Re: [PATCH v3] x86, mem: move memmove to out of line assembler

[Nick Desaulniers]

On Wed, Sep 28, 2022 at 12:06 PM Nick Desaulniers
<ndesaulniers@xxxxxxxxxx> wrote:
>
> On Wed, Sep 28, 2022 at 12:24 AM Rasmus Villemoes
> <linux@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > On 27/09/2022 23.02, Nick Desaulniers wrote:
> >
> > > +     /*
> > > +      * Handle data forward by movs.
> > > +      */
> > > +.p2align 4
> > > +.Lforward_movs:
> > > +     movl    -4(src, n), tmp0
> > > +     leal    -4(dest, n), tmp1
> > > +     shrl    $2, n
> > > +     rep     movsl
> > > +     movl    tmp0, (tmp1)
> > > +     jmp     .Ldone
> >
> > So in the original code, %1 was forced to be %esi and %2 was forced to
> > be %edi and they were initialized by src and dest. But here I fail to
> > see how those registers have been properly set up before the rep movs;
> > your names for those are tmp0 and tmp2. You have just loaded the last
> > word of the source to %edi, and AFAICT %esi aka tmp2 is entirely
> > uninitialized at this point (the only use is in L16_byteswap).
> >
> > I must be missing something. Please enlighten me.
>
> No, you're right.  It looks like rep movsl needs src in %esi and dest
> needs to be in %edi, so I can't reuse the input registers from
> -mregparm=3; a pair of movs is required.  A v4 is required.
>
> Probably should write a test for memcpy where n > magic constant 680.

This unit test hangs with v3 (and passes with my local v4 which I
haven't sent out yet):
```
index 62f8ffcbbaa3..c2e852762846 100644
--- a/lib/memcpy_kunit.c
+++ b/lib/memcpy_kunit.c
@@ -107,6 +107,8 @@ static void memcpy_test(struct kunit *test)
 #undef TEST_OP
 }

+static unsigned char larger_array [2048];
+
 static void memmove_test(struct kunit *test)
 {
 #define TEST_OP "memmove"
@@ -181,6 +183,20 @@ static void memmove_test(struct kunit *test)
        ptr = &overlap.data[2];
        memmove(ptr, overlap.data, 5);
        compare("overlapping write", overlap, overlap_expected);
+
+       /* Verify larger overlapping moves. */
+       larger_array[256] = 0xaa;
+       memmove(larger_array, larger_array + 256, 1024);
+       KUNIT_ASSERT_EQ(test, larger_array[0], 0xaa);
+       KUNIT_ASSERT_EQ(test, larger_array[256], 0x00);
+       KUNIT_ASSERT_NULL(test,
+               memchr(larger_array + 1, 0xaa, ARRAY_SIZE(larger_array) - 1));
```
I'll include the tests in my v4, including another for overlapping
memmove forwards.
-- 
Thanks,
~Nick Desaulniers