Skip to contentSkip to conversation

usb: avoid -Werror=address-of-packed-member

PATCH New Heinrich Schuchardt 4 participants 10 messages Dec 17, 2019 – Dec 31, 2019

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 usb: avoid -Werror=address-of-packed-member

With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur when passing a member of a packed structure to le16_to_cpus() on a big endian system (e.g. P2041RDB_defconfig).

Replace le16_to_cpus() by get_unaligned_le16(). Check defined(__BIG_ENDIAN) to avoid the introduction of unnecessary instructions on little endian systems as seen on aarch64.

---
 common/usb.c | 14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)
common/usb.c +10 -4
index d9bcb5a57e..de6f7a3bf4 100644
@@ -1078,10 +1078,16 @@ int usb_select_config(struct usb_device *dev)
 		return err;
 
 	/* correct le values */
-	le16_to_cpus(&dev->descriptor.bcdUSB);
-	le16_to_cpus(&dev->descriptor.idVendor);
-	le16_to_cpus(&dev->descriptor.idProduct);
-	le16_to_cpus(&dev->descriptor.bcdDevice);
+#if defined(__BIG_ENDIAN)
+	dev->descriptor.bcdUSB =
+		get_unaligned_le16(&dev->descriptor.bcdUSB);
+	dev->descriptor.idVendor =
+		get_unaligned_le16(&dev->descriptor.idVendor);
+	dev->descriptor.idProduct =
+		get_unaligned_le16(&dev->descriptor.idProduct);
+	dev->descriptor.bcdDevice =
+		get_unaligned_le16(&dev->descriptor.bcdDevice);
+#endif
 
 	/*
 	 * Kingston DT Ultimate 32GB USB 3.0 seems to be extremely sensitive

--
2.24.0

#1 lore raw

PATCH 1/1 usb: avoid -Werror=address-of-packed-member

Quoted 7 lines · Heinrich Schuchardt

With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur when passing a member of a packed structure to le16_to_cpus() on a big endian system (e.g. P2041RDB_defconfig).

Replace le16_to_cpus() by get_unaligned_le16(). Check defined(__BIG_ENDIAN) to avoid the introduction of unnecessary instructions on little endian systems as seen on aarch64.

I would expect the compiler would optimize such stuff out ? Can we do without the ifdef ?

[...]

#2 lore raw

PATCH 1/1 usb: avoid -Werror=address-of-packed-member

Quoted 11 lines · Marek Vasut
Quoted 7 lines · Heinrich Schuchardt

With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur when passing a member of a packed structure to le16_to_cpus() on a big endian system (e.g. P2041RDB_defconfig).

Replace le16_to_cpus() by get_unaligned_le16(). Check defined(__BIG_ENDIAN) to avoid the introduction of unnecessary instructions on little endian systems as seen on aarch64.

I would expect the compiler would optimize such stuff out ? Can we do without the ifdef ?

When compiling qemu_arm64_defconfig without the #ifdef the GCC 9.2.1 adds assembler instructions:

         /* correct le values */
         dev->descriptor.bcdUSB =
   48:   79020660        strh    w0, [x19, #258]
   4c:   39442660        ldrb    w0, [x19, #265]
   50:   2a002020        orr     w0, w1, w0, lsl #8
   54:   39442a61        ldrb    w1, [x19, #266]
                 get_unaligned_le16(&dev->descriptor.bcdUSB);
         dev->descriptor.idVendor =
   58:   79021260        strh    w0, [x19, #264]
   5c:   39442e60        ldrb    w0, [x19, #267]
   60:   2a002020        orr     w0, w1, w0, lsl #8
   64:   39443261        ldrb    w1, [x19, #268]
                 get_unaligned_le16(&dev->descriptor.idVendor);
         dev->descriptor.idProduct =
   68:   79021660        strh    w0, [x19, #266]
   6c:   39443660        ldrb    w0, [x19, #269]
   70:   2a002020        orr     w0, w1, w0, lsl #8
                 get_unaligned_le16(&dev->descriptor.idProduct);
         dev->descriptor.bcdDevice =
   74:   79021a60        strh    w0, [x19, #268]
         udelay(1000 * msec);
   78:   d2807d00        mov     x0, #0x3e8                      // #1000
   7c:   94000000        bl      0 <udelay>
          * one microframe duration here (1mS for USB 1.x , 125uS for
USB 2.0).
          */
         mdelay(1);

Best regards

Heinrich

#3 lore raw

PATCH 1/1 usb: avoid -Werror=address-of-packed-member

Quoted 12 lines · Heinrich Schuchardt
Quoted 11 lines · Marek Vasut
Quoted 9 lines · Heinrich Schuchardt

With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur when
passing a member of a packed structure to le16_to_cpus() on a big endian system (e.g. P2041RDB_defconfig).

Replace le16_to_cpus() by get_unaligned_le16(). Check defined(__BIG_ENDIAN)
to avoid the introduction of unnecessary instructions on little endian systems as seen on aarch64.

Quoted 5 lines
Quoted 2 lines

I would expect the compiler would optimize such stuff out ? Can we do without the ifdef ?

