From ba837323dbda2bca5a1c8a4c78092a88241dcfa3 Mon Sep 17 00:00:00 2001 From: Jakub Jelinek Date: Wed, 6 Oct 2021 10:40:12 +0200 Subject: [PATCH] openmp: Optimize for OpenMP atomics 2x__builtin_clear_padding+__builtin_memcmp if possible For the few long double types that do have padding bits, e.g. on x86 the clear_type_padding_in_mask computed mask is ff ff ff ff ff ff ff ff ff ff 00 00 for 32-bit and ff ff ff ff ff ff ff ff ff ff 00 00 00 00 00 00 for 64-bit. Instead of doing __builtin_clear_padding on both operands that will clear the last 2 or 6 bytes and then memcmp on the whole 12/16 bytes, we can just memcmp 10 bytes. The code also handles if the padding would be at the start or both at the start and end, but everything on byte boundaries only and non-padding bits being contiguous. This works around a tree-ssa-dse.c bug (but we need to fix it anyway, as libstdc++ won't do this and as it can deal with arbitrary types, it even can't do that generally). 2021-10-06 Jakub Jelinek PR tree-optimization/102571 * c-omp.c (c_finish_omp_atomic): Optimize the case where type has padding, but the non-padding bits are contiguous set of bytes by adjusting the memcmp call arguments instead of emitting __builtin_clear_padding and then comparing all the type's bytes. --- gcc/c-family/c-omp.c | 50 ++++++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 48 insertions(+), 2 deletions(-) diff --git a/gcc/c-family/c-omp.c b/gcc/c-family/c-omp.c index 2849fda..d8b9855 100644 --- a/gcc/c-family/c-omp.c +++ b/gcc/c-family/c-omp.c @@ -379,6 +379,8 @@ c_finish_omp_atomic (location_t loc, enum tree_code code, if (SCALAR_FLOAT_TYPE_P (cmptype) && !test) { bool clear_padding = false; + HOST_WIDE_INT non_padding_start = 0; + HOST_WIDE_INT non_padding_end = 0; if (BITS_PER_UNIT == 8 && CHAR_BIT == 8) { HOST_WIDE_INT sz = int_size_in_bytes (cmptype), i; @@ -392,6 +394,40 @@ c_finish_omp_atomic (location_t loc, enum tree_code code, clear_padding = true; break; } + if (clear_padding && buf[i] == 0) + { + /* Try to optimize. In the common case where + non-padding bits are all continuous and start + and end at a byte boundary, we can just adjust + the memcmp call arguments and don't need to + emit __builtin_clear_padding calls. */ + if (i == 0) + { + for (i = 0; i < sz; i++) + if (buf[i] != 0) + break; + if (i < sz && buf[i] == (unsigned char) ~0) + { + non_padding_start = i; + for (; i < sz; i++) + if (buf[i] != (unsigned char) ~0) + break; + } + else + i = 0; + } + if (i != 0) + { + non_padding_end = i; + for (; i < sz; i++) + if (buf[i] != 0) + { + non_padding_start = 0; + non_padding_end = 0; + break; + } + } + } } tree inttype = NULL_TREE; if (!clear_padding && tree_fits_uhwi_p (TYPE_SIZE (cmptype))) @@ -428,12 +464,22 @@ c_finish_omp_atomic (location_t loc, enum tree_code code, tmp2 = build4 (TARGET_EXPR, cmptype, tmp2, TREE_OPERAND (rhs1, 1), NULL, NULL); tmp2 = build1 (ADDR_EXPR, pcmptype, tmp2); + if (non_padding_start) + { + tmp1 = build2 (POINTER_PLUS_EXPR, pcmptype, tmp1, + size_int (non_padding_start)); + tmp2 = build2 (POINTER_PLUS_EXPR, pcmptype, tmp2, + size_int (non_padding_start)); + } tree fndecl = builtin_decl_explicit (BUILT_IN_MEMCMP); rhs1 = build_call_expr_loc (loc, fndecl, 3, tmp1, tmp2, - TYPE_SIZE_UNIT (cmptype)); + non_padding_end + ? size_int (non_padding_end + - non_padding_start) + : TYPE_SIZE_UNIT (cmptype)); rhs1 = build2 (EQ_EXPR, boolean_type_node, rhs1, integer_zero_node); - if (clear_padding) + if (clear_padding && non_padding_end == 0) { fndecl = builtin_decl_explicit (BUILT_IN_CLEAR_PADDING); tree cp1 = build_call_expr_loc (loc, fndecl, 1, tmp1); -- 2.7.4