[v2] tree-optimization/107389 - honor __builtin_assume_alignment at -O0
Commit Message
The following makes sure to set alignment information on the LHS
of __builtin_assume_alignment calls even when not optimizing so
uses as arguments to builtin functions like memcpy or __atomic_load_n
can be reflected at RTL expansion time.
Bootstrap and regtest running on x86_64-unknown-linux-gnu, OK?
Thanks,
Richard.
PR tree-optimization/107389
* gimple-low.cc (lower_builtin_assume_aligned): New.
(lower_stmt): Call it.
* gcc.dg/pr107389.c: New testcase.
---
gcc/gimple-low.cc | 41 +++++++++++++++++++++++++++++++++
gcc/testsuite/gcc.dg/pr107389.c | 13 +++++++++++
2 files changed, 54 insertions(+)
create mode 100644 gcc/testsuite/gcc.dg/pr107389.c
Comments
On Tue, Nov 08, 2022 at 01:53:48PM +0100, Richard Biener wrote:
> The following makes sure to set alignment information on the LHS
> of __builtin_assume_alignment calls even when not optimizing so
> uses as arguments to builtin functions like memcpy or __atomic_load_n
> can be reflected at RTL expansion time.
>
> Bootstrap and regtest running on x86_64-unknown-linux-gnu, OK?
>
> Thanks,
> Richard.
>
> PR tree-optimization/107389
> * gimple-low.cc (lower_builtin_assume_aligned): New.
> (lower_stmt): Call it.
>
> * gcc.dg/pr107389.c: New testcase.
> ---
> gcc/gimple-low.cc | 41 +++++++++++++++++++++++++++++++++
> gcc/testsuite/gcc.dg/pr107389.c | 13 +++++++++++
> 2 files changed, 54 insertions(+)
> create mode 100644 gcc/testsuite/gcc.dg/pr107389.c
>
> diff --git a/gcc/gimple-low.cc b/gcc/gimple-low.cc
> index 512aa9feada..f9bcb772163 100644
> --- a/gcc/gimple-low.cc
> +++ b/gcc/gimple-low.cc
> @@ -84,6 +84,7 @@ static void lower_try_catch (gimple_stmt_iterator *, struct lower_data *);
> static void lower_gimple_return (gimple_stmt_iterator *, struct lower_data *);
> static void lower_builtin_setjmp (gimple_stmt_iterator *);
> static void lower_builtin_posix_memalign (gimple_stmt_iterator *);
> +static void lower_builtin_assume_aligned (gimple_stmt_iterator *);
>
>
> /* Lower the body of current_function_decl from High GIMPLE into Low
> @@ -768,6 +769,13 @@ lower_stmt (gimple_stmt_iterator *gsi, struct lower_data *data)
> lower_builtin_posix_memalign (gsi);
> return;
> }
> + else if (DECL_FUNCTION_CODE (decl) == BUILT_IN_ASSUME_ALIGNED
> + && !optimize)
> + {
> + lower_builtin_assume_aligned (gsi);
> + data->cannot_fallthru = false;
> + return;
> + }
> }
>
> if (decl && (flags_from_decl_or_type (decl) & ECF_NORETURN))
> @@ -1310,6 +1318,39 @@ lower_builtin_posix_memalign (gimple_stmt_iterator *gsi)
> gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
> gsi_insert_after (gsi, gimple_build_label (noalign_label), GSI_NEW_STMT);
> }
> +
> +/* Lower calls to __builtin_assume_aligned when not optimizing. */
> +
> +static void
> +lower_builtin_assume_aligned (gimple_stmt_iterator *gsi)
> +{
> + gcall *call = as_a <gcall *> (gsi_stmt (*gsi));
> +
> + tree lhs = gimple_call_lhs (call);
> + if (!lhs || !POINTER_TYPE_P (TREE_TYPE (lhs)))
> + return;
I think nothing checks that TREE_CODE (lhs) == SSA_NAME,
that isn't guaranteed before ssa pass.
Otherwise LGTM.
Jakub
On Tue, 8 Nov 2022, Jakub Jelinek wrote:
> On Tue, Nov 08, 2022 at 01:53:48PM +0100, Richard Biener wrote:
> > The following makes sure to set alignment information on the LHS
> > of __builtin_assume_alignment calls even when not optimizing so
> > uses as arguments to builtin functions like memcpy or __atomic_load_n
> > can be reflected at RTL expansion time.
> >
> > Bootstrap and regtest running on x86_64-unknown-linux-gnu, OK?
> >
> > Thanks,
> > Richard.
> >
> > PR tree-optimization/107389
> > * gimple-low.cc (lower_builtin_assume_aligned): New.
> > (lower_stmt): Call it.
> >
> > * gcc.dg/pr107389.c: New testcase.
