__volatile__ for asms in unaligned.c

__volatile__ for asms in unaligned.c

[Kip Walker]

I just discovered that my compiler is scheduling the code in the asm
blocks in unaligned.c *before* the 'goto sigbus'.  My gcc is 3.1.1 with
some almost certainly unrelated local mods.

Anyway, is there a reason these aren't marked as volatile?  The gcc docs
have the scary comment "You can prevent an `asm' instruction from being
deleted, MOVED SIGNIFICANTLY, or combined, by writing the keyword
`volatile' after the`asm'."

Here's an example, for the lw_op case in the mips64 kernel:
---------
	case lw_op:
		if (verify_area(VERIFY_READ, addr, 4))
			goto sigbus;

		__asm__(
			"1:\tlwl\t%0, (%2)\n"
			"2:\tlwr\t%0, 3(%2)\n\t"
---------

Compiled with normal mips64 build flags (for SB1) was turned into:
---------
	### verify_area
        ld      $2,2400($28)
        daddu   $3,$5,4
        or      $3,$5,$3
        and     $2,$2,$3
        li      $4,-14                  # 0xfffffffffffffff2
        movz    $4,$0,$2
...
	### the asm code
        1:      lwl     $9, ($5)
2:      lwr     $9, 3($5)
        li      $3, 0
3:      .section        .fixup,"ax"
        4:      li      $3, -14
        j       3b

...
	### finally, the verify_area result check
        beq     $4,$0,$L1131

---------

Kip

Re: __volatile__ for asms in unaligned.c

[Ralf Baechle]

On Thu, Feb 20, 2003 at 02:43:58PM -0800, Kip Walker wrote:

> Anyway, is there a reason these aren't marked as volatile?  The gcc docs
> have the scary comment "You can prevent an `asm' instruction from being
> deleted, MOVED SIGNIFICANTLY, or combined, by writing the keyword
> `volatile' after the`asm'."

It's a valid code movement by gcc's no longer really new basic block
reordering thing.  I admit I'm pretty surprised this wasn't found before.

So I just added the volatiles.

  Ralf