Skip to contentSkip to conversation

stmmac: avoid ipq806x constant overflow warning

PATCH Superseded Arnd Bergmann 4 participants 8 messages Nov 12, 2015 – Nov 13, 2015

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 stmmac: avoid ipq806x constant overflow warning

Building dwmac-ipq806x on a 64-bit architecture produces a harmless warning from gcc:

stmmac/dwmac-ipq806x.c: In function 'ipq806x_gmac_probe':
include/linux/bitops.h:6:19: warning: overflow in implicit constant conversion [-Woverflow]
val = QSGMII_PHY_CDR_EN |
stmmac/dwmac-ipq806x.c:333:8: note: in expansion of macro 'QSGMII_PHY_CDR_EN'
#define QSGMII_PHY_CDR_EN BIT(0)
#define BIT(nr) (1UL << (nr))

The compiler warns about the fact that a 64-bit literal is passed into a function that takes a 32-bit argument. I could not fully understand why it warns despite the fact that this number is always small enough to fit, but changing the use of BIT() macros into the equivalent hexadecimal representation avoids the warning

Fixes b1c17215d718 ("stmmac: add ipq806x glue layer")

---
This came up on the arm64 allmodconfig build

drivers/net/ethernet/stmicro/stmmac/dwmac-ipq806x.c +5 -5
index 9d89bdbf029f..4abd9b0b542a 100644
@@ -77,11 +77,11 @@
 /* Only GMAC1/2/3 support SGMII and their CTL register are not contiguous */
 #define QSGMII_PHY_SGMII_CTL(x)			((x == 1) ? 0x134 : \
 						 (0x13c + (4 * (x - 2))))
-#define QSGMII_PHY_CDR_EN			BIT(0)
-#define QSGMII_PHY_RX_FRONT_EN			BIT(1)
-#define QSGMII_PHY_RX_SIGNAL_DETECT_EN		BIT(2)
-#define QSGMII_PHY_TX_DRIVER_EN			BIT(3)
-#define QSGMII_PHY_QSGMII_EN			BIT(7)
+#define QSGMII_PHY_CDR_EN			0x01u
+#define QSGMII_PHY_RX_FRONT_EN			0x02u
+#define QSGMII_PHY_RX_SIGNAL_DETECT_EN		0x04u
+#define QSGMII_PHY_TX_DRIVER_EN			0x08u
+#define QSGMII_PHY_QSGMII_EN			0x80u
 #define QSGMII_PHY_PHASE_LOOP_GAIN_OFFSET	12
 #define QSGMII_PHY_PHASE_LOOP_GAIN_MASK		0x7
 #define QSGMII_PHY_RX_DC_BIAS_OFFSET		18
maintainer
#1 lore raw

From: Arnd Bergmann <[email protected]>
Date: Thu, 12 Nov 2015 15:12:48 +0100

Quoted 10 lines

Building dwmac-ipq806x on a 64-bit architecture produces a harmless warning from gcc:

stmmac/dwmac-ipq806x.c: In function 'ipq806x_gmac_probe':
include/linux/bitops.h:6:19: warning: overflow in implicit constant conversion [-Woverflow]
val = QSGMII_PHY_CDR_EN |
stmmac/dwmac-ipq806x.c:333:8: note: in expansion of macro 'QSGMII_PHY_CDR_EN'
#define QSGMII_PHY_CDR_EN BIT(0)
#define BIT(nr) (1UL << (nr))

Quoted 8 lines

The compiler warns about the fact that a 64-bit literal is passed into a function that takes a 32-bit argument. I could not fully understand why it warns despite the fact that this number is always small enough to fit, but changing the use of BIT() macros into the equivalent hexadecimal representation avoids the warning

Fixes b1c17215d718 ("stmmac: add ipq806x glue layer")

I've seen this warning too on x86_64 and had been meaning to look into it, thanks for taking the initiative. :)

Moving away from using BIT() is somewhat disappointing, because we want to encourage people to use these macros.

I think it's also easier from a driver author and auditing perspective, you can see that something is being defined as bit X and then check the documentation for the chip to see if bit X is correct or not.

With the hex values there is more mental work and room for... mistakes.

