<div dir="ltr">Which cases does this pass handle which aren't otherwise optimized out by passes like GlobalValueNumbering or DeadCodeElimination?<div><br></div><div>Thanks,</div><div>Jake VanAdrighem</div></div><div class="gmail_extra"><br><div class="gmail_quote">On Thu, Sep 10, 2015 at 2:35 PM, Steve King <span dir="ltr"><<a href="mailto:steve@metrokings.com" target="_blank">steve@metrokings.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Hello LLVM,<br>
It seems this thread has gone cold. Is there some low risk way for<br>
the community to take the new pass for a test drive?<br>
Regards,<br>
-steve<br>
<div class="HOEnZb"><div class="h5"><br>
On Wed, Sep 2, 2015 at 8:27 PM, Steve King <<a href="mailto:steve@metrokings.com">steve@metrokings.com</a>> wrote:<br>
> On Wed, Sep 2, 2015 at 5:36 AM, James Molloy <<a href="mailto:james@jamesmolloy.co.uk">james@jamesmolloy.co.uk</a>> wrote:<br>
>> Hi,<br>
>><br>
>> Coremark really isn't a good enough test - have you run the LLVM test suite<br>
>> with this patch, and what were the performance differences?<br>
><br>
> For the test suite single source benches, the 235 tests improved<br>
> performance, 2 regressed and 705 were unchanged. That seems very<br>
> optimistic. Comparing consecutive runs with identical setting shows<br>
> there is a lot of noise in the performance data. Tips for stable<br>
> results would be appreciated.<br>
><br>
><br>
>> I'm still a bit confused about what pattern exactly this pass is supposed to<br>
>> trigger on. I understand the mechanics, but I still can't quite see what<br>
>> patterns it would be useful on. You've mentioned matrix multiply - how does<br>
>> this pass alter the IR?<br>
><br>
> Here's before and after IR for the matrix_mul example. Notice the two<br>
> bitcasts %1 and %2 generated in the for.cond.cleanup block. The L.E.V<br>
> pass converts these to scevgep values that already exist.<br>
><br>
> *** Code after LSR ***<br>
><br>
> ; Function Attrs: nounwind optsize<br>
> define void @matrix_mul(i32 %Size, i32* nocapture %Dst, i32* nocapture<br>
> readonly %Src, i32 %Val) #0 {<br>
> entry:<br>
> %cmp.25 = icmp eq i32 %Size, 0<br>
> br i1 %cmp.25, label %for.cond.cleanup, label %for.body.4.lr.ph.preheader<br>
><br>
> for.body.4.lr.ph.preheader: ; preds = %entry<br>
> %0 = shl i32 %Size, 2<br>
> br label %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a><br>
><br>
> <a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a>: ; preds =<br>
> %for.body.4.lr.ph.preheader, %for.cond.cleanup.3<br>
> %lsr.iv5 = phi i32* [ %Src, %for.body.4.lr.ph.preheader ], [ %2,<br>
> %for.cond.cleanup.3 ]<br>
> %lsr.iv1 = phi i32* [ %Dst, %for.body.4.lr.ph.preheader ], [ %1,<br>
> %for.cond.cleanup.3 ]<br>
> %Outer.026 = phi i32 [ %inc10, %for.cond.cleanup.3 ], [ 0,<br>
> %for.body.4.lr.ph.preheader ]<br>
> %lsr.iv56 = bitcast i32* %lsr.iv5 to i1*<br>
> %lsr.iv12 = bitcast i32* %lsr.iv1 to i1*<br>
> br label %for.body.4<br>
><br>
> for.cond.cleanup.loopexit: ; preds = %for.cond.cleanup.3<br>
> br label %for.cond.cleanup<br>
><br>
> for.cond.cleanup: ; preds =<br>
> %for.cond.cleanup.loopexit, %entry<br>
> ret void<br>
><br>
> for.cond.cleanup.3: ; preds = %for.body.4<br>
> %inc10 = add nuw nsw i32 %Outer.026, 1<br>
> %scevgep = getelementptr i1, i1* %lsr.iv12, i32 %0<br>
> %1 = bitcast i1* %scevgep to i32*<br>
> %scevgep7 = getelementptr i1, i1* %lsr.iv56, i32 %0<br>
> %2 = bitcast i1* %scevgep7 to i32*<br>
> %exitcond27 = icmp eq i32 %inc10, %Size<br>
> br i1 %exitcond27, label %for.cond.cleanup.loopexit, label %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a><br>
><br>
> for.body.4: ; preds =<br>
> %for.body.4, %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a><br>
> %lsr.iv8 = phi i32* [ %scevgep9, %for.body.4 ], [ %lsr.iv5,<br>
