[llvm-dev] Hoisting convergent function calls

[Woo, David J via llvm-dev]

Hi,
Is it a bug for optimization passes to hoist calls to convergent functions out of control flow? If I run SimplifyCFG on test/Transforms/GVNHoist/hoist-convergent.ll it will do this and EarlyCSE does it as well when run on the following LLVM IR:
define float @no_convergent_func_hoisting(float %d, float %min, float %max, float %a) {
entry:
  %div = fdiv float 1.000000e+00, %d
  %cmp = fcmp oge float %div, 0.000000e+00
  %sub1 = fsub float %max, %a
  %c = call float @convergent_func(float %sub1, float %div)
  br i1 %cmp, label %if.then, label %if.else
if.then:
  %mul2 = call float @convergent_func(float %sub1, float %div)
  br label %if.end
if.else:
  %mul6 = call float @convergent_func(float %sub1, float %div)
  br label %if.end
if.end:
  %tmax.0 = phi float [ %mul2, %if.then ], [ %mul6, %if.else ]
  %add = fadd float %tmax.0, %c
  ret float %add
}
declare float @convergent_func(float, float) #0
attributes #0 = { nounwind readnone convergent }
The LLVM language reference says this behavior isn't a bug since it only forbids making functions control-dependent on additional values, and this is removing a control dependence. However, the hoist-convergent test seems to indicate that convergent functions should not be hoisted, and this is necessary for implementing cross-invocation operations like OpenGL's ballotARB() since it could affect which invocations are active when the function is called.
Thanks,
David



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170818/cd352b72/attachment.html>