gdb/testsuite: Work around clang fails in gdb.base/watchpoint.exp
Checks
Context |
Check |
Description |
linaro-tcwg-bot/tcwg_gdb_check--master-arm |
success
|
Testing passed
|
linaro-tcwg-bot/tcwg_gdb_build--master-aarch64 |
success
|
Testing passed
|
linaro-tcwg-bot/tcwg_gdb_build--master-arm |
success
|
Testing passed
|
linaro-tcwg-bot/tcwg_gdb_check--master-aarch64 |
success
|
Testing passed
|
Commit Message
as mentioned in commit def86538a46f7ce6fbb215cfb184e23015b5d538, clang
doesn't use the CFA information for variable locations, making it so
software watchpoints get false hits when exiting a function. Differently
to that commit, however, gdb.base/watchpoint.exp also needs to be
explicitly continued so the inferior will remain in sync with what the
exp file expects, so this commit uses gdb_test_multiple to identify that
situation.
I also chose to keep the test passing in that scenario because the GDB
feature being tested, software watchpoints, is working as expected.
---
gdb/testsuite/gdb.base/watchpoint.exp | 82 +++++++++++++++++++++++++--
1 file changed, 76 insertions(+), 6 deletions(-)
Comments
Guinevere Larsen <blarsen@redhat.com> writes:
> as mentioned in commit def86538a46f7ce6fbb215cfb184e23015b5d538, clang
> doesn't use the CFA information for variable locations, making it so
> software watchpoints get false hits when exiting a function. Differently
> to that commit, however, gdb.base/watchpoint.exp also needs to be
> explicitly continued so the inferior will remain in sync with what the
> exp file expects, so this commit uses gdb_test_multiple to identify that
> situation.
>
> I also chose to keep the test passing in that scenario because the GDB
> feature being tested, software watchpoints, is working as expected.
I'm not sure I 100% agree with this conclusion. The test is placing on
a watchpoint that isn't expected to change ... yet we see the watchpoint
trigger. I'd say the test is failing.
I wonder if we could just set a flag when we see an unexpected stop, and
then xfail rather than pass once the watchpoint goes out of scope?
I think the right fix would be to implement the
gdbarch_stack_frame_destroyed_p method for amd64. Looking in
amd64-tdep.c I noticed we actually already have
amd64_stack_frame_destroyed_p, we just don't register it with the
gdbarch ... I guess there's some history here.
Anyway, the patch below (untested except for watchpoint.exp) seems to
remove the need for any changes to watchpoint.exp, what are your
thoughts?
> ---
> gdb/testsuite/gdb.base/watchpoint.exp | 82 +++++++++++++++++++++++++--
> 1 file changed, 76 insertions(+), 6 deletions(-)
>
> diff --git a/gdb/testsuite/gdb.base/watchpoint.exp b/gdb/testsuite/gdb.base/watchpoint.exp
> index 70864655c6d..5d8d6b9cb31 100644
> --- a/gdb/testsuite/gdb.base/watchpoint.exp
> +++ b/gdb/testsuite/gdb.base/watchpoint.exp
> @@ -30,6 +30,11 @@ if { [gdb_compile "${srcdir}/${subdir}/${srcfile}" "${binfile}" executable {deb
> return -1
> }
>
> +set using_clang 0
> +if {[test_compiler_info "clang-*"]} {
> + set using_clang 1
> +}
You can actually use 'true' and 'false' here, given this is a boolean.
> +
> # True if we're forcing no hardware watchpoints.
> set no_hw 0
>
> @@ -486,6 +491,20 @@ proc test_complex_watchpoint {} {
> }
> fail $test
> }
> + -re -wrap ".*Continuing.*\[Ww\]atchpoint.*" {
> + global no_hw
> + global using_clang
> + # Clang doesn't use the CFA, so software watchpoints get one
> + # false hit here. Detect if we're in that situation and
> + # ignore the false hit. For more info, see:
> + # https://github.com/llvm/llvm-project/issues/64390
> + if {$using_clang == 1 && $no_hw == 1} {
And here (and elsewhere) you can write:
if {$using_clang && $no_hw} {
which better reflects the boolean nature of these flags.
