| Message ID | 91eb94bc0b91c8d1bda7f8fd776d94113acf8cd3.1718347699.git.ps@pks.im (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | d4d364b2c7442a1dd5bec8ab818791ec8769c3e9 |
| Headers | show |
| Series | Introduce `USE_THE_REPOSITORY_VARIABLE` macro | expand |
Patrick Steinhardt <ps@pks.im> writes: > With the preceding commit, the hash array of object IDs is now fully > zero-padded even when the hash algorithm's output is smaller than the > array length. With that, we can now adapt both `oidcmp()` and `oideq()` > to unconditionally memcmp(3P) the whole array instead of depending on > the hash size. > > While it may feel inefficient to compare unused bytes for e.g. SHA-1, in > practice the compiler should now be able to produce code that is better > optimized both because we have no branch anymore, but also because the > size to compare is now known at compile time. Goldbolt spits out the > following assembly on an x86_64 platform with GCC 14.1 for the old and > new implementations of `oidcmp()`: > > oidcmp_old: > movsx rax, DWORD PTR [rdi+32] > test eax, eax > jne .L2 > mov rax, QWORD PTR the_repository[rip] > cmp QWORD PTR [rax+16], 32 > je .L6 > .L4: > mov edx, 20 > jmp memcmp > .L2: > lea rdx, [rax+rax*2] > lea rax, [rax+rdx*4] > lea rax, hash_algos[0+rax*8] > cmp QWORD PTR [rax+16], 32 > jne .L4 > .L6: > mov edx, 32 > jmp memcmp > > oidcmp_new: > mov edx, 32 > jmp memcmp > > The new implementation gets ridi of all the branches and effectively s/ridi/rid > only ends setting up `edx` for `memcmp()` and then calling it. > > And for `oideq()`: > > oideq_old: > movsx rcx, DWORD PTR [rdi+32] > mov rax, rdi > mov rdx, rsi > test ecx, ecx > jne .L2 > mov rcx, QWORD PTR the_repository[rip] > cmp QWORD PTR [rcx+16], 32 > mov rcx, QWORD PTR [rax] > je .L12 > .L4: > mov rsi, QWORD PTR [rax+8] > xor rcx, QWORD PTR [rdx] > xor rsi, QWORD PTR [rdx+8] > or rcx, rsi > je .L13 > .L8: > mov eax, 1 > test eax, eax > sete al > movzx eax, al > ret > .L2: > lea rsi, [rcx+rcx*2] > lea rcx, [rcx+rsi*4] > lea rcx, hash_algos[0+rcx*8] > cmp QWORD PTR [rcx+16], 32 > mov rcx, QWORD PTR [rax] > jne .L4 > .L12: > mov rsi, QWORD PTR [rax+8] > xor rcx, QWORD PTR [rdx] > xor rsi, QWORD PTR [rdx+8] > or rcx, rsi > jne .L8 > mov rcx, QWORD PTR [rax+16] > mov rax, QWORD PTR [rax+24] > xor rcx, QWORD PTR [rdx+16] > xor rax, QWORD PTR [rdx+24] > or rcx, rax > jne .L8 > xor eax, eax > .L14: > test eax, eax > sete al > movzx eax, al > ret > .L13: > mov edi, DWORD PTR [rdx+16] > cmp DWORD PTR [rax+16], edi > jne .L8 > xor eax, eax > jmp .L14 > > oideq_new: > mov rax, QWORD PTR [rdi] > mov rdx, QWORD PTR [rdi+8] > xor rax, QWORD PTR [rsi] > xor rdx, QWORD PTR [rsi+8] > or rax, rdx > je .L5 > .L2: > mov eax, 1 > xor eax, 1 > ret > .L5: > mov rax, QWORD PTR [rdi+16] > mov rdx, QWORD PTR [rdi+24] > xor rax, QWORD PTR [rsi+16] > xor rdx, QWORD PTR [rsi+24] > or rax, rdx > jne .L2 > xor eax, eax > xor eax, 1 > ret > > Interestingly, the compiler decides to split the comparisons into two so > that it first compares the lower half of the object ID for equality and > then the upper half. If the first check shows a difference, then we > wouldn't even end up comparing the second