Also I don't even understand the compiler's behavior, it's warning about QSGMII_PHY_CDR_EN but if you define only that to "0x1u" it still warns about QSGMII_PHY_CDR_EN.

The warning goes away only if you change all 5 BIT() uses.

This makes me like the change even less, something foul is going on here and I'd rather figure out what that is than install this patch.

Thanks.

#2 lore raw
Quoted 19 lines · On Thursday 12 November 2015 12:25:28 David Miller wrote:

From: Arnd Bergmann <[email protected]>
Date: Thu, 12 Nov 2015 15:12:48 +0100

Quoted 16 lines

Building dwmac-ipq806x on a 64-bit architecture produces a harmless warning from gcc:

stmmac/dwmac-ipq806x.c: In function 'ipq806x_gmac_probe':
include/linux/bitops.h:6:19: warning: overflow in implicit constant conversion [-Woverflow]
val = QSGMII_PHY_CDR_EN |
stmmac/dwmac-ipq806x.c:333:8: note: in expansion of macro 'QSGMII_PHY_CDR_EN'
#define QSGMII_PHY_CDR_EN BIT(0)
#define BIT(nr) (1UL << (nr))

The compiler warns about the fact that a 64-bit literal is passed into a function that takes a 32-bit argument. I could not fully understand why it warns despite the fact that this number is always small enough to fit, but changing the use of BIT() macros into the equivalent hexadecimal representation avoids the warning

Quoted 8 lines
Quoted 2 lines
Fixes b1c17215d718 ("stmmac: add ipq806x glue layer")

I've seen this warning too on x86_64 and had been meaning to look into it, thanks for taking the initiative.

Moving away from using BIT() is somewhat disappointing, because we want to encourage people to use these macros.

Ok, I never really liked that macro, so I didn't mind removing it. ;-)

I spent too much time working at IBM where all internal documentation uses bit numbers that call the MSB bit 0, and drivers use randomly either the IBM notation or the one that everyone else uses, so I always found the hex numbers way more intuitive.

Quoted 5 lines

Also I don't even understand the compiler's behavior, it's warning about QSGMII_PHY_CDR_EN but if you define only that to "0x1u" it still warns about QSGMII_PHY_CDR_EN.

The warning goes away only if you change all 5 BIT() uses.

Yes, I have spent the time to analyze the problem now and it all makes sense. A proper patch follows.

Arnd

#3 lore raw

PATCH v2 1/1 stmmac: avoid ipq806x constant overflow warning

Building dwmac-ipq806x on a 64-bit architecture produces a harmless warning from gcc:

stmmac/dwmac-ipq806x.c: In function 'ipq806x_gmac_probe':
include/linux/bitops.h:6:19: warning: overflow in implicit constant conversion [-Woverflow]
val = QSGMII_PHY_CDR_EN |
stmmac/dwmac-ipq806x.c:333:8: note: in expansion of macro 'QSGMII_PHY_CDR_EN'
#define QSGMII_PHY_CDR_EN BIT(0)
#define BIT(nr) (1UL << (nr))

This is a result of the type conversion rules in C, when we take the logical OR of multiple different types. In particular, we have and unsigned long

	QSGMII_PHY_CDR_EN == BIT(0) == (1ul << 0) == 0x0000000000000001ul

and a signed int

	0xC << QSGMII_PHY_TX_DRV_AMP_OFFSET == 0xc0000000

which together gives a signed long value

	0xffffffffc0000001l

and when this is passed into a function that takes an unsigned int type, gcc warns about the signed overflow and the loss of the upper 32-bits that are all ones.

This patch adds 'ul' type modifiers to the literal numbers passed in here, so now the expression remains an 'unsigned long' with the upper bits all zero, and that avoids the signed overflow and the warning.

