Link failures on arm32 all{yes,mod}config after e7a96b0b7d1669e95aa77c007e71e3c88c531cc4

[nathanchance]

So this is an interesting one...

After commit e7a96b0b7d16 ("pinctrl: aspeed: Add multiple pin group support for functions") in -next, arm all{yes,mod}config builds started failing at link time with a lot of errors like so:

arm-linux-gnueabi-ld: drivers/pinctrl/aspeed/pinctrl-aspeed-g6.o:(.data+0x1088): multiple definition of `sig_exprs_205_GPIOZ5'; drivers/pinctrl/aspeed/pinctrl-aspeed-g4.o:(.data+0x11d4): first defined here
arm-linux-gnueabi-ld: drivers/pinctrl/aspeed/pinctrl-aspeed-g6.o:(.data+0x10a0): multiple definition of `sig_exprs_206_GPIOZ6'; drivers/pinctrl/aspeed/pinctrl-aspeed-g4.o:(.data+0x11f8): first defined here
arm-linux-gnueabi-ld: drivers/pinctrl/aspeed/pinctrl-aspeed-g6.o:(.data+0x10b8): multiple definition of `sig_exprs_207_GPIOZ7'; drivers/pinctrl/aspeed/pinctrl-aspeed-g4.o:(.data+0x121c): first defined here 

This can be easily reproduced on next-20190806 and this patch (the KCFLAGS= is needed for an unrelated issue that I will try to deal with shortly):

$ make -s -j$(nproc) ARCH=arm CC=clang CROSS_COMPILE=arm-linux-gnueabi- KCFLAGS="-Wno-error=incompatible-pointer-types" O=out distclean allmodconfig vmlinux

This issue is not present with GCC. Dumping the symbol tables shows that GCC keeps the symbols local whereas Clang makes them global.

$ rg sig_exprs_20 ${TMP}
/tmp/tmp.39vwqYENfC/gcc-pinctrl-aspeed-g6.syms
2362:00001354 d sig_exprs_200_GPIOZ0
2363:0000135c d sig_exprs_200_SPI1CS1
2364:00001334 d sig_exprs_201_GPIOZ1
2365:0000133c d sig_exprs_201_SPI1ABR
2366:00001314 d sig_exprs_202_GPIOZ2
2367:0000131c d sig_exprs_202_SPI1WP
2368:000012f4 d sig_exprs_203_GPIOZ3
2369:000012fc d sig_exprs_203_SPI1CK
2370:000012d8 d sig_exprs_204_GPIOZ4
2371:000012e0 d sig_exprs_204_SPI1MOSI
2372:000012bc d sig_exprs_205_GPIOZ5
2373:000012c4 d sig_exprs_205_SPI1MISO
2374:00001298 d sig_exprs_206_GPIOZ6
2375:000012a8 d sig_exprs_206_SPI1DQ2
2376:000012a0 d sig_exprs_206_TXD13
2377:00001270 d sig_exprs_207_GPIOZ7
2378:00001278 d sig_exprs_207_RXD13
2379:00001280 d sig_exprs_207_SPI1DQ3
2380:00001238 d sig_exprs_208_GPIO18A0
2381:00001248 d sig_exprs_208_RGMII1TXCK
2382:00001240 d sig_exprs_208_RMII1RCLKO
2383:00001210 d sig_exprs_209_GPIO18A1
2384:00001220 d sig_exprs_209_RGMII1TXCTL
2385:00001218 d sig_exprs_209_RMII1TXEN
2386:00002e90 d sig_exprs_20_GPIOC4
2387:00002e98 d sig_exprs_20_RGMII3TXD2

