Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
Patch preapproved. Jakub Hi, Checked into trunk: http://gcc.gnu.org/ml/gcc-cvs/2013-06/msg00646.html Thanks, K
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Jun 19, 2013, at 1:38 AM, Richard Biener wrote: > On Wed, Jun 19, 2013 at 9:22 AM, Jakub Jelinek wrote: >> On Wed, Jun 19, 2013 at 11:12:21AM +0400, Igor Zamyatin wrote: >>> Right, as you did for other cases. It works here as well. >> >> Patch preapproved. > > I wonder how much code breaks these days when we enable -fno-common by > default? Not much. gcc as Apple shipped it, has always been no-common, and indeed the shared library scheme doesn't like common. There are a few test cases that would need -fcommon, but I don't think that is a big deal. Most oss I think is -fno-common friendly. I think gcc should default to c99, and I think c99 mode (and later) could use -fno-common by default. For pre c99 modes, I'd probably just leave it to the dust bin of history.
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Jun 19, 2013, at 1:44 AM, Jakub Jelinek wrote: > On Wed, Jun 19, 2013 at 10:38:47AM +0200, Richard Biener wrote: >> On Wed, Jun 19, 2013 at 9:22 AM, Jakub Jelinek wrote: >>> On Wed, Jun 19, 2013 at 11:12:21AM +0400, Igor Zamyatin wrote: Right, as you did for other cases. It works here as well. >>> >>> Patch preapproved. >> >> I wonder how much code breaks these days when we enable -fno-common by >> default? ... > > Somebody would need to try it ;). Been there done that. That experiment has been running for at least 10 years now… :-)
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Wed, Jun 19, 2013 at 10:38:47AM +0200, Richard Biener wrote: > On Wed, Jun 19, 2013 at 9:22 AM, Jakub Jelinek wrote: > > On Wed, Jun 19, 2013 at 11:12:21AM +0400, Igor Zamyatin wrote: > >> Right, as you did for other cases. It works here as well. > > > > Patch preapproved. > > I wonder how much code breaks these days when we enable -fno-common by > default? ... Somebody would need to try it ;). From vectorization POV, it surely would be better if -fno-common was the default. Jakub
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Wed, Jun 19, 2013 at 9:22 AM, Jakub Jelinek wrote: > On Wed, Jun 19, 2013 at 11:12:21AM +0400, Igor Zamyatin wrote: >> Right, as you did for other cases. It works here as well. > > Patch preapproved. I wonder how much code breaks these days when we enable -fno-common by default? ... Richard.
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Wed, Jun 19, 2013 at 11:12:21AM +0400, Igor Zamyatin wrote: > Right, as you did for other cases. It works here as well. Patch preapproved. Jakub
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
Right, as you did for other cases. It works here as well. Thanks, Igor On Wed, Jun 19, 2013 at 11:05 AM, Jakub Jelinek wrote: > On Wed, Jun 19, 2013 at 11:01:59AM +0400, Igor Zamyatin wrote: >> The change also affects vectorizer in avx case which could be seen for >> gcc.dg/tree-ssa/loop-19.c test. >> >> After the change report says >> >> loop-19_bad.c:16: note: === vect_analyze_data_refs_alignment === >> loop-19_bad.c:16: note: vect_compute_data_ref_alignment: >> loop-19_bad.c:16: note: can't force alignment of ref: a[j_9] >> loop-19_bad.c:16: note: vect_compute_data_ref_alignment: >> loop-19_bad.c:16: note: can't force alignment of ref: c[j_9] >> >> AFAICS first condition in ix86_data_alignment was true before the >> change so 256 was a return value. >> >> Do we need to tweak this test also? > > I'd add -fno-common to the test. > > Jakub
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Thu, Jun 13, 2013 at 11:37 AM, Alan Modra wrote: > Revised patch with offsettable_ok_by_alignment change, avoiding dumb > idea of using statement expressions. This one actually bootstraps and > passes regression testing. > > * config/rs6000/rs6000.h (enum data_align): New. > (LOCAL_ALIGNMENT, DATA_ALIGNMENT): Use rs6000_data_alignment. > (DATA_ABI_ALIGNMENT): Define. > (CONSTANT_ALIGNMENT): Correct comment. > * config/rs6000/rs6000-protos.h (rs6000_data_alignment): Declare. > * config/rs6000/rs6000.c (rs6000_data_alignment): New function. The revised patch, without the DECL_P part is okay. The original code produced the necessary alignment and neither of us can find any code in public packages that increases the alignment for PPC vector types. While there is the possibility that a user could encounter an object file produced by an older GCC with less strict alignment and a version of GCC with this fix would make an incorrect assumption, this does not seem very likely in practice. Thanks, David
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Fri, Jun 14, 2013 at 12:54:40PM +0200, Jakub Jelinek wrote: > On Fri, Jun 14, 2013 at 08:12:02PM +0930, Alan Modra wrote: > > I see your point, but for there to be a real problem we'd need > > a) A library exporting such a type with (supposed) increased > >alignment, and, > > b) gcc would need to make use of the increased alignment. > > > > (a) must be rare or non-existent or you'd think we would have had a > > bug report about lack of user alignment in vector typedefs. The code > > has been like this since 2001-11-07, so users have had a long time to > > discover it. (Of course, this is an argument for just ignoring the > > bug too.) > > It doesn't have to be an exported symbol from a library, it is enough to > compile some objects using one compiler and other objects using another > compiler, then link into the same library. OK. > Try (long) &x & 31; ? That &x & -32 not being optimized into &x > is guess a missed optimization. Huh, trust me to hit another bug. :) > Consider if you put: > typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); > vec_align x = { 0, 0, 0, 0 }; > into one TU and compile with gcc 4.8.1, then > typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); > extern vec_align x; > > long f1 (void) > { > return (long) &x & 31; > } > in another TU and compile with gcc trunk after your patch. I bet > it will be optimized into return 0; by the trunk + your patch compiler, > while the alignment will be actually just 16 byte. Right. Counterpoint is that gcc made exactly the same sort of error across TUs and even in the same TU prior to my change. eg. typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); vec_align x = { 0, 0, 0, 0 }; long f1 (void) { return (long)&x & 31; } int y __attribute__ ((vector_size(16), aligned(32))) = { 0, 0, 0, 0 }; long f2 (void) { return (long)&y & 31; } compiles to .L.f1: li 3,0 blr .. .L.f2: li 3,0 blr .. .globl y .lcomm y,16,32 .type y, @object .globl x .lcomm x,16,16 .type x, @object .ident "GCC: (GNU) 4.7.2 20120921 (Red Hat 4.7.2-2)" My implementation of rs6000_data_alignment therefore doesn't introduce a *new* ABI incompatibility. I question whether it is worth complicating rs6000_data_alignment, especially since your suggestion of using the older buggy alignment in DATA_ABI_ALIGNMENT then increasing in DATA_ALIGNMENT isn't as simple as it sounds. We're not talking about some fixed increase in DATA_ALIGNMENT but what we want is the value of alignment before DATA_ABI_ALIGNMENT. Perhaps that could be retrieved from TYPE_ALIGN (type) and MAX_OFILE_ALIGNMENT, but that would make our DATA_ALIGNMENT the only target to need such tricks. -- Alan Modra Australia Development Lab, IBM