half. > > In both cases, the new generated code is significantly shorter and has > way less branches. While I didn't benchmark the change, I'd be surprised > if the new code was slower. > This was nice to read, thanks for adding the ASM here. > Signed-off-by: Patrick Steinhardt <ps@pks.im> > --- > hash-ll.h | 10 ++++++++++ > hash.h | 20 -------------------- > 2 files changed, 10 insertions(+), 20 deletions(-) > > diff --git a/hash-ll.h b/hash-ll.h > index b72f84f4ae..b04fe12aef 100644 > --- a/hash-ll.h > +++ b/hash-ll.h > @@ -278,6 +278,16 @@ static inline void hashclr(unsigned char *hash, const struct git_hash_algo *algo > memset(hash, 0, algop->rawsz); > } > > +static inline int oidcmp(const struct object_id *oid1, const struct object_id *oid2) > +{ > + return memcmp(oid1->hash, oid2->hash, GIT_MAX_RAWSZ); > +} > + > +static inline int oideq(const struct object_id *oid1, const struct object_id *oid2) > +{ > + return !memcmp(oid1->hash, oid2->hash, GIT_MAX_RAWSZ); > +} > + > static inline void oidcpy(struct object_id *dst, const struct object_id *src) > { > memcpy(dst->hash, src->hash, GIT_MAX_RAWSZ); > diff --git a/hash.h b/hash.h > index e43e3d8b5a..ddc2e5ca47 100644 > --- a/hash.h > +++ b/hash.h > @@ -6,26 +6,6 @@ > > #define the_hash_algo the_repository->hash_algo > > -static inline int oidcmp(const struct object_id *oid1, const struct object_id *oid2) > -{ > - const struct git_hash_algo *algop; > - if (!oid1->algo) > - algop = the_hash_algo; > - else > - algop = &hash_algos[oid1->algo]; > - return hashcmp(oid1->hash, oid2->hash, algop); > -} > - > -static inline int oideq(const struct object_id *oid1, const struct object_id *oid2) > -{ > - const struct git_hash_algo *algop; > - if (!oid1->algo) > - algop = the_hash_algo; > - else > - algop = &hash_algos[oid1->algo]; > - return hasheq(oid1->hash, oid2->hash, algop); > -} > - > static inline int is_null_oid(const struct object_id *oid) > { > return oideq(oid, null_oid()); > -- > 2.45.2.457.g8d94cfb545.dirty
diff --git a/hash-ll.h b/hash-ll.h index b72f84f4ae..b04fe12aef 100644 --- a/hash-ll.h +++ b/hash-ll.h @@ -278,6 +278,16 @@ static inline void hashclr(unsigned char *hash, const struct git_hash_algo *algo memset(hash, 0, algop->rawsz); } +static inline int oidcmp(const struct object_id *oid1, const struct object_id *oid2) +{ + return memcmp(oid1->hash, oid2->hash, GIT_MAX_RAWSZ); +} + +static inline int oideq(const struct object_id *oid1, const struct object_id *oid2) +{ + return !memcmp(oid1->hash, oid2->hash, GIT_MAX_RAWSZ); +} + static inline void oidcpy(struct object_id *dst, const struct object_id *src) { memcpy(dst->hash, src->hash, GIT_MAX_RAWSZ); diff --git a/hash.h b/hash.h index e43e3d8b5a..ddc2e5ca47 100644 --- a/hash.h +++ b/hash.h @@ -6,26 +6,6 @@ #define the_hash_algo the_repository->hash_algo -static inline int oidcmp(const struct object_id *oid1, const struct object_id *oid2) -{ - const struct git_hash_algo *algop; - if (!oid1->algo) - algop = the_hash_algo; - else - algop = &hash_algos[oid1->algo]; - return hashcmp(oid1->hash, oid2->hash, algop); -} - -static inline int oideq(const struct object_id *oid1, const struct object_id *oid2) -{ - const struct git_hash_algo *algop; - if (!oid1->algo) - algop = the_hash_algo; - else - algop = &hash_algos[oid1->algo]; - return hasheq(oid1->hash, oid2->hash, algop); -} - static inline int is_null_oid(const struct object_id *oid) { return oideq(oid, null_oid());
