[from internal] Hi! This concerns a class of ICEs seen as of og10 branch with the "openacc: Middle-end worker-partitioning support" and "amdgcn: Enable OpenACC worker partitioning for AMD GCN" changes applied: On 2020-06-06T16:07:36+0100, Kwok Cheung Yeung wrote: > On 01/06/2020 8:48 pm, Kwok Cheung Yeung wrote: >> On 21/05/2020 10:23 pm, Kwok Cheung Yeung wrote: >>> These all have the same failure mode: >>> >>> during RTL pass: expand >>> [...]/libgomp/testsuite/libgomp.oacc-fortran/parallel-dims.f90: In function 'MAIN__._omp_fn.1': >>> [...]/libgomp/testsuite/libgomp.oacc-fortran/parallel-dims.f90:86: internal compiler error: in convert_memory_address_addr_space_1, at explow.c:302 >>> 0xc29f20 convert_memory_address_addr_space_1(scalar_int_mode, rtx_def*, unsigned char, bool, bool) >>> [...]/gcc/explow.c:302 >>> 0xc29f57 convert_memory_address_addr_space(scalar_int_mode, rtx_def*, unsigned char) >>> [...]/gcc/explow.c:404 >>> [...] >>> This occurs if the -ftree-slp-vectorize flag is specified (default at -O3). >> The problematic bit of Gimple code is this: >> >> .oacc_worker_o.44._120 = gangs_min_472; >> .oacc_worker_o.44._122 = workers_min_473; >> .oacc_worker_o.44._124 = vectors_min_474; >> .oacc_worker_o.44._126 = gangs_max_475; >> .oacc_worker_o.44._128 = workers_max_476; >> .oacc_worker_o.44._130 = vectors_max_477; >> .oacc_worker_o.44._132 = 0; >> >> With SLP vectorization enabled, it becomes this: >> >> _40 = {gangs_min_472, workers_min_473, vectors_min_474, gangs_max_475}; >> ... >> MEM [(int *)&.oacc_worker_o.44] = _40; >> .oacc_worker_o.44._128 = workers_max_476; >> .oacc_worker_o.44._130 = vectors_max_477; >> .oacc_worker_o.44._132 = 0; >> >> The optimization is trying to transform 4 separate assignments into a single >> memory operation. The trouble is that &o.acc_worker_o is an SImode pointer in >> AS4 (LDS), while the memory expression appears to be in the default memory >> space. The 'to' expression of the assignment is: >> >> > type > type > size >> unit-size >> align:32 warn_if_not_align:0 symtab:0 alias-set 1 canonical-type 0x7ffff73195e8 precision:32 min max >> pointer_to_this reference_to_this > >> TI >> size >> unit-size >> align:128 warn_if_not_align:0 symtab:0 alias-set 1 structural-equality nunits:4 >> pointer_to_this > >> >> arg:0 > type >> public unsigned DI >> size >> unit-size >> align:64 warn_if_not_align:0 symtab:0 alias-set 2 structural-equality> >> constant >> arg:0 >> addressable used static ignored BLK [...]/libgomp/testsuite/libgomp.oacc-fortran/parallel-dims.f90:86:0 >> >> size >> unit-size >> align:128 warn_if_not_align:0 >> (mem/c:BLK (symbol_ref:SI (".oacc_worker_o.44.14") [flags 0x2] ) [9 .oacc_worker_o.44+0 S28 A128 AS4])>> >> arg:1 constant 0>> >> >> In convert_memory_address_addr_space_1: >> >> #ifndef POINTERS_EXTEND_UNSIGNED >> gcc_assert (GET_MODE (x) == to_mode || GET_MODE (x) == VOIDmode); >> return x; >> #else /* defined(POINTERS_EXTEND_UNSIGNED) */ >> >> POINTERS_EXTEND_UNSIGNED is not defined, so it hits the assert. The expected >> to_mode is DI_mode, but x is SI_mode, so the assert fires. > I now have a fix for this. > > > MEM [(int *)&.oacc_worker_o.44] = _40; > > The ICE occurs because the SLP vectorization pass creates the new statement > using the type of the expression '&.oacc_worker_o.44', which is a pointer to a > component ref in the default address space. The expand pass gets confused > because it is handed an SImode pointer (for LDS) when it is expecting a DImode > pointer (for flat/global space). > > The underlying problem is that although .oacc_worker_o is in the correct address > space, the component ref .oacc_worker_o is not. I fixed this by propagating the > address space of .oacc_worker_o when the component ref is created. > static tree > oacc_build_component_ref (tree obj, tree field) > { > - tree ret = build3 (COMPONENT_REF, TREE_TYPE (field), obj, field, NULL); > + 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))); > + > + tree ret = build3 (COMPONENT_REF, field_type, obj, field, NULL); > if (TREE_THIS_VOLATILE (field)) > TREE_THIS_VOLATILE (ret) |= 1; > if (TREE_READONLY (field)) This code change has been included in the recent master branch commit e2a58ed6dc5293602d0d168475109caa81ad0f0d "openacc: Middle-end worker-partitioning support", which thus includes a 'gcc/omp-oacc-neuter-broadcast.cc:oacc_build_component_ref' that is slightly different from 'gcc/omp-low.c:omp_build_component_ref'. I'm confirming that with this reverted, we're seeing ICEs as follows: +FAIL: libgomp.oacc-fortran/gemm-2.f90 [...] -foffload=amdgcn-amdhsa -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions (internal compiler error) +FAIL: libgomp.oacc-fortran/gemm-2.f90 [...] -foffload=amdgcn-amdhsa -O3 -g (internal compiler error) +FAIL: libgomp.oacc-fortran/gemm.f90 [...] -foffload=amdgcn-amdhsa -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions (internal compiler error) +FAIL: libgomp.oacc-fortran/gemm.f90 [...] -foffload=amdgcn-amdhsa -O3 -g (internal compiler error) +FAIL: libgomp.oacc-fortran/optional-reduction.f90 [...] -foffload=amdgcn-amdhsa -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions (internal compiler error) +FAIL: libgomp.oacc-fortran/optional-reduction.f90 [...] -foffload=amdgcn-amdhsa -O3 -g (internal compiler error) +FAIL: libgomp.oacc-fortran/private-variables.f90 [...] -foffload=amdgcn-amdhsa -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions (internal compiler error) +FAIL: libgomp.oacc-fortran/private-variables.f90 [...] -foffload=amdgcn-amdhsa -O3 -g (internal compiler error) +FAIL: libgomp.oacc-fortran/reduction-1.f90 [...] -foffload=amdgcn-amdhsa -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions (internal compiler error) +FAIL: libgomp.oacc-fortran/reduction-1.f90 [...] -foffload=amdgcn-amdhsa -O3 -g (internal compiler error) +FAIL: libgomp.oacc-fortran/reduction-5.f90 [...] -foffload=amdgcn-amdhsa -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions (internal compiler error) +FAIL: libgomp.oacc-fortran/reduction-5.f90 [...] -foffload=amdgcn-amdhsa -O3 -g (internal compiler error) +FAIL: libgomp.oacc-fortran/reduction-6.f90 [...] -foffload=amdgcn-amdhsa -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions (internal compiler error) +FAIL: libgomp.oacc-fortran/reduction-6.f90 [...] -foffload=amdgcn-amdhsa -O3 -g (internal compiler error) 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). Would it make sense to "Re-unify 'omp_build_component_ref' and 'oacc_build_component_ref'", see attached? 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