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Fri, Jun 14, 2013 at 08:12:02PM +0930, Alan Modra wrote: > > As for the > > typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); > > > > vec_align x = { 0, 0, 0, 0 }; > > > > changes, that is ABI changing bugfix, so the question is, are you fine with > > breaking the ABI (between 4.8 and 4.9, or if you wanted to backport it to > > 4.8 too (I certainly plan to backport the non-ppc DATA_ABI_ALIGNMENT changes > > to 4.8.2, already am using it in our compilers))? The other option is > > to fix the ABI, but keep things backwards ABI compatible. That would be > > done by decreasing the alignment as it used to do before in > > DATA_ABI_ALIGNMENT, > > and increasing it to the desirable level only in DATA_ALIGNMENT. That has > > the effect that when emitting the decls into assembly e.g. the above will > > now be correctly 32 byte aligned, but accesses to such decl in compiler > > generated code will only assume that alignment if > > decl_binds_to_current_def_p, otherwise they will keep assuming the old > > (broken) lowered alignment. At least for 4.8 backport IMHO that would be a > > better idea (but of course would need big comment explaning it). > > I see your point, but for there to be a real problem we'd need > a) A library exporting such a type with (supposed) increased >alignment, and, > b) gcc would need to make use of the increased alignment. > > (a) must be rare or non-existent or you'd think we would have had a > bug report about lack of user alignment in vector typedefs. The code > has been like this since 2001-11-07, so users have had a long time to > discover it. (Of course, this is an argument for just ignoring the > bug too.) It doesn't have to be an exported symbol from a library, it is enough to compile some objects using one compiler and other objects using another compiler, then link into the same library. > (b) doesn't happen in the rs6000 backend as far as I'm aware. Do you > know whether there is some optimisation based on alignment in generic > parts of gcc? A quick test like Tons of them, the DECL_ALIGN value is used say by get_pointer_alignment, vectorizer assumptions, is added to MEM_ATTRS, so anything looking at alignment in RTL can optimize too. > typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); > vec_align x = { 0, 0, 0, 0 }; > > long f1 (void) > { > return (long) &x & -32; > } Try (long) &x & 31; ? That &x & -32 not being optimized into &x is guess a missed optimization. Consider if you put: typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); vec_align x = { 0, 0, 0, 0 }; into one TU and compile with gcc 4.8.1, then typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); extern vec_align x; long f1 (void) { return (long) &x & 31; } in another TU and compile with gcc trunk after your patch. I bet it will be optimized into return 0; by the trunk + your patch compiler, while the alignment will be actually just 16 byte. Jakub
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Fri, Jun 14, 2013 at 10:59:52AM +0200, Jakub Jelinek wrote: > On Fri, Jun 14, 2013 at 08:18:19AM +0930, Alan Modra wrote: > > It is handling !DECL_P trees, which must be local. I know I saw > > STRING_CST here when I wrote offsettable_ok_by_alignment, hence the > > use of CONSTANT_ALIGNMENT. I'm not so sure about the need for > > DATA_ALIGNMENT now, but if it was correct before then we ought to > > be using both DATA_ABI_ALIGNMENT and DATA_ALIGNMENT after your > > changes. > > Yeah, then it makes sense. Sorry for not looking up earlier that this is > the !DECL_P case. Your comment prodded me into looking at whether the !DECL_P code is needed in 4.9, and it looks like we never see !DECL_P trees.. Bootstrap and regression tests all langs on powerpc64 didn't hit a gcc_unreachable() I put there, both with -fsection-anchors and -fno-section-anchors. David, please consider that piece of the patch retracted. > As for the > typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); > > vec_align x = { 0, 0, 0, 0 }; > > changes, that is ABI changing bugfix, so the question is, are you fine with > breaking the ABI (between 4.8 and 4.9, or if you wanted to backport it to > 4.8 too (I certainly plan to backport the non-ppc DATA_ABI_ALIGNMENT changes > to 4.8.2, already am using it in our compilers))? The other option is > to fix the ABI, but keep things backwards ABI compatible. That would be > done by decreasing the alignment as it used to do before in > DATA_ABI_ALIGNMENT, > and increasing it to the desirable level only in DATA_ALIGNMENT. That has > the effect that when emitting the decls into assembly e.g. the above will > now be correctly 32 byte aligned, but accesses to such decl in compiler > generated code will only assume that alignment if > decl_binds_to_current_def_p, otherwise they will keep assuming the old > (broken) lowered alignment. At least for 4.8 backport IMHO that would be a > better idea (but of course would need big comment explaning it). I see your point, but for there to be a real problem we'd need a) A library exporting such a type with (supposed) increased alignment, and, b) gcc would need to make use of the increased alignment. (a) must be rare or non-existent or you'd think we would have had a bug report about lack of user alignment in vector typedefs. The code has been like this since 2001-11-07, so users have had a long time to discover it. (Of course, this is an argument for just ignoring the bug too.) (b) doesn't happen in the rs6000 backend as far as I'm aware. Do you know whether there is some optimisation based on alignment in generic parts of gcc? A quick test like typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); vec_align x = { 0, 0, 0, 0 }; long f1 (void) { return (long) &x & -32; } static int y __attribute__ ((aligned(32))); long f2 (void) { return (long) &y & -32; } shows the "& -32" in both functions isn't optimised away. -- Alan Modra Australia Development Lab, IBM
