<div dir="ltr">I understand that GEPs do not access memory.<div>They do a (possibly expensive) address calculation,</div><div>perhaps adding a few values to a label and leaving the result in a register.</div><div>Getting a label into a register is (to me) just like loading a 64-bit integer value</div><div>into a register. It can happen in many places and it can cost a few instructions</div><div>and several bytes. I'd like to see such things commoned and hoisted when possible.</div><div><br></div><div>I agree with you that I may be fighting the spirit of the IR.</div><div><br></div><div>I also agree with Daniel that the classic solution is to lower and re-optimize.</div><div>Indeed, I tried it, sort of, by rewriting GEPs into a sequence of adds and multiplies,</div><div>with bitcasts to gets the types right. This helps in some places, but the</div><div>re-canonicalization defeats my efforts in others places.</div><div><br></div><div>So now I feel like I'm going to have to write my own LICM and VN passes.</div><div>I can do it and it won't take that long, but I had imagined that I could</div><div>pick and choose from the tools already available to achieve the desired effect.</div><div>Seems not, and I'm surprised.</div><div><br></div><div>Preston</div><div><br></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Wed, Dec 9, 2015 at 1:10 PM, Mehdi Amini <span dir="ltr"><<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word"><br><div><span class=""><blockquote type="cite"><div>On Dec 9, 2015, at 11:16 AM, Preston Briggs <<a href="mailto:briggs@reservoir.com" target="_blank">briggs@reservoir.com</a>> wrote:</div><br><div><div dir="ltr">A GEP can represent a potentially large tree of instructions.<div>Seems like all the sub-trees are hidden from optimization;</div><div>that is, I never see licm or value numbering doing anything with them.</div><div>If I rewrite the GEPs as lots of little adds and multiplies,</div><div>then opt will do a better job (I speculate this happens during lowering).</div><div><br></div><div>One of the computations that's hidden in the GEP in my example</div><div>is the non-zero effort required to get the value of a label into a register.</div><div>I'd like to expose that effort to the optimizer; instead, it seems</div><div>hidden, papered over with the GEP.</div><div><br></div><div>Your point about putting labels in global variables seems to work</div><div>if I do it by hand. Kind of embarassing though, don't you think,</div><div>introducing an indirection to achieve better code?</div></div></div></blockquote><div><br></div></span><div>Isn’t this indirection just how your target work? </div><div>It seems to me that this indirection just achieve exactly what you want: a more risc-like and less implicit representation!</div><div><br></div><div>As far as I can tell, accessing a “label” does not require a load at the IR level, (and indeed it does not on most targets). It is baked in everywhere in the code that operates on the IR. You will fight the model all the time if you try to circumvent that. </div><div>For instance the bitcast of the global label I used in my first answer is a compile time constant expression for LLVM.</div><div>On the same topic, integer literals are assumed to be inlined in the code, we don’t expose any load in the IR to operate on a literal value.</div><div><br></div><div>Also, you are relying on an implicit load (only on your target) for global labels, which I believe is not valid IR. For LLVM IR, a GEP is nothing more than pointer arithmetic.</div><div>Per LangRef:</div><div><br></div><div> The ‘getelementptr‘ instruction is used to get the address of a subelement of an aggregate data structure. It performs address calculation only and does not access memory. The instruction can also be used to calculate a vector of such addresses.<br><br>Note the *does not access memory* part (now, you may argue that the “load” you are emitting is not accessing memory exposed in the optimizer memory model and this does not apply).</div><div><br></div><div>— </div><span class="HOEnZb"><font color="#888888"><div>Mehdi</div></font></span><div><div class="h5"><div><br></div><div><br></div><div><br></div><br><blockquote type="cite"><div><div dir="ltr"><div><br></div><div>Thanks,</div><div>Preston</div><div><br></div><div><br></div><div><br></div><div><br></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Wed, Dec 9, 2015 at 9:17 AM, Mehdi Amini <span dir="ltr"><<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word"><br><div><span><blockquote type="cite"><div>On Dec 9, 2015, at 9:00 AM, Preston Briggs <<a href="mailto:briggs@reservoir.com" target="_blank">briggs@reservoir.com</a>> wrote:</div><br><div><div dir="ltr">I suppose your view is reasonable, and perhaps common.