target/5381: compiler does not issue emmx

[David Kastrup]

The following reply was made to PR target/5381; it has been noted by GNATS.

From: David.Kastrup@t-online.de (David Kastrup)
To: Richard Henderson <rth@redhat.com>
Cc: gcc-gnats@gcc.gnu.org
Subject: Re: target/5381: compiler does not issue emmx
Date: 15 Jan 2002 09:07:37 +0100

 Richard Henderson <rth@redhat.com> writes:
 
 > Not a bug -- pmulhuw is an SSE instruction, not part of
 > the base MMX instruction set.
 
 Oops, sorry.  Darn those spec sheets that don't bother mentioning
 this.  Probably MMX2 or something.
 
 But here to the second problem that I consider real:
 
 > > Anyhow, as you can see from the generated code, the compiler fails to
 > > emit an emmx instruction at the end of the routine, leading to possible
 > > floating point errors after return.
 > 
 > Again, not a bug.  The user is responsible for handling emmx.
 > 
 > If the compiler were deciding to emit MMX code on its own, 
 > then yes I would agree that it must clean up after itself.
 > But this is not the case here.
 
 The problem is that the compiler is allocating the MMX registers used,
 might be rearranging expressions including floating point operands in
 between, and is possibly doing optimizations that might move the
 __builtin_ functions.  It is the use of the MMX registers that
 necessitates the emmx instruction before the next floating point
 instruction.  Since the compiler is the one in control that clobbers
 the registers, and since it is the one that knows when it will next
 need the floating point stack and will have to flush any remaining MMX
 registers back to memory, it stands to reason that it should also be
 responsible for issuing the emmx instruction.  If the user put emmx
 instructions in at a place where the compiler decided to leave a
 variable in an MMX register for optimization purposes, things get
 ugly.
 
 I would even go as far as to claim that the compiler should issue an
 emmx instruction if it has assigned mmx registers due to a "y"
 contraint on assembly language templates: after all, it is only after
 the template finishes that any output operands can be moved into
 memory locations or reused for the a possibly coming up MMX
 instruction, so putting an emmx instruction into your assembly
 template yourself would definitely be impossible.  Adding another
 assembly template afterwards and hoping that the compiler will have
 cleared out registers before its execution is just crossing fingers
 and second-guessing.
 
 As long as the compiler is in charge of assigning the registers, it
 should also be in charge of telling the CPU when it is done.
 
 -- 
 David Kastrup, Kriemhildstr. 15, 44793 Bochum
 Email: David.Kastrup@t-online.de