> + send_gdb "cont\n"
> + exp_continue
> + } else {
> + fail $gdb_test_name
> + }
> + }
> }
>
> gdb_continue_to_breakpoint "func2 breakpoint here, second time"
> @@ -501,8 +520,25 @@ proc test_complex_watchpoint {} {
> "trigger1 partially local watch"
> gdb_test "cont" "Continuing.*\[Ww\]atchpoint .*: local_a . ival5.*" \
> "trigger2 partially local watch"
> - gdb_test "cont" "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" \
> - "self-delete partially local watch"
> + gdb_test_multiple "cont" "self-delete partially local watch" {
> + -re -wrap "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" {
> + pass $gdb_test_name
> + }
> + -re -wrap ".*Continuing.*\[Ww\]atchpoint.*" {
> + global no_hw
> + global using_clang
> + # Clang doesn't use the CFA, so software watchpoints get one
> + # false hit here. Detect if we're in that situation and
> + # ignore the false hit. For more info, see:
> + # https://github.com/llvm/llvm-project/issues/64390
> + if {$using_clang == 1 && $no_hw == 1} {
> + send_gdb "cont\n"
> + exp_continue
> + } else {
> + fail $gdb_test_name
> + }
> + }
> + }
>
> # We should be in "func2" again now. Test a watch of a
> # static (non-stack-based) local. Since this has scope
> @@ -535,8 +571,25 @@ proc test_complex_watchpoint {} {
> "set local watch in recursive call"
> gdb_test "cont" "Continuing.*\[Ww\]atchpoint .*: local_x.*New value = 2.*" \
> "trigger local watch in recursive call"
> - gdb_test "cont" "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" \
> - "self-delete local watch in recursive call"
> + gdb_test_multiple "cont" "self-delete local watch in recursive call" {
> + -re -wrap "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" {
> + pass $gdb_test_name
> + }
> + -re -wrap ".*Continuing.*\[Ww\]atchpoint.*" {
> + global no_hw
> + global using_clang
> + # Clang doesn't use the CFA, so software watchpoints get one
> + # false hit here. Detect if we're in that situation and
> + # ignore the false hit. For more info, see:
> + # https://github.com/llvm/llvm-project/issues/64390
> + if {$using_clang == 1 && $no_hw == 1} {
> + send_gdb "cont\n"
> + exp_continue
> + } else {
> + fail $gdb_test_name
> + }
> + }
> + }
> }
>
> # Repeat the preceding test, but this time use "recurser::local_x" as
> @@ -551,8 +604,25 @@ proc test_complex_watchpoint {} {
> "set local watch in recursive call with explicit scope"
> gdb_test "cont" "Continuing.*\[Ww\]atchpoint .*: recurser::local_x.*New value = 2.*" \
> "trigger local watch with explicit scope in recursive call"
> - gdb_test "cont" "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" \
> - "self-delete local watch with explicit scope in recursive call (2)"
> + gdb_test_multiple "cont" "self-delete local watch with explicit scope in recursive call (2)" {
> + -re -wrap "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" {
> + pass $gdb_test_name
> + }
> + -re -wrap ".*Continuing.*\[Ww\]atchpoint.*" {
> + global no_hw
> + global using_clang
> + # Clang doesn't use the CFA, so software watchpoints get one
> + # false hit here. Detect if we're in that situation and
> + # ignore the false hit. For more info, see:
> + # https://github.com/llvm/llvm-project/issues/64390
> + if {$using_clang == 1 && $no_hw == 1} {
> + send_gdb "cont\n"
> + exp_continue
> + } else {
> + fail $gdb_test_name
> + }
> + }
> + }
> }
>
> # Disable everything so we can finish the program at full speed
> --
> 2.41.0
Thanks,
Andrew
---
diff --git a/gdb/amd64-tdep.c b/gdb/amd64-tdep.c
index e6feee677b3..15120d55976 100644
--- a/gdb/amd64-tdep.c
+++ b/gdb/amd64-tdep.c
@@ -2888,7 +2888,7 @@ static const struct frame_base amd64_frame_base =
/* Normal frames, but in a function epilogue. */
-/* Implement the stack_frame_destroyed_p gdbarch method.
+/* Implement core of the stack_frame_destroyed_p gdbarch method.
The epilogue is defined here as the 'ret' instruction, which will
follow any instruction such as 'leave' or 'pop %ebp' that destroys
@@ -2908,6 +2908,33 @@ amd64_stack_frame_destroyed_p (struct gdbarch *gdbarch, CORE_ADDR pc)
return 1;
}
+/* Implement the gdbarch_stack_frame_destroyed_p method.
+
+ This wrapper delegates to amd64_stack_frame_destroyed_p_1 for compilers
+ that we know get the debug information for stack local variables wrong
+ (i.e. Clang) during the function epilogue.
+
+ For other compilers (i.e. not Clang) we trust the debug information. */
+
+static int
+amd64_stack_frame_destroyed_p_1 (struct gdbarch *gdbarch, CORE_ADDR pc)
+{
+ struct compunit_symtab *cust = find_pc_compunit_symtab (pc);
+
+ /* LLVM backend (Clang/Flang) doesn't use CFA based locations for stack
+ local variables. As a consequence, when we enter the function
+ epilogue GDB will calculate the wrong location for stack local
+ variables, and watchpoints will trigger. */
+ if (cust != nullptr
+ && cust->producer () != nullptr
+ && producer_is_llvm (cust->producer ()))
+ return amd64_stack_frame_destroyed_p_1 (gdbarch, pc);
+
+ /* For other producers, trust the debug information. */
+ return 0;
+}
+
+
static int
amd64_epilogue_frame_sniffer_1 (const struct frame_unwind *self,
frame_info_ptr this_frame,
@@ -2938,7 +2965,7 @@ amd64_epilogue_frame_sniffer_1 (const struct frame_unwind *self,
}
/* Check whether we're in an epilogue. */
- return amd64_stack_frame_destroyed_p (gdbarch, pc);
+ return amd64_stack_frame_destroyed_p_1 (gdbarch, pc);
}
static int
@@ -3310,6 +3337,8 @@ amd64_init_abi (struct gdbarch_info info, struct gdbarch *gdbarch,
set_gdbarch_gen_return_address (gdbarch, amd64_gen_return_address);
+ set_gdbarch_stack_frame_destroyed_p (gdbarch, amd64_stack_frame_destroyed_p);
+
/* SystemTap variables and functions. */
set_gdbarch_stap_integer_prefixes (gdbarch, stap_integer_prefixes);
set_gdbarch_stap_register_prefixes (gdbarch, stap_register_prefixes);
@@ -30,6 +30,11 @@ if { [gdb_compile "${srcdir}/${subdir}/${srcfile}" "${binfile}" executable {deb
