eqfui odd inconsistency

[ihnorton]

This came up in #8402. There seems to be an inconsistency between eqfui results at the toplevel vs. in a function. This does appear to be fixed in llvm-svn. But, it seems odd enough to deserve its own issue (or perhaps there is some explanation that will help me figure out the failure on ARM -- and the fact that the result there is inverted).

On linux and windows x86_64/llvm3.3 builds (on both master and 0.3.2), I see the following:

julia> typemax(UInt64) != 2.0^64
true
julia> f() = typemax(UInt64) != 2.0^64
f (generic function with 1 method)
julia> f()
false

We assert the first result in a numbers test, so presumably the same should hold inside of a function...

The IR for julia_!=... is identical on all platforms / llvm versions:
julia> @code_llvm typemax(Uint64) != 2.0^64

define i1 @"julia_!=41538"(i64, double) {
top:
  %2 = uitofp i64 %0 to double, !dbg !8
  %3 = fptoui double %2 to i64, !dbg !8
  %notlhs = fcmp une double %2, %1, !dbg !8
  %notrhs = icmp ne i64 %3, %0, !dbg !8
  %4 = or i1 %notlhs, %notrhs, !dbg !8
  ret i1 %4, !dbg !8
}

(the comparison is folded to a bool for the function version)

[PallHaraldsson]

Just came from #8402 that is closed. If it depended on this and already fixed should this be closed?

[ihnorton]

No. This depends on LLVM3.5/6/7 upgrade. As tagged.

On Fri, May 15, 2015 at 3:17 PM, Páll Haraldsson notifications@github.com
wrote:

Just came from #8402 #8402
that is closed. If it depended on this and already fixed should this be
closed?

—
Reply to this email directly or view it on GitHub
#9085 (comment).