[llvm-dev] persuading licm to do the right thing
Preston Briggs via llvm-dev
llvm-dev at lists.llvm.org
Wed Dec 9 09:00:58 PST 2015
I suppose your view is reasonable, and perhaps common.
My own "taste" has always preferred machine-independent code
that is as simple as possible, so GEPs reduced to nothing more than an
add, etc, i.e., quite risc-like. Then optimize it to reduce the total number
of operations (as best we can), then raise the level during instruction
selection, taking advantage of available instructions.
I guess my whole scheme of using opt in this context is probably wrong
headed.
Thanks
On Wed, Dec 9, 2015 at 8:45 AM, Mehdi Amini <mehdi.amini at apple.com> wrote:
>
> On Dec 9, 2015, at 7:58 AM, Preston Briggs <briggs at reservoir.com> wrote:
>
> I'm trying to make the IR "better", in a machine-independent fashion,
> without having to do any lowering.
>
>
> The question is “would the IR be more canonical” with the representation
> you suggest? Why would the optimizer benefit from this representation
> instead of the current one in general?
> Right now this GEP reads as an offset from a constant global, which seems
> pretty optimal to me.
>
> My impression is that when you reach a point where the “better” is target
> specific, this is part of the lowering (I’m using lowering in the sense
> that you go away from the canonical representation the optimizer expects).
> I believe it is pretty common that targets need to do this kind of lowering.
>
> —
> Mehdi
>
>
>
> I've written code that rewrites GEPs as simple adds and multiplies,
> which helps a lot, but there's still some sort of re-canonicalization
> that's getting in my way. Is there perhaps a way to suppress it?
>
>
> Thanks,
> Preston
>
>
> On Wed, Dec 9, 2015 at 7:47 AM, Mehdi Amini <mehdi.amini at apple.com> wrote:
>
>> I guess is has to be done as part of the lowering for such a target,
>> either during CodegenPrepare or during something like MachineLICM.
>>
>> —
>> Mehdi
>>
>>
>>
>> On Dec 9, 2015, at 7:13 AM, Preston Briggs <briggs at reservoir.com> wrote:
>>
>> On some targets with limited addressing modes,
>> getting that 64-bit relocatable but loop-invariant value into a register
>> requires several instructions. I'd like those several instruction outside
>> the loop, where they belong.
>>
>> Yes, my experience is that something (I assume instcombine)
>> recanonicalizes.
>>
>> Thanks,
>> Preston
>>
>>
>> On Tue, Dec 8, 2015 at 11:21 PM, Mehdi Amini <mehdi.amini at apple.com>
>> wrote:
>>
>>> Hi Preston,
>>>
>>> On Dec 8, 2015, at 10:56 PM, Preston Briggs via llvm-dev <
>>> llvm-dev at lists.llvm.org> wrote:
>>>
>>> When I compile two different modules using
>>>
>>> clang -O -S -emit-llvm
>>>
>>>
>>> I get different .ll files, no surprise.
>>>
>>> The first looks like
>>>
>>> double *v;
>>>
>>> double zap(long n) {
>>> double sum = 0;
>>> for (long i = 0; i < n; i++)
>>> sum += v[i];
>>> return sum;
>>> }
>>>
>>>
>>> yielding
>>>
>>> @v = common global double* null, align 8
>>>
>>> ; Function Attrs: nounwind readonly uwtable
>>> define double @zap(i64 %n) #0 {
>>> entry:
>>> %cmp4 = icmp sgt i64 %n, 0
>>> br i1 %cmp4, label %for.body.lr.ph, label %for.end
>>>
>>> for.body.lr.ph: ; preds = %entry
>>> %0 = load double** @v, align 8, !tbaa !1
>>> br label %for.body
>>>
>>> for.body: ; preds = %for.body, %
>>> for.body.lr.ph
>>> %i.06 = phi i64 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
>>> %sum.05 = phi double [ 0.000000e+00, %for.body.lr.ph ], [ %add,
>>> %for.body ]
>>> %arrayidx = getelementptr inbounds double* %0, i64 %i.06
>>> %1 = load double* %arrayidx, align 8, !tbaa !5
>>> %add = fadd double %sum.05, %1
>>> %inc = add nsw i64 %i.06, 1
>>>
>>> %exitcond = icmp eq i64 %inc, %n
>>> br i1 %exitcond, label %for.end, label %for.body
>>>
>>> for.end: ; preds = %for.body,
>>> %entry
>>> %sum.0.lcssa = phi double [ 0.000000e+00, %entry ], [ %add, %for.body ]
>>> ret double %sum.0.lcssa
>>> }
>>>
>>>