return -1
}
+set using_clang 0
+if {[test_compiler_info "clang-*"]} {
+ set using_clang 1
+}
+
# True if we're forcing no hardware watchpoints.
set no_hw 0
@@ -486,6 +491,20 @@ proc test_complex_watchpoint {} {
}
fail $test
}
+ -re -wrap ".*Continuing.*\[Ww\]atchpoint.*" {
+ global no_hw
+ global using_clang
+ # Clang doesn't use the CFA, so software watchpoints get one
+ # false hit here. Detect if we're in that situation and
+ # ignore the false hit. For more info, see:
+ # https://github.com/llvm/llvm-project/issues/64390
+ if {$using_clang == 1 && $no_hw == 1} {
+ send_gdb "cont\n"
+ exp_continue
+ } else {
+ fail $gdb_test_name
+ }
+ }
}
gdb_continue_to_breakpoint "func2 breakpoint here, second time"
@@ -501,8 +520,25 @@ proc test_complex_watchpoint {} {
"trigger1 partially local watch"
gdb_test "cont" "Continuing.*\[Ww\]atchpoint .*: local_a . ival5.*" \
"trigger2 partially local watch"
- gdb_test "cont" "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" \
- "self-delete partially local watch"
+ gdb_test_multiple "cont" "self-delete partially local watch" {
+ -re -wrap "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" {
+ pass $gdb_test_name
+ }
+ -re -wrap ".*Continuing.*\[Ww\]atchpoint.*" {
+ global no_hw
+ global using_clang
+ # Clang doesn't use the CFA, so software watchpoints get one
+ # false hit here. Detect if we're in that situation and
+ # ignore the false hit. For more info, see:
+ # https://github.com/llvm/llvm-project/issues/64390
+ if {$using_clang == 1 && $no_hw == 1} {
+ send_gdb "cont\n"
+ exp_continue
+ } else {
+ fail $gdb_test_name
+ }
+ }
+ }
# We should be in "func2" again now. Test a watch of a
# static (non-stack-based) local. Since this has scope
@@ -535,8 +571,25 @@ proc test_complex_watchpoint {} {
"set local watch in recursive call"
gdb_test "cont" "Continuing.*\[Ww\]atchpoint .*: local_x.*New value = 2.*" \
"trigger local watch in recursive call"
- gdb_test "cont" "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" \
- "self-delete local watch in recursive call"
+ gdb_test_multiple "cont" "self-delete local watch in recursive call" {
+ -re -wrap "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" {
+ pass $gdb_test_name
+ }
+ -re -wrap ".*Continuing.*\[Ww\]atchpoint.*" {
+ global no_hw
+ global using_clang
+ # Clang doesn't use the CFA, so software watchpoints get one
+ # false hit here. Detect if we're in that situation and
+ # ignore the false hit. For more info, see:
+ # https://github.com/llvm/llvm-project/issues/64390
+ if {$using_clang == 1 && $no_hw == 1} {
+ send_gdb "cont\n"
+ exp_continue
+ } else {
+ fail $gdb_test_name
+ }
+ }
+ }
}
# Repeat the preceding test, but this time use "recurser::local_x" as
@@ -551,8 +604,25 @@ proc test_complex_watchpoint {} {
"set local watch in recursive call with explicit scope"
gdb_test "cont" "Continuing.*\[Ww\]atchpoint .*: recurser::local_x.*New value = 2.*" \
"trigger local watch with explicit scope in recursive call"
- gdb_test "cont" "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" \
- "self-delete local watch with explicit scope in recursive call (2)"
+ gdb_test_multiple "cont" "self-delete local watch with explicit scope in recursive call (2)" {
+ -re -wrap "Continuing.*\[Ww\]atchpoint .* deleted because the program has left the block in.*which its expression is valid.*" {
+ pass $gdb_test_name
+ }
+ -re -wrap ".*Continuing.*\[Ww\]atchpoint.*" {
+ global no_hw
+ global using_clang
+ # Clang doesn't use the CFA, so software watchpoints get one
+ # false hit here. Detect if we're in that situation and
+ # ignore the false hit. For more info, see:
+ # https://github.com/llvm/llvm-project/issues/64390
+ if {$using_clang == 1 && $no_hw == 1} {
+ send_gdb "cont\n"
+ exp_continue
+ } else {
+ fail $gdb_test_name
+ }
+ }
+ }
}
# Disable everything so we can finish the program at full speed