[PATCH] Print bfloat16 DWARF types correctly

[PATCH] Print bfloat16 DWARF types correctly

[Luis Machado]

Even if the DWARF information contains a bfloat16 base type (__bf16), a
variable of such type will still be printed using the IEEE half float format,
which is wrong.

This patch teaches GDB how to pick the bfloat16 format for __bf16 types in
DWARF (based on the base type name) and uses IEEE half float for all the other
16-bit float formats.

Tested on aarch64-linux/x86_64-linux.

OK?

gdb/ChangeLog:

YYYY-MM-DD  Luis Machado  <luis.machado@linaro.org>

	* arch-utils.c (default_floatformat_for_type): Handle bfloat16.

gdb/testsuite:

YYYY-MM-DD  Luis Machado  <luis.machado@linaro.org>

	* gdb.dwarf2/dw2-bfloat16.exp: New file.
---
 gdb/arch-utils.c                          |  8 ++-
 gdb/testsuite/gdb.dwarf2/dw2-bfloat16.exp | 82 +++++++++++++++++++++++
 2 files changed, 89 insertions(+), 1 deletion(-)
 create mode 100644 gdb/testsuite/gdb.dwarf2/dw2-bfloat16.exp

diff --git a/gdb/arch-utils.c b/gdb/arch-utils.c
index 0017e706ef1..e1d5afd6039 100644
--- a/gdb/arch-utils.c
+++ b/gdb/arch-utils.c
@@ -279,7 +279,13 @@ default_floatformat_for_type (struct gdbarch *gdbarch,
 {
   const struct floatformat **format = NULL;
 
-  if (len == gdbarch_half_bit (gdbarch))
+  /* Check if this is a bfloat16 type.  It has the same size as the
+     IEEE half float type, so we use the base type name to tell them
+     apart.  */
+  if (name != nullptr && strcmp (name, "__bf16") == 0
+      && len == gdbarch_bfloat16_bit (gdbarch))
+    format = gdbarch_bfloat16_format (gdbarch);
+  else if (len == gdbarch_half_bit (gdbarch))
     format = gdbarch_half_format (gdbarch);
   else if (len == gdbarch_float_bit (gdbarch))
     format = gdbarch_float_format (gdbarch);
diff --git a/gdb/testsuite/gdb.dwarf2/dw2-bfloat16.exp b/gdb/testsuite/gdb.dwarf2/dw2-bfloat16.exp
new file mode 100644
index 00000000000..b7c394a3a3d
--- /dev/null
+++ b/gdb/testsuite/gdb.dwarf2/dw2-bfloat16.exp
@@ -0,0 +1,82 @@
+# Copyright 2021 Free Software Foundation, Inc.
+
+# This program is free software; you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation; either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program.  If not, see <http://www.gnu.org/licenses/>.
+#
+# This test checks if GDB can identify bfloat16 and IEEE half float types and
+# print them correctly.
+#
+load_lib dwarf.exp
+
+# This test can only be run on targets which support DWARF-2 and use gas.
+if {![dwarf2_support]} {
+    return 0
+}
+
+standard_testfile main.c -dw.S
+
+# Make some DWARF for the test.
+# 0x4049 is the equivalent of 3.141 for bfloat16.
+# 0x4248 is the equivalent of 3.1406 for IEEE half float.
+set asm_file [standard_output_file $srcfile2]
+Dwarf::assemble $asm_file {
+    global srcdir subdir srcfile
+
+    cu {} {
+	DW_TAG_compile_unit {
+                {DW_AT_language @DW_LANG_C}
+                {DW_AT_name     $srcfile}
+                {DW_AT_comp_dir /tmp}
+        } {
+	    declare_labels bf16_type fp16_type \
+		bf16_var fp16_var
+
+            bf16_type: DW_TAG_base_type {
+                {DW_AT_name __bf16}
+		{encoding @DW_ATE_float}
+                {DW_AT_byte_size 2 DW_FORM_sdata}
+            }
+
+            fp16_type: DW_TAG_base_type {
+                {DW_AT_name __fp16}
+		{encoding @DW_ATE_float}
+                {DW_AT_byte_size 2 DW_FORM_sdata}
+            }
+
+	    bf16_var: DW_TAG_variable {
+		{DW_AT_name "bf16_1"}
+		{DW_AT_type :${bf16_type}}
+		{DW_AT_const_value 0x4049 DW_FORM_sdata}
+	    }
+
+	    fp16_var: DW_TAG_variable {
+		{DW_AT_name "fp16_1"}
+		{DW_AT_type :${fp16_type}}
+		{DW_AT_const_value 0x4248 DW_FORM_sdata}
+	    }
+	}
+    }
+}
+
+if { [prepare_for_testing "failed to prepare" ${testfile} \
+	  [list $srcfile $asm_file] {nodebug}] } {
+    return -1
+}
+
+if ![runto_main] {
+    return -1
+}
+
+# Make sure we can print both types correctly.
+gdb_test "print bf16_1" " = 3.141"
+gdb_test "print fp16_1" " = 3.1406"
-- 
2.25.1

RE: [PATCH] Print bfloat16 DWARF types correctly

[Willgerodt, Felix]

>gdb/ChangeLog:
>
>YYYY-MM-DD  Luis Machado  <luis.machado@linaro.org>
>
>	* arch-utils.c (default_floatformat_for_type): Handle bfloat16.
>
>gdb/testsuite:
>
>YYYY-MM-DD  Luis Machado  <luis.machado@linaro.org>
>
>	* gdb.dwarf2/dw2-bfloat16.exp: New file.