>>> and the second looks like
>>>
>>> double v[10000];
>>>
>>> double zap(long n) {
>>> double sum = 0;
>>> for (long i = 0; i < n; i++)
>>> sum += v[i];
>>> return sum;
>>> }
>>>
>>>
>>> yielding
>>>
>>> ; ModuleID = 'z.c'
>>> target datalayout =
>>> "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-f128:128:128-n8:16:32:64-S128"
>>> target triple = "x86_64-unknown-linux-gnu"
>>>
>>> @v = common global [10000 x double] zeroinitializer, align 16
>>>
>>> ; Function Attrs: nounwind readonly uwtable
>>> define double @zap(i64 %n) #0 {
>>> entry:
>>> %cmp4 = icmp sgt i64 %n, 0
>>> br i1 %cmp4, label %for.body, label %for.end
>>>
>>> for.body: ; preds = %entry,
>>> %for.body
>>> %i.06 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
>>> %sum.05 = phi double [ %add, %for.body ], [ 0.000000e+00, %entry ]
>>> %arrayidx = getelementptr inbounds [10000 x double]* @v, i64 0, i64
>>> %i.06
>>> %0 = load double* %arrayidx, align 8, !tbaa !1
>>> %add = fadd double %sum.05, %0
>>> %inc = add nsw i64 %i.06, 1
>>> %exitcond = icmp eq i64 %inc, %n
>>> br i1 %exitcond, label %for.end, label %for.body
>>>
>>> for.end: ; preds = %for.body,
>>> %entry
>>> %sum.0.lcssa = phi double [ 0.000000e+00, %entry ], [ %add, %for.body ]
>>> ret double %sum.0.lcssa
>>> }
>>>
>>> attributes #0 = { nounwind readonly uwtable "less-precise-fpmad"="false"
>>> "no-frame-pointer-elim"="false" "no-infs-fp-math"="false"
>>> "no-nans-fp-math"="false" "stack-protector-buffer-size"="8"
>>> "unsafe-fp-math"="false" "use-soft-float"="false" }
>>>
>>> !llvm.ident = !{!0}
>>>
>>> !0 = metadata !{metadata !"Clang Front-End version 3.4.1
>>> (tags/RELEASE_34/final)"}
>>> !1 = metadata !{metadata !2, metadata !2, i64 0}
>>> !2 = metadata !{metadata !"double", metadata !3, i64 0}
>>> !3 = metadata !{metadata !"omnipotent char", metadata !4, i64 0}
>>> !4 = metadata !{metadata !"Simple C/C++ TBAA"}
>>>
>>>
>>> (I included all the metadata and such for the 2nd case, on the off
>>> chance it matters.)
>>>
>>> Is there any way I can convince licm (or something) to rip open the GEP
>>> and hoist the reference to @v outside the loop, similar to the first
>>> example?
>>>
>>>
>>>
>>> I believe that in the second case, there is no need to load the address
>>> of v as it is constant. However you have a constant address to an array,
>>> which is represented by [10000 x double]* @v in the IR, which requires to
>>> use the two-level GEP.
>>>
>>> You “could” manage to represent it this way:
>>>
>>> define double @zap(i64 %n) #0 {
>>> entry:
>>> %cmp6 = icmp sgt i64 %n, 0
>>> %hoisted = bitcast [10000 x double]* @v to double*
>>> br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
>>>
>>> for.body.preheader: ; preds = %entry
>>> br label %for.body
>>>
>>> for.cond.cleanup.loopexit: ; preds = %for.body
>>> %add.lcssa = phi double [ %add, %for.body ]
>>> br label %for.cond.cleanup
>>>
>>> for.cond.cleanup: ; preds =
>>> %for.cond.cleanup.loopexit, %entry
>>> %sum.0.lcssa = phi double [ 0.000000e+00, %entry ], [ %add.lcssa,
>>> %for.cond.cleanup.loopexit ]
>>> ret double %sum.0.lcssa
>>>
>>> for.body: ; preds =
>>> %for.body.preheader, %for.body
>>> %i.08 = phi i64 [ %inc, %for.body ], [ 0, %for.body.preheader ]
>>> %sum.07 = phi double [ %add, %for.body ], [ 0.000000e+00,
>>> %for.body.preheader ]
>>> %arrayidx = getelementptr double, double* %hoisted, i64 %i.08
>>> %0 = load double, double* %arrayidx, align 8, !tbaa !2
>>> %add = fadd double %sum.07, %0
>>> %inc = add nuw nsw i64 %i.08, 1
>>> %exitcond = icmp eq i64 %inc, %n
>>> br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body
>>> }
>>>
>>>
>>> However instcombine will recanonicalize it like it was originally.
>>>
>>> Since it is a GEP that operate on a constant address, this shouldn’t
>>> matter, why would you want to split this?
>>>
>>> Best,
>>>
>>> —
>>> Mehdi
>>>
>>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20151209/9b036d71/attachment.html>
More information about the llvm-dev
mailing list