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Fri, Jun 14, 2013 at 08:18:19AM +0930, Alan Modra wrote: > On Thu, Jun 13, 2013 at 05:42:17PM +0200, Jakub Jelinek wrote: > > On Fri, Jun 14, 2013 at 01:07:01AM +0930, Alan Modra wrote: > > > @@ -5774,10 +5818,11 @@ offsettable_ok_by_alignment (rtx op, HOST_WIDE_INT > > >type = TREE_TYPE (decl); > > > > > >dalign = TYPE_ALIGN (type); > > > + dalign = DATA_ABI_ALIGNMENT (type, dalign); > > >if (CONSTANT_CLASS_P (decl)) > > > dalign = CONSTANT_ALIGNMENT (decl, dalign); > > >else > > > - dalign = DATA_ALIGNMENT (decl, dalign); > > > + dalign = DATA_ALIGNMENT (type, dalign); > > > > > >if (dsize == 0) > > > { > > > > What is this code trying to do? Shouldn't it just use DECL_ALIGN > > which should be set to the right value from get_variable_alignment? > > I mean, if !decl_binds_to_current_def_p (decl), then using DATA_ALIGNMENT > > or CONSTANT_ALIGNMENT (for anything but actually emitting the var into > > object, or just as an optimization hint that very likely the decl will be > > aligned enough, but not guaranteed), which are optimization, is wrong > > (an ABI problem). > > It is handling !DECL_P trees, which must be local. I know I saw > STRING_CST here when I wrote offsettable_ok_by_alignment, hence the > use of CONSTANT_ALIGNMENT. I'm not so sure about the need for > DATA_ALIGNMENT now, but if it was correct before then we ought to > be using both DATA_ABI_ALIGNMENT and DATA_ALIGNMENT after your > changes. Yeah, then it makes sense. Sorry for not looking up earlier that this is the !DECL_P case. As for the typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); vec_align x = { 0, 0, 0, 0 }; changes, that is ABI changing bugfix, so the question is, are you fine with breaking the ABI (between 4.8 and 4.9, or if you wanted to backport it to 4.8 too (I certainly plan to backport the non-ppc DATA_ABI_ALIGNMENT changes to 4.8.2, already am using it in our compilers))? The other option is to fix the ABI, but keep things backwards ABI compatible. That would be done by decreasing the alignment as it used to do before in DATA_ABI_ALIGNMENT, and increasing it to the desirable level only in DATA_ALIGNMENT. That has the effect that when emitting the decls into assembly e.g. the above will now be correctly 32 byte aligned, but accesses to such decl in compiler generated code will only assume that alignment if decl_binds_to_current_def_p, otherwise they will keep assuming the old (broken) lowered alignment. At least for 4.8 backport IMHO that would be a better idea (but of course would need big comment explaning it). Jakub
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Thu, Jun 13, 2013 at 05:42:17PM +0200, Jakub Jelinek wrote: > On Fri, Jun 14, 2013 at 01:07:01AM +0930, Alan Modra wrote: > > @@ -5774,10 +5818,11 @@ offsettable_ok_by_alignment (rtx op, HOST_WIDE_INT > >type = TREE_TYPE (decl); > > > >dalign = TYPE_ALIGN (type); > > + dalign = DATA_ABI_ALIGNMENT (type, dalign); > >if (CONSTANT_CLASS_P (decl)) > > dalign = CONSTANT_ALIGNMENT (decl, dalign); > >else > > - dalign = DATA_ALIGNMENT (decl, dalign); > > + dalign = DATA_ALIGNMENT (type, dalign); > > > >if (dsize == 0) > > { > > What is this code trying to do? Shouldn't it just use DECL_ALIGN > which should be set to the right value from get_variable_alignment? > I mean, if !decl_binds_to_current_def_p (decl), then using DATA_ALIGNMENT > or CONSTANT_ALIGNMENT (for anything but actually emitting the var into > object, or just as an optimization hint that very likely the decl will be > aligned enough, but not guaranteed), which are optimization, is wrong > (an ABI problem). It is handling !DECL_P trees, which must be local. I know I saw STRING_CST here when I wrote offsettable_ok_by_alignment, hence the use of CONSTANT_ALIGNMENT. I'm not so sure about the need for DATA_ALIGNMENT now, but if it was correct before then we ought to be using both DATA_ABI_ALIGNMENT and DATA_ALIGNMENT after your changes. -- Alan Modra Australia Development Lab, IBM
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Fri, Jun 14, 2013 at 01:07:01AM +0930, Alan Modra wrote: > @@ -5774,10 +5818,11 @@ offsettable_ok_by_alignment (rtx op, HOST_WIDE_INT >type = TREE_TYPE (decl); > >dalign = TYPE_ALIGN (type); > + dalign = DATA_ABI_ALIGNMENT (type, dalign); >if (CONSTANT_CLASS_P (decl)) > dalign = CONSTANT_ALIGNMENT (decl, dalign); >else > - dalign = DATA_ALIGNMENT (decl, dalign); > + dalign = DATA_ALIGNMENT (type, dalign); > >if (dsize == 0) > { What is this code trying to do? Shouldn't it just use DECL_ALIGN which should be set to the right value from get_variable_alignment? I mean, if !decl_binds_to_current_def_p (decl), then using DATA_ALIGNMENT or CONSTANT_ALIGNMENT (for anything but actually emitting the var into object, or just as an optimization hint that very likely the decl will be aligned enough, but not guaranteed), which are optimization, is wrong (an ABI problem). Jakub
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Thu, Jun 13, 2013 at 05:10:51PM +0930, Alan Modra wrote: > On Wed, Jun 12, 2013 at 12:52:03PM -0500, Edmar Wienskoski wrote: > > The e500v2 (SPE) hardware is such that if the address of vector (double > > world > > load / stores) are not double world aligned the instruction will trap. > > > > So this alignment is not optional. > > Vector type alignment is also specified by the ppc64 abi. I think we > want the following. Note that DATA_ALIGNMENT has been broken for > vectors right from the initial vector support (and the error was > copied for e500 double). For example > > typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); > vec_align x = { 0, 0, 0, 0 }; > > currently loses the extra alignment. Fixed by never decreasing > alignment in DATA_ABI_ALIGNMENT. Testing in progress. OK to > apply assuming bootstrap is good? (I think I need a change in > offsettable_ok_by_alignment too. I'll do that in a separate patch.) Revised patch with offsettable_ok_by_alignment change, avoiding dumb idea of using statement expressions. This one actually bootstraps and passes regression testing. * config/rs6000/rs6000.h (enum data_align): New. (LOCAL_ALIGNMENT, DATA_ALIGNMENT): Use rs6000_data_alignment. (DATA_ABI_ALIGNMENT): Define. (CONSTANT_ALIGNMENT): Correct comment. * config/rs6000/rs6000-protos.h (rs6000_data_alignment): Declare. * config/rs6000/rs6000.c (rs6000_data_alignment): New function. (offsettable_ok_by_alignment): Align by DATA_ABI_ALIGNMENT. Pass "type" not "decl" to DATA_ALIGNMENT. Index: gcc/config/rs6000/rs6000.h === --- gcc/config/rs6000/rs6000.h (revision 200055) +++ gcc/config/rs6000/rs6000.h (working copy) @@ -813,12 +813,6 @@ extern unsigned rs6000_pointer_size; /* No data type wants to be aligned rounder than this. */ #define BIGGEST_ALIGNMENT 128 -/* A C expression to compute the alignment for a variables in the - local store. TYPE is the data type, and ALIGN is the alignment - that the object would ordinarily have. */ -#define LOCAL_ALIGNMENT(TYPE, ALIGN) \ - DATA_ALIGNMENT (TYPE, ALIGN) - /* Alignment of field after `int : 0' in a structure. */ #define EMPTY_FIELD_BOUNDARY 32 @@ -828,8 +822,15 @@ extern unsigned rs6000_pointer_size; /* A bit-field declared as `int' forces `int' alignment for the struct. */ #define PCC_BITFIELD_TYPE_MATTERS 1 -/* Make strings word-aligned so strcpy from constants will be faster. - Make vector constants quadword aligned. */ +enum data_align { align_abi, align_opt, align_both }; + +/* A C expression to compute the alignment for a variables in the + local store. TYPE is the data type, and ALIGN is the alignment + that the object would ordinarily have. */ +#define LOCAL_ALIGNMENT(TYPE, ALIGN) \ + rs6000_data_alignment (TYPE, ALIGN, align_both) + +/* Make strings word-aligned so strcpy from constants will be faster. */ #define CONSTANT_ALIGNMENT(EXP, ALIGN) \ (TREE_CODE (EXP) == STRING_CST\ && (STRICT_ALIGNMENT || !optimize_size) \ @@ -837,21 +838,14 @@ extern unsigned rs6000_pointer_size; ? BITS_PER_WORD \ : (ALIGN)) -/* Make arrays of chars word-aligned for the same reasons. - Align vectors to 128 bits. Align SPE vectors and E500 v2 doubles to +/* Make arrays of chars word-aligned for the same reasons. */ +#define DATA_ALIGNMENT(TYPE, ALIGN) \ + rs6000_data_alignment (TYPE, ALIGN, align_opt) + +/* Align vectors to 128 bits. Align SPE vectors and E500 v2 doubles to 64 bits. */ -#define DATA_ALIGNMENT(TYPE, ALIGN)\ - (TREE_CODE (TYPE) == VECTOR_TYPE \ - ? (((TARGET_SPE && SPE_VECTOR_MODE (TYPE_MODE (TYPE))) \ - || (TARGET_PAIRED_FLOAT && PAIRED_VECTOR_MODE (TYPE_MODE (TYPE \ - ? 64 : 128) \ - : ((TARGET_E500_DOUBLE \ - && TREE_CODE (TYPE) == REAL_TYPE \ - && TYPE_MODE (TYPE) == DFmode) \ - ? 64 \ - : (TREE_CODE (TYPE) == ARRAY_TYPE \ -&& TYPE_MODE (TREE_TYPE (TYPE)) == QImode \ -&& (ALIGN) < BITS_PER_WORD) ? BITS_PER_WORD : (ALIGN))) +#define DATA_ABI_ALIGNMENT(TYPE, ALIGN) \ + rs6000_data_alignment (TYPE, ALIGN, align_abi) /* Nonzero if move instructions will actually fail to work when given unaligned data. */ Index: gcc/config/rs6000/rs6000-protos.h === --- gcc/config/rs
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Wed, Jun 12, 2013 at 12:52:03PM -0500, Edmar Wienskoski wrote: > The e500v2 (SPE) hardware is such that if the address of vector (double world > load / stores) are not double world aligned the instruction will trap. > > So this alignment is not optional. Vector type alignment is also specified by the ppc64 abi. I think we want the following. Note that DATA_ALIGNMENT has been broken for vectors right from the initial vector support (and the error was copied for e500 double). For example typedef int vec_align __attribute__ ((vector_size(16), aligned(32))); vec_align x = { 0, 0, 0, 0 }; currently loses the extra alignment. Fixed by never decreasing alignment in DATA_ABI_ALIGNMENT. Testing in progress. OK to apply assuming bootstrap is good? (I think I need a change in offsettable_ok_by_alignment too. I'll do that in a separate patch.) * config/rs6000/rs6000.h (DATA_ABI_ALIGNMENT): Define. (DATA_ALIGNMENT): Remove alignment already covered by above. (LOCAL_ALIGNMENT): Use both DATA_ABI_ALIGNMENT and DATA_ALIGNMENT. Index: gcc/config/rs6000/rs6000.h === --- gcc/config/rs6000/rs6000.h (revision 200055) +++ gcc/config/rs6000/rs6000.h (working copy) @@ -817,7 +817,8 @@ extern unsigned rs6000_pointer_size; local store. TYPE is the data type, and ALIGN is the alignment that the object would ordinarily have. */ #define LOCAL_ALIGNMENT(TYPE, ALIGN) \ - DATA_ALIGNMENT (TYPE, ALIGN) + ({unsigned int _align = DATA_ABI_ALIGNMENT (TYPE, ALIGN);\ +DATA_ALIGNMENT (TYPE, _align); }) /* Alignment of field after `int : 0' in a structure. */ #define EMPTY_FIELD_BOUNDARY 32 @@ -837,21 +838,26 @@ extern unsigned rs6000_pointer_size; ? BITS_PER_WORD \ : (ALIGN)) -/* Make arrays of chars word-aligned for the same reasons. - Align vectors to 128 bits. Align SPE vectors and E500 v2 doubles to +/* Make arrays of chars word-aligned for the same reasons. */ +#define DATA_ALIGNMENT(TYPE, ALIGN)\ + ((TREE_CODE (TYPE) == ARRAY_TYPE \ +&& TYPE_MODE (TREE_TYPE (TYPE)) == QImode \ +&& (ALIGN) < BITS_PER_WORD) ? BITS_PER_WORD : (ALIGN)) + +/* Align vectors to 128 bits. Align SPE vectors and E500 v2 doubles to 64 bits. */ -#define DATA_ALIGNMENT(TYPE, ALIGN)\ +#define DATA_ABI_ALIGNMENT(TYPE, ALIGN) \ (TREE_CODE (TYPE) == VECTOR_TYPE \ ? (((TARGET_SPE && SPE_VECTOR_MODE (TYPE_MODE (TYPE))) \ || (TARGET_PAIRED_FLOAT && PAIRED_VECTOR_MODE (TYPE_MODE (TYPE \ - ? 64 : 128) \ - : ((TARGET_E500_DOUBLE \ - && TREE_CODE (TYPE) == REAL_TYPE \ - && TYPE_MODE (TYPE) == DFmode) \ - ? 64 \ - : (TREE_CODE (TYPE) == ARRAY_TYPE \ -&& TYPE_MODE (TREE_TYPE (TYPE)) == QImode \ -&& (ALIGN) < BITS_PER_WORD) ? BITS_PER_WORD : (ALIGN))) + ? ((ALIGN) < 64 ? 64 : (ALIGN)) \ + : ((ALIGN) < 128 ? 128 : (ALIGN))) \ + : (TARGET_E500_DOUBLE \ + && TREE_CODE (TYPE) == REAL_TYPE \ + && TYPE_MODE (TYPE) == DFmode\ + && (ALIGN) < 64) \ + ? 64 \ + : (ALIGN)) /* Nonzero if move instructions will actually fail to work when given unaligned data. */ -- Alan Modra Australia Development Lab, IBM
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
The e500v2 (SPE) hardware is such that if the address of vector (double world load / stores) are not double world aligned the instruction will trap. So this alignment is not optional. Edmar On Fri, Jun 7, 2013 at 3:43 PM, Richard Henderson wrote: > On 06/07/2013 12:25 PM, Jakub Jelinek wrote: >> This PR is about DATA_ALIGNMENT macro increasing alignment of some decls >> for optimization purposes beyond ABI mandated levels. It is fine to emit >> the vars aligned as much as we want for optimization purposes, but if we >> can't be sure that references to that decl bind to the definition we >> increased the alignment on (e.g. common variables, or -fpic code without >> hidden visibility, weak vars etc.), we can't assume that alignment. > > When the linker merges common blocks, it chooses both maximum size and maximum > alignment. Thus for any common block for which we can prove the block must > reside in the module (any executable, or hidden common in shared object), we > can go ahead and use the increased alignment. > > It's only in shared objects with non-hidden common blocks that we have a > problem, since in that case we may resolve the common block via copy reloc to > a > memory block in another module. > > So while decl_binds_to_current_def_p is a good starting point, I think we can > do a little better with common blocks. Which ought to take care of those > vectorization regressions you mention. > >> @@ -966,8 +966,12 @@ align_variable (tree decl, bool dont_out >>align = MAX_OFILE_ALIGNMENT; >> } >> >> - /* On some machines, it is good to increase alignment sometimes. */ >> - if (! DECL_USER_ALIGN (decl)) >> + /* On some machines, it is good to increase alignment sometimes. >> + But as DECL_ALIGN is used both for actually emitting the variable >> + and for code accessing the variable as guaranteed alignment, we >> + can only increase the alignment if it is a performance optimization >> + if the references to it must bind to the current definition. */ >> + if (! DECL_USER_ALIGN (decl) && decl_binds_to_current_def_p (decl)) >> { >> #ifdef DATA_ALIGNMENT >>unsigned int data_align = DATA_ALIGNMENT (TREE_TYPE (decl), align); >> @@ -988,12 +992,69 @@ align_variable (tree decl, bool dont_out >> } >> #endif >> } >> +#ifdef DATA_ABI_ALIGNMENT >> + else if (! DECL_USER_ALIGN (decl)) >> +{ >> + unsigned int data_align = DATA_ABI_ALIGNMENT (TREE_TYPE (decl), >> align); >> + /* For backwards compatibility, don't assume the ABI alignment for >> + TLS variables. */ >> + if (! DECL_THREAD_LOCAL_P (decl) || data_align <= BITS_PER_WORD) >> + align = data_align; >> +} >> +#endif > > This structure would seem to do the wrong thing if DATA_ABI_ALIGNMENT is > defined, but DATA_ALIGNMENT isn't. And while I realize you documented it, I > don't like the restriction that D_A /must/ return something larger than D_A_A. > All that means is that in complex cases D_A will have to call D_A_A itself. > > I would think that it would be better to rearrange as > > if (!D_U_A) > { > #ifdef D_A_A > align = ... > #endif > #ifdef D_A > if (d_b_t_c_d_p) > align = ... > #endif > } > > Why the special case for TLS? If we really want that special case surely that > test should go into D_A_A itself, and not here in generic code. > >> Bootstrapped/regtested on x86_64-linux and i686-linux. No idea about other >> targets, I've kept them all using DATA_ALIGNMENT, which is considered >> optimization increase only now, if there is some ABI mandated alignment >> increase on other targets, that should be done in DATA_ABI_ALIGNMENT as >> well as DATA_ALIGNMENT. > > I've had a brief look over the instances of D_A within the tree atm. Most of > them carry the cut-n-paste comment "for the same reasons". These I believe > never intended an ABI change, and were really only interested in optimization. > > But these I think require a good hard look to see if they really intended an > ABI alignment: > > c6x comment explicitly mentions abi > criscompiler options for alignment -- systemwide or local? > mmixcomment mentions GETA instruction > s390comment mentions LARL instruction > rs6000 SPE and E500 portion of the alignment non-optional? > > Relevant port maintainers CCed. > > > r~
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
Thanks!
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Mon, Jun 10, 2013 at 08:44:05PM -0400, DJ Delorie wrote: > > > @@ -986,12 +1053,10 @@ align_variable (tree decl, bool dont_out > > if (! DECL_THREAD_LOCAL_P (decl) || const_align <= BITS_PER_WORD) > > align = const_align; > > } > > -#endif > > } > > +#endif > > I think this change in get_variable_align() is wrong; it results in > unbalanced braces inside an #ifdef, if the #ifdef body is not included > (i.e. CONSTANT_ALIGNMENT not defined), the compile fails... You are right, fixed thusly, committed as obvious: 2013-06-11 Jakub Jelinek PR target/56564 * varasm.c (get_variable_align): Move #endif to the right place. --- gcc/varasm.c(revision 199933) +++ gcc/varasm.c(revision 199934) @@ -1053,8 +1053,8 @@ get_variable_align (tree decl) if (! DECL_THREAD_LOCAL_P (decl) || const_align <= BITS_PER_WORD) align = const_align; } -} #endif +} return align; } Jakub
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
> @@ -986,12 +1053,10 @@ align_variable (tree decl, bool dont_out > if (! DECL_THREAD_LOCAL_P (decl) || const_align <= BITS_PER_WORD) > align = const_align; > } > -#endif > } > +#endif I think this change in get_variable_align() is wrong; it results in unbalanced braces inside an #ifdef, if the #ifdef body is not included (i.e. CONSTANT_ALIGNMENT not defined), the compile fails...
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Mon, Jun 10, 2013 at 11:44 AM, Jakub Jelinek wrote: > On Mon, Jun 10, 2013 at 07:51:54AM -0700, Richard Henderson wrote: >> On 06/07/2013 02:14 PM, Jakub Jelinek wrote: >> >> > When the linker merges common blocks, it chooses both maximum size and >> >> > maximum >> >> > alignment. Thus for any common block for which we can prove the block >> >> > must >> >> > reside in the module (any executable, or hidden common in shared >> >> > object), we >> >> > can go ahead and use the increased alignment. >> > But consider say: >> > one TU: >> > struct S { char buf[15]; } s __attribute__((aligned (32))); >> > >> > another TU: >> > char c = 7; >> > struct S { char buf[15]; } s = { { 1, 2 } }; >> > char d = 8; >> > int main () { return 0; } >> > >> > (the aligned(32) is there just to simulate the DATA_ALIGNMENT optimization >> > increase). Linker warns about this (thus the question is if we want to >> > increase the alignment for optimization on commons at all) and doesn't >> > align >> > it. >> > >> >> Oh, right. I hadn't considered commons unifying with non-common variables, >> and the failure of commoning in that case. I'd mostly been thinking of >> uninitialized Fortran-like common blocks, where it is more common for the >> blocks to have nothing in common but the name. > > Ok, here is what I've committed to trunk (will wait for some time before > backporting). As discussed with Honza on IRC, decl_binds_to_current_def_p > will need further fixing, it does the wrong thing for extern int var > __attribute__((visibility ("hidden"))) or hidden DECL_COMMON symbols. > > And, we don't have any feedback about SPE/E500 rs6000 - DATA_ALIGNMENT vs. > DATA_ABI_ALIGNMENT yet. SPE/e500 support was written by Aldy. He or someone from Freescale needs to comment. - David
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Mon, Jun 10, 2013 at 07:51:54AM -0700, Richard Henderson wrote: > On 06/07/2013 02:14 PM, Jakub Jelinek wrote: > >> > When the linker merges common blocks, it chooses both maximum size and > >> > maximum > >> > alignment. Thus for any common block for which we can prove the block > >> > must > >> > reside in the module (any executable, or hidden common in shared > >> > object), we > >> > can go ahead and use the increased alignment. > > But consider say: > > one TU: > > struct S { char buf[15]; } s __attribute__((aligned (32))); > > > > another TU: > > char c = 7; > > struct S { char buf[15]; } s = { { 1, 2 } }; > > char d = 8; > > int main () { return 0; } > > > > (the aligned(32) is there just to simulate the DATA_ALIGNMENT optimization > > increase). Linker warns about this (thus the question is if we want to > > increase the alignment for optimization on commons at all) and doesn't align > > it. > > > > Oh, right. I hadn't considered commons unifying with non-common variables, > and the failure of commoning in that case. I'd mostly been thinking of > uninitialized Fortran-like common blocks, where it is more common for the > blocks to have nothing in common but the name. Ok, here is what I've committed to trunk (will wait for some time before backporting). As discussed with Honza on IRC, decl_binds_to_current_def_p will need further fixing, it does the wrong thing for extern int var __attribute__((visibility ("hidden"))) or hidden DECL_COMMON symbols. And, we don't have any feedback about SPE/E500 rs6000 - DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT yet. 2013-06-10 Jakub Jelinek PR target/56564 * varasm.c (align_variable): Don't use DATA_ALIGNMENT