try_store_by_multiple_pieces was added not long ago, enabling variable-sized memset to be expanded inline when the worst-case in-range constant length would, using conditional blocks with powers of two to cover all possibilities of length and alignment.
This patch extends the memset expansion to start with a loop, so as to still take advantage of known alignment even with long lengths, but without necessarily adding store blocks for every power of two. This makes it possible for any memset call to be expanded, even if storing a single byte per iteration. Surely efficient implementations of memset can do better, with a pre-loop to increase alignment, but that would likely be excessive for inline expansions of memset. Still, in some cases, users prefer to inline memset, even if it's not as performant, or when it's known to be performant in ways the compiler can't tell, to avoid depending on a C runtime library. With this flag, global or per-function optimizations may enable inline expansion of memset to be selectively requested, while the infrastructure for that may enable us to introduce per-target tuning to enable such looping even advantageous, even if not explicitly requested. I realize this is late for new features in this cycle; I`d be happy to submit it again later, but I wonder whether there's any interest in this feature, or any objections to it. FWIW, I've regstrapped this on x86_64-linux-gnu, and also tested earlier versions of this patch in earlier GCC branches with RISC-v crosses. Is this ok for GCC 14? Maybe even simple enough for GCC 13, considering it's disabled by default? TIA, for gcc/ChangeLog * builtins.cc (try_store_by_multiple_pieces): Support starting with a loop. * common.opt (finline-memset-loops): New. * doc/invoke.texi (-finline-memset-loops): Add. for gcc/testsuite/ChangeLog * gcc.dg/torture/inline-mem-set-1.c: New. --- gcc/builtins.cc | 50 ++++++++++++++++++++++- gcc/common.opt | 4 ++ gcc/doc/invoke.texi | 13 ++++++ gcc/testsuite/gcc.dg/torture/inline-mem-set-1.c | 14 ++++++ 4 files changed, 77 insertions(+), 4 deletions(-) create mode 100644 gcc/testsuite/gcc.dg/torture/inline-mem-set-1.c diff --git a/gcc/builtins.cc b/gcc/builtins.cc index 02c4fefa86f48..388bae58ce49e 100644 --- a/gcc/builtins.cc +++ b/gcc/builtins.cc @@ -4361,9 +4361,37 @@ try_store_by_multiple_pieces (rtx to, rtx len, unsigned int ctz_len, if (max_bits >= 0) xlenest += ((HOST_WIDE_INT_1U << max_bits) * 2 - (HOST_WIDE_INT_1U << ctz_len)); + bool max_loop = false; if (!can_store_by_pieces (xlenest, builtin_memset_read_str, &valc, align, true)) - return false; + { + if (!flag_inline_memset_loops) + return false; + while (--max_bits >= sctz_len) + { + xlenest = ((HOST_WIDE_INT_1U << max_bits) * 2 + - (HOST_WIDE_INT_1U << ctz_len)); + if (can_store_by_pieces (xlenest + blksize, + builtin_memset_read_str, + &valc, align, true)) + { + max_loop = true; + break; + } + if (!blksize) + continue; + if (can_store_by_pieces (xlenest, + builtin_memset_read_str, + &valc, align, true)) + { + blksize = 0; + max_loop = true; + break; + } + } + if (!max_loop) + return false; + } by_pieces_constfn constfun; void *constfundata; @@ -4405,6 +4433,7 @@ try_store_by_multiple_pieces (rtx to, rtx len, unsigned int ctz_len, the least significant bit possibly set in the length. */ for (int i = max_bits; i >= sctz_len; i--) { + rtx_code_label *loop_label = NULL; rtx_code_label *label = NULL; blksize = HOST_WIDE_INT_1U << i; @@ -4423,14 +4452,24 @@ try_store_by_multiple_pieces (rtx to, rtx len, unsigned int ctz_len, else if ((max_len & blksize) == 0) continue; + if (max_loop && i == max_bits) + { + loop_label = gen_label_rtx (); + emit_label (loop_label); + /* Since we may run this multiple times, don't assume we + know anything about the offset. */ + clear_mem_offset (to); + } + /* Issue a store of BLKSIZE bytes. */ + bool update_needed = i != sctz_len || loop_label; to = store_by_pieces (to, blksize, constfun, constfundata, align, true, - i != sctz_len ? RETURN_END : RETURN_BEGIN); + update_needed ? RETURN_END : RETURN_BEGIN); /* Adjust REM and PTR, unless this is the last iteration. */ - if (i != sctz_len) + if (update_needed) { emit_move_insn (ptr, force_operand (XEXP (to, 0), NULL_RTX)); to = replace_equiv_address (to, ptr); @@ -4438,6 +4477,11 @@ try_store_by_multiple_pieces (rtx to, rtx len, unsigned int ctz_len, emit_move_insn (rem, force_operand (rem_minus_blksize, NULL_RTX)); } + if (loop_label) + emit_cmp_and_jump_insns (rem, GEN_INT (blksize), GE, NULL, + ptr_mode, 1, loop_label, + profile_probability::likely ()); + if (label) { emit_label (label); diff --git a/gcc/common.opt b/gcc/common.opt index 562d73d7f552a..c28af170be896 100644 --- a/gcc/common.opt +++ b/gcc/common.opt @@ -1874,6 +1874,10 @@ finline-atomics Common Var(flag_inline_atomics) Init(1) Optimization Inline __atomic operations when a lock free instruction sequence is available. +finline-memset-loops +Common Var(flag_inline_memset_loops) Init(0) Optimization +Inline memset even if it requires loops. + fcf-protection Common RejectNegative Alias(fcf-protection=,full) diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi index da9ad1068fbf6..19f436ad46385 100644 --- a/gcc/doc/invoke.texi +++ b/gcc/doc/invoke.texi @@ -548,7 +548,8 @@ Objective-C and Objective-C++ Dialects}. -fgcse-sm -fhoist-adjacent-loads -fif-conversion @gol -fif-conversion2 -findirect-inlining @gol -finline-functions -finline-functions-called-once -finline-limit=@var{n} @gol --finline-small-functions -fipa-modref -fipa-cp -fipa-cp-clone @gol +-finline-memset-loops -finline-small-functions @gol +-fipa-modref -fipa-cp -fipa-cp-clone @gol -fipa-bit-cp -fipa-vrp -fipa-pta -fipa-profile -fipa-pure-const @gol -fipa-reference -fipa-reference-addressable @gol -fipa-stack-alignment -fipa-icf -fira-algorithm=@var{algorithm} @gol @@ -11960,6 +11961,16 @@ in its own right. Enabled at levels @option{-O1}, @option{-O2}, @option{-O3} and @option{-Os}, but not @option{-Og}. +@item -finline-memset-loops +@opindex finline-memset-loops +Expand @code{memset} calls inline, even when the length is variable or +big enough as to require looping. This may enable the compiler to take +advantage of known alignment and length multipliers, but it will often +generate code that is less efficient than performant implementations of +@code{memset}, and grow code size so much that even a less performant +@code{memset} may run faster due to better use of the code cache. This +option is disabled by default. + @item -fearly-inlining @opindex fearly-inlining Inline functions marked by @code{always_inline} and functions whose body seems diff --git a/gcc/testsuite/gcc.dg/torture/inline-mem-set-1.c b/gcc/testsuite/gcc.dg/torture/inline-mem-set-1.c new file mode 100644 index 0000000000000..73bd1025f191f --- /dev/null +++ b/gcc/testsuite/gcc.dg/torture/inline-mem-set-1.c @@ -0,0 +1,14 @@ +/* { dg-do compile } */ +/* { dg-options "-finline-memset-loops -gno-record-gcc-switches -fno-lto" } */ + +void *zero (unsigned long long (*p)[32], int n) +{ + return __builtin_memset (p, 0, n * sizeof (*p)); +} + +void *ones (char (*p)[128], int n) +{ + return __builtin_memset (p, -1, n * sizeof (*p)); +} + +/* { dg-final { scan-assembler-not "memset" } } */ -- Alexandre Oliva, happy hacker https://FSFLA.org/blogs/lxo/ Free Software Activist GNU Toolchain Engineer Disinformation flourishes because many people care deeply about injustice but very few check the facts. Ask me about <https://stallmansupport.org>