Fixes b1c17215d718 ("stmmac: add ipq806x glue layer")
drivers/net/ethernet/stmicro/stmmac/dwmac-ipq806x.c +5 -5
index 9d89bdbf029f..82de68b1a452 100644
@@ -337,11 +337,11 @@ static int ipq806x_gmac_probe(struct platform_device *pdev)
 			     QSGMII_PHY_RX_SIGNAL_DETECT_EN |
 			     QSGMII_PHY_TX_DRIVER_EN |
 			     QSGMII_PHY_QSGMII_EN |
-			     0x4 << QSGMII_PHY_PHASE_LOOP_GAIN_OFFSET |
-			     0x3 << QSGMII_PHY_RX_DC_BIAS_OFFSET |
-			     0x1 << QSGMII_PHY_RX_INPUT_EQU_OFFSET |
-			     0x2 << QSGMII_PHY_CDR_PI_SLEW_OFFSET |
-			     0xC << QSGMII_PHY_TX_DRV_AMP_OFFSET);
+			     0x4ul << QSGMII_PHY_PHASE_LOOP_GAIN_OFFSET |
+			     0x3ul << QSGMII_PHY_RX_DC_BIAS_OFFSET |
+			     0x1ul << QSGMII_PHY_RX_INPUT_EQU_OFFSET |
+			     0x2ul << QSGMII_PHY_CDR_PI_SLEW_OFFSET |
+			     0xCul << QSGMII_PHY_TX_DRV_AMP_OFFSET);
 	}
 
 	plat_dat->has_gmac = true;
maintainer Applied
#4 lore raw

From: Arnd Bergmann <[email protected]>
Date: Thu, 12 Nov 2015 22:03:40 +0100

Quoted 6 lines
drivers/net/ethernet/stmicro/stmmac/dwmac-ipq806x.c +0 -2
@@ -337,11 +337,11 @@ static int ipq806x_gmac_probe(struct platform_device *pdev)
 			     QSGMII_PHY_RX_SIGNAL_DETECT_EN |
 			     QSGMII_PHY_TX_DRIVER_EN |
 			     QSGMII_PHY_QSGMII_EN |
-			     0x4 << QSGMII_PHY_PHASE_LOOP_GAIN_OFFSET |
-			     0x3 << QSGMII_PHY_RX_DC_BIAS_OFFSET |

drivers/net/ethernet/stmicro/stmmac/dwmac-ipq806x.c static int ipq806x_gmac_probe(struct platform_device *pdev) · line 346

Quoted 10 lines
drivers/net/ethernet/stmicro/stmmac/dwmac-ipq806x.c +5 -3
-			     0x1 << QSGMII_PHY_RX_INPUT_EQU_OFFSET |
-			     0x2 << QSGMII_PHY_CDR_PI_SLEW_OFFSET |
-			     0xC << QSGMII_PHY_TX_DRV_AMP_OFFSET);
+			     0x4ul << QSGMII_PHY_PHASE_LOOP_GAIN_OFFSET |
+			     0x3ul << QSGMII_PHY_RX_DC_BIAS_OFFSET |
+			     0x1ul << QSGMII_PHY_RX_INPUT_EQU_OFFSET |
+			     0x2ul << QSGMII_PHY_CDR_PI_SLEW_OFFSET |
+			     0xCul << QSGMII_PHY_TX_DRV_AMP_OFFSET);
 	}
 

Indeed, this looks so much better.

Applied.

Thanks for looking more deeply into this!

#5 lore raw
In reply to Arnd Bergmann
Quoted 24 lines · Arnd Bergmann

Building dwmac-ipq806x on a 64-bit architecture produces a harmless warning from gcc:

stmmac/dwmac-ipq806x.c: In function 'ipq806x_gmac_probe':
include/linux/bitops.h:6:19: warning: overflow in implicit constant conversion [-Woverflow]
val = QSGMII_PHY_CDR_EN |
stmmac/dwmac-ipq806x.c:333:8: note: in expansion of macro 'QSGMII_PHY_CDR_EN'
#define QSGMII_PHY_CDR_EN BIT(0)
#define BIT(nr) (1UL << (nr))

This is a result of the type conversion rules in C, when we take the logical OR of multiple different types. In particular, we have and unsigned long

        QSGMII_PHY_CDR_EN == BIT(0) == (1ul << 0) == 0x0000000000000001ul

and a signed int

        0xC << QSGMII_PHY_TX_DRV_AMP_OFFSET == 0xc0000000

which together gives a signed long value

        0xffffffffc0000001l
Quoted 7 lines

and when this is passed into a function that takes an unsigned int type, gcc warns about the signed overflow and the loss of the upper 32-bits that are all ones.

This patch adds 'ul' type modifiers to the literal numbers passed in here, so now the expression remains an 'unsigned long' with the upper bits all zero, and that avoids the signed overflow and the warning.

