The previous code was brittle and missed cases such as the mipn32
variants which for some reason has the 64 bit syscalls. This leads to
a number of binary targets having deps lines like:

  all.clang-sanitizer/mipsn32el-linux-user/linux-user/signal.d
  140:  /home/alex/lsrc/qemu.git/linux-user/mips64/syscall_nr.h \
  455:/home/alex/lsrc/qemu.git/linux-user/mips64/syscall_nr.h:

  all.clang-sanitizer/mipsn32el-linux-user/linux-user/syscall.d
  146:  /home/alex/lsrc/qemu.git/linux-user/mips64/syscall_nr.h \
  485:/home/alex/lsrc/qemu.git/linux-user/mips64/syscall_nr.h:

which in turn would trigger the re-generation of syscall_nr.h in the
source tree (thanks to generic %/syscall_nr.h rules). The previous
code attempts to clean it out but misses edge cases but fails.

After spending a day trying to understand how this was happening I'm
unconvinced that there are not other such breakages possible with this
"caching". As we add more auto-generated code to the build it is likely
to trip up again. Apply a hammer to the problem.

Fixes: 91e5998f18 (which fixes 5f29856b852d and 4d6a835dea47)
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
---
 configure | 15 ++++++---------
 1 file changed, 6 insertions(+), 9 deletions(-)

diff --git a/configure b/configure
index d2d93ae51e47..81ef123fb187 100755
--- a/configure
+++ b/configure
@@ -1950,23 +1950,20 @@ EOF
 exit 0
 fi
 
-# Remove old dependency files to make sure that they get properly regenerated
-rm -f */config-devices.mak.d
-
 # Remove syscall_nr.h to be sure they will be regenerated in the build
 # directory, not in the source directory
 for arch in alpha hppa m68k xtensa sh4 microblaze arm ppc s390x sparc sparc64 \
     i386 x86_64 mips mips64 ; do
     # remove the file if it has been generated in the source directory
     rm -f "${source_path}/linux-user/${arch}/syscall_nr.h"
-    # remove the dependency files
-    for target in ${arch}*-linux-user ; do
-        test -d "${target}" && find "${target}" -type f -name "*.d" \
-             -exec grep -q "${source_path}/linux-user/${arch}/syscall_nr.h" {} \; \
-             -print | while read file ; do rm "${file}" "${file%.d}.o" ; done
-    done
 done
 
+# Clean out all old dependency files. As more files are generated we
+# run the risk of old dependencies triggering generation in the wrong
+# places. Previous brittle attempts to be surgical tend to miss edge
+# cases leading to wasted time and much confusion.
+find -type f -name "*.d" -exec rm -f {} \;
+
 if test -z "$python"
 then
     error_exit "Python not found. Use --python=/path/to/python"
-- 
2.20.1

On 7/8/20 10:53 AM, Alex Bennée wrote:
> The previous code was brittle and missed cases such as the mipn32
> variants which for some reason has the 64 bit syscalls. This leads to
> a number of binary targets having deps lines like:
> 
>   all.clang-sanitizer/mipsn32el-linux-user/linux-user/signal.d
>   140:  /home/alex/lsrc/qemu.git/linux-user/mips64/syscall_nr.h \
>   455:/home/alex/lsrc/qemu.git/linux-user/mips64/syscall_nr.h:
> 
>   all.clang-sanitizer/mipsn32el-linux-user/linux-user/syscall.d
>   146:  /home/alex/lsrc/qemu.git/linux-user/mips64/syscall_nr.h \
>   485:/home/alex/lsrc/qemu.git/linux-user/mips64/syscall_nr.h:
> 
> which in turn would trigger the re-generation of syscall_nr.h in the
> source tree (thanks to generic %/syscall_nr.h rules). The previous
> code attempts to clean it out but misses edge cases but fails.
> 
> After spending a day trying to understand how this was happening I'm
> unconvinced that there are not other such breakages possible with this
> "caching". As we add more auto-generated code to the build it is likely
> to trip up again. Apply a hammer to the problem.
> 
> Fixes: 91e5998f18 (which fixes 5f29856b852d and 4d6a835dea47)
> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
> ---
>  configure | 15 ++++++---------
>  1 file changed, 6 insertions(+), 9 deletions(-)

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>

r~