[PATCH dwarves 1/2] dwarf_loader: Skip the argument register the arm64 ABI leaves as a hole

[PATCH dwarves 1/2] dwarf_loader: Skip the argument register the arm64 ABI leaves as a hole

[Yonghong Song]

arm64 requires an argument whose alignment is twice the register size to
start on an even-numbered argument register, so such an argument arriving
when the next free register is an odd one leaves that register unused:

  u64 f_odd(u64 a, __int128 v, u64 b);

passes a in x0, v in x2:x3 -- skipping x1 -- and b in x4.  The x86-64 ABI
has no such rule and packs v into rsi:rdx instead.

Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
---
 dwarf_loader.c | 53 +++++++++++++++++++++++++++++++++++++++++++++++---
 dwarves.h      |  1 +
 2 files changed, 51 insertions(+), 3 deletions(-)

diff --git a/dwarf_loader.c b/dwarf_loader.c
index 61ef52f..81c2076 100644
--- a/dwarf_loader.c
+++ b/dwarf_loader.c
@@ -1501,6 +1501,23 @@ static bool arch__agg_use_two_regs(const GElf_Ehdr *ehdr)
 	}
 }
 
+/*
+ * Some ABIs require an argument whose alignment is twice the register size to
+ * start on an even-numbered argument register, leaving a hole when the next
+ * free register is an odd one. For example, on arm64,
+ *	u64 f(u64 a, __int128 v, u64 b)
+ * passes a in x0, v in x2:x3 -- skipping x1 -- and b in x4.
+ */
+static bool arch__arg_align_two_regs(const GElf_Ehdr *ehdr)
+{
+	switch (ehdr->e_machine) {
+	case EM_AARCH64:
+		return true;
+	default:
+		return false;
+	}
+}
+
 static struct template_type_param *template_type_param__new(Dwarf_Die *die, struct cu *cu, struct conf_load *conf)
 {
 	struct template_type_param *ttparm = tag__alloc(cu, sizeof(*ttparm));
@@ -3634,6 +3651,28 @@ static int parameter__abi_slots(const struct parameter *parm, const struct cu *c
 	return slots > 0 ? slots : 1;
 }
 
+static int parameter__abi_reg_align(const struct parameter *parm, const struct cu *cu)
+{
+	struct tag *type;
+
+	if (!cu->arg_align_two_regs || parm->type_byte_size <= cu->addr_size)
+		return 1;
+
+	type = tag__strip_typedefs_and_modifiers(&parm->tag, cu);
+	if (type == NULL)
+		return 1;
+
+	return tag__natural_alignment(type, cu) > cu->addr_size ? 2 : 1;
+}
+
+static int parameter__align_reg_idx(const struct parameter *parm, int reg_idx,
+				    const struct cu *cu)
+{
+	int align = parameter__abi_reg_align(parm, cu);
+
+	return (reg_idx + align - 1) & ~(align - 1);
+}
+
 static bool parameter__has_piece_info(const struct parameter *parm)
 {
 	return parm->first_reg_fields || parm->second_reg_fields;
@@ -3653,7 +3692,7 @@ static bool ftype__next_parameter_preserves_slots(struct ftype *ftype, struct pa
 	if (!next || next->loc_reg == PARAMETER_UNKNOWN_REG)
 		return false;
 
-	next_reg_idx = reg_idx + slots;
+	next_reg_idx = parameter__align_reg_idx(next, reg_idx + slots, cu);
 	return next_reg_idx < cu->nr_register_params &&
 	       next->loc_reg == cu->register_params[next_reg_idx];
 }
@@ -3702,6 +3741,7 @@ static void function__match_clang_parameter_locations(struct ftype *ftype, struc
 		if (pos->passed_in_memory)
 			continue;
 
+		reg_idx = parameter__align_reg_idx(pos, reg_idx, cu);
 		if (reg_idx >= cu->nr_register_params)
 			break;
 
@@ -3732,11 +3772,17 @@ static void function__analyze_parameter_locations(struct function *fn, struct cu
 
 	ftype__for_each_parameter(ftype, pos) {
 		bool consumes_register = true;
-		bool regs_available = reg_idx < cu->nr_register_params;
+		bool regs_available;
 		int slots = parameter__abi_slots(pos, cu);
-		int expected_reg = regs_available ? cu->register_params[reg_idx] : -1;
+		int expected_reg;
 		int reg_slots = pos->passed_in_memory ? 1 : slots;
 