FWIW, the 64-bitness of BIT() on 64-bit platforms is also causing subtle warnings in other places, e.g. when inverting them to create bit mask, cfr.
https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=a9efeca613a8fe5281d7c91f5c8c9ea46f2312f6

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert at linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds

#6 lore raw
Quoted 42 lines · Geert Uytterhoeven
Quoted 34 lines · Arnd Bergmann

Building dwmac-ipq806x on a 64-bit architecture produces a harmless warning from gcc:

stmmac/dwmac-ipq806x.c: In function 'ipq806x_gmac_probe':
include/linux/bitops.h:6:19: warning: overflow in implicit constant conversion [-Woverflow]
val = QSGMII_PHY_CDR_EN |
stmmac/dwmac-ipq806x.c:333:8: note: in expansion of macro 'QSGMII_PHY_CDR_EN'
#define QSGMII_PHY_CDR_EN BIT(0)
#define BIT(nr) (1UL << (nr))

This is a result of the type conversion rules in C, when we take the
logical OR of multiple different types. In particular, we have and unsigned long

        QSGMII_PHY_CDR_EN == BIT(0) == (1ul << 0) ==

0x0000000000000001ul

and a signed int

        0xC << QSGMII_PHY_TX_DRV_AMP_OFFSET == 0xc0000000

which together gives a signed long value

        0xffffffffc0000001l

and when this is passed into a function that takes an unsigned int type,
gcc warns about the signed overflow and the loss of the upper 32 -bits that
are all ones.

Quoted 5 lines

This patch adds 'ul' type modifiers to the literal numbers passed in
here, so now the expression remains an 'unsigned long' with the upper
bits all zero, and that avoids the signed overflow and the warning.

Quoted 8 lines

FWIW, the 64-bitness of BIT() on 64-bit platforms is also causing subtle
warnings in other places, e.g. when inverting them to create bit mask, cfr.
https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commi
t/?id=a9efeca613a8fe5281d7c91f5c8c9ea46f2312f6

Gr{oetje,eeting}s,

I still think specific length BIT macros
can be useful.

https://lkml.org/lkml/2015/10/16/852

#7 lore raw

Hi Joe,

Quoted 50 lines · Joe Perches
Quoted 42 lines · Geert Uytterhoeven
Quoted 34 lines · Arnd Bergmann

Building dwmac-ipq806x on a 64-bit architecture produces a harmless warning from gcc:

stmmac/dwmac-ipq806x.c: In function 'ipq806x_gmac_probe':
include/linux/bitops.h:6:19: warning: overflow in implicit constant conversion [-Woverflow]
val = QSGMII_PHY_CDR_EN |
stmmac/dwmac-ipq806x.c:333:8: note: in expansion of macro 'QSGMII_PHY_CDR_EN'
#define QSGMII_PHY_CDR_EN BIT(0)
#define BIT(nr) (1UL << (nr))

This is a result of the type conversion rules in C, when we take the
logical OR of multiple different types. In particular, we have and unsigned long

        QSGMII_PHY_CDR_EN == BIT(0) == (1ul << 0) ==

0x0000000000000001ul

and a signed int

        0xC << QSGMII_PHY_TX_DRV_AMP_OFFSET == 0xc0000000

which together gives a signed long value

        0xffffffffc0000001l

and when this is passed into a function that takes an unsigned int type,
gcc warns about the signed overflow and the loss of the upper 32 -bits that
are all ones.

Quoted 5 lines

This patch adds 'ul' type modifiers to the literal numbers passed in
here, so now the expression remains an 'unsigned long' with the upper
bits all zero, and that avoids the signed overflow and the warning.

Quoted 6 lines

FWIW, the 64-bitness of BIT() on 64-bit platforms is also causing subtle
warnings in other places, e.g. when inverting them to create bit mask, cfr.
https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commi
t/?id=a9efeca613a8fe5281d7c91f5c8c9ea46f2312f6

Quoted 4 lines

I still think specific length BIT macros
can be useful.

https://lkml.org/lkml/2015/10/16/852

Yeah!

I only recently started liking the BIT() macro (before I preferred hex, too). Perhaps because Renesas datasheets use bit numbers all over the place ;-)

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert at linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds