Re: [PATCH][RFC] Re-write LTO type merging again, do tree merging
> On Fri, 14 Jun 2013, Jan Hubicka wrote: > > > > > > > > > Ok, not streaming and comparing TREE_USED gets it improved to > > > > > > I will try to gather better data tomorrow. My mozilla build died on disk > > > space, > > > but according to stats we are now at about 7GB of GGC memory after > > > merging. > > > I was playing with the following patch that implements testing whether > > > types > > > are same in my (probably naive and wrong) understanding of ODR rule in C++ > > > > So i can confirm that we now need 3GB of TMP space instead of 8GB with > > earlier > > version of patch. I will compare to mainline tomorrow, but I think it is > > about the same. > > phase opt and generate : 96.39 ( 9%) usr 40.45 (45%) sys 136.91 (12%) > > wall 271042 kB ( 7%) ggc > > phase stream in : 457.87 (43%) usr 8.38 ( 9%) sys 466.44 (40%) > > wall 3798844 kB (93%) ggc > > phase stream out: 509.39 (48%) usr 40.82 (46%) sys 550.88 (48%) > > wall7149 kB ( 0%) ggc > > ipa cp : 13.62 ( 1%) usr 5.00 ( 6%) sys 18.61 ( 2%) > > wall 425204 kB (10%) ggc > > ipa inlining heuristics : 60.52 ( 6%) usr 36.15 (40%) sys 96.71 ( 8%) > > wall 1353370 kB (33%) ggc > > ipa lto decl in : 346.94 (33%) usr 5.49 ( 6%) sys 352.60 (31%) > > wall7042 kB ( 0%) ggc > > ipa lto decl out: 481.19 (45%) usr 23.28 (26%) sys 504.68 (44%) > > wall 0 kB ( 0%) ggc > > TOTAL :1063.6789.65 1154.26 > > 4078436 kB > > > > So we are still bound by streaming. I am running -flto-report overnight. > > > > My ODR patch finds 36377 matches and also weird looking mismatches of type: > > > size > bitsizetype> constant 1024> > > unit size > sizetype> constant 128> > > align 64 symtab 0 alias set -1 canonical type 0x7fbd30f0bc78 > > fields > type > size > > unit size > > align 16 symtab 0 alias set -1 canonical type 0x7fbd41566540 > > precision 16 min max > 0x7fbd41555fc0 65535>> > > unsigned nonlocal HI file /usr/include/bits/socket.h line 189 col 0 > > size unit size > 2> > > align 16 offset_align 128 > > offset > > bit offset context > > > > chain > 0x7fbd415667e0 long unsigned int> > > unsigned nonlocal DI file /usr/include/bits/socket.h line 190 > > col 0 > > size > > unit size > > align 64 offset_align 128 offset > > bit offset context > 0x7fbd30d46dc8 sockaddr_storage> chain > __ss_padding>>> context > > chain > > > > size > bitsizetype> constant 1024> > > unit size > sizetype> constant 128> > > align 64 symtab 0 alias set -1 canonical type 0x7fbd30f0bc78 > > fields > type > size > > unit size > > align 16 symtab 0 alias set -1 canonical type 0x7fbd41566540 > > precision 16 min max > 0x7fbd41555fc0 65535>> > > unsigned HI file /usr/include/bits/socket.h line 189 col 0 size > > unit size > > align 16 offset_align 128 > > offset > > bit offset context > > > > chain > 0x7fbd415667e0 long unsigned int> > > unsigned DI file /usr/include/bits/socket.h line 190 col 0 > > size > > unit size > > align 64 offset_align 128 offset > > bit offset context > 0x7fbd30f0bc78 sockaddr_storage> chain > __ss_padding>>> context > > pointer_to_this chain > 0x7fbd30ea9398 D.938243>> > > > > that mismatch because we run into following difference: > > > type > size > > unit size > > align 64 symtab 0 alias set -1 canonical type 0x7fbd30f0bc78 > > fields > 0x7fbd3b98c000 sa_family_t> > > unsigned nonlocal HI file /usr/include/bits/socket.h line 189 > > col 0 > > size > > unit size > > align 16 offset_align 128 > > offset > > bit offset