Hi! On 2021-08-20T09:51:36+0200, Richard Biener wrote: > On Thu, Aug 19, 2021 at 10:14 PM Thomas Schwinge > wrote: >> Richard, maybe you have an opinion here, in particular about my >> "SLP vectorizer" comment below? Please see >> >> for the full context. >> >> On 2021-08-16T10:21:04+0200, Jakub Jelinek wrote: >> > On Mon, Aug 16, 2021 at 10:08:42AM +0200, Thomas Schwinge wrote: >> >> /* Build COMPONENT_REF and set TREE_THIS_VOLATILE and TREE_READONLY on it >> >> as appropriate. */ >> >> >> >> tree >> >> omp_build_component_ref (tree obj, tree field) >> >> { >> >> + tree field_type = TREE_TYPE (field); >> >> + tree obj_type = TREE_TYPE (obj); >> >> + if (!ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (obj_type))) >> >> + field_type >> >> + = build_qualified_type (field_type, >> >> + KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (obj_type))); >> >> (For later reference: "Kwok's new code" here is to propagate to >> 'field_type' any non-generic address space of 'obj_type'.) >> >> |> Concerning the current 'gcc/omp-low.c:omp_build_component_ref', for the >> |> current set of offloading testcases, we never see a >> |> '!ADDR_SPACE_GENERIC_P' there, so the address space handling doesn't seem >> |> to be necessary there (but also won't do any harm: no-op). >> > >> > Are you sure this can't trigger? >> > Say >> > extern int __seg_fs a; >> > >> > void >> > foo (void) >> > { >> > #pragma omp parallel private (a) >> > a = 2; >> > } >> >> That test case doesn't run into 'omp_build_component_ref' at all, >> but I'm attaching an altered and extended variant that does, >> "Add 'libgomp.c/address-space-1.c'". OK to push to master branch? >> >> In this case, 'omp_build_component_ref' called via host compilation >> 'pass_lower_omp', it's the 'field_type' that has 'address-space-1', not >> 'obj_type', so indeed Kwok's new code is a no-op: >> >> (gdb) call debug_tree(field_type) >> > type > size >> unit-size >> align:32 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff7686498 precision:32 min max >> pointer_to_this > >> unsigned DI >> size constant 64> >> unit-size constant 8> >> align:64 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff7686b28> >> >> (gdb) call debug_tree(obj_type) >> > size constant 64> >> unit-size constant 8> >> align:64 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff7686bd0 >> fields > type >> unsigned DI size unit-size >> align:64 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff7686b28> >> unsigned DI /home/thomas/shared/gcc/omp/as.c:4:14 size unit-size >> align:64 warn_if_not_align:0 offset_align 128 >> offset >> bit-offset context > reference_to_this > >> >> The case that Kwok's new code handles, however, is when 'obj_type' has a >> non-generic address space, and then propagates that one to 'field_type'. >> >> For a similar OpenACC example, 'omp_build_component_ref' called via GCN >> offloading compilation 'pass_omp_oacc_neuter_broadcast', we've got >> without Kwok's new code: >> >> (gdb) call debug_tree(field_type) >> > size constant 8> >> unit-size constant 1> >> align:8 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff7550b28 precision:1 min max > >> >> (gdb) call debug_tree(obj_type) >> > size constant 8> >> unit-size constant 1> >> align:8 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff7631000 >> fields > type unit-size >> align:8 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff7550b28 precision:1 min max > >> unsigned QI :0:0 size unit-size >> align:8 warn_if_not_align:0 offset_align 64 >> offset >> bit-offset context > >> pointer_to_this > >> >> ..., and with Kwok's new