It doesn't interfere with the avx512_bf16 support I added last year, as that is a register view only (no variables). I guess that is why you cc'ed me.

I don't know enough about it to judge if "__bf16" is the only string in use by different compilers. Looks ok to me other than that, but I am not an approver.

Regards,
Felix
Intel Deutschland GmbH
Registered Address: Am Campeon 10, 85579 Neubiberg, Germany
Tel: +49 89 99 8853-0, www.intel.de <http://www.intel.de>
Managing Directors: Christin Eisenschmid, Sharon Heck, Tiffany Doon Silva  
Chairperson of the Supervisory Board: Nicole Lau
Registered Office: Munich
Commercial Register: Amtsgericht Muenchen HRB 186928

Re: [PATCH] Print bfloat16 DWARF types correctly

[Luis Machado]

On 4/14/21 12:45 PM, Willgerodt, Felix wrote:
>> gdb/ChangeLog:
>>
>> YYYY-MM-DD  Luis Machado  <luis.machado@linaro.org>
>>
>> 	* arch-utils.c (default_floatformat_for_type): Handle bfloat16.
>>
>> gdb/testsuite:
>>
>> YYYY-MM-DD  Luis Machado  <luis.machado@linaro.org>
>>
>> 	* gdb.dwarf2/dw2-bfloat16.exp: New file.
> 
> 
> It doesn't interfere with the avx512_bf16 support I added last year, as that is a register view only (no variables). I guess that is why you cc'ed me.

I cc-ed you in case there was anything obviously wrong from x86's side. 
It sounds like there isn't, so that's good. :-)

> 
> I don't know enough about it to judge if "__bf16" is the only string in use by different compilers. Looks ok to me other than that, but I am not an approver.
That's fine. Thanks for checking.

Re: [PATCH] Print bfloat16 DWARF types correctly

[Tom Tromey]

>>>>> "Luis" == Luis Machado via Gdb-patches <gdb-patches@sourceware.org> writes:

Luis> YYYY-MM-DD  Luis Machado  <luis.machado@linaro.org>

Luis> 	* arch-utils.c (default_floatformat_for_type): Handle bfloat16.

It's unfortunate to have to do it this way, but nevertheless I think the
patch is OK.  Thank you.

Tom

Re: [PATCH] Print bfloat16 DWARF types correctly

[Tom Tromey]

>>>>> ">" == Willgerodt, Felix via Gdb-patches <gdb-patches@sourceware.org> writes:

>> I don't know enough about it to judge if "__bf16" is the only string
>> in use by different compilers.

Yeah, though it's easy to add more if we discover them.

Tom

Re: [PATCH] Print bfloat16 DWARF types correctly

[Luis Machado]

On 4/15/21 1:35 PM, Tom Tromey wrote:
>>>>>> "Luis" == Luis Machado via Gdb-patches <gdb-patches@sourceware.org> writes:
> 
> Luis> YYYY-MM-DD  Luis Machado  <luis.machado@linaro.org>
> 
> Luis> 	* arch-utils.c (default_floatformat_for_type): Handle bfloat16.
> 
> It's unfortunate to have to do it this way, but nevertheless I think the
> patch is OK.  Thank you.

It is a bit iffy, given other compilers may name it differently.

I didn't find anything in the DWARF spec that differentiates float types 
through a different encoding field. That would've been the best choice I 
think.

Re: [PATCH] Print bfloat16 DWARF types correctly

[Tom Tromey]

Luis> I didn't find anything in the DWARF spec that differentiates float
Luis> types through a different encoding field. That would've been the best
Luis> choice I think.

Yeah.  We could add an extension of course, assuming the compilers are
changeable.  However then we'd probably still want to support existing
executables.

Tom

Re: [PATCH] Print bfloat16 DWARF types correctly

[Luis Machado]

On 4/15/21 2:10 PM, Tom Tromey wrote:
> Luis> I didn't find anything in the DWARF spec that differentiates float
> Luis> types through a different encoding field. That would've been the best
> Luis> choice I think.
> 
> Yeah.  We could add an extension of course, assuming the compilers are
> changeable.  However then we'd probably still want to support existing
> executables.

I've pushed this now. Thanks.

Re: [PATCH] Print bfloat16 DWARF types correctly

[Luis Machado]

On 4/15/21 2:10 PM, Tom Tromey wrote:
> Luis> I didn't find anything in the DWARF spec that differentiates float
> Luis> types through a different encoding field. That would've been the best
> Luis> choice I think.
> 
> Yeah.  We could add an extension of course, assuming the compilers are
> changeable.  However then we'd probably still want to support existing
> executables.

FTR, I've raised this with the compiler folks and we're discussing how 
this should look like in DWARF. There's discussion going on about a 
possible DWARF extension proposal.