+		if (!pos->passed_in_memory)
+			reg_idx = parameter__align_reg_idx(pos, reg_idx, cu);
+
+		regs_available = reg_idx < cu->nr_register_params;
+		expected_reg = regs_available ? cu->register_params[reg_idx] : -1;
+
 		if (pos->has_loc) {
 			if (true_sig_enabled && pos->loc_const_value) {
 				pos->optimized = 1;
@@ -4467,6 +4513,7 @@ static int cu__set_common(struct cu *cu, struct conf_load *conf,
 	cu->little_endian = ehdr.e_ident[EI_DATA] == ELFDATA2LSB;
 	cu->nr_register_params = arch__nr_register_params(&ehdr);
 	cu->agg_use_two_regs = arch__agg_use_two_regs(&ehdr);
+	cu->arg_align_two_regs = arch__arg_align_two_regs(&ehdr);
 	arch__set_register_params(&ehdr, cu);
 	return 0;
 }
diff --git a/dwarves.h b/dwarves.h
index df77f1e..70adbbf 100644
--- a/dwarves.h
+++ b/dwarves.h
@@ -304,6 +304,7 @@ struct cu {
 	uint8_t		 little_endian:1;
 	uint8_t		 producer_clang:1;
 	uint8_t		 agg_use_two_regs:1;	/* An aggregate like {long a; long b;} */
+	uint8_t		 arg_align_two_regs:1;	/* An over-aligned arg starts on an even register */
 	uint8_t		 nr_register_params;
 	int		 register_params[ARCH_MAX_REGISTER_PARAMS];
 	int		 functions_saved;
-- 
2.53.0-Meta

[PATCH dwarves 2/2] tests: tests: Add test for 16-byte aligned arguments on arm64

[Yonghong Song]

Cover the arm64 even-register rule for 16-byte aligned arguments:
f_even() and f_stack() where the rule does not bite, f_odd() and
f_odd_tail() where it leaves a register hole, and f_box() where the
alignment comes from a struct member rather than from the parameter
type itself.

The test results on arm64:
  $ VERBOSE=1 ./clang_parm_align16.sh
  Validation of BTF encoding of over-aligned arguments.
     BTF: u64 f_even(u64 a, u64 b, __int128 v);
     BTF: u64 f_odd(u64 a, __int128 v, u64 b);
     BTF: u64 f_odd_tail(u64 a, __int128 v, u64 b, u64 c, u64 d);
     BTF: u64 f_stack(u64 a, u64 b, u64 c, u64 d, u64 e, u64 f, u64 g, __int128 v);
     BTF: u64 f_box(u64 a, struct box s, u64 b);
  Test ./clang_parm_align16.sh passed

Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
---
 tests/clang_parm_align16.sh | 77 +++++++++++++++++++++++++++++++++++++
 tests/test_lib.sh           |  3 +-
 2 files changed, 78 insertions(+), 2 deletions(-)
 create mode 100755 tests/clang_parm_align16.sh

