[PATCH] gdb/hip: Prepare for DWARF-assembler-based HIP testcases

[PATCH] gdb/hip: Prepare for DWARF-assembler-based HIP testcases

[Pedro Alves]

Compiling a testcase with debug info (by passing passing "debug" as
gdb_compile option) has the effect of passing -g to the compiler.  -g
in turn enables an LLVM option that is essential for debugging:
'-mllvm -amdgpu-spill-cfi-saved-regs'.

That option affects code generation, and some features of the DWARF
assembler machinery rely on code generated by the compiler without -g
to be the exact same as the code generated with -g.

This patch addresses that by always compiling with the LLVM option
that -g would enable.

Change-Id: I64a938edf97c85d0ea9fc34b37071daa560de7e7
---
 gdb/testsuite/lib/gdb.exp | 32 ++++++++++++++++++++++++++++----
 1 file changed, 28 insertions(+), 4 deletions(-)

diff --git a/gdb/testsuite/lib/gdb.exp b/gdb/testsuite/lib/gdb.exp
index a19e8ba6729..a40c87c6727 100644
--- a/gdb/testsuite/lib/gdb.exp
+++ b/gdb/testsuite/lib/gdb.exp
@@ -6745,11 +6745,35 @@ proc gdb_compile {source dest type options} {
 	lappend new_options "early_flags=-fno-stack-protector"
     }
 
-    # hipcc defaults to -O2, so add -O0 to early flags for the hip language.
-    # If "optimize" is also requested, another -O flag (e.g. -O2) will be added
-    # to the flags, overriding this -O0.
     if {[lsearch -exact $options hip] != -1} {
-	lappend new_options "early_flags=-O0"
+	# The -g option enables an LLVM option that is essential for
+	# debugging (-mllvm -amdgpu-spill-cfi-saved-regs), but that
+	# unfortunately also affects code generation.  We don't want
+	# to hardcode -g here though, because some tests want to test
+	# debugging without debug info, and yet others want to compile
+	# their binary without debug info or with -gline-info-only,
+	# and then use the DWARF assembler to generate their own debug
+	# info.  Some features of the DWARF assembler rely on nodebug
+	# and debug binaries having the exact same generated code, so,
+	# explicitly compile with the LLVM option that -g would
+	# enable, and leave adding -g to the testcase, via the "debug"
+	# option, as normal.
+	#
+	# Disable -Wunused-command-line-argument to avoid warnings
+	# like:
+	#
+	#  clang: warning: argument unused during compilation: \
+	#    '-mllvm=-amdgpu-spill-cfi-saved-regs' [-Wunused-command-line-argument]
+	#
+	# These happen when we're compiling an executable _and_ all
+	# the input files are .o files, like: hipcc foo.o -o foo ...
+	#
+	# HIPCC defaults to -O2, so add -O0 to early flags for the HIP
+	# language.  If "optimize" is also requested, another -O flag
+	# (e.g. -O2) will be added to the flags, overriding this -O0.
+	lappend new_options "early_flags=-O0\
+			     -mllvm=-amdgpu-spill-cfi-saved-regs\
+			     -Wno-unused-command-line-argument"
     }
 
     # Because we link with libraries using their basename, we may need

base-commit: 57d9dfcf39cd3c2897bb76c2dbd7074f065685df
-- 
2.54.0

Re: [PATCH] gdb/hip: Prepare for DWARF-assembler-based HIP testcases

[Lancelot SIX]

On 07/07/2026 12:15, Pedro Alves wrote:
> Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.
> 
> 
> Compiling a testcase with debug info (by passing passing "debug" as
> gdb_compile option) has the effect of passing -g to the compiler.  -g
> in turn enables an LLVM option that is essential for debugging:
> '-mllvm -amdgpu-spill-cfi-saved-regs'.
> 
> That option affects code generation, and some features of the DWARF
> assembler machinery rely on code generated by the compiler without -g
> to be the exact same as the code generated with -g.
> 
> This patch addresses that by always compiling with the LLVM option
> that -g would enable.