code the 'address-space-4' of 'obj_type' is >> propagated to 'field_type': >> >> (gdb) call debug_tree(field_type) >> > size constant 8> >> unit-size constant 1> >> align:8 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff7631540 precision:1 min max > >> >> I'm not familiar enough with these bits to tell whether Kwok's new code >> is the right solution to this problem -- or if, for example, the problem >> is rather in the SLP vectorizer, where the ICE seems to ultimately >> emerge? >> >> Without (ICEs later) vs. with (works) Kwok's new code, we see the >> 'a.xamdgcn-amdhsa.mkoffload.175t.slp1' dump change as follows (word-diff, >> only additional '', occasionally): >> >> [...] >> {++} vector(2) long int * vectp.58; >> {++} vector(2) long int * vectp_.oacc_worker_o.57; >> {++} vector(2) int * vectp.56; >> {++} vector(2) int * vectp_.oacc_worker_o.55; >> [...] >> {++} long int * _104; >> [...] >> {++} long int * _108; >> [...] >> void * _350; >> [...] >> _350 = __builtin_gcn_single_copy_start (&.oacc_worker_o.6); >> [...] >> MEM <{++} vector(2) long int> [(long int *)&.oacc_worker_o.6] = _101; >> _108 = &.oacc_worker_o.6._22 + 16; >> MEM <{++} vector(2) long int> [(long int *)_108] = _100; >> _104 = &.oacc_worker_o.6._22 + 32; >> [...] >> >> For example, with Kwok's new code, '_108' (' long int *') >> is cast into '(long int *)' -- presumably synthesized in the SLP >> vectorizer? Is that correct or shouldn't that cast also include >> ''? >> >> I see a similar issue has been fixed a while ago: r245772 (Git commit >> c7d97b2846c5647a81548caa3264d77c0a595010) for PR79723 >> "Another case of dropped gs: prefix", changing >> 'gcc/tree-vect-stmts.c:get_vectype_for_scalar_type_and_size' as follows: >> >> + /* Re-attach the address-space qualifier if we canonicalized the scalar >> + type. */ >> + if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) >> + return build_qualified_type >> + (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); >> + >> return vectype; >> >> (It looks a bit like the address space handling is quite fragile in GCC's >> 'tree' types/interfaces? Do we have ideas about how to make that more >> robust, less "bolt-on"?) > > If in doubt always look at what RTL expansion does - it looks like > set_mem_attributes expects the address-space qualifier to be > present on the type or in case it is passed an object, on the > type of the base, or in case of a dereference, on the pointed-to > type of the pointer (and yes, that does look somewhat fragile). > > So it looks like the patch you refer to shouldn't fix anything and > >> + /* Re-attach the address-space qualifier if we canonicalized the scalar >> + type. */ >> + if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) >> + return build_qualified_type >> + (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); > > looks incomplete. What you'd need to look for is MEM_REFs built > by the vectorizer and the address-space information on the pointers, > like generated from vect_create_data_ref_ptr. It might also be that > data-ref analysis / SCEV looks through address-space qualifier changing > casts and thus we pick up the wrong address-space in the end. Aah, more GCC pieces to learn about ;-) -- thanks for the pointers! > What's the testcase that ICEs on trunk? You'll need a GCN offloading build with the attached "[WIP] Reproduce GCN address space vs. SLP vectorization ICEs", run 'make check-target-libgomp', and observe a number of ICEs like: during RTL pass: expand [...]/libgomp.oacc-c-c++-common/loop-red-gwv-1.c: In function 'main._omp_fn.0': [...]