On Wed, Jun 17, 2026 at 6:00 AM Naveen
<[email protected]> wrote:
>
> gimple_fold_builtin_memset currently folds constant-length memset calls into
> scalar stores only when the destination is an ADDR_EXPR. Calls through an
> arbitrary pointer destination such as a function parameter are left as library
> calls. Additionally, non-constant fill values are rejected unconditionally
> even for the length-one case where no byte replication is needed.
>
> The patch adds a dedicated folding path for length-one memset calls through
> arbitrary pointer destinations. The approach follows the mode-determination
> pattern already used in gimple_fold_builtin_memory_op: the store mode and type
> are derived via int_mode_for_size, bitwise_mode_for_size and
> bitwise_type_for_mode. Destination alignment is validated against
> targetm.slow_unaligned_access and movmisalign_optab before proceeding.
>
> The fold is gated by check_bounds_or_overlap to avoid suppressing diagnostics
> from the access-warning pass. Volatile qualifiers are propagated from the
> pointed-to type and under-aligned destinations are handled via
> build_aligned_type. Both constant and non-constant fill values are supported.
> A constant fill is converted directly via fold_convert and a non-constant fill
> is converted to the byte store type via gimple_convert. The store is emitted
> as
> a MEM_REF at offset zero using a ref-all character pointer type following the
> same convention used elsewhere in gimple-fold.cc.
> The existing ADDR_EXPR-based folding path for multi-byte constant memsets is
> preserved unchanged.
>
> Attempting to lift the ADDR_EXPR restriction for multi-byte lengths caused a
> broad set of regressions across multiple passes:
> When the memset call is folded early into stores before the access-warning
> pass
> runs -Wstringop-overflow diagnostics are suppressed.
> Early folding through arbitrary pointer destinations interferes with
> __builtin_object_size and __builtin_dynamic_object_size analysis.
> -fanalyzer pass inspects memset calls for safety analysis. Folding them to raw
> MEM_REF stores prematurely removes the semantic information the analyzer
> depends
> on for reporting violations.
But would all those regressions not also appear if you'd replace the constant
that's likely in those testcases with a variable value? But possibly coverage
for size == 1 isn't big?
memcpy folding has
/* Detect out-of-bounds accesses without issuing warnings.
Avoid folding out-of-bounds copies but to avoid false
positives for unreachable code defer warning until after
DCE has worked its magic.
-Wrestrict is still diagnosed. */
if (int warning = check_bounds_or_overlap (as_a <gcall *>(stmt),
dest, src, len, len,
false, false))
if (warning != OPT_Wrestrict)
return false;
did you investigate doing sth similar?
Sorry for slowing you down here ...
Richard.
>
> gcc/ChangeLog:
> PR tree-optimization/102202
> * gimple-fold.cc (gimple_fold_builtin_memset): Hoist tree_to_uhwi
> (len)
> before the INTEGER_CST check. Add a folding path for length-one
> memset
> calls through arbitrary pointer destinations supporting both constant
> and
> non-constant fill values following the mode-determination pattern of
> gimple_fold_builtin_memory_op.
>
> gcc/testsuite/ChangeLog:
> PR tree-optimization/102202
> * gcc.dg/pr102202-fold.c: New test.
