Input constraints for inline assembler are not used for CSE

[Quuxplusone]

Bugzilla Link	PR9517
Status	NEW
Importance	P normal
Reported by	Jörg Sonnenberger (joerg@NetBSD.org)
Reported on	2011-03-20 16:41:11 -0700
Last modified on	2018-07-23 13:23:22 -0700
Version	trunk
Hardware	PC Linux
CC	echristo@gmail.com, efriedma@quicinc.com, llvm-bugs@lists.llvm.org, llvm-dev@ndave.org
Fixed by commit(s)
Attachments
Blocks
Blocked by
See also

Consider the following program:

static inline void
outb(unsigned short port, unsigned char data)
{
        __asm volatile("outb %%al,%w1" : : "a" (data), "d" (port));
}

extern unsigned short base;

void test(void)
{
        outb(base + 16, 0x0);
        outb(base + 16, 0x1);
        outb(base + 16, 0x2);
        outb(base + 16, 0x3);
        outb(base + 16, 0x4);
        outb(base + 16, 0x5);
        outb(base + 16, 0x6);
        outb(base + 16, 0x7);
        outb(base + 16, 0x8);
}

This should be compiled into something like:

movzwl base, %edx
add 0x10, %edx
xorl %eax, %eax
outb %al,%dx
movb $1, %al
outb %al,%dx
movb $2, %al
outb %al,%dx
...

or even better in terms of code size:

movzwl base, %edx
add 0x10, %edx
xorl %eax, %eax
outb %al,%dx
incl %eax
outb %al,%dx
incl %eax
outb %al,%dx
...

But currently it gets compiled to:

movzwl base, %edx
add 0x10, %edx
xorl %eax, %eax
outb %al,%dx
movzwl base, %edx
add 0x10, %edx
movb $1, %al
outb %al,%dx
movzwl base, %edx
add 0x10, %edx
movb $2, %al
outb %al,%dx
...

Input constraints are not clobbered and therefore, they don't have to be
reloaded. The reload overhead in this case is significant.

[Quuxplusone]

The issue here is actually that the loads are not being CSE'ed; alias analysis is being overly conservative, in that it assumes the inline asm can modify "base".

[Quuxplusone]

(In reply to comment #1)
> The issue here is actually that the loads are not being CSE'ed; alias analysis
> is being overly conservative, in that it assumes the inline asm can modify
> "base".

I wasn't sure about that. It shouldn't assume that since there is no clobber
"memory" specified.

[Quuxplusone]

It shouldn't, but it does. There doesn't happen to be any convenient way to represent that the inline asm has side effects, but doesn't modify any memory visible to the compiler.

[Quuxplusone]

The volatile already expresses "this has side effects".

[Quuxplusone]

Still applies as of r200573.

[Quuxplusone]

https://reviews.llvm.org/D49691