__compiletime_assert_184 in intel_workarounds.c

[nickdesaulniers]

via ci:

ld.lld-11: error: undefined symbol: __compiletime_assert_184
>>> referenced by intel_workarounds.c
>>>               gpu/drm/i915/gt/intel_workarounds.o:(intel_engine_init_workarounds) in archive drivers/built-in.a
>>> referenced by intel_workarounds.c
>>>               gpu/drm/i915/gt/intel_workarounds.o:(intel_engine_init_workarounds) in archive drivers/built-in.a

[nickdesaulniers]

It looks like _MASKED_FIELD is related, the first 2 BCP related asserts are failing.

  263 #define _MASKED_FIELD(mask, value) ({            \                              
  264   if (__builtin_constant_p(mask))            \                                  
  265     BUILD_BUG_ON_MSG(((mask) & 0xffff0000), "Incorrect mask"); \                
  266   if (__builtin_constant_p(value))           \                                  
  267     BUILD_BUG_ON_MSG((value) & 0xffff0000, "Incorrect value"); \                
  268   if (__builtin_constant_p(mask) && __builtin_constant_p(value))     \          
  269     BUILD_BUG_ON_MSG((value) & ~(mask),        \                                
  270          "Incorrect value for mask");      \                                    
  271   __MASKED_FIELD(mask, value); })

via _MASKED_BIT_ENABLE via the call to wa_masked_en(wal, UNSLICE_UNIT_LEVEL_CLKGATE2, VSUNIT_CLKGATE_DIS_TGL); in rcs_engine_wa_init. I haven't tested REG_BIT's BUILD_BUG_ON_ZERO.

[nickdesaulniers]

Seems to be from commit 50148a2 ("drm/i915/tgl: Move and restrict Wa_1408615072").
https://lore.kernel.org/intel-gfx/d5d0609ac6a81f940d91bf93a11506b0af0ce631.camel@intel.com/

[nickdesaulniers]

Not sure this is the precise issue, but with creduce spat out:

a() {
  int b;
  if (__builtin_constant_p(b))
    __compiletime_assert_184();
}

as a difference between GCC and Clang. cc @jyknight @gwelymernans
intel_workarounds.i.orig.txt
test.sh.txt

[nickdesaulniers]

So I think part of why this happens is that clang is inlining wa_masked_en into rcs_engine_wa_init, which gives it further visibility into the constant check. (when I manually inline all callees of wa_masked_en into it, we can still repro. Once I manually inline wa_masked_en into rcs_engine_wa_init, then the differences between compilers disappear). GCC does not make the same inlining decision, so the __builtin_constant_p check fails in GCC, and not in Clang. Thus GCC doesn't run the BUILD_BUG_ON_MSG validation.

So I think there's likely a bug in the source code; it's just that GCC made different inlining decisions, so the __builtin_constant_p check in _MASKED_FIELD (called via _MASKED_BIT_ENABLE via wa_masked_en doesn't run).

I just need to figure out what the value of

(unsigned int)( ((((1UL))) << (19)) + ( (int)( sizeof(struct { int:( -!!(( sizeof(int) == sizeof(      *(8 ? ((void *)((long)(19) * 0l)) : (int *)8))) && ((19) < 0 || (19) > 31))); }))));

is, and why it fails the BUILD_BUG_ON_MSG.

Looks like both compilers agree that it's 0x80000. https://godbolt.org/z/w6ErLH

So the checks: (mask) & 0xffff0000) and (value) & 0xffff0000 (_MASKED_BIT_ENABLE passes the same value twice for mask and value parameters of _MASKED_FIELD) will fail.

[nickdesaulniers]

Author contacted: https://lore.kernel.org/intel-gfx/20200305190811.3009-1-ndesaulniers@google.com/

[nickdesaulniers]

submitted: https://lore.kernel.org/intel-gfx/20200305202435.1284242-1-matthew.d.roper@intel.com/T/#u

[nathanchance]

Fixed: https://git.kernel.org/next/linux-next/c/dbe748cd3af4a7c264a94e3e7c56a084dbd0164d