> %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a> ]<br>
> %lsr.iv3 = phi i32* [ %scevgep4, %for.body.4 ], [ %lsr.iv1,<br>
> %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a> ]<br>
> %lsr.iv = phi i32 [ %lsr.iv.next, %for.body.4 ], [ %Size, %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a> ]<br>
> %3 = load i32, i32* %lsr.iv8, align 4, !tbaa !1<br>
> %mul5 = mul i32 %3, %Val<br>
> store i32 %mul5, i32* %lsr.iv3, align 4, !tbaa !1<br>
> %lsr.iv.next = add i32 %lsr.iv, -1<br>
> %scevgep4 = getelementptr i32, i32* %lsr.iv3, i32 1<br>
> %scevgep9 = getelementptr i32, i32* %lsr.iv8, i32 1<br>
> %exitcond = icmp eq i32 %lsr.iv.next, 0<br>
> br i1 %exitcond, label %for.cond.cleanup.3, label %for.body.4<br>
> }<br>
><br>
><br>
> *** Code after Loop Exit Values Optimization **<br>
><br>
> ; Function Attrs: nounwind optsize<br>
> define void @matrix_mul(i32 %Size, i32* nocapture %Dst, i32* nocapture<br>
> readonly %Src, i32 %Val) #0 {<br>
> entry:<br>
> %cmp.25 = icmp eq i32 %Size, 0<br>
> br i1 %cmp.25, label %for.cond.cleanup, label %for.body.4.lr.ph.preheader<br>
><br>
> for.body.4.lr.ph.preheader: ; preds = %entry<br>
> br label %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a><br>
><br>
> <a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a>: ; preds =<br>
> %for.body.4.lr.ph.preheader, %for.cond.cleanup.3<br>
> %lsr.iv5 = phi i32* [ %Src, %for.body.4.lr.ph.preheader ], [<br>
> %scevgep9, %for.cond.cleanup.3 ]<br>
> %lsr.iv1 = phi i32* [ %Dst, %for.body.4.lr.ph.preheader ], [<br>
> %scevgep4, %for.cond.cleanup.3 ]<br>
> %Outer.026 = phi i32 [ %inc10, %for.cond.cleanup.3 ], [ 0,<br>
> %for.body.4.lr.ph.preheader ]<br>
> br label %for.body.4<br>
><br>
> for.cond.cleanup.loopexit: ; preds = %for.cond.cleanup.3<br>
> br label %for.cond.cleanup<br>
><br>
> for.cond.cleanup: ; preds =<br>
> %for.cond.cleanup.loopexit, %entry<br>
> ret void<br>
><br>
> for.cond.cleanup.3: ; preds = %for.body.4<br>
> %inc10 = add nuw nsw i32 %Outer.026, 1<br>
> %exitcond27 = icmp eq i32 %inc10, %Size<br>
> br i1 %exitcond27, label %for.cond.cleanup.loopexit, label %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a><br>
><br>
> for.body.4: ; preds =<br>
> %for.body.4, %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a><br>
> %lsr.iv8 = phi i32* [ %scevgep9, %for.body.4 ], [ %lsr.iv5,<br>
> %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a> ]<br>
> %lsr.iv3 = phi i32* [ %scevgep4, %for.body.4 ], [ %lsr.iv1,<br>
> %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a> ]<br>
> %lsr.iv = phi i32 [ %lsr.iv.next, %for.body.4 ], [ %Size, %<a href="http://for.body.4.lr.ph" rel="noreferrer" target="_blank">for.body.4.lr.ph</a> ]<br>
> %0 = load i32, i32* %lsr.iv8, align 4, !tbaa !1<br>
> %mul5 = mul i32 %0, %Val<br>
> store i32 %mul5, i32* %lsr.iv3, align 4, !tbaa !1<br>
> %lsr.iv.next = add i32 %lsr.iv, -1<br>
> %scevgep4 = getelementptr i32, i32* %lsr.iv3, i32 1<br>
> %scevgep9 = getelementptr i32, i32* %lsr.iv8, i32 1<br>
> %exitcond = icmp eq i32 %lsr.iv.next, 0<br>
> br i1 %exitcond, label %for.cond.cleanup.3, label %for.body.4<br>
><br>
><br>
>> What value is it avoiding being recomputed?<br>
> I'm not precisely sure, but it's residue from LSR. The pass checks<br>
> all computable SCEV values when a loop exits and in this case found<br>
> GEPs with the same value.<br>
><br>
>> How does this pass affect register pressure?<br>
>> Also, your example just removes a mov and an add - the push/pops are just<br>
>> register allocation (unless your pass in fact *reduces* register pressure?)<br>
><br>
> Right, the computation eliminated is the mov and add. Register<br>
> savings is a byproduct.<br>
><br>
> Regards,<br>
> -steve<br>
</div></div></blockquote></div><br></div>