On Thu, Apr 14, 2011 at 5:48 PM, H.J. Lu <hjl.to...@gmail.com> wrote: > On Thu, Apr 14, 2011 at 8:26 AM, Michael Matz <m...@suse.de> wrote: >> Hi, >> >> On Thu, 14 Apr 2011, Richard Guenther wrote: >> >>> > + if (align > DECL_ALIGN (decl)) >>> > + DECL_ALIGN (decl) = align; >>> >>> Shouldn't this unconditionally set DECL_ALIGN in case >>> LOCAL_DECL_ALINGMENT returns something smaller? >> >> Decreasing alignment of DECLs points to a problem elsewhere, so perhaps an >> assert that this doesn't happen is better. Decreasing is a problem >> because it's not conservative: there might have been code generated >> already assuming the once larger alignment that then possibly breaks if it >> turns out the alignment is actually smaller. > > ia32 may decrease local variable alignment: > > /* Don't do dynamic stack realignment for long long objects with > -mpreferred-stack-boundary=2. */ > if (!TARGET_64BIT > && align == 64 > && ix86_preferred_stack_boundary < 64 > && (mode == DImode || (type && TYPE_MODE (type) == DImode)) > && (!type || !TYPE_USER_ALIGN (type)) > && (!decl || !DECL_USER_ALIGN (decl))) > align = 32; > > I am running bootstrap/test on Linux/x86-64 and Linux/ia32.
That's broken. It may cause explicit alignment checks to be bogously optimized away. Richard. > -- > H.J. >
