This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
[PATCH][gdb/tdep] Fix inferior call arg passing for amd64
[ was: Re: [PATCH] gdb/testsuite: Fix typos in infcall-nested-structs.c ]
On 11-10-2019 14:19, Tom de Vries wrote:
> On 10-10-2019 23:07, Tom de Vries wrote:
>> On 10-10-2019 22:26, Tom de Vries wrote:
>>> On 10-10-2019 20:30, Tom de Vries wrote:
>>>> On 10-10-2019 19:24, Andreas Arnez wrote:
>>>>> On Thu, Oct 10 2019, Tom de Vries wrote:
>>>>>
>>>>>> I see these new failures on x86_64-linux:
>>>>>> ...
>>>>>> FAIL: gdb.base/infcall-nested-structs.exp: l=c++: types-tc-tf: p/d
>>>>>> check_arg_struct_02_01 (ref_val_struct_02_01)
>>>>>> FAIL: gdb.base/infcall-nested-structs.exp: l=c++: types-ts-tf: p/d
>>>>>> check_arg_struct_02_01 (ref_val_struct_02_01)
>>>>>> FAIL: gdb.base/infcall-nested-structs.exp: l=c++: types-ti-tf: p/d
>>>>>> check_arg_struct_02_01 (ref_val_struct_02_01)
>>>>>> ...
>>>>>
>>>>> Maybe the test case caught a real bug then, right? Or do you see a
>>>>> problem with the test case?
>>>>
>>>> I think it's a real bug.
>>>>
>>>> I've minimized the types-ti-tf FAIL to:
>>>> ...
>>>> $ cat test.c
>>>> typedef int ti;
>>>> typedef float tf;
>>>> struct struct_02_01
>>>> {
>>>> struct { } es1;
>>>> struct {
>>>> struct {
>>>> ti a;
>>>> tf b;
>>>> } s1;
>>>> } s2;
>>>> };
>>>>
>>>> struct struct_02_01 ref_val_struct_02_01 = {
>>>> {},
>>>> {
>>>> {
>>>> 'a',
>>>> 'b'
>>>> }
>>>> }
>>>> };
>>>>
>>>> int cmp_struct_02_01 (struct struct_02_01 a, struct struct_02_01 b)
>>>> { return a.s2.s1.a == b.s2.s1.a && a.s2.s1.b == b.s2.s1.b; }
>>>>
>>>> int
>>>> check_arg_struct_02_01 (struct struct_02_01 arg) {
>>>> return cmp_struct_02_01 (arg, ref_val_struct_02_01);
>>>> }
>>>>
>>>> int
>>>> main (void)
>>>> {
>>>> return check_arg_struct_02_01 (ref_val_struct_02_01);
>>>> }
>>>> $ g++ test.c -g
>>>> $ ./a.out; echo $?
>>>> 1
>>>> $ gdb a.out -batch -ex start -ex "p check_arg_struct_02_01
>>>> (ref_val_struct_02_01)"
>>>> Temporary breakpoint 1 at 0x400563: file test.c, line 35.
>>>>
>>>> Temporary breakpoint 1, main () at test.c:35
>>>> 35 return check_arg_struct_02_01 (ref_val_struct_02_01);
>>>> $1 = 0
>>>> ...
>>>>
>>>
>>> The discrepancy is that the code generated by gcc passes the struct in
>>> registers %rdi and %xmm0, but amd64_push_arguments classifies the struct as:
>>> ...
>>> (gdb) p theclass
>>> $57 = {AMD64_INTEGER, AMD64_INTEGER}
>>> ...
>>> and therefore passes it in %rdi and %rsi.
>>>
>>
>> I've simplified the test-case a bit further, and filed as:
>> https://sourceware.org/bugzilla/show_bug.cgi?id=25096.
>>
>
> I've submitted a fix for PR24104 (
> https://sourceware.org/ml/gdb-patches/2019-10/msg00293.html ) which
> marks these 3 FAILs as KFAILs.
>
> Thanks,
> - Tom
>
And here's a follow-up patch that fixes PR25096.
OK for trunk?
Thanks,
- Tom
[gdb/tdep] Fix inferior call arg passing for amd64
With "[gdb/tdep] Fix 'Unexpected register class' assert in
amd64_push_arguments" as proposed here (
https://sourceware.org/ml/gdb-patches/2019-10/msg00293.html ) applied, we have
12 KFAILS for PR tdep/25096.
A minimal version of the failure looks like this. Consider test.c:
...
struct s { int c; struct { int a; float b; } s1; };
struct s ref = { 0, { 'a', 'b' } };
int __attribute__((noinline,noclone)) check (struct s arg)
{ return arg.s1.a == 'a' && arg.s1.b == 'b' && arg.c == 0; }
int main (void)
{ return check (ref); }
...
When calling 'check (ref)' from main, we have '1' as expected:
...
$ g++ test.c -g ; ./a.out ; echo $?
1
...
But when calling 'check (ref)' from the gdb prompt, we get '0':
...
$ gdb a.out -batch -ex start -ex "p check (ref)"
Temporary breakpoint 1 at 0x400518: file test.c, line 8.