When compiling qemu_arm64_defconfig without the #ifdef the GCC 9.2.1 adds assembler instructions:

Why ?

#4 lore raw

PATCH 1/1 usb: avoid -Werror=address-of-packed-member

Quoted 13 lines · Marek Vasut
Quoted 12 lines · Heinrich Schuchardt
Quoted 11 lines · Marek Vasut
Quoted 9 lines · Heinrich Schuchardt

With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur when
passing a member of a packed structure to le16_to_cpus() on a big endian system (e.g. P2041RDB_defconfig).

Replace le16_to_cpus() by get_unaligned_le16(). Check defined(__BIG_ENDIAN)
to avoid the introduction of unnecessary instructions on little endian systems as seen on aarch64.

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

I would expect the compiler would optimize such stuff out ? Can we do without the ifdef ?

When compiling qemu_arm64_defconfig without the #ifdef the GCC 9.2.1 adds assembler instructions:

Why ?

I am not a GCC developer. I simply observed that GCC currently cannot optimize this away on its own. That is why I added the #ifdef.

Identifying that bit operations like << 8 in __get_unaligned_le16() have zero effect is not an easy task when developing a compiler. You would have to follow the flow of every bit.

Best regards

Heinrich

#5 lore raw

PATCH 1/1 usb: avoid -Werror=address-of-packed-member

Quoted 18 lines · Heinrich Schuchardt
Quoted 17 lines · Marek Vasut
Quoted 16 lines · Heinrich Schuchardt
Quoted 14 lines · Marek Vasut
Quoted 10 lines · Heinrich Schuchardt

With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur when
passing a member of a packed structure to le16_to_cpus() on a big endian
system (e.g. P2041RDB_defconfig).

Replace le16_to_cpus() by get_unaligned_le16(). Check defined(__BIG_ENDIAN)
to avoid the introduction of unnecessary instructions on little endian systems as seen on aarch64.

I would expect the compiler would optimize such stuff out ? Can we do without the ifdef ?

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

When compiling qemu_arm64_defconfig without the #ifdef the GCC 9.2.1 adds assembler instructions:

Why ?

I am not a GCC developer. I simply observed that GCC currently cannot optimize this away on its own. That is why I added the #ifdef.

Are we now adding workarounds instead of solving issues properly?

Quoted 3 lines

Identifying that bit operations like << 8 in __get_unaligned_le16() have zero effect is not an easy task when developing a compiler. You would have to follow the flow of every bit.

Maybe the fix is then to somehow optimize the get_unaligned_le16() to help the compiler ?

#6 lore raw

PATCH 1/1 usb: avoid -Werror=address-of-packed-member

Quoted 28 lines · Marek Vasut
Quoted 18 lines · Heinrich Schuchardt
Quoted 17 lines · Marek Vasut
Quoted 16 lines · Heinrich Schuchardt
Quoted 14 lines · Marek Vasut
Quoted 10 lines · Heinrich Schuchardt

With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur when
passing a member of a packed structure to le16_to_cpus() on a big endian
system (e.g. P2041RDB_defconfig).

Replace le16_to_cpus() by get_unaligned_le16(). Check defined(__BIG_ENDIAN)
to avoid the introduction of unnecessary instructions on little endian systems as seen on aarch64.

I would expect the compiler would optimize such stuff out ? Can we do without the ifdef ?

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

When compiling qemu_arm64_defconfig without the #ifdef the GCC 9.2.1 adds assembler instructions:

Why ?

I am not a GCC developer. I simply observed that GCC currently cannot optimize this away on its own. That is why I added the #ifdef.

Quoted 8 lines

Are we now adding workarounds instead of solving issues properly?

Quoted 3 lines

Identifying that bit operations like << 8 in __get_unaligned_le16() have zero effect is not an easy task when developing a compiler. You would have to follow the flow of every bit.

Maybe the fix is then to somehow optimize the get_unaligned_le16() to help the compiler ?

Inside get_unaligned_le16() it is not known that we will be reassigning to the same memory location. So I cannot imagine what to improve here.

You could invent new functions for in place byte swapping. But that would only move the #ifdef to a different place.

Best regards

Heinrich

#7 lore raw

PATCH 1/1 usb: avoid -Werror=address-of-packed-member

Quoted 36 lines · Heinrich Schuchardt
Quoted 35 lines · Marek Vasut
Quoted 19 lines · Heinrich Schuchardt
Quoted 18 lines · Marek Vasut
Quoted 17 lines · Heinrich Schuchardt
Quoted 15 lines · Marek Vasut
Quoted 11 lines · Heinrich Schuchardt

With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur when
passing a member of a packed structure to le16_to_cpus() on a big endian
system (e.g. P2041RDB_defconfig).

Replace le16_to_cpus() by get_unaligned_le16(). Check defined(__BIG_ENDIAN)
to avoid the introduction of unnecessary instructions on little endian
systems as seen on aarch64.

I would expect the compiler would optimize such stuff out ? Can we do without the ifdef ?

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

When compiling qemu_arm64_defconfig without the #ifdef the GCC 9.2.1 adds assembler instructions:

Why ?