context > > chain > 0x7fbd30e99000 __ss_align>> context > D.967968> > > chain > > > public VOID file /usr/include/bits/socket.h line 187 col 0 > > align 8 context > > > > > > > I am not sure what means that one type has more TYPE_DECLs stacked than the > > other. > > I think that's the usual case of two nearly identical types in the > _same_ SCC, linked through one of the types TYPE_DECL DECL_ORIGINAL_TYPE. > The C++ frontend likes to produce those ... Hmm, so i guess I should walk into ORIGINAL_TYPE when seeing TYPE_DECL? Honza
Re: [PATCH][RFC] Re-write LTO type merging again, do tree merging
On Fri, 14 Jun 2013, Jan Hubicka wrote: > > > > > > Ok, not streaming and comparing TREE_USED gets it improved to > > > > I will try to gather better data tomorrow. My mozilla build died on disk > > space, > > but according to stats we are now at about 7GB of GGC memory after merging. > > I was playing with the following patch that implements testing whether types > > are same in my (probably naive and wrong) understanding of ODR rule in C++ > > So i can confirm that we now need 3GB of TMP space instead of 8GB with earlier > version of patch. I will compare to mainline tomorrow, but I think it is > about the same. > phase opt and generate : 96.39 ( 9%) usr 40.45 (45%) sys 136.91 (12%) > wall 271042 kB ( 7%) ggc > phase stream in : 457.87 (43%) usr 8.38 ( 9%) sys 466.44 (40%) > wall 3798844 kB (93%) ggc > phase stream out: 509.39 (48%) usr 40.82 (46%) sys 550.88 (48%) > wall7149 kB ( 0%) ggc > ipa cp : 13.62 ( 1%) usr 5.00 ( 6%) sys 18.61 ( 2%) > wall 425204 kB (10%) ggc > ipa inlining heuristics : 60.52 ( 6%) usr 36.15 (40%) sys 96.71 ( 8%) > wall 1353370 kB (33%) ggc > ipa lto decl in : 346.94 (33%) usr 5.49 ( 6%) sys 352.60 (31%) > wall7042 kB ( 0%) ggc > ipa lto decl out: 481.19 (45%) usr 23.28 (26%) sys 504.68 (44%) > wall 0 kB ( 0%) ggc > TOTAL :1063.6789.65 1154.26 > 4078436 kB > > So we are still bound by streaming. I am running -flto-report overnight. > > My ODR patch finds 36377 matches and also weird looking mismatches of type: > size bitsizetype> constant 1024> > unit size sizetype> constant 128> > align 64 symtab 0 alias set -1 canonical type 0x7fbd30f0bc78 > fields type size > unit size > align 16 symtab 0 alias set -1 canonical type 0x7fbd41566540 > precision 16 min max 0x7fbd41555fc0 65535>> > unsigned nonlocal HI file /usr/include/bits/socket.h line 189 col 0 > size unit size > align 16 offset_align 128 > offset > bit offset context > > chain 0x7fbd415667e0 long unsigned int> > unsigned nonlocal DI file /usr/include/bits/socket.h line 190 col > 0 > size > unit size > align 64 offset_align 128 offset > bit offset context 0x7fbd30d46dc8 sockaddr_storage> chain __ss_padding>>> context > chain > > size bitsizetype> constant 1024> > unit size sizetype> constant 128> > align 64 symtab 0 alias set -1 canonical type 0x7fbd30f0bc78 > fields type size > unit size > align 16 symtab 0 alias set -1 canonical type 0x7fbd41566540 > precision 16 min max 0x7fbd41555fc0 65535>> > unsigned HI file /usr/include/bits/socket.h line 189 col 0 size > unit size > align 16 offset_align 128 > offset > bit offset context > > chain 0x7fbd415667e0 long unsigned int> > unsigned DI file /usr/include/bits/socket.h line 190 col 0 > size > unit size > align 64 offset_align 128 offset > bit offset context 0x7fbd30f0bc78 sockaddr_storage> chain __ss_padding>>> context > pointer_to_this chain 0x7fbd30ea9398 D.938243>> > > that mismatch because