2022-04-12 11:42:18

by Sergei Trofimovich

[permalink] [raw]
Subject: gcc inserts __builtin_popcount, causes 'modpost: "__popcountdi2" ... amdgpu.ko] undefined'

Current linux-5.17.1 on fresh gcc-12 fails to build with errors like:

ERROR: modpost: "__popcountdi2" [drivers/net/ethernet/broadcom/bnx2x/bnx2x.ko] undefined!
ERROR: modpost: "__popcountdi2" [drivers/gpu/drm/amd/amdgpu/amdgpu.ko] undefined!

It is occasionally seen by others on previous gcc versions as well:

https://lkml.org/lkml/2021/7/11/261
https://lkml.org/lkml/2018/10/24/403

'__popcountdi2' are inserted by gcc for code like the following
from 'drivers/gpu/drm/amd/display/modules/hdcp/hdcp1_execution.c':

static inline enum mod_hdcp_status validate_bksv(struct mod_hdcp *hdcp)
{
uint64_t n = 0;
uint8_t count = 0;
u8 bksv[sizeof(n)] = { };

memcpy(bksv, hdcp->auth.msg.hdcp1.bksv, sizeof(hdcp->auth.msg.hdcp1.bksv));
n = *(uint64_t *)bksv;

/* Here gcc inserts 'count = __builtin_popcount(n);' */
while (n) {
count++;
n &= (n - 1);
}

return (count == 20) ? MOD_HDCP_STATUS_SUCCESS :
MOD_HDCP_STATUS_HDCP1_INVALID_BKSV;
}

Note that gcc can insert it regardless of -mno-* options.

How should linux.git handle it? A few options come to mind:

- Perhaps use libgcc.a directly.
- Just implement '__popcountdi2'. Example definition from libgcc:
https://gcc.gnu.org/git/?p=gcc.git;a=blob;f=libgcc/libgcc2.c;hb=HEAD#l846
- Or workaround it with -fno-builtin-popcount in Makefiles.

CCing people who can help routing it and/or deciding on the fix:
[email protected], Joe Perches, [email protected],
Jakub Jelinek, Segher Boessenkool, Thomas Gleixner,Peter Zijlstra, Andy
Lutomirski.

WDYT?

Thanks!

--

Sergei


2022-04-12 12:57:46

by Jakub Jelinek

[permalink] [raw]
Subject: Re: gcc inserts __builtin_popcount, causes 'modpost: "__popcountdi2" ... amdgpu.ko] undefined'

On Mon, Apr 11, 2022 at 10:08:15PM +0100, Sergei Trofimovich wrote:
> Current linux-5.17.1 on fresh gcc-12 fails to build with errors like:
>
> ERROR: modpost: "__popcountdi2" [drivers/net/ethernet/broadcom/bnx2x/bnx2x.ko] undefined!
> ERROR: modpost: "__popcountdi2" [drivers/gpu/drm/amd/amdgpu/amdgpu.ko] undefined!
>
> It is occasionally seen by others on previous gcc versions as well:
>
> https://lkml.org/lkml/2021/7/11/261
> https://lkml.org/lkml/2018/10/24/403
>
> '__popcountdi2' are inserted by gcc for code like the following
> from 'drivers/gpu/drm/amd/display/modules/hdcp/hdcp1_execution.c':
>
> static inline enum mod_hdcp_status validate_bksv(struct mod_hdcp *hdcp)
> {
> uint64_t n = 0;
> uint8_t count = 0;
> u8 bksv[sizeof(n)] = { };
>
> memcpy(bksv, hdcp->auth.msg.hdcp1.bksv, sizeof(hdcp->auth.msg.hdcp1.bksv));
> n = *(uint64_t *)bksv;
>
> /* Here gcc inserts 'count = __builtin_popcount(n);' */
> while (n) {
> count++;
> n &= (n - 1);
> }
>
> return (count == 20) ? MOD_HDCP_STATUS_SUCCESS :
> MOD_HDCP_STATUS_HDCP1_INVALID_BKSV;
> }
>
> Note that gcc can insert it regardless of -mno-* options.

Just FYI, this has been added in GCC 9 for https://gcc.gnu.org/PR82479
If the kernel implements its own __popcount?i2, it can perhaps with
runtime patching use hw instructions when those are available and
fallback to the generic version.

Jakub

2022-04-14 13:32:25

by Sergei Trofimovich

[permalink] [raw]
Subject: Re: gcc inserts __builtin_popcount, causes 'modpost: "__popcountdi2" ... amdgpu.ko] undefined'

On Mon, Apr 11, 2022 at 10:08:15PM +0100, Sergei Trofimovich wrote:
> Current linux-5.17.1 on fresh gcc-12 fails to build with errors like:
>
> ERROR: modpost: "__popcountdi2" [drivers/net/ethernet/broadcom/bnx2x/bnx2x.ko] undefined!
> ERROR: modpost: "__popcountdi2" [drivers/gpu/drm/amd/amdgpu/amdgpu.ko] undefined!
>
> It is occasionally seen by others on previous gcc versions as well:
>
> https://lkml.org/lkml/2021/7/11/261
> https://lkml.org/lkml/2018/10/24/403
>
> '__popcountdi2' are inserted by gcc for code like the following
> from 'drivers/gpu/drm/amd/display/modules/hdcp/hdcp1_execution.c':
>
> static inline enum mod_hdcp_status validate_bksv(struct mod_hdcp *hdcp)
> {
> uint64_t n = 0;
> uint8_t count = 0;
> u8 bksv[sizeof(n)] = { };
>
> memcpy(bksv, hdcp->auth.msg.hdcp1.bksv, sizeof(hdcp->auth.msg.hdcp1.bksv));
> n = *(uint64_t *)bksv;
>
> /* Here gcc inserts 'count = __builtin_popcount(n);' */
> while (n) {
> count++;
> n &= (n - 1);
> }
>
> return (count == 20) ? MOD_HDCP_STATUS_SUCCESS :
> MOD_HDCP_STATUS_HDCP1_INVALID_BKSV;
> }
>
> Note that gcc can insert it regardless of -mno-* options.
>
> How should linux.git handle it? A few options come to mind:
>
> - Perhaps use libgcc.a directly.
> - Just implement '__popcountdi2'. Example definition from libgcc:
> https://gcc.gnu.org/git/?p=gcc.git;a=blob;f=libgcc/libgcc2.c;hb=HEAD#l846
> - Or workaround it with -fno-builtin-popcount in Makefiles.
>
> CCing people who can help routing it and/or deciding on the fix:
> [email protected], Joe Perches, [email protected],
> Jakub Jelinek, Segher Boessenkool, Thomas Gleixner,Peter Zijlstra, Andy
> Lutomirski.

There is now a discussion in gcc bugzilla:

https://gcc.gnu.org/PR105253

--

Sergei