Working on IRA regressions I found a bug which occurs only on IRA
branch.  The reason of the bug is a wrong descriptions of SSSE3
pmaddubsw insn which permits to exchange operands by the reload.  It
is described as commutative operation but it is not.  The insn is
defined as

[a0 a1 ...] pmaddubsw [b0 b1 ...] = [satsw(a0b0+a1b1) satsw(a2b2+a3b3) ...]

where satsw is a saturation to sign value, ax is zero extended and bx
is sign extended.  If signed extension of ax gives a negative value
and bx does not, the result will be different if a and b are
exchanged.  So

a pmaddubsw b != b pmaddubsw a

Here the patch fixing the bug.  Is it ok to commit into the mainline?

I failed create a test which exposes the bug on the mainline.
It is only exposed on IRA branch (test gcc.target/i386/ssse3-pmaddubsw.c).

2008-03-10  Vladimir Makarov  <vmakarov@redhat.com>

	* config/i386/sse.md (ssse3_pmaddubswv8hi3, ssse3_pmaddubswv4hi3):
	Remove commutativity hint.