we run into following difference: > type size > unit size > align 64 symtab 0 alias set -1 canonical type 0x7fbd30f0bc78 > fields 0x7fbd3b98c000 sa_family_t> > unsigned nonlocal HI file /usr/include/bits/socket.h line 189 col > 0 > size > unit size > align 16 offset_align 128 > offset > bit offset context > chain 0x7fbd30e99000 __ss_align>> context D.967968> > chain > > public VOID file /usr/include/bits/socket.h line 187 col 0 > align 8 context > > > > I am not sure what means that one type has more TYPE_DECLs stacked than the > other. I think that's the usual case of two nearly identical types in the _same_ SCC, linked through one of the types TYPE_DECL DECL_ORIGINAL_TYPE. The C++ frontend likes to produce those ... They are accounted as the "are in the same SCC" in the stats (for the latest updated patch): [WPA] Old merging code merges an additional 41493 types of which 21633 are in the same SCC with their prevailing variant (415188 and 324371 associated trees) I'm looking into other things now and for example see TYPE_ALIAS_SET issues (we do stream only alias sets -1 and 0, but whether -1 got transformed to 0 at compile-time depends on whether get_alias_set was called on it). That's one case where we can either forcefully call get_alias_set at streaming time, introduce a predicate type_would_have_alias_set_zero that doesn't run the whole get_alias_set machinery, or ignore this for merging and make sure to merge -1 and 0 as 0 (the whol
Re: [PATCH][RFC] Re-write LTO type merging again, do tree merging
> > > > > > Ok, not streaming and comparing TREE_USED gets it improved to > > > > I will try to gather better data tomorrow. My mozilla build died on disk > > space, > > but according to stats we are now at about 7GB of GGC memory after merging. > > I was playing with the following patch that implements testing whether types > > are same in my (probably naive and wrong) understanding of ODR rule in C++ > > So i can confirm that we now need 3GB of TMP space instead of 8GB with earlier > version of patch. I will compare to mainline tomorrow, but I think it is > about the same. > phase opt and generate : 96.39 ( 9%) usr 40.45 (45%) sys 136.91 (12%) > wall 271042 kB ( 7%) ggc > phase stream in : 457.87 (43%) usr 8.38 ( 9%) sys 466.44 (40%) > wall 3798844 kB (93%) ggc > phase stream out: 509.39 (48%) usr 40.82 (46%) sys 550.88 (48%) > wall7149 kB ( 0%) ggc > ipa cp : 13.62 ( 1%) usr 5.00 ( 6%) sys 18.61 ( 2%) > wall 425204 kB (10%) ggc > ipa inlining heuristics : 60.52 ( 6%) usr 36.15 (40%) sys 96.71 ( 8%) > wall 1353370 kB (33%) ggc > ipa lto decl in : 346.94 (33%) usr 5.49 ( 6%) sys 352.60 (31%) > wall7042 kB ( 0%) ggc > ipa lto decl out: 481.19 (45%) usr 23.28 (26%) sys 504.68 (44%) > wall 0 kB ( 0%) ggc > TOTAL :1063.6789.65 1154.26 > 4078436 kB > > So we are still bound by streaming. I am running -flto-report overnight. [WPA] read 43363300 SCCs of average size 2.264113 [WPA] 98179403 tree bodies read in total [WPA] tree SCC table: size 16777213, 6422251 elements, collision ratio: 0.811639 [WPA] tree SCC max chain length 88 (size 1) [WPA] Compared 16544560 SCCs, 275298 collisions (0.016640) [WPA] Merged 16458553 SCCs [WPA] Merged 46453870 tree bodies [WPA] Merged 9535385 types [WPA] 6771259 types prevailed (21348860 associated trees) [WPA] Old merging code merges an additional 1759918 types of which 379059 are in the same SCC with their prevailing variant (19696849 and 15301625 associated trees) [WPA] GIMPLE canonical