or CONSTANT_ALIGNMENT if !decl_binds_to_current_def_p (decl). Use DATA_ABI_ALIGNMENT for that case instead if defined. (get_variable_align): New function. (get_variable_section, emit_bss, emit_common, assemble_variable_contents, place_block_symbol): Use get_variable_align instead of DECL_ALIGN. (assemble_noswitch_variable): Add align argument, use it instead of DECL_ALIGN. (assemble_variable): Adjust caller. Use get_variable_align instead of DECL_ALIGN. * config/i386/i386.h (DATA_ALIGNMENT): Adjust x86_data_alignment caller. (DATA_ABI_ALIGNMENT): Define. * config/i386/i386-protos.h (x86_data_alignment): Adjust prototype. * config/i386/i386.c (x86_data_alignment): Add opt argument. If opt is false, only return the psABI mandated alignment increase. * config/c6x/c6x.h (DATA_ALIGNMENT): Renamed to... (DATA_ABI_ALIGNMENT): ... this. * config/mmix/mmix.h (DATA_ALIGNMENT): Renamed to... (DATA_ABI_ALIGNMENT): ... this. * config/mmix/mmix.c (mmix_data_alignment): Adjust function comment. * config/s390/s390.h (DATA_ALIGNMENT): Renamed to... (DATA_ABI_ALIGNMENT): ... this. * doc/tm.texi.in (DATA_ABI_ALIGNMENT): Document. * doc/tm.texi: Regenerated. * gcc.target/i386/pr56564-1.c: New test. * gcc.target/i386/pr56564-2.c: New test. * gcc.target/i386/pr56564-3.c: New test. * gcc.target/i386/pr56564-4.c: New test. * gcc.target/i386/avx256-unaligned-load-4.c: Add -fno-common. * gcc.target/i386/avx256-unaligned-store-1.c: Likewise. * gcc.target/i386/avx256-unaligned-store-3.c: Likewise. * gcc.target/i386/avx256-unaligned-store-4.c: Likewise. * gcc.target/i386/vect-sizes-1.c: Likewise. * gcc.target/i386/memcpy-1.c: Likewise. * gcc.dg/vect/costmodel/i386/costmodel-vect-31.c (tmp): Initialize. * gcc.dg/vect/costmodel/x86_64/costmodel-vect-31.c (tmp): Likewise. --- gcc/config/c6x/c6x.h.jj 2013-02-26 16:39:34.0 +0100 +++ gcc/config/c6x/c6x.h2013-06-10 17:36:44.850082918 +0200 @@ -134,7 +134,7 @@ extern c6x_cpu_t c6x_arch; Really only externally visible arrays must be aligned this way, as only those are directly visible from another compilation unit. But we don't have that information available here. */ -#define DATA_ALIGNMENT(TYPE, ALIGN)\ +#define DATA_ABI_ALIGNMENT(TYPE, ALIGN) \ (((ALIGN) < BITS_PER_UNIT * 8 && TREE_CODE (TYPE) == ARRAY_TYPE) \ ? BITS_PER_UNIT * 8 : (ALIGN)) --- gcc/config/mmix/mmix.h.jj 2013-01-11 09:03:16.0 +0100 +++ gcc/config/mmix/mmix.h 2013-06-10 17:36:05.585730695 +0200 @@ -164,7 +164,7 @@ struct GTY(()) machine_function /* Copied from elfos.h. */ #define MAX_OFILE_ALIGNMENT (32768 * 8) -#define DATA_ALIGNMENT(TYPE, BASIC_ALIGN) \ +#define DATA_ABI_ALIGNMENT(TYPE, BASIC_ALIGN) \ mmix_data_alignment (TYPE, BASIC_ALIGN) #define CONSTANT_ALIGNMENT(CONSTANT, BASIC_ALIGN) \ --- gcc/config/mmix/mmix.c.jj 2013-03-26 10:03:58.0 +0100 +++ gcc/config/mmix/mmix.c 2013-06-10 17:36
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On 06/07/2013 02:14 PM, Jakub Jelinek wrote: >> > When the linker merges common blocks, it chooses both maximum size and >> > maximum >> > alignment. Thus for any common block for which we can prove the block must >> > reside in the module (any executable, or hidden common in shared object), >> > we >> > can go ahead and use the increased alignment. > But consider say: > one TU: > struct S { char buf[15]; } s __attribute__((aligned (32))); > > another TU: > char c = 7; > struct S { char buf[15]; } s = { { 1, 2 } }; > char d = 8; > int main () { return 0; } > > (the aligned(32) is there just to simulate the DATA_ALIGNMENT optimization > increase). Linker warns about this (thus the question is if we want to > increase the alignment for optimization on commons at all) and doesn't align > it. > Oh, right. I hadn't considered commons unifying with non-common variables, and the failure of commoning in that case. I'd mostly been thinking of uninitialized Fortran-like common blocks, where it is more common for the blocks to have nothing in common but the name. r~
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
Richard Henderson wrote: > s390 comment mentions LARL instruction On s390(x) it is indeed an ABI requirement that all global symbols are at least 2-aligned. (Note that we skip that alignment requirement if a symbol is marked as attribute((aligned(1)), but that attribute must then be present for every use, too.) Bye, Ulrich -- Dr. Ulrich Weigand GNU Toolchain for Linux on System z and Cell BE ulrich.weig...@de.ibm.com
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On 06/10/2013 12:55 PM, Jakub Jelinek wrote: > On Mon, Jun 10, 2013 at 12:51:05PM +0200, Bernd Schmidt wrote: >> On 06/07/2013 10:43 PM, Richard Henderson wrote: >>> But these I think require a good hard look to see if they really intended an >>> ABI alignment: >>> >>> c6x comment explicitly mentions abi >> >> The ABI specifies a minimum alignment for arrays. > > Thus after the patch c6x.h (DATA_ALIGNMENT) should be renamed to > DATA_ABI_ALIGNMENT, right? I think so. Bernd
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Mon, Jun 10, 2013 at 12:51:05PM +0200, Bernd Schmidt wrote: > On 06/07/2013 10:43 PM, Richard Henderson wrote: > > But these I think require a good hard look to see if they really intended an > > ABI alignment: > > > > c6x comment explicitly mentions abi > > The ABI specifies a minimum alignment for arrays. Thus after the patch c6x.h (DATA_ALIGNMENT) should be renamed to DATA_ABI_ALIGNMENT, right? Jakub
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On 06/07/2013 10:43 PM, Richard Henderson wrote: > But these I think require a good hard look to see if they really intended an > ABI alignment: > > c6x comment explicitly mentions abi The ABI specifies a minimum alignment for arrays. Bernd
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Fri, Jun 07, 2013 at 11:14:19PM +0200, Jakub Jelinek wrote: > > This structure would seem to do the wrong thing if DATA_ABI_ALIGNMENT is > > defined, but DATA_ALIGNMENT isn't. And while I realize you documented it, I > > don't like the restriction that D_A /must/ return something larger than > > D_A_A. > > All that means is that in complex cases D_A will have to call D_A_A itself. > > Yeah, I guess I can rearrange it. The reason I wrote it that way was to > avoid an extra function call, but that is probably not big enough overhead. Here is the code rearranged so that DATA_ABI_ALIGNMENT is independent of DATA_ALIGNMENT. The rest of stuff is kept as is. As for the commons getting bigger alignment than the ABI has for them, I'm afraid the linker usually has no option but to warn and don't do anything. Because if the non-common definition that is supposed to win over the common one isn't sufficiently aligned (and it could be aligned just to the ABI mandated boundary), that definition could be already in the middle of say .data or other section and so the linker doesn't have the luxury of aligning it individually. E.g. vect_can_force_dr_alignment_p has been changed some time ago to: /* We cannot change alignment of common or external symbols as another translation unit may contain a definition with lower alignment. The rules of common symbol linking mean that the definition will override the common symbol. The same is true for constant pool entries which may be shared and are not properly merged by LTO. */ if (DECL_EXTERNAL (decl) || DECL_COMMON (decl) || DECL_IN_CONSTANT_POOL (decl)) return false; but at that point we haven't changed align_variable. Thus perhaps we want in align_variable handle DECL_COMMON the same way as we handle TLS with > word alignment. 