diff --git a/tests/clang_parm_align16.sh b/tests/clang_parm_align16.sh
new file mode 100755
index 0000000..484aeb6
--- /dev/null
+++ b/tests/clang_parm_align16.sh
@@ -0,0 +1,77 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0-only
+
+source test_lib.sh
+
+outdir=$(make_tmpdir)
+
+# Comment this out to save test data.
+trap cleanup EXIT
+
+title_log "Validation of BTF encoding of over-aligned arguments."
+
+align16="${outdir}/align16"
+CC=$(which clang 2>/dev/null)
+
+if [[ -z "$CC" ]]; then
+	info_log "skip: clang not available"
+	test_skip
+fi
+
+arch=$(uname -m)
+if [[ "$arch" != "aarch64" ]]; then
+	info_log "skip: test is arm64 only, running on $arch"
+	test_skip
+fi
+
+# arm64 makes an argument whose alignment is 16 start on an even-numbered
+# argument register, so f_odd() passes a in x0, v in x2:x3 -- leaving x1 as a
+# hole -- and b in x4.  pahole has to account for that hole, otherwise every
+# parameter after v looks like it is in an unexpected register and the whole
+# function is dropped from BTF.
+cat > ${align16}.c << EOF
+typedef unsigned long long u64;
+struct box { __int128 v; };
+
+__attribute__((noinline)) u64 f_even(u64 a, u64 b, __int128 v)
+{ return a + b + (u64)v; }
+
+__attribute__((noinline)) u64 f_odd(u64 a, __int128 v, u64 b)
+{ return a + b + (u64)v; }
+
+__attribute__((noinline)) u64 f_odd_tail(u64 a, __int128 v, u64 b, u64 c, u64 d)
+{ return a + b + c + d + (u64)v; }
+
+__attribute__((noinline)) u64 f_stack(u64 a, u64 b, u64 c, u64 d, u64 e, u64 f,
+				      u64 g, __int128 v)
+{ return a + b + c + d + e + f + g + (u64)v; }
+
+__attribute__((noinline)) u64 f_box(u64 a, struct box s, u64 b)
+{ return a + b + (u64)s.v; }
+
+u64 (*keep[])() = { (u64(*)())f_even, (u64(*)())f_odd, (u64(*)())f_odd_tail,
+		    (u64(*)())f_stack, (u64(*)())f_box };
+EOF
+
+${CC} -g -O2 -c -o ${align16}.o ${align16}.c 2>/dev/null
+if [[ $? -ne 0 ]]; then
+	info_log "skip: clang could not compile ${align16}.c"
+	test_skip
+fi
+
+LLVM_OBJCOPY=objcopy pahole -J --btf_features=consistent_func ${align16}.o
+if [[ $? -ne 0 ]]; then
+	error_log "Could not encode BTF for ${align16}.o"
+	test_fail
+fi
+
+for fn in f_even f_odd f_odd_tail f_stack f_box; do
+	encoded=$(pfunct --all --format_path=btf ${align16}.o | grep " ${fn}(")
+	verbose_log "BTF: $encoded"
+	if [[ -z "$encoded" ]]; then
+		error_log "${fn}() is missing from BTF"
+		test_fail
+	fi
+done
+
+test_pass
diff --git a/tests/test_lib.sh b/tests/test_lib.sh
index e2a7218..e2d0b0c 100755
--- a/tests/test_lib.sh
+++ b/tests/test_lib.sh
@@ -661,8 +661,7 @@ check_bpftool_btf_support()
 cleanup()
 {
 	if [ -n "$outdir" ] && [ -d "$outdir" ]; then
-		rm ${outdir}/*
-		rmdir $outdir
+		rm -rf "$outdir"
 	fi
 	return 0
 }
-- 
2.53.0-Meta

Re: [PATCH dwarves 1/2] dwarf_loader: Skip the argument register the arm64 ABI leaves as a hole

[Yonghong Song]

On 9/10/26 9:09 PM, Yonghong Song wrote:
> arm64 requires an argument whose alignment is twice the register size to
> start on an even-numbered argument register, so such an argument arriving
> when the next free register is an odd one leaves that register unused:
>
>    u64 f_odd(u64 a, __int128 v, u64 b);
>
> passes a in x0, v in x2:x3 -- skipping x1 -- and b in x4.  The x86-64 ABI
> has no such rule and packs v into rsi:rdx instead.

The kernel patch (bpf: Support by-value struct and __int128 arguments)
   https://lore.kernel.org/bpf/20260911154914.2004336-1-yonghong.song@linux.dev/
depends on this pahole patch. See kernel patch 15.

It would be great if this patch can be reviewed soon.

