On Wed, Nov 26, 2025 at 10:53 AM Uros Bizjak <[email protected]> wrote:
> > > > For
> > > >
> > > > volatile unsigned char u8;
> > > >
> > > > void test (void)
> > > > {
> > > > u8 = u8 + u8;
> > > > u8 = u8 - u8;
> > > > }
> > > >
> > > > When volatile store is allowed, we generate
> > > >
> > > > (insn 8 7 9 2 (parallel [
> > > > (set (mem/v/c:QI (symbol_ref:DI ("u8") [flags 0x2]
> > > > <var_decl 0x7fe9719d6e40 u8>) [0 u8+0 S1 A8])
> > > > (ashift:QI (mem/v/c:QI (symbol_ref:DI ("u8") [flags
> > > > 0x2] <var_decl 0x7fe9719d6e40 u8>) [0 u8+0 S1 A8])
> > > > (const_int 1 [0x1])))
> > > > (clobber (reg:CC 17 flags))
> > > > ])
> > > > "/export/gnu/import/git/gitlab/x86-gcc-test/gcc/testsuite/gcc.dg/pr86617.c":7:6
> > > > 1139 {*ashlqi3_1}
> > > > (expr_list:REG_UNUSED (reg:CC 17 flags)
> > > > (nil)))
> > > >
> > > > on x86 which leads to
> > > >
> > > > salb u8(%rip)
> > > >
> > > > instead of 2 loads. Without the instruction, we don't know if a
> > > > memory reference
> > > > should be allowed for stores.
> > >
> > > combine pass doesn't handle volatiles correctly in all cases, this is
> > > the reason I think this propagation should be done in late-combine. We
> > > are interested only in propagations of memory loads/stores into
> > > instructions, and late-combine does exactly that.
> > >
> > > Uros.
> >
> > late combine doesn't try to combine memory references at all.
So, by enabling propagation of volatile defs in late-combine pass,
your above testcase compiles to
movzbl u8(%rip), %eax
addb u8(%rip), %al
movb %al, u8(%rip)
movzbl u8(%rip), %eax
subb u8(%rip), %al
movb %al, u8(%rip)
ret
which is better than the current:
movzbl u8(%rip), %eax
movzbl u8(%rip), %edx
addl %edx, %eax
movb %al, u8(%rip)
movzbl u8(%rip), %eax
movzbl u8(%rip), %edx
subl %edx, %eax
movb %al, u8(%rip)
ret
but still worse than clang's:
movzbl u8(%rip), %eax
addb %al, u8(%rip)
movzbl u8(%rip), %eax
subb u8(%rip), %al
movb %al, u8(%rip)
retq
(clang is able to propagate *to* volatile output and form RMW insn).
By improving late-combine pass to allow volatile defs (and checking
that volatile def gets propagated to *exactly one* place), we can
already substantially improve generated code for e.g. linux kernel.
There are many volatile reads that can be combined with a follow-up
use in the kernel, but RMW insns are relatively rarely used.
Uros.