>
> Signed-off-by: Naveen <[email protected]>
> ---
> gcc/gimple-fold.cc | 76 +++++++++++++++++++++++++++-
> gcc/testsuite/gcc.dg/pr102202-fold.c | 25 +++++++++
> 2 files changed, 99 insertions(+), 2 deletions(-)
> create mode 100644 gcc/testsuite/gcc.dg/pr102202-fold.c
>
> diff --git a/gcc/gimple-fold.cc b/gcc/gimple-fold.cc
> index fab4886b0f2..856bc1f9e7f 100644
> --- a/gcc/gimple-fold.cc
> +++ b/gcc/gimple-fold.cc
> @@ -1478,10 +1478,83 @@ gimple_fold_builtin_memset (gimple_stmt_iterator
> *gsi, tree c, tree len)
> if (! tree_fits_uhwi_p (len))
> return false;
>
> + length = tree_to_uhwi (len);
> +
> + tree dest = gimple_call_arg (stmt, 0);
> + if (TREE_CODE (c) != INTEGER_CST && length == 1
> + && POINTER_TYPE_P (TREE_TYPE (dest)))
> + {
> + /* Detect out-of-bounds accesses without issuing warnings. Avoid
> + folding calls that should be diagnosed later by the access warning
> + pass. */
> + if (check_bounds_or_overlap (as_a <gcall *>(stmt),
> + dest, NULL_TREE, len,
> + NULL_TREE, false, false))
> + return false;
> +
> + unsigned int dest_align = get_pointer_alignment (dest);
> + scalar_int_mode imode;
> + machine_mode mode;
> + if (!int_mode_for_size (BITS_PER_UNIT, 0).exists (&imode)
> + || !bitwise_mode_for_size (BITS_PER_UNIT).exists (&mode)
> + || !known_eq (GET_MODE_BITSIZE (mode), BITS_PER_UNIT)
> + /* If the destination pointer is not aligned we must be able to emit
> + an unaligned store. */
> + || (dest_align < GET_MODE_ALIGNMENT (mode)
> + && targetm.slow_unaligned_access (mode, dest_align)
> + && (optab_handler (movmisalign_optab, mode)
> + == CODE_FOR_nothing)))
> + return false;
> +
> + etype = bitwise_type_for_mode (mode);
> + if (etype == NULL_TREE)
> + return false;
> +
> + if (dest_align < GET_MODE_ALIGNMENT (mode))
> + etype = build_aligned_type (etype, dest_align);
> +
> + tree ptype = TREE_TYPE (TREE_TYPE (dest));
> + if (TYPE_VOLATILE (ptype))
> + etype = build_qualified_type (etype, TYPE_QUAL_VOLATILE);
> +
> + location_t loc = gimple_location (stmt);
> + tree cval_tree;
> + if (TREE_CODE (c) == INTEGER_CST)
> + cval_tree = fold_convert (etype, c);
> + else
> + cval_tree = gimple_convert (gsi, true, GSI_SAME_STMT, loc, etype, c);
> +
> + /* Build accesses at offset zero with a ref-all character type. */
> + tree off0
> + = build_int_cst (build_pointer_type_for_mode (char_type_node,
> + ptr_mode, true), 0);
> + tree var = fold_build2_loc (loc, MEM_REF, etype, dest, off0);
> + gimple *store = gimple_build_assign (var, cval_tree);
> + gimple_move_vops (store, stmt);
> + gimple_set_location (store, loc);
> + copy_warning (store, stmt);
> +
> + tree lhs = gimple_call_lhs (stmt);
> + if (!lhs)
> + {
> + gsi_replace (gsi, store, false);
> + return true;
> + }
> +
> + gsi_insert_before (gsi, store, GSI_SAME_STMT);
> + tree ret = dest;
> + if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (dest)))
> + ret = gimple_convert (gsi, true, GSI_SAME_STMT, loc,
> + TREE_TYPE (lhs), dest);
> + gimple *asgn = gimple_build_assign (lhs, ret);
> + gsi_replace (gsi, asgn, false);
> +
> + return true;
> + }
> +
> if (TREE_CODE (c) != INTEGER_CST)
> return false;
>
> - tree dest = gimple_call_arg (stmt, 0);
> tree var = dest;
> if (TREE_CODE (var) != ADDR_EXPR)
> return false;
> @@ -1502,7 +1575,6 @@ gimple_fold_builtin_memset (gimple_stmt_iterator *gsi,
> tree c, tree len)
> if (! var_decl_component_p (var))
> return false;
>
> - length = tree_to_uhwi (len);
> if (GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (etype)) != length
> || (GET_MODE_PRECISION (SCALAR_INT_TYPE_MODE (etype))
> != GET_MODE_BITSIZE (SCALAR_INT_TYPE_MODE (etype)))
> diff --git a/gcc/testsuite/gcc.dg/pr102202-fold.c
> b/gcc/testsuite/gcc.dg/pr102202-fold.c
> new file mode 100644
> index 00000000000..3c1a0ab4fa9
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/pr102202-fold.c
> @@ -0,0 +1,25 @@
> +/* PR tree-optimization/102202 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-optimized" } */
> +
> +void *
> +g5 (char *d, int c)
> +{
> + return __builtin_memset (d, c, 1);
> +}
> +
> +void
> +g6 (char *d, int c)
> +{
> + __builtin_memset (d, c, 1);
> +}
> +
> +void *
> +g7 (int *p, int c)
> +{
> + return __builtin_memset (p, c, 1);
> +}
> +
> +/* { dg-final { scan-tree-dump-not "__builtin_memset" "optimized" } } */
> +/* { dg-final { scan-tree-dump-times {\(unsigned char\) c_[0-9]+\(D\)} 3
> "optimized" } } */
> +/* { dg-final { scan-tree-dump-times {MEM[^;\n\r]*= _[0-9]+;} 3 "optimized"
> } } */
> --
> 2.34.1
>