/libgomp.oacc-c-c++-common/loop-red-gwv-1.c:19:9: internal compiler error: in convert_memory_address_addr_space_1, at explow.c:301 [...] mkoffload: fatal error: build-gcc/gcc/x86_64-pc-linux-gnu-accel-amdgcn-amdhsa-gcc returned 1 exit status '-O1 -ftree-slp-vectorize' would be sufficient to trigger that one. Run with '-save-temps -v', see the '[...]/build-gcc-offload-amdgcn-amdhsa/gcc/lto1' command ICE: #0 fancy_abort (file=file@entry=0x182e418 "[...]/source-gcc/gcc/explow.c", line=line@entry=301, function=function@entry=0x182e960 "convert_memory_address_addr_space_1") at [...]/source-gcc/gcc/diagnostic.c:1961 #1 0x00000000007ef690 in convert_memory_address_addr_space_1 (to_mode=..., x=x@entry=0x7ffff764fa08, as=as@entry=0 '\000', in_const=in_const@entry=false, no_emit=no_emit@entry=false) at [...]/source-gcc/gcc/explow.c:301 #2 0x00000000007ef6cb in convert_memory_address_addr_space (to_mode=..., x=0x7ffff764fa08, as=as@entry=0 '\000') at [...]/source-gcc/gcc/explow.c:423 #3 0x0000000000812f48 in expand_expr_addr_expr (modifier=EXPAND_SUM, tmode=E_DImode, target=0x0, exp=0x7ffff764a520) at [...]/source-gcc/gcc/expr.c:8535 #4 expand_expr_real_1 (exp=0x7ffff764a520, target=, tmode=, modifier=EXPAND_SUM, alt_rtl=0x0, inner_reference_p=) at [...]/source-gcc/gcc/expr.c:11741 #5 0x0000000000813139 in expand_expr (modifier=EXPAND_SUM, mode=E_VOIDmode, target=0x0, exp=0x7ffff764a520) at [...]/source-gcc/gcc/expr.h:301 #6 expand_expr_real_1 (exp=0x7ffff7649d48, target=, tmode=E_VOIDmode, modifier=EXPAND_WRITE, alt_rtl=0x0, inner_reference_p=) at [...]/source-gcc/gcc/expr.c:10887 #7 0x000000000082475a in expand_expr (modifier=EXPAND_WRITE, mode=E_VOIDmode, target=0x0, exp=0x7ffff7649d48) at [...]/source-gcc/gcc/expr.h:301 #8 expand_assignment (to=to@entry=0x7ffff7649d48, from=from@entry=0x7ffff763a7e0, nontemporal=) at [...]/source-gcc/gcc/expr.c:5732 #9 0x00000000006c807d in expand_gimple_stmt_1 (stmt=stmt@entry=0x7ffff7646aa0) at [...]/source-gcc/gcc/cfgexpand.c:3944 #10 0x00000000006c95c7 in expand_gimple_stmt (stmt=stmt@entry=0x7ffff7646aa0) at [...]/source-gcc/gcc/cfgexpand.c:4040 #11 0x00000000006ce884 in expand_gimple_basic_block (bb=0x7ffff7635dd0, disable_tail_calls=disable_tail_calls@entry=false) at [...]/source-gcc/gcc/cfgexpand.c:6082 #12 0x00000000006d13de in (anonymous namespace)::pass_expand::execute (this=, fun=) at [...]/source-gcc/gcc/cfgexpand.c:6808 [...] (gdb) up #1 0x00000000007ef690 in convert_memory_address_addr_space_1 (to_mode=..., x=x@entry=0x7ffff764fa08, as=as@entry=0 '\000', in_const=in_const@entry=false, no_emit=no_emit@entry=false) at [...]/source-gcc/gcc/explow.c:301 301 gcc_assert (GET_MODE (x) == to_mode || GET_MODE (x) == VOIDmode); (gdb) list 296 rtx x, addr_space_t as ATTRIBUTE_UNUSED, 297 bool in_const ATTRIBUTE_UNUSED, 298 bool no_emit ATTRIBUTE_UNUSED) 299 { 300 #ifndef POINTERS_EXTEND_UNSIGNED 301 gcc_assert (GET_MODE (x) == to_mode || GET_MODE (x) == VOIDmode); 302 return x; 303 #else /* defined(POINTERS_EXTEND_UNSIGNED) */ 304 scalar_int_mode pointer_mode, address_mode, from_mode; 305 rtx temp; (gdb) call debug_rtx(x) (symbol_ref:SI (".oacc_worker_o.13.6") [flags 0x2] ) (gdb) print x->mode $1 = E_SImode (gdb) print to_mode $2 = {m_mode = E_DImode} (gdb) up #2 0x00000000007ef6cb in convert_memory_address_addr_space (to_mode=..., x=0x7ffff764fa08, as=as@entry=0 '\000') at [...]/source-gcc/gcc/explow.c:423 423 return convert_memory_address_addr_space_1 (to_mode, x, as, false, false); (gdb) up #3 0x0000000000812f48 in expand_expr_addr_expr (modifier=EXPAND_SUM, tmode=E_DImode, target=0x0, exp=0x7ffff764a520) at [...]