>
> Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
> ---
>   dwarf_loader.c | 53 +++++++++++++++++++++++++++++++++++++++++++++++---
>   dwarves.h      |  1 +
>   2 files changed, 51 insertions(+), 3 deletions(-)
>
> diff --git a/dwarf_loader.c b/dwarf_loader.c
> index 61ef52f..81c2076 100644
> --- a/dwarf_loader.c
> +++ b/dwarf_loader.c
> @@ -1501,6 +1501,23 @@ static bool arch__agg_use_two_regs(const GElf_Ehdr *ehdr)
>   	}
>   }
>   
> +/*
> + * Some ABIs require an argument whose alignment is twice the register size to
> + * start on an even-numbered argument register, leaving a hole when the next
> + * free register is an odd one. For example, on arm64,
> + *	u64 f(u64 a, __int128 v, u64 b)
> + * passes a in x0, v in x2:x3 -- skipping x1 -- and b in x4.
> + */
> +static bool arch__arg_align_two_regs(const GElf_Ehdr *ehdr)
> +{
> +	switch (ehdr->e_machine) {
> +	case EM_AARCH64:
> +		return true;
> +	default:
> +		return false;
> +	}
> +}
> +
>   static struct template_type_param *template_type_param__new(Dwarf_Die *die, struct cu *cu, struct conf_load *conf)
>   {
>   	struct template_type_param *ttparm = tag__alloc(cu, sizeof(*ttparm));
> @@ -3634,6 +3651,28 @@ static int parameter__abi_slots(const struct parameter *parm, const struct cu *c
>   	return slots > 0 ? slots : 1;
>   }
>   
> +static int parameter__abi_reg_align(const struct parameter *parm, const struct cu *cu)
> +{
> +	struct tag *type;
> +
> +	if (!cu->arg_align_two_regs || parm->type_byte_size <= cu->addr_size)
> +		return 1;
> +
> +	type = tag__strip_typedefs_and_modifiers(&parm->tag, cu);
> +	if (type == NULL)
> +		return 1;
> +
> +	return tag__natural_alignment(type, cu) > cu->addr_size ? 2 : 1;
> +}
> +
> +static int parameter__align_reg_idx(const struct parameter *parm, int reg_idx,
> +				    const struct cu *cu)
> +{
> +	int align = parameter__abi_reg_align(parm, cu);
> +
> +	return (reg_idx + align - 1) & ~(align - 1);
> +}
> +
>   static bool parameter__has_piece_info(const struct parameter *parm)
>   {
>   	return parm->first_reg_fields || parm->second_reg_fields;
> @@ -3653,7 +3692,7 @@ static bool ftype__next_parameter_preserves_slots(struct ftype *ftype, struct pa
>   	if (!next || next->loc_reg == PARAMETER_UNKNOWN_REG)
>   		return false;
>   
> -	next_reg_idx = reg_idx + slots;
> +	next_reg_idx = parameter__align_reg_idx(next, reg_idx + slots, cu);
>   	return next_reg_idx < cu->nr_register_params &&
>   	       next->loc_reg == cu->register_params[next_reg_idx];
>   }
> @@ -3702,6 +3741,7 @@ static void function__match_clang_parameter_locations(struct ftype *ftype, struc
>   		if (pos->passed_in_memory)
>   			continue;
>   
> +		reg_idx = parameter__align_reg_idx(pos, reg_idx, cu);
>   		if (reg_idx >= cu->nr_register_params)
>   			break;
>   
> @@ -3732,11 +3772,17 @@ static void function__analyze_parameter_locations(struct function *fn, struct cu
>   
>   	ftype__for_each_parameter(ftype, pos) {
>   		bool consumes_register = true;
> -		bool regs_available = reg_idx < cu->nr_register_params;
> +		bool regs_available;
>   		int slots = parameter__abi_slots(pos, cu);
> -		int expected_reg = regs_available ? cu->register_params[reg_idx] : -1;
> +		int expected_reg;
>   		int reg_slots = pos->passed_in_memory ? 1 : slots;
>   
> +		if (!pos->passed_in_memory)
> +			reg_idx = parameter__align_reg_idx(pos, reg_idx, cu);
> +
> +		regs_available = reg_idx < cu->nr_register_params;
> +		expected_reg = regs_available ? cu->register_params[reg_idx] : -1;
> +
>   		if (pos->has_loc) {
>   			if (true_sig_enabled && pos->loc_const_value) {
>   				pos->optimized = 1;
> @@ -4467,6 +4513,7 @@ static int cu__set_common(struct cu *cu, struct conf_load *conf,
>   	cu->little_endian = ehdr.e_ident[EI_DATA] == ELFDATA2LSB;
>   	cu->nr_register_params = arch__nr_register_params(&ehdr);
>   	cu->agg_use_two_regs = arch__agg_use_two_regs(&ehdr);
> +	cu->arg_align_two_regs = arch__arg_align_two_regs(&ehdr);
>   	arch__set_register_params(&ehdr, cu);
>   	return 0;
>   }
> diff --git a/dwarves.h b/dwarves.h
> index df77f1e..70adbbf 100644
> --- a/dwarves.h
> +++ b/dwarves.h
> @@ -304,6 +304,7 @@ struct cu {
>   	uint8_t		 little_endian:1;
>   	uint8_t		 producer_clang:1;
>   	uint8_t		 agg_use_two_regs:1;	/* An aggregate like {long a; long b;} */
> +	uint8_t		 arg_align_two_regs:1;	/* An over-aligned arg starts on an even register */
>   	uint8_t		 nr_register_params;
>   	int		 register_params[ARCH_MAX_REGISTER_PARAMS];
>   	int		 functions_saved;