2013-06-08 Jakub Jelinek PR target/56564 * varasm.c (align_variable): Don't use DATA_ALIGNMENT or CONSTANT_ALIGNMENT if !decl_binds_to_current_def_p (decl). Use DATA_ABI_ALIGNMENT for that case instead if defined. (get_variable_align): New function. (get_variable_section, emit_bss, emit_common, assemble_variable_contents, place_block_symbol): Use get_variable_align instead of DECL_ALIGN. (assemble_noswitch_variable): Add align argument, use it instead of DECL_ALIGN. (assemble_variable): Adjust caller. Use get_variable_align instead of DECL_ALIGN. * config/i386/i386.h (DATA_ALIGNMENT): Adjust x86_data_alignment caller. (DATA_ABI_ALIGNMENT): Define. * config/i386/i386-protos.h (x86_data_alignment): Adjust prototype. * config/i386/i386.c (x86_data_alignment): Add opt argument. If opt is false, only return the psABI mandated alignment increase. * doc/tm.texi.in (DATA_ABI_ALIGNMENT): Document. * doc/tm.texi: Regenerated. * gcc.target/i386/pr56564-1.c: New test. * gcc.target/i386/pr56564-2.c: New test. * gcc.target/i386/pr56564-3.c: New test. * gcc.target/i386/pr56564-4.c: New test. * gcc.target/i386/avx256-unaligned-load-4.c: Add -fno-common. * gcc.target/i386/avx256-unaligned-store-1.c: Likewise. * gcc.target/i386/avx256-unaligned-store-3.c: Likewise. * gcc.target/i386/avx256-unaligned-store-4.c: Likewise. * gcc.target/i386/vect-sizes-1.c: Likewise. * gcc.target/i386/memcpy-1.c: Likewise. * gcc.dg/vect/costmodel/i386/costmodel-vect-31.c (tmp): Initialize. * gcc.dg/vect/costmodel/x86_64/costmodel-vect-31.c (tmp): Likewise. --- gcc/varasm.c.jj 2013-06-07 13:17:17.0 +0200 +++ gcc/varasm.c2013-06-08 16:53:40.717372488 +0200 @@ -966,13 +966,80 @@ align_variable (tree decl, bool dont_out align = MAX_OFILE_ALIGNMENT; } - /* On some machines, it is good to increase alignment sometimes. */ if (! DECL_USER_ALIGN (decl)) { +#ifdef DATA_ABI_ALIGNMENT + unsigned int data_abi_align + = DATA_ABI_ALIGNMENT (TREE_TYPE (decl), align); + /* For backwards compatibility, don't assume the ABI alignment for +TLS variables. */ + if (! DECL_THREAD_LOCAL_P (decl) || data_abi_align <= BITS_PER_WORD) + align = data_abi_align; +#endif + + /* On some machines, it is good to increase alignment sometimes. +But as DECL_ALIGN is used both for actually emitting the variable +and for code accessing the variable as guaranteed alignment, we +can only increase the alignment if it is a performance optimization +if the references to it must bind to the current definition. */ + if (decl_binds_to_current_def_p (decl)) + { +#ifdef DATA_ALIGNMENT + unsigned int data_align = DATA_ALIGNMENT (TREE_TYPE (decl), align); + /* Don't increase alignment too much for TLS variables - TLS space +is too precious. */ + if (! DECL_THREAD
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Fri, Jun 07, 2013 at 06:56:34PM -0400, Hans-Peter Nilsson wrote: > > criscompiler options for alignment -- systemwide or local? > > No, DATA_ALIGNMENT in cris.h is not intended as an ABI > indication, but as an optimization when emitting data. > (This was the way to do it at the time. Has this changed?) > > The ABI is as indicated by BIGGEST_ALIGNMENT: 8 (bits; one > byte). Nothing is guaranteed (to the data referer) to have a > bigger alignment - unless otherwise indicated by attribute-align. > > (Unfortunately I can't change BIGGEST_ALIGNMENT to indicate that > atomic variables require "natural alignment", or actually not to > straddle a cache-boundary, as increasing BIGGEST_ALIGNMENT makes > GCC change the ABI. But that's a slightly different issue.) Right now it is unfortunately part of ABI, which is something the patch attempts to cure in a backwards compatible way (i.e., data will be still alignment the way it used to be, but code will no longer assume it is that much aligned, unless it is DATA_ABI_ALIGNMENT). > > > mmixcomment mentions GETA instruction > > Yep, data must be at least 32-bit-aligned so addresses can be > formed with a GETA insn. BIGGEST_ALIGNMENT is 64 though and > STRICT_ALIGNMENT; natural alignment is required for proper > interpretation as the low bits are ignored. Then mmix would probably want to define DATA_ABI_ALIGNMENT instead of DATA_ALIGNMENT after the patch (or both). Jakub
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Fri, 7 Jun 2013, Richard Henderson wrote: > I've had a brief look over the instances of D_A within the tree atm. Most of > them carry the cut-n-paste comment "for the same reasons". These I believe > never intended an ABI change, and were really only interested in optimization. > > But these I think require a good hard look to see if they really intended an > ABI alignment: I'm not sure what is about to change how? > cris compiler options for alignment -- systemwide or local? No, DATA_ALIGNMENT in cris.h is not intended as an ABI indication, but as an optimization when emitting data. (This was the way to do it at the time. Has this changed?) The ABI is as indicated by BIGGEST_ALIGNMENT: 8 (bits; one byte). Nothing is guaranteed (to the data referer) to have a bigger alignment - unless otherwise indicated by attribute-align. (Unfortunately I can't change BIGGEST_ALIGNMENT to indicate that atomic variables require "natural alignment", or actually not to straddle a cache-boundary, as increasing BIGGEST_ALIGNMENT makes GCC change the ABI. But that's a slightly different issue.) > mmix comment mentions GETA instruction Yep, data must be at least 32-bit-aligned so addresses can be formed with a GETA insn. BIGGEST_ALIGNMENT is 64 though and STRICT_ALIGNMENT; natural alignment is required for proper interpretation as the low bits are ignored. brgds, H-P
