Closed kimwalisch closed 8 years ago
You'd think so, but (at least on x86) the mask is actually not emitted at all, because the shift instruction only looks at the bottom bits. So the mask currently costs 0 instructions.
This isn't true for the vector case, so we can leave this open to track that.
Are you sure about this?
I compiled both versions using GCC -O2 -S and the 2nd version t >> (denom->more & LIBDIVIDE_64_SHIFT_MASK)
contains many andl $63, %ecx
instructions whereas the 1st version t >> denom->more
contains none!
Here is the assembly diff of both versions (1st version t >> denom->more
, 2nd version t >> (denom->more & LIBDIVIDE_64_SHIFT_MASK)
):
$ diff S2_easy_libdivide_new.s S2_easy_libdivide_old.s
333a334
> andl $63, %ecx
356a358
> andl $63, %ecx
398a401
> andl $63, %ecx
791a795
> andl $63, %ecx
811,815c815,816
< leaq (%rcx,%rdx), %rax
< movq 48(%rsp), %rcx
< subq %rax, %rcx
< shrq %rcx
< addq %rcx, %rax
---
> movq 48(%rsp), %rax
> addq %rcx, %rdx
816a818,821
> subq %rdx, %rax
> shrq %rax
> addq %rdx, %rax
> andl $63, %ecx
867a873
> andl $63, %ecx
1409a1416
> andl $63, %ecx
1429,1433c1436,1437
< leaq (%rcx,%rdx), %rax
< movq 48(%rsp), %rcx
< subq %rax, %rcx
< shrq %rcx
< addq %rcx, %rax
---
> movq 48(%rsp), %rax
> addq %rcx, %rdx
1434a1439,1442
> subq %rdx, %rax
> shrq %rax
> addq %rdx, %rax
> andl $63, %ecx
1485a1494
> andl $63, %ecx
So I still think you could save one andl $63, %ecx
instruction in libdivide_u64_branchfree_do()
. But I also think that it would not be faster because today's out of order CPUs can execute up to 4 independent instructions simultaneously.
Ugh, you're right, clang is getting it right, but gcc is being stupid here. Ok, I guess we ought to fix this. Nice catch!
Fixed by #13
I have compiled my code using the very latest branchfree libdivide.h
using GCC 5.2 and -std=c++11
and got some warnings which have been introduced by my Save 1 instruction in libdivide_u64_branchfree_do()
code changes:
g++ -std=c++11 -DPACKAGE_NAME=\"primecount\" -DPACKAGE_TARNAME=\"primecount\" -DPACKAGE_VERSION=\"3.1\" "-DPACKAGE_STRING=\"primecount 3.1\"" -DPACKAGE_BUGREPORT=\"kim.walisch@gmail.com\" -DPACKAGE_URL=\"\" -DPACKAGE=\"primecount\" -DVERSION=\"3.1\" -DSTDC_HEADERS=1 -DHAVE_SYS_TYPES_H=1 -DHAVE_SYS_STAT_H=1 -DHAVE_STDLIB_H=1 -DHAVE_STRING_H=1 -DHAVE_MEMORY_H=1 -DHAVE_STRINGS_H=1 -DHAVE_INTTYPES_H=1 -DHAVE_STDINT_H=1 -DHAVE_UNISTD_H=1 -DHAVE_DLFCN_H=1 -DLT_OBJDIR=\".libs/\" -DNDEBUG=1 -DHAVE___BUILTIN_POPCOUNT=1 -DHAVE___BUILTIN_POPCOUNTLL=1 -DHAVE___INT128_T=1 -DHAVE_LIBDIVIDE=1 -I. -I./include -fopenmp -mpopcnt -DHAVE_POPCNT=1 -O2 -Iprimesieve-5.6.0/include -MT src/libprimecount_la-phi_libdivide.lo -MD -MP -MF src/.deps/libprimecount_la-phi_libdivide.Tpo -c src/phi_libdivide.cpp -fPIC -DPIC -o src/.libs/libprimecount_la-phi_libdivide.o
In file included from src/phi_libdivide.cpp:22:0:
./include/libdivide.h: In function ‘{anonymous}::libdivide::libdivide_u32_branchfree_t {anonymous}::libdivide::libdivide_u32_branchfree_gen(uint32_t)’:
./include/libdivide.h:774:66: warning: narrowing conversion of ‘(((int)tmp.{anonymous}::libdivide::libdivide_u32_t::more) & 31)’ from ‘int’ to ‘uint8_t {aka unsigned char}’ inside { } [-Wnarrowing]
struct libdivide_u32_branchfree_t ret = {tmp.magic, tmp.more & LIBDIVIDE_32_SHIFT_MASK};
^
./include/libdivide.h: In function ‘uint32_t {anonymous}::libdivide::libdivide_u32_branchfree_recover(const {anonymous}::libdivide::libdivide_u32_branchfree_t*)’:
./include/libdivide.h:833:67: warning: narrowing conversion of ‘(int)(((unsigned char)denom->{anonymous}::libdivide::libdivide_u32_branchfree_t::more) | 64u)’ from ‘int’ to ‘uint8_t {aka unsigned char}’ inside { } [-Wnarrowing]
struct libdivide_u32_t denom_u32 = {denom->magic, denom->more | LIBDIVIDE_ADD_MARKER};
^
./include/libdivide.h: In function ‘{anonymous}::libdivide::libdivide_u64_branchfree_t {anonymous}::libdivide::libdivide_u64_branchfree_gen(uint64_t)’:
./include/libdivide.h:969:66: warning: narrowing conversion of ‘(((int)tmp.{anonymous}::libdivide::libdivide_u64_t::more) & 63)’ from ‘int’ to ‘uint8_t {aka unsigned char}’ inside { } [-Wnarrowing]
struct libdivide_u64_branchfree_t ret = {tmp.magic, tmp.more & LIBDIVIDE_64_SHIFT_MASK};
^
./include/libdivide.h: In function ‘uint64_t {anonymous}::libdivide::libdivide_u64_branchfree_recover(const {anonymous}::libdivide::libdivide_u64_branchfree_t*)’:
./include/libdivide.h:1039:67: warning: narrowing conversion of ‘(int)(((unsigned char)denom->{anonymous}::libdivide::libdivide_u64_branchfree_t::more) | 64u)’ from ‘int’ to ‘uint8_t {aka unsigned char}’ inside { } [-Wnarrowing]
struct libdivide_u64_t denom_u64 = {denom->magic, denom->more | LIBDIVIDE_ADD_MARKER};
In order to fix the warnings we must add a (uint8_t)
cast at lines 774, 833, 969 and 1039.
For example:
struct libdivide_u64_t denom_u64 = {denom->magic, (uint8_t)(denom->more | LIBDIVIDE_ADD_MARKER)};
I guess you could decrease the instruction count by 1 in libdivide_u64_branchfree_do() by using:
instead of:
This would require a different initialization algorithm for libdivide_u64_branchfree_do(). I am not sure if it is worth it.