<div dir="ltr"><div><div style="font-size:12.8px">> - is there any target for which fp division does not have side effects? <br></div><br>Yes - all of them. This goes back to the fact that clang/llvm do not support changing the FPENV:<br><a href="https://bugs.llvm.org/show_bug.cgi?id=8100" target="_blank">https://bugs.llvm.org/show_<wbr>bug.cgi?id=8100</a><br><br></div><div>There has been some effort to change that recently, so maybe this is (partly) fixed? (cc'ing Andy Kaylor)<br><br></div><div>These reports may also provide info / answers / work-arounds:<br></div><div><a href="https://bugs.llvm.org/show_bug.cgi?id=6050" target="_blank">https://bugs.llvm.org/show_<wbr>bug.cgi?id=6050</a><br><a href="https://bugs.llvm.org/show_bug.cgi?id=24343">https://bugs.llvm.org/show_bug.cgi?id=24343</a><br></div><div><div><a href="https://bugs.llvm.org/show_bug.cgi?id=24818" target="_blank">https://bugs.llvm.org/show_<wbr>bug.cgi?id=24818</a><br><div class="gmail_extra"><br><br><div class="gmail_quote">On Wed, Mar 15, 2017 at 3:41 AM, Samuel Antão via llvm-dev <span dir="ltr"><<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><span style="font-size:12.8px">Hi all,</span><div style="font-size:12.8px"><br></div><div style="font-size:12.8px">I came across an issue caused by the Call-Graph Simplify Pass. Here is a a small repro:</div><div style="font-size:12.8px"><br></div><div style="font-size:12.8px">```</div><div style="font-size:12.8px"><div>define double @foo(double %a1, double %a2, double %a3) #0 {</div><div>entry:</div><div> %a_mul = fmul double %a1, %a2</div><div> %a_cmp = fcmp ogt double %a3, %a_mul</div><div> br i1 %a_cmp, label %a.then, label %a.end</div><div><br></div><div>a.then: </div><div> %a_div = fdiv double %a_mul, %a3</div><div> br label %a.end</div><div><br></div><div>a.end:</div><div> %a_factor = phi double [ %a_div, %a.then ], [ 1.000000e+00, %entry ]</div><div> ret double %a_factor</div><div>}</div></div><div style="font-size:12.8px">```</div><div style="font-size:12.8px">Here, the conditional is guarding a possible division by zero. However if I run CGSimplify on this I get:</div><div style="font-size:12.8px">```</div><div style="font-size:12.8px"><div>define double @foo(double %a1, double %a2, double %a3) local_unnamed_addr #0 {</div><div>entry:</div><div> %a_mul = fmul double %a1, %a2</div><div> %a_cmp = fcmp olt double %a_mul, %a3</div><div> %a_div = fdiv double %a_mul, %a3</div><div> %a_factor = select i1 %a_cmp, double %a_div, double 1.000000e+00</div><div> ret double %a_factor</div><div>}</div></div><div style="font-size:12.8px">``` </div><div style="font-size:12.8px">This will cause a FP arithmetic exception, and the application will get a SIGFPE signal. The code that drives the change in the optimizer relies on `llvm::isSafeToSpeculativelyEx<wbr>ecute` to decide whether the division should be performed speculatively. Right now, this function returns true. One could do something similar to integer divisions and add a case there (this solved the issue for me):</div><div style="font-size:12.8px">```</div><div style="font-size:12.8px"><div>diff --git a/lib/Analysis/ValueTracking.c<wbr>pp b/lib/Analysis/ValueTracking.c<wbr>pp</div><div>index 1761dac..c61fefd 100644</div><div>--- a/lib/Analysis/ValueTracking.c<wbr>pp</div><div>+++ b/lib/Analysis/ValueTracking.c<wbr>pp</div><div>@@ -3352,6 +3352,21 @@ bool llvm::isSafeToSpeculativelyExe<wbr>cute(const Value *V,</div><div> // The numerator *might* be MinSignedValue.</div><div> return false;</div><div> }</div><div>+ case Instruction::FDiv:</div><div>+ case Instruction::FRem:{</div><div>+ const Value *Denominator = Inst->getOperand(1);</div><div>+ // x / y is undefined if y == 0</div><div>+ // The denominator is not a constant, so there is nothing we can do to prove</div><div>+ // it is non-zero.</div><div>+ if (auto *VV = dyn_cast<ConstantVector>(Denom<wbr>inator))</div><div>+ Denominator = VV->getSplatValue();</div><div>+ if (!isa<ConstantFP>(Denominator)<wbr>)</div><div>+ return false;</div><div>+ // The denominator is a zero constant - we can't speculate here.</div><div>+ if (m_AnyZero().match(Denominator<wbr>))</div><div>+ return false;</div><div>+ return true;</div><div>+ }</div><div> case Instruction::Load: {</div><div> const LoadInst *LI = cast<LoadInst>(Inst);</div><div> if (!LI->isUnordered() ||</div></div><div style="font-size:12.8px">```</div><div style="font-size:12.8px">I did my tests using a powerpc64le target, but I couldn't find any target specific login involved in this transform. In any case, I wanted to drop the questions: </div><div style="font-size:12.8px"><br></div><div style="font-size:12.8px">- is there any target that would benefit from speculative fp divisions? </div><div style="font-size:12.8px">- is there any target for which fp division does not have side effects? </div><div style="font-size:12.8px"><br></div><div style="font-size:12.8px">If not, I can go ahead and post the patch above for review.</div><div style="font-size:12.8px"><br></div><div style="font-size:12.8px">Many thanks!</div><div style="font-size:12.8px">Sam</div></div></div>
<br>______________________________<wbr>_________________<br>
LLVM Developers mailing list<br>
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/<wbr>mailman/listinfo/llvm-dev</a><br>
<br></blockquote></div><br></div></div></div></div>