<div>My own "taste" has always preferred machine-independent code</div><div>that is as simple as possible, so GEPs reduced to nothing more than an</div><div>add, etc, i.e., quite risc-like. Then optimize it to reduce the total number</div><div>of operations (as best we can), then raise the level during instruction</div><div>selection, taking advantage of available instructions.</div></div></div></blockquote><div><br></div></span><div>I’m not sure I see something related to risc-like here, it seems to me that your problem is not GEP vs ADD but rather that you want to expose a mode where global addresses need to be loaded and can’t be referenced directly.</div><div>(Unless I misunderstood the problem which is very possible as well)</div><div><br></div><div>Maybe you could do this with a transformation that would put all the global variable addresses in a global array and reference them through the array. That’s the only workaround I could see.</div><div><br></div><div>— </div><span><font color="#888888"><div>Mehdi</div></font></span><div><div><div><br></div><div><br></div><br><blockquote type="cite"><div><div dir="ltr"><div><br></div><div>I guess my whole scheme of using opt in this context is probably wrong headed.</div><div><br></div><div>Thanks</div><div><br></div><div><br></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Wed, Dec 9, 2015 at 8:45 AM, Mehdi Amini <span dir="ltr"><<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word"><br><div><span><blockquote type="cite"><div>On Dec 9, 2015, at 7:58 AM, Preston Briggs <<a href="mailto:briggs@reservoir.com" target="_blank">briggs@reservoir.com</a>> wrote:</div><br><div><div dir="ltr">I'm trying to make the IR "better", in a machine-independent fashion,<div>without having to do any lowering.</div></div></div></blockquote><div><br></div></span><div><div>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?</div></div><div>Right now this GEP reads as an offset from a constant global, which seems pretty optimal to me.</div><div><br></div><div><div><div>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.</div><div><br></div></div></div><div>— </div><span><font color="#888888"><div>Mehdi</div></font></span><div><div><div><br></div><br><blockquote type="cite"><div><div dir="ltr"><div><br></div><div>I've written code that rewrites GEPs as simple adds and multiplies,</div><div>which helps a lot, but there's still some sort of re-canonicalization</div><div>that's getting in my way. Is there perhaps a way to suppress it?</div></div></div></blockquote><blockquote type="cite"><div><div dir="ltr"><div><br></div><div>Thanks,</div><div>Preston</div><div><br></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Wed, Dec 9, 2015 at 7:47 AM, Mehdi Amini <span dir="ltr"><<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word">I guess is has to be done as part of the lowering for such a target, either during CodegenPrepare or during something like MachineLICM.<div><br></div><div>— </div><span><font color="#888888"><div>Mehdi</div></font></span><div><div><div><br></div><div><div><br></div><div><br><div><blockquote type="cite"><div>On Dec 9, 2015, at 7:13 AM, Preston Briggs <<a href="mailto:briggs@reservoir.com" target="_blank">briggs@reservoir.com</a>> wrote:</div><br><div><div dir="ltr">On some targets with limited addressing modes,<div>getting that 64-bit relocatable but loop-invariant value into a register</div><div>requires several instructions. I'd like those several instruction outside</div><div>the loop, where they belong.</div><div><br></div><div>Yes, my experience is that something (I assume instcombine) recanonicalizes.</div><div><br></div><div>Thanks,</div><div>Preston</div><div><br></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Tue, Dec 8, 2015 at 11:21 PM, Mehdi Amini <span dir="ltr"><<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word">Hi Preston,<div><div><div><br><div><blockquote type="cite"><div>On Dec 8, 2015, at 10:56 PM, Preston Briggs via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> wrote:</div><br><div><div dir="ltr">When I compile two different modules using<div><br></div><blockquote style="margin:0px 0px 0px 40px;border:none;padding:0px"><div><font face="monospace, monospace">clang -O -S -emit-llvm</font></div></blockquote><div><br></div><div>I get different .ll files, no surprise.