/tmp/tmp.39vwqYENfC/clang-pinctrl-aspeed-g6.syms
2362:00001038 D sig_exprs_200_GPIOZ0
2363:00001030 D sig_exprs_200_SPI1CS1
2364:00001048 D sig_exprs_201_GPIOZ1
2365:00001040 D sig_exprs_201_SPI1ABR
2366:00001058 D sig_exprs_202_GPIOZ2
2367:00001050 D sig_exprs_202_SPI1WP
2368:00001068 D sig_exprs_203_GPIOZ3
2369:00001060 D sig_exprs_203_SPI1CK
2370:00001078 D sig_exprs_204_GPIOZ4
2371:00001070 D sig_exprs_204_SPI1MOSI
2372:00001088 D sig_exprs_205_GPIOZ5
2373:00001080 D sig_exprs_205_SPI1MISO
2374:000010a0 D sig_exprs_206_GPIOZ6
2375:00001090 D sig_exprs_206_SPI1DQ2
2376:00001098 D sig_exprs_206_TXD13
2377:000010b8 D sig_exprs_207_GPIOZ7
2378:000010b0 D sig_exprs_207_RXD13
2379:000010a8 D sig_exprs_207_SPI1DQ3
2380:000010d0 D sig_exprs_208_GPIO18A0
2381:000010c0 D sig_exprs_208_RGMII1TXCK
2382:000010c8 D sig_exprs_208_RMII1RCLKO
2383:000010e8 D sig_exprs_209_GPIO18A1
2384:000010d8 D sig_exprs_209_RGMII1TXCTL
2385:000010e0 D sig_exprs_209_RMII1TXEN
2386:000001e8 D sig_exprs_20_GPIOC4
2387:000001e0 D sig_exprs_20_RGMII3TXD2

That commit is a handful to take in but it looks like this is the problematic section according to creduce:

/**
 * Create an expression symbol alias from (signal, group) to (pin, signal).
 *
 * @pin: The pin number
 * @sig: The signal name
 * @group: The name of the group of which the pin is a member that is
 *         associated with the function's signal
 *
 * Using an alias in this way enables detection of copy/paste errors (defining
 * the signal for a group multiple times) whilst enabling multiple pin groups
 * to exist for a signal without intrusive side-effects on defining the list of
 * signals available on a pin.
 */
#define SIG_EXPR_LIST_ALIAS(pin, sig, group) \
	static const struct aspeed_sig_expr *\
		SIG_EXPR_LIST_SYM(pin, sig)[ARRAY_SIZE(SIG_EXPR_LIST_SYM(sig, group))] \
		__attribute__((alias(istringify(SIG_EXPR_LIST_SYM(sig, group)))))

which spits out:

$ cat test.i
a, b = &a;
static c[0] __attribute__((alias("b")));
d = c;

$ arm-linux-gnueabi-gcc -O2 -c -o test.o test.i
...

$ arm-linux-gnueabi-nm test.o
00000004 C a
00000004 D b
00000004 d c
00000000 D d

$ clang -O2 -no-integrated-as --target=arm-linux-gnueabi --prefix=/home/nathan/cbl/usr/bin/ --gcc-toolchain=/home/nathan/cbl/usr -c -o test.o test.i
...

$ arm-linux-gnueabi-nm test.o
00000004 C a
00000000 D b
00000000 D c
00000004 D d

Interestingness test and full files available here.

[nickdesaulniers]

So it looks like static and __attribute__((alias("")) is busted in Clang; I swear I've heard this issue before.

[nathanchance]

https://llvm.org/pr42377

[ihalip]

I think this happens because in clang all aliases are created with ExternalLinkage. Changing this line to use InternalLinkage instead, I can see that the output symbol is local, as expected.

It's obvious to me some checks need to be added to select the right linkage type.

[nickdesaulniers]

https://reviews.llvm.org/D66492

[nathanchance]

That patch fixes this issue and (tentatively, haven't done a full run) doesn't introduce any immediately apparent regressions, thanks!

[nickdesaulniers]

https://reviews.llvm.org/rL369705

[nickdesaulniers]

@arndb is asking about this in clang-9 in https://llvm.org/pr43243
cc @zmodem

[zmodem]

I've merged the fix to release_90 and replied on the bug. Thanks!