/source-gcc/gcc/expr.c:8535 8535 result = convert_memory_address_addr_space (new_tmode, result, as); (gdb) call debug_tree(exp) unit-size align:32 warn_if_not_align:0 symtab:0 alias-set 4 canonical-type 0x7ffff75505e8 precision:32 min max pointer_to_this > public unsigned DI size unit-size align:64 warn_if_not_align:0 symtab:0 alias-set 1 structural-equality> constant arg:0 unit-size align:32 warn_if_not_align:0 symtab:0 alias-set 5 canonical-type 0x7ffff76215e8 fields pointer_to_this > addressable used static ignored BLK source-gcc/libgomp/testsuite/libgomp.oacc-c-c++-common/loop-red-gwv-1.c:19:9 size unit-size align:128 warn_if_not_align:0 (mem/c:BLK (symbol_ref:SI (".oacc_worker_o.13.6") [flags 0x2] ) [5 .oacc_worker_o.13+0 S8 A128 AS4])>> In 'arg:0' of 'exp' note 'address-space-4' (expected): 'ADDR_SPACE_LDS' (per 'gcc/config/gcn/gcn.h:gcn_address_spaces'). With the attached "[WIP] [GCN] '+#define POINTERS_EXTEND_UNSIGNED 1'", we instead fail as follows: ./a.xamdgcn-amdhsa.mkoffload.2.s:92:23: error: invalid modifier 'rel32@lo' (no symbols present) s_add_u32 s2, s2, 32@rel32@lo+4 ^ ./a.xamdgcn-amdhsa.mkoffload.2.s:92:23: error: failed parsing operand. s_add_u32 s2, s2, 32@rel32@lo+4 ^ ./a.xamdgcn-amdhsa.mkoffload.2.s:93:24: error: invalid modifier 'rel32@hi' (no symbols present) s_addc_u32 s3, s3, 32@rel32@hi+4 ^ ./a.xamdgcn-amdhsa.mkoffload.2.s:93:24: error: failed parsing operand. s_addc_u32 s3, s3, 32@rel32@hi+4 ^ mkoffload: fatal error: build-gcc/gcc/x86_64-pc-linux-gnu-accel-amdgcn-amdhsa-gcc returned 1 exit status ..., so it's not that simple. (I have no clue whether 'POINTERS_EXTEND_UNSIGNED' would make sense for GCN -- but thought it was worth a quick try.) Grüße Thomas >> I did add a few 'assert's for non-generic address space to >> 'gcc/tree-vect*', but have not yet located where things may be going >> wrong. >> >> >> > I think keeping the qual addr space here is the wrong thing to do, >> > it should keep the other quals and clear the address space instead, >> > the whole struct is going to be in generic addres space, isn't it? >> >> Correct for 'omp_build_component_ref' called via host compilation >> 'pass_lower_omp', but in the case of 'omp_build_component_ref' called via >> GCN offloading compilation 'pass_omp_oacc_neuter_broadcast', 'obj_type' >> has a non-generic address space. >> >> However, regarding the former comment -- shouldn't we force generic >> address space for all 'tree' types read in via LTO streaming for >> offloading compilation? I assume that (in the general case) address >> spaces are never compatible between host and offloading compilation? >> For the attached "Add 'libgomp.c/address-space-1.c'", propagating the >> '__seg_fs' address space across the offloading boundary (assuming I did >> interpret the dumps correctly) doesn't seem to cause any problems, but >> maybe it's problematic for other cases? (This is, however, a separate >> issue from what I'm discussing here.) >> >> >> >> + tree ret = build3 (COMPONENT_REF, field_type, obj, field, NULL); >> >> + if (TREE_THIS_VOLATILE (field)) >> >> + TREE_THIS_VOLATILE (ret) |= 1; >> >> + if (TREE_READONLY (field)) >> >> + TREE_READONLY (ret) |= 1; >> > >> > When touching these two, shouldn't it be better written as >> > = 1; instead of |= 1; ? For a bitfield... >> >> Yes, that was just copied from the original >> 'gcc/omp-general.c:omp_build_component_ref' -- but happy to simplify >> that, of course. >> >> >> Grüße >> Thomas >> >> >> ----------------- >> Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955 ----------------- Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955