type table: size 131071, 77875 elements, 6771394 searches, 1528380 collisions (ratio: 0.225711) [WPA] GIMPLE canonical type hash table: size 16777213, 6771339 elements, 23174504 searches, 21075518 collisions (ratio: 0.909427) [LTRANS] read 228296 SCCs of average size 11.882460 [LTRANS] 2712718 tree bodies read in total [LTRANS] GIMPLE canonical type table: size 16381, 7025 elements, 704670 searches, 24040 collisions (ratio: 0.034115) [LTRANS] GIMPLE canonical type hash table: size 1048573, 704613 elements, 2269381 searches, 2021919 collisions (ratio: 0.890956) We manage to get stuck in one of ltranses on LRA LRA hard reg assignment : 476.07 (44%) usr 0.03 ( 0%) sys 476.08 (44%) wall 0 kB ( 0%) ggc 2860712.1151 lto1 alloc_page(unsigned int) 3564 1.5094 lto1 record_reg_classes(int, int, rtx_def**, machine_mode*, char const**, rtx_def*, reg_class*) 3235 1.3700 libc-2.11.1.so _int_malloc 3056 1.2942 lto1 ggc_set_mark(void const*) 2646 1.1206 lto1 gt_ggc_mx_lang_tree_node(void*) 2539 1.0753 lto1 bitmap_set_bit(bitmap_head_def*, int) 2333 0.9880 opreport /usr/bin/opreport 2210 0.9359 lto1 for_each_rtx_1(rtx_def*, int, int (*)(rtx_def**, void*), void*) 2133 0.9033 lto1 constrain_operands(int) 2128 0.9012 lto1 lookup_page_table_entry(void const*) 1586 0.6717 lto1 preprocess_constraints() While GGC memory is now under 7GB after type streaming and we GGC just once in WPA, the TOP usage still goes to about 12GB. With the ODR patch there are 424 devirtualizations happening during WPA and some extra (do not have stats for) during ltrans. Honza
Re: [PATCH][RFC] Re-write LTO type merging again, do tree merging
> > > > Ok, not streaming and comparing TREE_USED gets it improved to > > I will try to gather better data tomorrow. My mozilla build died on disk > space, > but according to stats we are now at about 7GB of GGC memory after merging. > I was playing with the following patch that implements testing whether types > are same in my (probably naive and wrong) understanding of ODR rule in C++ So i can confirm that we now need 3GB of TMP space instead of 8GB with earlier version of patch. I will compare to mainline tomorrow, but I think it is about the same. phase opt and generate : 96.39 ( 9%) usr 40.45 (45%) sys 136.91 (12%) wall 271042 kB ( 7%) ggc phase stream in : 457.87 (43%) usr 8.38 ( 9%) sys 466.44 (40%) wall 3798844 kB (93%) ggc phase stream out: 509.39 (48%) usr 40.82 (46%) sys 550.88 (48%) wall 7149 kB ( 0%) ggc ipa cp : 13.62 ( 1%) usr 5.00 ( 6%) sys 18.61 ( 2%) wall 425204 kB (10%) ggc ipa inlining heuristics : 60.52 ( 6%) usr 36.15 (40%) sys 96.71 ( 8%) wall 1353370 kB (33%) ggc ipa lto decl in : 346.94 (33%) usr 5.49 ( 6%) sys 352.60 (31%) wall 7042 kB ( 0%) ggc ipa lto decl out: 481.19 (45%) usr 23.28 (26%) sys 504.68 (44%) wall 0 kB ( 0%) ggc TOTAL :1063.6789.65 1154.26 4078436 kB So we are still bound by streaming. I am running -flto-report overnight. My ODR patch finds 36377 matches and also weird looking mismatches of type: constant 1024> unit size constant 128> align 64 symtab 0 alias set -1 canonical type 0x7fbd30f0bc78 fields unit size align 16 symtab 0 alias set -1 canonical type 0x7fbd41566540 precision 16 min max > unsigned nonlocal HI file /usr/include/bits/socket.h line 189 col 0 size unit size align 16 offset_align 128 offset bit offset context chain unsigned nonlocal DI file /usr/include/bits/socket.h line 190 col 0 size unit size align 64 offset_align 128 offset bit offset context chain >> context