</div><div><br></div><div>The first looks like</div><div><br></div><blockquote style="margin:0px 0px 0px 40px;border:none;padding:0px"><div><div><font face="monospace, monospace">double *v;</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">double zap(long n) {</font></div></div><div><div><font face="monospace, monospace"> double sum = 0;</font></div></div><div><div><font face="monospace, monospace"> for (long i = 0; i < n; i++)</font></div></div><div><div><font face="monospace, monospace"> sum += v[i];</font></div></div><div><div><font face="monospace, monospace"> return sum;</font></div></div><div><div><font face="monospace, monospace">}</font></div></div></blockquote><div><br></div><div>yielding</div><div><br></div><blockquote style="margin:0px 0px 0px 40px;border:none;padding:0px"><div><div><font face="monospace, monospace">@v = common global double* null, align 8</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">; Function Attrs: nounwind readonly uwtable</font></div></div><div><div><font face="monospace, monospace">define double @zap(i64 %n) #0 {</font></div></div><div><div><font face="monospace, monospace">entry:</font></div></div><div><div><font face="monospace, monospace"> %cmp4 = icmp sgt i64 %n, 0</font></div></div><div><div><font face="monospace, monospace"> br i1 %cmp4, label %<a href="http://for.body.lr.ph/" target="_blank">for.body.lr.ph</a>, label %for.end</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace"><a href="http://for.body.lr.ph/" target="_blank">for.body.lr.ph</a>: ; preds = %entry</font></div></div><div><div><font face="monospace, monospace"> %0 = load double** @v, align 8, !tbaa !1</font></div></div><div><div><font face="monospace, monospace"> br label %for.body</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">for.body: ; preds = %for.body, %<a href="http://for.body.lr.ph/" target="_blank">for.body.lr.ph</a></font></div></div><div><div><font face="monospace, monospace"> %i.06 = phi i64 [ 0, %<a href="http://for.body.lr.ph/" target="_blank">for.body.lr.ph</a> ], [ %inc, %for.body ]</font></div></div><div><div><font face="monospace, monospace"> %sum.05 = phi double [ 0.000000e+00, %<a href="http://for.body.lr.ph/" target="_blank">for.body.lr.ph</a> ], [ %add, %for.body ]</font></div></div><div><div><font face="monospace, monospace"> %arrayidx = getelementptr inbounds double* %0, i64 %i.06</font></div></div><div><div><font face="monospace, monospace"> %1 = load double* %arrayidx, align 8, !tbaa !5</font></div></div><div><div><font face="monospace, monospace"> %add = fadd double %sum.05, %1</font></div></div><div><div><font face="monospace, monospace"> %inc = add nsw i64 %i.06, 1</font></div></div><div><div><font face="monospace, monospace"> </font></div></div><div><div><font face="monospace, monospace">%exitcond = icmp eq i64 %inc, %n</font></div></div><div><div><font face="monospace, monospace"> br i1 %exitcond, label %for.end, label %for.body</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">for.end: ; preds = %for.body, %entry</font></div></div><div><div><font face="monospace, monospace"> %sum.0.lcssa = phi double [ 0.000000e+00, %entry ], [ %add, %for.body ]</font></div></div><div><div><font face="monospace, monospace"> ret double %sum.0.lcssa</font></div></div><div><div><font face="monospace, monospace">}</font></div></div></blockquote><div><br></div><div>and the second looks like</div><div><br></div><blockquote style="margin:0px 0px 0px 40px;border:none;padding:0px"><div><div><font face="monospace, monospace">double v[10000];</font></div></div><div><div><br></div></div></blockquote><blockquote style="margin:0px 0px 0px 40px;border:none;padding:0px"><div><div><font face="monospace, monospace">double zap(long n) {</font></div></div><div><div><font face="monospace, monospace"> double sum = 0;</font></div></div><div><div><font face="monospace, monospace"> for (long i = 0; i < n; i++)</font></div></div><div><div><font face="monospace, monospace"> sum += v[i];</font></div></div><div><div><font face="monospace, monospace"> return sum;</font></div></div><div><div><font face="monospace, monospace">}</font></div></div></blockquote><div><br></div><div>yielding</div><div><br></div><blockquote style="margin:0 0 0 40px;border:none;padding:0px"><div><div><font face="monospace, monospace">; ModuleID = 'z.c'</font></div></div><div><div><font face="monospace, monospace">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"</font></div></div><div><div><font face="monospace, monospace">target triple = "x86_64-unknown-linux-gnu"</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">@v = common global [10000 x double] zeroinitializer, align 16</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">; Function Attrs: nounwind readonly uwtable</font></div></div><div><div><font face="monospace, monospace">define double @zap(i64 %n) #0 {</font></div></div><div><div><font face="monospace, monospace">entry:</font></div></div><div><div><font