I am not a GCC developer. I simply observed that GCC currently cannot optimize this away on its own. That is why I added the #ifdef.

Are we now adding workarounds instead of solving issues properly?

Quoted 3 lines

Identifying that bit operations like << 8 in __get_unaligned_le16() have zero effect is not an easy task when developing a compiler. You would have to follow the flow of every bit.

Quoted 8 lines
Quoted 2 lines

Maybe the fix is then to somehow optimize the get_unaligned_le16() to help the compiler ?

Inside get_unaligned_le16() it is not known that we will be reassigning to the same memory location. So I cannot imagine what to improve here.

You could invent new functions for in place byte swapping. But that would only move the #ifdef to a different place.

Isn't there already such a function in Linux ?

Also, why would you need the ifdef if the compiler would now know that the operation is noop ?

#8 lore raw

PATCH 1/1 usb: avoid -Werror=address-of-packed-member

Quoted 43 lines · Marek Vasut
Quoted 42 lines · Heinrich Schuchardt
Quoted 29 lines · Marek Vasut
Quoted 19 lines · Heinrich Schuchardt
Quoted 18 lines · Marek Vasut
Quoted 17 lines · Heinrich Schuchardt
Quoted 15 lines · Marek Vasut
Quoted 11 lines · Heinrich Schuchardt

With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur when
passing a member of a packed structure to le16_to_cpus() on a big endian
system (e.g. P2041RDB_defconfig).

Replace le16_to_cpus() by get_unaligned_le16(). Check defined(__BIG_ENDIAN)
to avoid the introduction of unnecessary instructions on little endian
systems as seen on aarch64.

I would expect the compiler would optimize such stuff out ? Can we do without the ifdef ?

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

When compiling qemu_arm64_defconfig without the #ifdef the GCC 9.2.1 adds assembler instructions:

Why ?

I am not a GCC developer. I simply observed that GCC currently cannot optimize this away on its own. That is why I added the #ifdef.

Quoted 8 lines

Are we now adding workarounds instead of solving issues properly?

Quoted 3 lines

Identifying that bit operations like << 8 in __get_unaligned_le16() have zero effect is not an easy task when developing a compiler. You would have to follow the flow of every bit.

Maybe the fix is then to somehow optimize the get_unaligned_le16() to help the compiler ?

Inside get_unaligned_le16() it is not known that we will be reassigning to the same memory location. So I cannot imagine what to improve here.

Quoted 7 lines
Quoted 2 lines

You could invent new functions for in place byte swapping. But that would only move the #ifdef to a different place.

Isn't there already such a function in Linux ?

Also, why would you need the ifdef if the compiler would now know that the operation is noop ?

This particular patch looks to me exactly why we want to follow the Linux Kernel and disable this particular warning for GCC like we already do for LLVM.

Signature · 8 lines
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20191230/02ff3b3b/attachment.sig>
#9 lore raw

PATCH 1/1 usb: avoid -Werror=address-of-packed-member

Quoted 48 lines · Tom Rini
Quoted 46 lines · Marek Vasut
Quoted 44 lines · Heinrich Schuchardt
Quoted 29 lines · Marek Vasut
Quoted 19 lines · Heinrich Schuchardt
Quoted 18 lines · Marek Vasut
Quoted 17 lines · Heinrich Schuchardt
Quoted 15 lines · Marek Vasut
Quoted 11 lines · Heinrich Schuchardt

With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur when
passing a member of a packed structure to le16_to_cpus() on a big endian
system (e.g. P2041RDB_defconfig).

Replace le16_to_cpus() by get_unaligned_le16(). Check defined(__BIG_ENDIAN)
to avoid the introduction of unnecessary instructions on little endian
systems as seen on aarch64.

I would expect the compiler would optimize such stuff out ? Can we do without the ifdef ?

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

When compiling qemu_arm64_defconfig without the #ifdef the GCC 9.2.1 adds assembler instructions:

Why ?

I am not a GCC developer. I simply observed that GCC currently cannot optimize this away on its own. That is why I added the #ifdef.

Quoted 8 lines

Are we now adding workarounds instead of solving issues properly?

Quoted 3 lines

Identifying that bit operations like << 8 in __get_unaligned_le16() have zero effect is not an easy task when developing a compiler. You would have to follow the flow of every bit.

Maybe the fix is then to somehow optimize the get_unaligned_le16() to help the compiler ?

Inside get_unaligned_le16() it is not known that we will be reassigning to the same memory location. So I cannot imagine what to improve here.

You could invent new functions for in place byte swapping. But that would only move the #ifdef to a different place.

Quoted 8 lines
Quoted 4 lines

Isn't there already such a function in Linux ?

Also, why would you need the ifdef if the compiler would now know that the operation is noop ?

This particular patch looks to me exactly why we want to follow the Linux Kernel and disable this particular warning for GCC like we already do for LLVM.

Agree. I think we can follow
Linux commit 6f303d60534c46aa1a239f29c321f95c83dda748

Signature · 2 lines
Best Regards
Masahiro Yamada