Kbuild: enable -Wmaybe-uninitialized warning for "make W=1"
authorArnd Bergmann <arnd@arndb.de>
Thu, 10 Nov 2016 16:44:44 +0000 (17:44 +0100)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 11 Nov 2016 16:45:08 +0000 (08:45 -0800)
Traditionally, we have always had warnings about uninitialized variables
enabled, as this is part of -Wall, and generally a good idea [1], but it
also always produced false positives, mainly because this is a variation
of the halting problem and provably impossible to get right in all cases
[2].

Various people have identified cases that are particularly bad for false
positives, and in commit e74fc973b6e5 ("Turn off -Wmaybe-uninitialized
when building with -Os"), I turned off the warning for any build that
was done with CC_OPTIMIZE_FOR_SIZE.  This drastically reduced the number
of false positive warnings in the default build but unfortunately had
the side effect of turning the warning off completely in 'allmodconfig'
builds, which in turn led to a lot of warnings (both actual bugs, and
remaining false positives) to go in unnoticed.

With commit 877417e6ffb9 ("Kbuild: change CC_OPTIMIZE_FOR_SIZE
definition") enabled the warning again for allmodconfig builds in v4.7
and in v4.8-rc1, I had finally managed to address all warnings I get in
an ARM allmodconfig build and most other maybe-uninitialized warnings
for ARM randconfig builds.

However, commit 6e8d666e9253 ("Disable "maybe-uninitialized" warning
globally") was merged at the same time and disabled it completely for
all configurations, because of false-positive warnings on x86 that I had
not addressed until then.  This caused a lot of actual bugs to get
merged into mainline, and I sent several dozen patches for these during
the v4.9 development cycle.  Most of these are actual bugs, some are for
correct code that is safe because it is only called under external
constraints that make it impossible to run into the case that gcc sees,
and in a few cases gcc is just stupid and finds something that can
obviously never happen.

I have now done a few thousand randconfig builds on x86 and collected
all patches that I needed to address every single warning I got (I can
provide the combined patch for the other warnings if anyone is
interested), so I hope we can get the warning back and let people catch
the actual bugs earlier.

This reverts the change to disable the warning completely and for now
brings it back at the "make W=1" level, so we can get it merged into
mainline without introducing false positives.  A follow-up patch enables
it on all levels unless some configuration option turns it off because
of false-positives.

Link: https://rusty.ozlabs.org/?p=232
Link: https://gcc.gnu.org/wiki/Better_Uninitialized_Warnings
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Makefile
arch/arc/Makefile
scripts/Makefile.extrawarn
scripts/Makefile.ubsan

index f97f786de58dd8497db3599279d9e2c3d40e121b..06e2b73978e8fe596907d15858e8bf7d2aa6783a 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -370,7 +370,7 @@ LDFLAGS_MODULE  =
 CFLAGS_KERNEL  =
 AFLAGS_KERNEL  =
 LDFLAGS_vmlinux =
-CFLAGS_GCOV    = -fprofile-arcs -ftest-coverage -fno-tree-loop-im
+CFLAGS_GCOV    = -fprofile-arcs -ftest-coverage -fno-tree-loop-im -Wno-maybe-uninitialized
 CFLAGS_KCOV    := $(call cc-option,-fsanitize-coverage=trace-pc,)
 
 
@@ -620,7 +620,6 @@ ARCH_CFLAGS :=
 include arch/$(SRCARCH)/Makefile
 
 KBUILD_CFLAGS  += $(call cc-option,-fno-delete-null-pointer-checks,)
-KBUILD_CFLAGS  += $(call cc-disable-warning,maybe-uninitialized,)
 KBUILD_CFLAGS  += $(call cc-disable-warning,frame-address,)
 
 ifdef CONFIG_LD_DEAD_CODE_DATA_ELIMINATION
@@ -629,15 +628,18 @@ KBUILD_CFLAGS     += $(call cc-option,-fdata-sections,)
 endif
 
 ifdef CONFIG_CC_OPTIMIZE_FOR_SIZE
-KBUILD_CFLAGS  += -Os
+KBUILD_CFLAGS  += -Os $(call cc-disable-warning,maybe-uninitialized,)
 else
 ifdef CONFIG_PROFILE_ALL_BRANCHES
-KBUILD_CFLAGS  += -O2
+KBUILD_CFLAGS  += -O2 $(call cc-disable-warning,maybe-uninitialized,)
 else
 KBUILD_CFLAGS   += -O2
 endif
 endif
 
+KBUILD_CFLAGS += $(call cc-ifversion, -lt, 0409, \
+                       $(call cc-disable-warning,maybe-uninitialized,))
+
 # Tell gcc to never replace conditional load with a non-conditional one
 KBUILD_CFLAGS  += $(call cc-option,--param=allow-store-data-races=0)
 
index 864adad522803cfdec31983d062514d7c50f19c3..25f81a1db9f933b853e968de8b274a9abe4d770f 100644 (file)
@@ -68,7 +68,9 @@ cflags-$(CONFIG_ARC_DW2_UNWIND)               += -fasynchronous-unwind-tables $(cfi)
 ifndef CONFIG_CC_OPTIMIZE_FOR_SIZE
 # Generic build system uses -O2, we want -O3
 # Note: No need to add to cflags-y as that happens anyways
-ARCH_CFLAGS += -O3
+#
+# Disable the false maybe-uninitialized warings gcc spits out at -O3
+ARCH_CFLAGS += -O3 $(call cc-disable-warning,maybe-uninitialized,)
 endif
 
 # small data is default for elf32 tool-chain. If not usable, disable it
index 53449a6ff6aa7de3c5c868f0645f6000b6f551e6..7fc2c5a3dbbe8a4012bf4468514b9d2af6a7d313 100644 (file)
@@ -36,6 +36,7 @@ warning-2 += -Wshadow
 warning-2 += $(call cc-option, -Wlogical-op)
 warning-2 += $(call cc-option, -Wmissing-field-initializers)
 warning-2 += $(call cc-option, -Wsign-compare)
+warning-2 += $(call cc-option, -Wmaybe-uninitialized)
 
 warning-3 := -Wbad-function-cast
 warning-3 += -Wcast-qual
@@ -59,6 +60,8 @@ endif
 KBUILD_CFLAGS += $(warning)
 else
 
+KBUILD_CFLAGS += $(call cc-disable-warning, maybe-uninitialized)
+
 ifeq ($(cc-name),clang)
 KBUILD_CFLAGS += $(call cc-disable-warning, initializer-overrides)
 KBUILD_CFLAGS += $(call cc-disable-warning, unused-value)
index dd779c40c8e6af713c0ad7039bc929b0c6c2944a..3b1b13818d594f9ffa9dcb15206314ab349325a7 100644 (file)
@@ -17,4 +17,8 @@ endif
 ifdef CONFIG_UBSAN_NULL
       CFLAGS_UBSAN += $(call cc-option, -fsanitize=null)
 endif
+
+      # -fsanitize=* options makes GCC less smart than usual and
+      # increase number of 'maybe-uninitialized false-positives
+      CFLAGS_UBSAN += $(call cc-option, -Wno-maybe-uninitialized)
 endif