Hi Pedro,

That looks good to me.

This change lives in lib/gdb.exp, but really only applies when building 
hip testcases.  If you are happy for this to fall under the amdgpu 
umbrella, then

Approved-by: Lancelot Six <lancelot.six@amd.com> (amdgpu)

Otherwise, you can change this to a Reviewed-by.

Best,
Lancelot.

> 
> Change-Id: I64a938edf97c85d0ea9fc34b37071daa560de7e7
> ---
>   gdb/testsuite/lib/gdb.exp | 32 ++++++++++++++++++++++++++++----
>   1 file changed, 28 insertions(+), 4 deletions(-)
> 
> diff --git a/gdb/testsuite/lib/gdb.exp b/gdb/testsuite/lib/gdb.exp
> index a19e8ba6729..a40c87c6727 100644
> --- a/gdb/testsuite/lib/gdb.exp
> +++ b/gdb/testsuite/lib/gdb.exp
> @@ -6745,11 +6745,35 @@ proc gdb_compile {source dest type options} {
>          lappend new_options "early_flags=-fno-stack-protector"
>       }
> 
> -    # hipcc defaults to -O2, so add -O0 to early flags for the hip language.
> -    # If "optimize" is also requested, another -O flag (e.g. -O2) will be added
> -    # to the flags, overriding this -O0.
>       if {[lsearch -exact $options hip] != -1} {
> -       lappend new_options "early_flags=-O0"
> +       # The -g option enables an LLVM option that is essential for
> +       # debugging (-mllvm -amdgpu-spill-cfi-saved-regs), but that
> +       # unfortunately also affects code generation.  We don't want
> +       # to hardcode -g here though, because some tests want to test
> +       # debugging without debug info, and yet others want to compile
> +       # their binary without debug info or with -gline-info-only,
> +       # and then use the DWARF assembler to generate their own debug
> +       # info.  Some features of the DWARF assembler rely on nodebug
> +       # and debug binaries having the exact same generated code, so,
> +       # explicitly compile with the LLVM option that -g would
> +       # enable, and leave adding -g to the testcase, via the "debug"
> +       # option, as normal.
> +       #
> +       # Disable -Wunused-command-line-argument to avoid warnings
> +       # like:
> +       #
> +       #  clang: warning: argument unused during compilation: \
> +       #    '-mllvm=-amdgpu-spill-cfi-saved-regs' [-Wunused-command-line-argument]
> +       #
> +       # These happen when we're compiling an executable _and_ all
> +       # the input files are .o files, like: hipcc foo.o -o foo ...
> +       #
> +       # HIPCC defaults to -O2, so add -O0 to early flags for the HIP
> +       # language.  If "optimize" is also requested, another -O flag
> +       # (e.g. -O2) will be added to the flags, overriding this -O0.
> +       lappend new_options "early_flags=-O0\
> +                            -mllvm=-amdgpu-spill-cfi-saved-regs\
> +                            -Wno-unused-command-line-argument"
>       }
> 
>       # Because we link with libraries using their basename, we may need
> 
> base-commit: 57d9dfcf39cd3c2897bb76c2dbd7074f065685df
> --
> 2.54.0
> 

Re: [PATCH] gdb/hip: Prepare for DWARF-assembler-based HIP testcases

[Pedro Alves]

On 2026-07-08 18:53, Lancelot SIX wrote:

> Hi Pedro,
> 
> That looks good to me.
> 
> This change lives in lib/gdb.exp, but really only applies when building hip testcases.  If you are happy for this to fall under the amdgpu umbrella, then
> 
> Approved-by: Lancelot Six <lancelot.six@amd.com> (amdgpu)

Yes, I think this is the right one.  I merged it now, thanks!

Pedro Alves