> > ---
> > gcc/gimple-low.cc | 41 +++++++++++++++++++++++++++++++++
> > gcc/testsuite/gcc.dg/pr107389.c | 13 +++++++++++
> > 2 files changed, 54 insertions(+)
> > create mode 100644 gcc/testsuite/gcc.dg/pr107389.c
> >
> > diff --git a/gcc/gimple-low.cc b/gcc/gimple-low.cc
> > index 512aa9feada..f9bcb772163 100644
> > --- a/gcc/gimple-low.cc
> > +++ b/gcc/gimple-low.cc
> > @@ -84,6 +84,7 @@ static void lower_try_catch (gimple_stmt_iterator *, struct lower_data *);
> > static void lower_gimple_return (gimple_stmt_iterator *, struct lower_data *);
> > static void lower_builtin_setjmp (gimple_stmt_iterator *);
> > static void lower_builtin_posix_memalign (gimple_stmt_iterator *);
> > +static void lower_builtin_assume_aligned (gimple_stmt_iterator *);
> >
> >
> > /* Lower the body of current_function_decl from High GIMPLE into Low
> > @@ -768,6 +769,13 @@ lower_stmt (gimple_stmt_iterator *gsi, struct lower_data *data)
> > lower_builtin_posix_memalign (gsi);
> > return;
> > }
> > + else if (DECL_FUNCTION_CODE (decl) == BUILT_IN_ASSUME_ALIGNED
> > + && !optimize)
> > + {
> > + lower_builtin_assume_aligned (gsi);
> > + data->cannot_fallthru = false;
> > + return;
> > + }
> > }
> >
> > if (decl && (flags_from_decl_or_type (decl) & ECF_NORETURN))
> > @@ -1310,6 +1318,39 @@ lower_builtin_posix_memalign (gimple_stmt_iterator *gsi)
> > gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
> > gsi_insert_after (gsi, gimple_build_label (noalign_label), GSI_NEW_STMT);
> > }
> > +
> > +/* Lower calls to __builtin_assume_aligned when not optimizing. */
> > +
> > +static void
> > +lower_builtin_assume_aligned (gimple_stmt_iterator *gsi)
> > +{
> > + gcall *call = as_a <gcall *> (gsi_stmt (*gsi));
> > +
> > + tree lhs = gimple_call_lhs (call);
> > + if (!lhs || !POINTER_TYPE_P (TREE_TYPE (lhs)))
> > + return;
>
> I think nothing checks that TREE_CODE (lhs) == SSA_NAME,
> that isn't guaranteed before ssa pass.
Yep, noticed in the testsuite and fixed.
Richard.
@@ -84,6 +84,7 @@ static void lower_try_catch (gimple_stmt_iterator *, struct lower_data *);
static void lower_gimple_return (gimple_stmt_iterator *, struct lower_data *);
static void lower_builtin_setjmp (gimple_stmt_iterator *);
static void lower_builtin_posix_memalign (gimple_stmt_iterator *);
+static void lower_builtin_assume_aligned (gimple_stmt_iterator *);
/* Lower the body of current_function_decl from High GIMPLE into Low
@@ -768,6 +769,13 @@ lower_stmt (gimple_stmt_iterator *gsi, struct lower_data *data)
lower_builtin_posix_memalign (gsi);
return;
}
+ else if (DECL_FUNCTION_CODE (decl) == BUILT_IN_ASSUME_ALIGNED
+ && !optimize)
+ {
+ lower_builtin_assume_aligned (gsi);
+ data->cannot_fallthru = false;
+ return;
+ }
}
if (decl && (flags_from_decl_or_type (decl) & ECF_NORETURN))
@@ -1310,6 +1318,39 @@ lower_builtin_posix_memalign (gimple_stmt_iterator *gsi)
gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
gsi_insert_after (gsi, gimple_build_label (noalign_label), GSI_NEW_STMT);
}
+
+/* Lower calls to __builtin_assume_aligned when not optimizing. */
+
+static void
+lower_builtin_assume_aligned (gimple_stmt_iterator *gsi)
+{
+ gcall *call = as_a <gcall *> (gsi_stmt (*gsi));
+
+ tree lhs = gimple_call_lhs (call);
+ if (!lhs || !POINTER_TYPE_P (TREE_TYPE (lhs)))
+ return;
+
+ tree align = gimple_call_arg (call, 1);
+ tree misalign = (gimple_call_num_args (call) > 2
+ ? gimple_call_arg (call, 2) : NULL_TREE);
+ if (!tree_fits_uhwi_p (align)
+ || (misalign && !tree_fits_uhwi_p (misalign)))
+ return;
+
+ unsigned aligni = TREE_INT_CST_LOW (align);
+ unsigned misaligni = misalign ? TREE_INT_CST_LOW (misalign) : 0;
+ if (aligni <= 1
+ || (aligni & (aligni - 1)) != 0
+ || (misaligni & ~(aligni - 1)) != 0)
+ return;
+
+ /* For lowering we simply transfer alignment information to the
+ result and leave the call otherwise unchanged, it will be elided
+ at RTL expansion time. */
+ ptr_info_def *pi = get_ptr_info (lhs);
+ set_ptr_info_alignment (pi, aligni, misaligni);
+ gsi_next (gsi);
+}
/* Record the variables in VARS into function FN. */
new file mode 100644
@@ -0,0 +1,13 @@
+/* { dg-do compile } */
+/* { dg-options "-fdump-tree-optimized-alias" } */
+
+unsigned foo (void *p)
+{
+ unsigned i;
+ __builtin_memcpy (&i, __builtin_assume_aligned (p, 4), sizeof (unsigned));
+ return i;
+}
+
+/* Even when not optimizing we should have alignment info on the temporary
+ feeding the memcpy. */
+/* { dg-final { scan-tree-dump "ALIGN = 4" "optimized" } } */