Temporary breakpoint 1, main () at test.c:8
8 { return check (ref); }
$1 = 0
...
The layout of struct s is this:
- the field c occupies 4 bytes at offset 0,
- the s1.a field occupies 4 bytes at offset 4, and
- the s1.b field occupies 4 bytes at offset 8.
When compiling at -O2, we can see from the disassembly of main:
...
4003f0: 48 8b 3d 31 0c 20 00 mov 0x200c31(%rip),%rdi \
# 601028 <ref>
4003f7: f3 0f 10 05 31 0c 20 movss 0x200c31(%rip),%xmm0 \
# 601030 <ref+0x8>
4003fe: 00
4003ff: e9 ec 00 00 00 jmpq 4004f0 <_Z5check1s>
...
that check is called with fields c and s1.a passed in %rdi, and s1.b passed
in %xmm0.
However, the classification in theclass (a variable representing the first and
second eightbytes, to put it in SYSV X86_64 psABI terms) in
amd64_push_arguments is incorrect:
...
(gdb) p theclass
$1 = {AMD64_INTEGER, AMD64_INTEGER}
...
and therefore the struct is passed using %rdi and %rsi instead of using %rdi
and %xmm0, which explains the failure.
The reason that we're misclassifying the argument in amd64_classify_aggregate
has to do with how nested struct are handled.
Rather than using fields c and s1.a for the first eightbyte, and using field
s1.b for the second eightbyte, instead field c is used for the first
eightbyte, and fields s1.a and s1.b are classified together in an intermediate
eightbyte, which is then used to merge with both the first and second
eightbyte.
Fix this by factoring out a new function amd64_classify_aggregate_field, and
letting it recursively handle fields of nested structs.
Tested on x86_64-linux.
Tested with g++ 4.8.5, 7.4.1, 8.3.1, 9.2.1.
Tested with clang++ 5.0.2 (which requires removing additional_flags=-Wno-psabi
and adding additional_flags=-Wno-deprecated).
gdb/ChangeLog:
2019-10-11 Tom de Vries <tdevries@suse.de>
PR tdep/25096
* amd64-tdep.c (amd64_classify_aggregate_field): Factor out of ...
(amd64_classify_aggregate): ... here.
(amd64_classify_aggregate_field): Handled fiels of nested structs
recursively.
gdb/testsuite/ChangeLog:
2019-10-11 Tom de Vries <tdevries@suse.de>
PR tdep/25096
* gdb.base/infcall-nested-structs.exp: Remove PR25096 KFAILs.
---
gdb/amd64-tdep.c | 114 +++++++++++++---------
gdb/testsuite/gdb.base/infcall-nested-structs.exp | 8 --
2 files changed, 67 insertions(+), 55 deletions(-)
diff --git a/gdb/amd64-tdep.c b/gdb/amd64-tdep.c
index 9006ec0167a..67a4d3f11a9 100644
--- a/gdb/amd64-tdep.c
+++ b/gdb/amd64-tdep.c
@@ -579,6 +579,72 @@ amd64_has_unaligned_fields (struct type *type)
return false;
}
+/* Classify field I of TYPE starting at BITOFFSET according to the rules for
+ structures and union types, and store the result in THECLASS. */
+
+static void
+amd64_classify_aggregate_field (struct type *type, int i,
+ enum amd64_reg_class theclass[2],
+ unsigned int bitoffset)
+{
+ struct type *subtype = check_typedef (TYPE_FIELD_TYPE (type, i));
+ int bitpos = bitoffset + TYPE_FIELD_BITPOS (type, i);
+ int pos = bitpos / 64;
+ enum amd64_reg_class subclass[2];
+ int bitsize = TYPE_FIELD_BITSIZE (type, i);
+ int endpos;
+
+ if (bitsize == 0)
+ bitsize = TYPE_LENGTH (subtype) * 8;
+ endpos = (bitpos + bitsize - 1) / 64;
+
+ /* Ignore static fields, or empty fields, for example nested
+ empty structures.*/
+ if (field_is_static (&TYPE_FIELD (type, i)) || bitsize == 0)
+ return;
+
+ if (TYPE_CODE (subtype) == TYPE_CODE_STRUCT
+ || TYPE_CODE (subtype) == TYPE_CODE_UNION)
+ {
+ /* Each field of an object is classified recursively. */
+ int j;
+ for (j = 0; j < TYPE_NFIELDS (subtype); j++)
+ amd64_classify_aggregate_field (subtype, j, theclass, bitpos);
+ return;
+ }
+
+ gdb_assert (pos == 0 || pos == 1);
+
+ amd64_classify (subtype, subclass);
+ theclass[pos] = amd64_merge_classes (theclass[pos], subclass[0]);
+ if (bitsize <= 64 && pos == 0 && endpos == 1)
+ /* This is a bit of an odd case: We have a field that would
+ normally fit in one of the two eightbytes, except that
+ it is placed in a way that this field straddles them.
+ This has been seen with a structure containing an array.
