Native code generation error for simplified, possibly frem related composite creation

[waTeim]

This is a followup from the question Odd type construction behavior on Stack Overflow; although that particular problem doesn't appear to exist in 0.4, I think this is because in 0.4, mod is a function call rather than inlined.

However, the following related problem still persists in 0.4

Decl

immutable Vec3
   x::Float32
   y::Float32
   z::Float32
end

f1(a::Vec3, b::Vec3) = Vec3(a.x % b.x, a.y % b.y, a.z % b.z)

Using it

julia> a = Vec3(1,1,1)
Vec3(1.0f0,1.0f0,1.0f0)

julia> b = Vec3(2,2,2)
Vec3(2.0f0,2.0f0,2.0f0)

julia> f1(a,b)
Vec3(1.0f0,2.0f0,NaN32)

Re-decl and use

Similar to the SO example, modifying slightly "fixes" things.

julia> f2(a::Vec3, b::Vec3) = Vec3(a.x % b.x, a.y % b.y, a.z % b.z + 1)
f1 (generic function with 1 method)

julia> f2(a,b)
Vec3(1.0f0,1.0f0,2.0f0)

f3(a::Vec3, b::Vec3) = Vec3(a.x % b.x, a.y % b.y,  2)
f3 (generic function with 1 method)

julia> f3(a,b)
Vec3(1.0f0,1.0f0,2.0f0)

Code LLVM and native output

julia> code_llvm(f1,(Vec3,Vec3))
define %Vec3 @"julia_f1;62077"(%Vec3, %Vec3) {
top:
  %2 = extractvalue %Vec3 %0, 0, !dbg !1949
  %3 = extractvalue %Vec3 %1, 0, !dbg !1949
  %4 = frem float %2, %3, !dbg !1949
  %5 = insertvalue %Vec3 undef, float %4, 0, !dbg !1949
  %6 = extractvalue %Vec3 %0, 1, !dbg !1949
  %7 = extractvalue %Vec3 %1, 1, !dbg !1949
  %8 = frem float %6, %7, !dbg !1949
  %9 = insertvalue %Vec3 %5, float %8, 1, !dbg !1949
  %10 = extractvalue %Vec3 %0, 2, !dbg !1949
  %11 = extractvalue %Vec3 %1, 2, !dbg !1949
  %12 = frem float %10, %11, !dbg !1949
  %13 = insertvalue %Vec3 %9, float %12, 2, !dbg !1949, !julia_type !1951
  ret %Vec3 %13, !dbg !1949
}
julia> code_native(f1,(Vec3,Vec3))
  .section  __TEXT,__text,regular,pure_instructions
Filename: none
Source line: 1
  push  RBP
  mov   RBP, RSP
  movabs    RAX, 140735600044048
Source line: 1
  vmovaps   XMM0, XMM2
  vmovaps   XMM1, XMM5
  pop   RBP
  jmp   RAX

Platform

Platform Info:
  System: Darwin (x86_64-apple-darwin13.3.0)
  CPU: Intel(R) Core(TM) i7-4850HQ CPU @ 2.30GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas
  LIBM: libopenlibm
  LLVM: libLLVM-3.3

[simonster]

Appears to work properly with LLVM SVN.

[tkelman]

There are too many conflicts for me to understand how to backport 13c83db correctly. Can someone else resolve them and open a PR against release-0.4?

[vtjnash]

See 889e31f, if that passes CI

[tkelman]

backported in #13182