chain > constant 1024> unit size constant 128> align 64 symtab 0 alias set -1 canonical type 0x7fbd30f0bc78 fields unit size align 16 symtab 0 alias set -1 canonical type 0x7fbd41566540 precision 16 min max > unsigned HI file /usr/include/bits/socket.h line 189 col 0 size unit size align 16 offset_align 128 offset bit offset context chain unsigned DI file /usr/include/bits/socket.h line 190 col 0 size unit size align 64 offset_align 128 offset bit offset context chain >> context pointer_to_this chain > that mismatch because we run into following difference: unit size align 64 symtab 0 alias set -1 canonical type 0x7fbd30f0bc78 fields unsigned nonlocal HI file /usr/include/bits/socket.h line 189 col 0 size unit size align 16 offset_align 128 offset bit offset context chain > context chain > public VOID file /usr/include/bits/socket.h line 187 col 0 align 8 context > I am not sure what means that one type has more TYPE_DECLs stacked than the other. Honza
Re: [PATCH][RFC] Re-write LTO type merging again, do tree merging
> > Ok, not streaming and comparing TREE_USED gets it improved to I will try to gather better data tomorrow. My mozilla build died on disk space, but according to stats we are now at about 7GB of GGC memory after merging. I was playing with the following patch that implements testing whether types are same in my (probably naive and wrong) understanding of ODR rule in C++ It prints type pairs that seems same and then it verifies that they are having same names and they are in same namespaces and records. On Javascript there are 5000 types found same by devirtualization code this way that are not having the same MAIN VARIANT. I gess those trees may be good starting point for you to look why they are not merged. I suppose that once we have maintenable code base we can get into more aggressive merging in special cases. Requiring trees to be exactly same is a good default behaviour. We however may take advantage of extra knowledge. FE may tag types/decls that are subject to ODR rule and for those we can reduce the hash to be based only on name+context and we can even output sane diagnostic on mismatches. Simiarly I think it would help a lot if we proactively merged !can_prevail_p decls with matching types into those that can prevail by hashing PUBLIC decls only by their assembler name. Merging those should subsequently allow collapsing the types that are otherwise kept separate just because associated vtables are having differences in EXTERNAL and PUBLIC flags on the methods and such. Index: tree.c === --- tree.c (revision 200064) +++ tree.c (working copy) @@ -11618,6 +11711,91 @@ lhd_gcc_personality (void) return gcc_eh_personality_decl; } +/* For languages with One Definition Rule, work out if + decls are actually the same even if the tree representation + differs. This handles only decls appearing in TYPE_NAME + and TYPE_CONTEXT. That is NAMESPACE_DECL, TYPE_DECL, + RECORD_TYPE and IDENTIFIER_NODE. */ + +static bool +decls_same_for_odr (tree decl1, tree decl2) +{ + if (decl1 == decl2) +return true; + if (!decl1 || !decl2) +{ + fprintf (stderr, "Nesting mismatch\n"); + debug_tree (decl1); + debug_tree (decl2); + return false; +} + if (TREE_CODE (decl1) != TREE_CODE (decl2)) +{ + fprintf (stderr, "Code mismatch\n"); + debug_tree (decl1); + debug_tree (decl2); + return false; +} + if (TREE_CODE (decl1) == TRANSLATION_UNIT_DECL) +return true; + if (TREE_CODE (decl1) != NAMESPACE_DECL + && TREE_CODE (decl1) != RECORD_TYPE + && TREE_CODE (decl1) != TYPE_DECL) +{ + fprintf (stderr, "Decl type mismatch\n"); + debug_tree (decl1); + return false; +} + if (!DECL_NAME (decl1)) +{ + fprintf (stderr, "Anonymous; name mysmatch\n"); + debug_tree (decl1); + return false; +} + if (!decls_same_for_odr (DECL_NAME (decl1), DECL_NAME (decl2))) +return false; + return decls_same_for_odr (DECL_CONTEXT (decl1), +DECL_CONTEXT (decl2)); +} + +/* For languages with One Definition Rule, work out if + types are same even if the tree representation differs. + This is non-trivial for LTO where minnor differences in + the type representation may have prevented type merging + to merge two copies of otherwise equivalent type. */ + +static bool +types_same_for_odr (tree type1, tree type2) +{ + type1 = TYPE_MAIN_VARIANT (type1); + type2 = TYPE_MAIN_VARIANT (type2); + if (type1 == type2) +return true; + if (!type1 || !type2) +return false; + + /* If types are not structuraly same, do not bother to contnue. + Match in the remainder of code would mean ODR violation. */ + if (!types_compatible_p (type1, type2)) +return false; + + debug_tree (type1); + debug_tree (type2); + if (!TYPE_NAME (type1)) +{ + fprintf (stderr, "Anonymous; name mysmatch\n"); + return false; +} + if (!decls_same_for_odr (TYPE_NAME (type1), TYPE_NAME (type2))) +return false; + if (!decls_same_for_odr (TYPE_CONTEXT (type1), TYPE_CONTEXT (type2))) +return false; + fprintf (stderr, "type match!\n"); + gcc_assert (in_lto_p); + + return true; +} + /* Try to find a base info of BINFO that would have its field decl at offset OFFSET within the BINFO type and which is of EXPECTED_TYPE. If it can be found, return, otherwise return NULL_TREE. */ @@ -11633,8 +11811,8 @@ get_binfo_at_offset (tree binfo, HOST_WI tree fld; int i; - if (TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (expected_type)) - return binfo; + if (types_same_for_odr (type, expected_type)) +return binfo; if (offset < 0) return NULL_TREE; @@ -11663,7 +11841,7 @@ get_binfo_at_offset (tree binfo, HOST_WI { tree base_binfo, found_binfo = NULL_TREE; for (i = 0; BINFO_BASE_ITERATE (binfo, i, base_binfo)
Re: [PATCH][RFC] Re-write LTO type merging again, do tree merging
On Wed, 12 Jun 2013, Richard Biener wrote: > > The following patch re-writes LTO type merging completely with the > goal to move as much work as possible to the compile step, away > from WPA time. At the same time the merging itself gets very > conservative but also more general - it now merges arbitrary trees, > not only types, but only if they are bit-identical and have the > same outgoing tree pointers. > > Especially the latter means that we now have to merge SCCs of trees > together and either take the whole SCC as prevailing or throw it > away. Moving work to the compile step means that we compute > SCCs and their hashes there, re-organizing streaming to stream > tree bodies as SCC blocks together with the computed hash. > > When we ask the streamer to output a tree T then it now has > to DFS walk all tree pointers, collecting SCCs of not yet > streamed trees and output them like the following: > > { LTO_tree_scc, N, hash, entry_len, >{ header1, header2, ... headerN }, >{ bits1, refs1, bits2, refs2, ... bitsN, refsN } } > { LTO_tree_scc, 1, hash, header, bits, refs } > { LTO_tree_scc, M, hash, entry_len, >{ header1, header2, ... headerM }, >{ bits1, refs1, bits2, refs2, ... bitsM, refsM } } > LTO_tree_pickle_reference to T > > with tree references in refsN