+
+ The ABI is a bit unclear in this case, but we assume that
+ this field's class (stored in subclass[0]) must also be merged
+ into class[1]. In other words, our field has a piece stored
+ in the second eight-byte, and thus its class applies to
+ the second eight-byte as well.
+
+ In the case where the field length exceeds 8 bytes,
+ it should not be necessary to merge the field class
+ into class[1]. As LEN > 8, subclass[1] is necessarily
+ different from AMD64_NO_CLASS. If subclass[1] is equal
+ to subclass[0], then the normal class[1]/subclass[1]
+ merging will take care of everything. For subclass[1]
+ to be different from subclass[0], I can only see the case
+ where we have a SSE/SSEUP or X87/X87UP pair, which both
+ use up all 16 bytes of the aggregate, and are already
+ handled just fine (because each portion sits on its own
+ 8-byte). */
+ theclass[1] = amd64_merge_classes (theclass[1], subclass[0]);
+ if (pos == 0)
+ theclass[1] = amd64_merge_classes (theclass[1], subclass[1]);
+}
+
/* Classify TYPE according to the rules for aggregate (structures and
arrays) and union types, and store the result in CLASS. */
@@ -619,53 +685,7 @@ amd64_classify_aggregate (struct type *type, enum amd64_reg_class theclass[2])
|| TYPE_CODE (type) == TYPE_CODE_UNION);
for (i = 0; i < TYPE_NFIELDS (type); i++)
- {
- struct type *subtype = check_typedef (TYPE_FIELD_TYPE (type, i));
- int pos = TYPE_FIELD_BITPOS (type, i) / 64;
- enum amd64_reg_class subclass[2];
- int bitsize = TYPE_FIELD_BITSIZE (type, i);
- int endpos;
-
- if (bitsize == 0)
- bitsize = TYPE_LENGTH (subtype) * 8;
- endpos = (TYPE_FIELD_BITPOS (type, i) + bitsize - 1) / 64;
-
- /* Ignore static fields, or empty fields, for example nested
- empty structures.*/
- if (field_is_static (&TYPE_FIELD (type, i)) || bitsize == 0)
- continue;
-
- gdb_assert (pos == 0 || pos == 1);
-
- amd64_classify (subtype, subclass);
- theclass[pos] = amd64_merge_classes (theclass[pos], subclass[0]);
- if (bitsize <= 64 && pos == 0 && endpos == 1)
- /* This is a bit of an odd case: We have a field that would
- normally fit in one of the two eightbytes, except that
- it is placed in a way that this field straddles them.
- This has been seen with a structure containing an array.
-
- The ABI is a bit unclear in this case, but we assume that
- this field's class (stored in subclass[0]) must also be merged
- into class[1]. In other words, our field has a piece stored
- in the second eight-byte, and thus its class applies to
- the second eight-byte as well.
-
- In the case where the field length exceeds 8 bytes,
- it should not be necessary to merge the field class
- into class[1]. As LEN > 8, subclass[1] is necessarily
- different from AMD64_NO_CLASS. If subclass[1] is equal
- to subclass[0], then the normal class[1]/subclass[1]
- merging will take care of everything. For subclass[1]
- to be different from subclass[0], I can only see the case
- where we have a SSE/SSEUP or X87/X87UP pair, which both
- use up all 16 bytes of the aggregate, and are already
- handled just fine (because each portion sits on its own
- 8-byte). */
- theclass[1] = amd64_merge_classes (theclass[1], subclass[0]);
- if (pos == 0)
- theclass[1] = amd64_merge_classes (theclass[1], subclass[1]);
- }
+ amd64_classify_aggregate_field (type, i, theclass, 0);
}
/* 4. Then a post merger cleanup is done: */
diff --git a/gdb/testsuite/gdb.base/infcall-nested-structs.exp b/gdb/testsuite/gdb.base/infcall-nested-structs.exp
index 957eb31bdc2..982149f42c1 100644
--- a/gdb/testsuite/gdb.base/infcall-nested-structs.exp
+++ b/gdb/testsuite/gdb.base/infcall-nested-structs.exp
@@ -132,10 +132,6 @@ proc run_tests { lang types } {
continue
}
- if { $lang == "c++" && $name == "struct_02_01"
- && [regexp "^types-(tf-t(c|s|i)|t(c|s|i)-tf)" $types match] } {
- setup_kfail gdb/25096 "x86_64-*-linux*"
- }
gdb_test "p/d check_arg_${name} (ref_val_${name})" "= 1"
set refval [ get_valueof "" "ref_val_${name}" "" ]
@@ -147,10 +143,6 @@ proc run_tests { lang types } {
set answer [ get_valueof "" "rtn_str_${name} ()" "XXXX"]
verbose -log "Answer: ${answer}"
- if { $lang == "c++" && $name == "struct_02_01"
- && [regexp "^types-(tf-t(c|s|i)|t(c|s|i)-tf)" $types match] } {
- setup_kfail gdb/25096 "x86_64-*-linux*"
- }
gdb_assert [string eq ${answer} ${refval}] ${test}
} else {
unresolved $test