Different floating point results with -msse2 on 32bit Linux

I noticed strange behaviour of Conversions test on 32bit machines. Here's a slightly reduced version of that test:

{-# LANGUAGE BangPatterns #-}

-- | Tests that conversions between various primitive types (e.g.
-- Word, Double, etc) doesn't allocate.
module Main (main) where

import Data.Word

-- Repeatedly convert Words to Doubles
loop :: Word -> Float
loop n = go 0 0.0
  where
    go i !acc | i < n = go (i+1) (acc + fromIntegral i)
              | otherwise = acc

main :: IO ()
main = do
    print (loop 1000000 :: Float)

This test is expected to produce:

[t-jastol@cam-05-unx : /5playpen/t-jastol/ghc-validate] inplace/bin/ghc-stage2 -O -fforce-recomp testsuite/tests/perf/should_run/Conversions.hs
[1 of 1] Compiling Main             ( testsuite/tests/perf/should_run/Conversions.hs, testsuite/tests/perf/should_run/Conversions.o )
Linking testsuite/tests/perf/should_run/Conversions ...
[t-jastol@cam-05-unx : /5playpen/t-jastol/ghc-validate] ./testsuite/tests/perf/should_run/Conversions
4.9994036e11

My optimization of self-recursive tail loops (d61c3ac1) change results of that program to be slightly different:

4.999995e11

I believe my optimisation shouldn't affect this, but clearly it does.

Now here's the interesting part: if I generate SSE assembly by adding -msse2 flag, then I get expected results:

4.9994036e11

I'm not sure if we should get different results for different instruction sets. I'm putting it up as a ticket so perhaps we could investigate this more one day if we consider it relevant.