always being LTO_tree_pickle_references > instead of starting a new tree inline. That results in at most > N extra LTO_tree_pickle_references for N streamed trees, together > with the LTO_tree_scc wrapping overhead this causes a slight > increase in LTO object size (around 10% last time I measured, which > was before some additional optimization went in). > > The overhead also happens on the LTRANS file producing side > which now has to do the DFS walk and stream the extra data. > It doesn't do the hashing though as on the LTRANS consumer > side no merging is performed. > > The patch preserves the core of the old merging code to compare > with the new code and output some statistics. That means that > if you build with -flto-report[-wpa] you get an additional > compile-time and memory overhead. > > For reference here are the stats when LTO bootstrapping for > stage2 cc1: > > WPA statistics > [WPA] read 2494507 SCCs of average size 2.380067 > [WPA] 5937095 tree bodies read in total > [WPA] tree SCC table: size 524287, 286280 elements, collision ratio: > 0.806376 > [WPA] tree SCC max chain length 11 (size 1) > [WPA] Compared 403361 SCCs, 6226 collisions (0.015435) > [WPA] Merged 399980 SCCs > [WPA] Merged 2438250 tree bodies > [WPA] Merged 192475 types > [WPA] 195422 types prevailed > [WPA] Old merging code merges an additional 54582 types of which 21083 are > in the same SCC with their prevailing variant > > this says that we've streamed in 5937095 tree bodies in > 2494507 SCCs (so the average SCC size is small), of those > we were able to immediately ggc_free 399980 SCCs because they > already existed in identical form (16% of the SCCs, 41% of the trees > and 49% of the types). The old merging code forced the merge > of an additional 54582 types (but 21083 of them it merged with > a type that is in the same SCC, that is, it changed the shape > of the SCC and collapsed parts of it - something that is > suspicious). > > The patch was LTO bootstrapped (testing currently running) on > x86_64-unknown-linux-gnu and I've built SPEC2k6 with -Ofast -g -flto > and did a test run of the binaries which shows that > currently 471.omnetpp, 483.xalancbmk and 447.dealII fail > (471.omnetpp segfaults in __cxxabiv1::__dynamic_cast) - these > fails were introduced quite recently likely due to the improved > FUNCTION_DECL and VAR_DECL merging and the cgraph fixup Honza did. The following incremental patch fixes that. Index: trunk/gcc/lto-symtab.c === --- trunk.orig/gcc/lto-symtab.c 2013-06-12 16:47:38.0 +0200 +++ trunk/gcc/lto-symtab.c 2013-06-12 17:00:12.664126423 +0200 @@ -96,9 +96,6 @@ lto_varpool_replace_node (struct varpool ipa_clone_referring ((symtab_node)prevailing_node, &vnode->symbol.ref_list); - /* Be sure we can garbage collect the initializer. */ - if (DECL_INITIAL (vnode->symbol.decl)) -DECL_INITIAL (vnode->symbol.decl) = error_mark_node; /* Finally remove the replaced node. */ varpool_remove_node (vnode); } Index: trunk/gcc/varpool.c === --- trunk.orig/gcc/varpool.c2013-06-12 13:13:06.0 +0200 +++ trunk/gcc/varpool.c 2013-06-12 17:01:46.088248807 +0200 @@ -77,15 +77,8 @@ varpool_remove_node (struct varpool_node /* Renove node initializer when it is no longer needed. */ void -varpool_remove_initializer (struct varpool_node *node) +varpool_remove_initializer (struct varpool_node *) { - if (DECL_INITIAL (node->symbol.decl) - && !DECL_IN_CONSTANT_POOL (node->symbol.decl) - /* Keep vtables for BINFO folding. */ - && !DECL_