _pp,12,2
> > .section __DATA,__data
> > .globl _scscx
> > .p2align 3
> > _scscx:
> > .long _pp
> >
> >
> > In this case the compiler optimises the load of _scscx into a register
> and reuses its value instead of loading the variable multiple times. This
> results in a cleaner and more optimal code, specially when compared with
> the first case.
> >
> > I would like to understand why this happens, and whether there’s a way
> (or workaround) to improve it?
> >
> > Should I file a bug report for that?
> >
> > Thanks.
> >
> > Joan
> >
> >
> >
> >
> >
> >
> >
> >
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://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/20190808/7278d9f6/attachment.html>