From f019c49e380f764c1ead36fe3602184804292711 Mon Sep 17 00:00:00 2001 From: Daniel Dragan Date: Wed, 24 Oct 2012 16:15:51 -0400 Subject: optimize memory wrap croaks, often used in MEM_WRAP_CHECK Most perls are built with PERL_MALLOC_WRAP. This causes MEM_WRAP_CHECK macro to perform some checks on the requested allocation size in macro Newx. The checks are performed at the caller, not in the callee (for me on Win32 perl the callee in Newx is Perl_safesysmalloc) of Newx. If the check fails a "Perl_croak_nocontext("%s",PL_memory_wrap)" is done. In x86 machine code, "if(bad_alloc) Perl_croak_nocontext("%s",PL_memory_wrap); will be written as "cond jmp ahead ~15 bytes", "push const pointer", "push const pointer", "call const pointer". For each Newx where the allocation amount was not a constant (constant folding would remove the croak memory wrap branch compleatly), the branch takes 15-19 bytes depending on x86 compiler. There are about 80 Newx'es in the interp (win32 dynamic linking perl) that do the memory wrap check and have a "Perl_croak_nocontext("%s",PL_memory_wrap)" in them after all optimizations by the compiler. This patch reduces the memory wrap branch from 15-19 to 5 bytes on x86. Since croak_memory_wrap is a static and a noreturn, a compiler with IPO may optimize the whole branch to "cond jmp 32 bits relative" at each callsite. A less optimal complier may do "cond jmp 8 bits relative (jump past the "call S_croak_memory_wrap" instruction), then "call S_croak_memory_wrap". Both ways are better than the current situation. The reason why croak_memory_wrap is a static and not an export is that the compiler has more opportunity to optimize/reduce the impact of the memory wrap branch at the call site if the target is in the same image rather than in a different image, which would require using the platform specific dynamic linking mechanism/export table/etc, which often requires a new stack frame per ABI of the platform. If a dynamic linked XS module does not use S_croak_memory_wrap it will be removed from the image by the C compiler. If it is included in the XS image, it is a very small block of code and a 3 byte string litteral. A CPU cache line is typically 32 or 64 bytes and a memory read is typically 16. Cutting the instructions by 10 to 16 bytes out of "hot code" (10 of the ~80 call sites are pp_*) is a worthy goal. In a few places the memory wrap croak is used explictly, not from a MEM_WRAP_CHECK, this patch converts those to use the static. If PERL_MALLOC_WRAP is undef, there are still a couple uses of croak memory wrap, so do not keep S_croak_memory_wrap in a ifdef PERL_MALLOC_WRAP. Also see http://www.nntp.perl.org/group/perl.perl5.porters/2012/10/msg194383.html and [perl #115456]. --- handy.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'handy.h') diff --git a/handy.h b/handy.h index 841dd9066c..3b18a0f72c 100644 --- a/handy.h +++ b/handy.h @@ -1141,13 +1141,13 @@ PoisonWith(0xEF) for catching access to freed memory. * overly eager compilers that will bleat about e.g. * (U16)n > (size_t)~0/sizeof(U16) always being false. */ #ifdef PERL_MALLOC_WRAP -#define MEM_WRAP_CHECK(n,t) MEM_WRAP_CHECK_1(n,t,PL_memory_wrap) +#define MEM_WRAP_CHECK(n,t) \ + (void)(sizeof(t) > 1 && ((MEM_SIZE)(n)+0.0) > MEM_SIZE_MAX/sizeof(t) && (croak_memory_wrap(),0)) #define MEM_WRAP_CHECK_1(n,t,a) \ (void)(sizeof(t) > 1 && ((MEM_SIZE)(n)+0.0) > MEM_SIZE_MAX/sizeof(t) && (Perl_croak_nocontext("%s",(a)),0)) #define MEM_WRAP_CHECK_(n,t) MEM_WRAP_CHECK(n,t), -#define PERL_STRLEN_ROUNDUP(n) ((void)(((n) > MEM_SIZE_MAX - 2 * PERL_STRLEN_ROUNDUP_QUANTUM) ? (Perl_croak_nocontext("%s",PL_memory_wrap),0):0),((n-1+PERL_STRLEN_ROUNDUP_QUANTUM)&~((MEM_SIZE)PERL_STRLEN_ROUNDUP_QUANTUM-1))) - +#define PERL_STRLEN_ROUNDUP(n) ((void)(((n) > MEM_SIZE_MAX - 2 * PERL_STRLEN_ROUNDUP_QUANTUM) ? (croak_memory_wrap(),0):0),((n-1+PERL_STRLEN_ROUNDUP_QUANTUM)&~((MEM_SIZE)PERL_STRLEN_ROUNDUP_QUANTUM-1))) #else #define MEM_WRAP_CHECK(n,t) -- cgit v1.2.1