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On Fri, Jun 07, 2013 at 01:43:27PM -0700, Richard Henderson wrote: > On 06/07/2013 12:25 PM, Jakub Jelinek wrote: > > This PR is about DATA_ALIGNMENT macro increasing alignment of some decls > > for optimization purposes beyond ABI mandated levels. It is fine to emit > > the vars aligned as much as we want for optimization purposes, but if we > > can't be sure that references to that decl bind to the definition we > > increased the alignment on (e.g. common variables, or -fpic code without > > hidden visibility, weak vars etc.), we can't assume that alignment. > > When the linker merges common blocks, it chooses both maximum size and maximum > alignment. Thus for any common block for which we can prove the block must > reside in the module (any executable, or hidden common in shared object), we > can go ahead and use the increased alignment. But consider say: one TU: struct S { char buf[15]; } s __attribute__((aligned (32))); another TU: char c = 7; struct S { char buf[15]; } s = { { 1, 2 } }; char d = 8; int main () { return 0; } (the aligned(32) is there just to simulate the DATA_ALIGNMENT optimization increase). Linker warns about this (thus the question is if we want to increase the alignment for optimization on commons at all) and doesn't align it. > This structure would seem to do the wrong thing if DATA_ABI_ALIGNMENT is > defined, but DATA_ALIGNMENT isn't. And while I realize you documented it, I > don't like the restriction that D_A /must/ return something larger than D_A_A. > All that means is that in complex cases D_A will have to call D_A_A itself. Yeah, I guess I can rearrange it. The reason I wrote it that way was to avoid an extra function call, but that is probably not big enough overhead. > I would think that it would be better to rearrange as > > if (!D_U_A) > { > #ifdef D_A_A > align = ... > #endif > #ifdef D_A > if (d_b_t_c_d_p) > align = ... > #endif > } > > Why the special case for TLS? If we really want that special case surely that > test should go into D_A_A itself, and not here in generic code. I special case TLS for backwards ABI compatibility with GCC <= 4.8.1, because we ignored DATA_ALIGNMENT for TLS vars, even if what it returned wasn't just an optimization, but ABI requirement. So, if you have: __thread char a[16] = { 1, 2 }; in one shared library and __thread char a[16] = { 1, 2 }; use (a); in another shared library, compile one with GCC 4.8.1 e.g. and the other one with 4.9.0, if use (a) would assume the psABI mandated 16 byte alignment, but the actual definition was from 4.8.1 compiled object and was only 1 byte aligned, it could crash. Also, neither DATA_ALIGNMENT nor DATA_ABI_ALIGNMENT sees the decl itself, it just looks at type and previously assumed alignment. > > Bootstrapped/regtested on x86_64-linux and i686-linux. No idea about other > > targets, I've kept them all using DATA_ALIGNMENT, which is considered > > optimization increase only now, if there is some ABI mandated alignment > > increase on other targets, that should be done in DATA_ABI_ALIGNMENT as > > well as DATA_ALIGNMENT. > > I've had a brief look over the instances of D_A within the tree atm. Most of > them carry the cut-n-paste comment "for the same reasons". These I believe > never intended an ABI change, and were really only interested in optimization. Thanks for looking into this. > But these I think require a good hard look to see if they really intended an > ABI alignment: > > c6x comment explicitly mentions abi > cris compiler options for alignment -- systemwide or local? > mmix comment mentions GETA instruction > s390 comment mentions LARL instruction > rs6000SPE and E500 portion of the alignment non-optional? > > Relevant port maintainers CCed. Jakub
Re: [PATCH] DATA_ALIGNMENT vs. DATA_ABI_ALIGNMENT (PR target/56564)
On 06/07/2013 12:25 PM, Jakub Jelinek wrote: > This PR is about DATA_ALIGNMENT macro increasing alignment of some decls > for optimization purposes beyond ABI mandated levels. It is fine to emit > the vars aligned as much as we want for optimization purposes, but if we > can't be sure that references to that decl bind to the definition we > increased the alignment on (e.g. common variables, or -fpic code without > hidden visibility, weak vars etc.), we can't assume that alignment. When the linker merges common blocks, it chooses both maximum size and maximum alignment. Thus for any common block for which we can prove the block must reside in the module (any executable, or hidden common in shared object), we can go ahead and use the increased alignment. It's only in shared objects with non-hidden common blocks that we have a problem, since in that case we may resolve the common block via copy reloc to a memory block in another module. So while decl_binds_to_current_def_p is a good starting point, I think we can do a little better with common blocks. Which ought to take care of those vectorization regressions you mention. > @@ -966,8 +966,12 @@ align_variable (tree decl, bool dont_out >align = MAX_OFILE_ALIGNMENT; > } > > - /* On some machines, it is good to increase alignment sometimes. */ > - if (! DECL_USER_ALIGN (decl)) > + /* On some machines, it is good to increase alignment sometimes. > + But as DECL_ALIGN is used both for actually emitting the variable > + and for code accessing the variable as guaranteed alignment, we > + can only increase the alignment if it is a performance optimization > + if the references to it must bind to the current definition. */ > + if (! DECL_USER_ALIGN (decl) && decl_binds_to_current_def_p (decl)) > { > #ifdef DATA_ALIGNMENT >unsigned int data_align = DATA_ALIGNMENT (TREE_TYPE (decl), align); > @@ -988,12 +992,69 @@ align_variable (tree decl, bool dont_out > } > #endif > } > +#ifdef DATA_ABI_ALIGNMENT > + else if (! DECL_USER_ALIGN (decl)) > +{ > + unsigned int data_align = DATA_ABI_ALIGNMENT (TREE_TYPE (decl), align); > + /* For backwards compatibility, don't assume the ABI alignment for > + TLS variables. */ > + if (! DECL_THREAD_LOCAL_P (decl) || data_align <= BITS_PER_WORD) > + align = data_align; > +} > +#endif This structure would seem to do the wrong thing if DATA_ABI_ALIGNMENT is defined, but DATA_ALIGNMENT isn't. And while I realize you documented it, I don't like the restriction that D_A /must/ return something larger than D_A_A. All that means is that in complex cases D_A will have to call D_A_A itself. I would think that it would be better to rearrange as if (!D_U_A) { #ifdef D_A_A align = ... #endif #ifdef D_A if (d_b_t_c_d_p) align = ... #endif } Why the special case for TLS? If we really want that special case surely that test should go into D_A_A itself, and not here in generic code. > Bootstrapped/regtested on x86_64-linux and i686-linux. No idea about other > targets, I've kept them all using DATA_ALIGNMENT, which is considered > optimization increase only now, if there is some ABI mandated alignment > increase on other targets, that should be done in DATA_ABI_ALIGNMENT as > well as DATA_ALIGNMENT. I've had a brief look over the instances of D_A within the tree atm. Most of them carry the cut-n-paste comment "for the same reasons". These I believe never intended an ABI change, and were really only interested in optimization. But these I think require a good hard look to see if they really intended an ABI alignment: c6x comment explicitly mentions abi criscompiler options for alignment -- systemwide or local? mmixcomment mentions GETA instruction s390comment mentions LARL instruction rs6000 SPE and E500 portion of the alignment non-optional? Relevant port maintainers CCed. r~