face="monospace, monospace"> %cmp4 = icmp sgt i64 %n, 0</font></div></div><div><div><font face="monospace, monospace"> br i1 %cmp4, label %for.body, label %for.end</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">for.body: ; preds = %entry, %for.body</font></div></div><div><div><font face="monospace, monospace"> %i.06 = phi i64 [ %inc, %for.body ], [ 0, %entry ]</font></div></div><div><div><font face="monospace, monospace"> %sum.05 = phi double [ %add, %for.body ], [ 0.000000e+00, %entry ]</font></div></div><div><div><font face="monospace, monospace"> %arrayidx = getelementptr inbounds [10000 x double]* @v, i64 0, i64 %i.06</font></div></div><div><div><font face="monospace, monospace"> %0 = load double* %arrayidx, align 8, !tbaa !1</font></div></div><div><div><font face="monospace, monospace"> %add = fadd double %sum.05, %0</font></div></div><div><div><font face="monospace, monospace"> %inc = add nsw i64 %i.06, 1</font></div></div><div><div><font face="monospace, monospace"> %exitcond = icmp eq i64 %inc, %n</font></div></div><div><div><font face="monospace, monospace"> br i1 %exitcond, label %for.end, label %for.body</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">for.end: ; preds = %for.body, %entry</font></div></div><div><div><font face="monospace, monospace"> %sum.0.lcssa = phi double [ 0.000000e+00, %entry ], [ %add, %for.body ]</font></div></div><div><div><font face="monospace, monospace"> ret double %sum.0.lcssa</font></div></div><div><div><font face="monospace, monospace">}</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">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" }</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">!llvm.ident = !{!0}</font></div></div><div><div><font face="monospace, monospace"><br></font></div></div><div><div><font face="monospace, monospace">!0 = metadata !{metadata !"Clang Front-End version 3.4.1 (tags/RELEASE_34/final)"}</font></div></div><div><div><font face="monospace, monospace">!1 = metadata !{metadata !2, metadata !2, i64 0}</font></div></div><div><div><font face="monospace, monospace">!2 = metadata !{metadata !"double", metadata !3, i64 0}</font></div></div><div><div><font face="monospace, monospace">!3 = metadata !{metadata !"omnipotent char", metadata !4, i64 0}</font></div></div><div><div><font face="monospace, monospace">!4 = metadata !{metadata !"Simple C/C++ TBAA"}</font></div></div></blockquote><div><br></div><div>(I included all the metadata and such for the 2nd case, on the off chance it matters.)</div><div><br></div><div>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?</div></div></div></blockquote><div><br></div></div><br></div></div><div>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. </div></div><div><br></div><div>You “could” manage to represent it this way:</div><div><br></div><div><span>define double @zap(i64 %n) #0 {<br>entry:<br></span> %cmp6 = icmp sgt i64 %n, 0<br> %hoisted = bitcast [10000 x double]* @v to double*<br> br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup<br><br>for.body.preheader: ; preds = %entry<br> br label %for.body<br><br>for.cond.cleanup.loopexit: ; preds = %for.body<br> %add.lcssa = phi double [ %add, %for.body ]<br> br label %for.cond.cleanup<br><br>for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry<br> %sum.0.lcssa = phi double [ 0.000000e+00, %entry ], [ %add.lcssa, %for.cond.cleanup.loopexit ]<br> ret double %sum.0.lcssa<br><br>for.body: ; preds = %for.body.preheader, %for.body<br> %i.08 = phi i64 [ %inc, %for.body ], [ 0, %for.body.preheader ]<br> %sum.07 = phi double [ %add, %for.body ], [ 0.000000e+00, %for.body.preheader ]<br> %arrayidx = getelementptr double, double* %hoisted, i64 %i.08<br> %0 = load double, double* %arrayidx, align 8, !tbaa !2<br> %add = fadd double %sum.07, %0<br> %inc = add nuw nsw i64 %i.08, 1<span><br> %exitcond = icmp eq i64 %inc, %n<br></span> br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body<br>}</div><div><br></div><div><br></div><div>However instcombine will recanonicalize it like it was originally.</div><div><br></div><div>Since it is a GEP that operate on a constant address, this shouldn’t matter, why would you want to split this? </div><div><br></div><div>Best,</div><div><br></div><div>— </div><span><font color="#888888"><div>Mehdi</div><div><br></div></font></span></div></blockquote></div><br></div>
</div></blockquote></div><br></div></div></div></div></div></blockquote></div><br></div>
</div></blockquote></div></div></div><br></div></blockquote></div><br></div>
</div></blockquote></div></div></div><br></div></blockquote></div><br></div>
</div></blockquote></div></div></div><br></div></blockquote></div><br></div>