With the preceding commit, the hash array of object IDs is now fully zero-padded even when the hash algorithm's output is smaller than the array length. With that, we can now adapt both `oidcmp()` and `oideq()` to unconditionally memcmp(3P) the whole array instead of depending on the hash size. While it may feel inefficient to compare unused bytes for e.g. SHA-1, in practice the compiler should now be able to produce code that is better optimized both because we have no branch anymore, but also because the size to compare is now known at compile time. Goldbolt spits out the following assembly on an x86_64 platform with GCC 14.1 for the old and new implementations of `oidcmp()`: oidcmp_old: movsx rax, DWORD PTR [rdi+32] test eax, eax jne .L2 mov rax, QWORD PTR the_repository[rip] cmp QWORD PTR [rax+16], 32 je .L6 .L4: mov edx, 20 jmp memcmp .L2: lea rdx, [rax+rax*2] lea rax, [rax+rdx*4] lea rax, hash_algos[0+rax*8] cmp QWORD PTR [rax+16], 32 jne .L4 .L6: mov edx, 32 jmp memcmp oidcmp_new: mov edx, 32 jmp memcmp The new implementation gets ridi of all the branches and effectively only ends setting up `edx` for `memcmp()` and then calling it. And for `oideq()`: oideq_old: movsx rcx, DWORD PTR [rdi+32] mov rax, rdi mov rdx, rsi test ecx, ecx jne .L2 mov rcx, QWORD PTR the_repository[rip] cmp QWORD PTR [rcx+16], 32 mov rcx, QWORD PTR [rax] je .L12 .L4: mov rsi, QWORD PTR [rax+8] xor rcx, QWORD PTR [rdx] xor rsi, QWORD PTR [rdx+8] or rcx, rsi je .L13 .L8: mov eax, 1 test eax, eax sete al movzx eax, al ret .L2: lea rsi, [rcx+rcx*2] lea rcx, [rcx+rsi*4] lea rcx, hash_algos[0+rcx*8] cmp QWORD PTR [rcx+16], 32 mov rcx, QWORD PTR [rax] jne .L4 .L12: mov rsi, QWORD PTR [rax+8] xor rcx, QWORD PTR [rdx] xor rsi, QWORD PTR [rdx+8] or rcx, rsi jne .L8 mov rcx, QWORD PTR [rax+16] mov rax, QWORD PTR [rax+24] xor rcx, QWORD PTR [rdx+16] xor rax, QWORD PTR [rdx+24] or rcx, rax jne .L8 xor eax, eax .L14: test eax, eax sete al movzx eax, al ret .L13: mov edi, DWORD PTR [rdx+16] cmp DWORD PTR [rax+16], edi jne .L8 xor eax, eax jmp .L14 oideq_new: mov rax, QWORD PTR [rdi] mov rdx, QWORD PTR [rdi+8] xor rax, QWORD PTR [rsi] xor rdx, QWORD PTR [rsi+8] or rax, rdx je .L5 .L2: mov eax, 1 xor eax, 1 ret .L5: mov rax, QWORD PTR [rdi+16] mov rdx, QWORD PTR [rdi+24] xor rax, QWORD PTR [rsi+16] xor rdx, QWORD PTR [rsi+24] or rax, rdx jne .L2 xor eax, eax xor eax, 1 ret Interestingly, the compiler decides to split the comparisons into two so that it first compares the lower half of the object ID for equality and then the upper half. If the first check shows a difference, then we wouldn't even end up comparing the second half. In both cases, the new generated code is significantly shorter and has way less branches. While I didn't benchmark the change, I'd be surprised if the new code was slower. Signed-off-by: Patrick Steinhardt <ps@pks.im> --- hash-ll.h | 10 ++++++++++ hash.h | 20 -------------------- 2 files changed, 10 insertions(+), 20 deletions(-)