convert some typeof tags to small integers #49556
Conversation
|
@nanosoldier |
|
@nanosoldier |
| @@ -989,7 +1048,7 @@ STATIC_INLINE jl_value_t *jl_svecset( | |||
| #else | |||
| STATIC_INLINE jl_value_t *jl_svecref(void *t JL_PROPAGATES_ROOT, size_t i) JL_NOTSAFEPOINT | |||
| { | |||
| assert(jl_typeis(t,jl_simplevector_type)); | |||
| assert(jl_typetagis(t,jl_simplevector_tag << 4)); | |||
There was a problem hiding this comment.
Move the << 4 into the macro?
| @@ -1055,13 +1114,13 @@ STATIC_INLINE jl_value_t *jl_array_ptr_set( | |||
| STATIC_INLINE uint8_t jl_array_uint8_ref(void *a, size_t i) JL_NOTSAFEPOINT | |||
| { | |||
| assert(i < jl_array_len(a)); | |||
| assert(jl_typeis(a, jl_array_uint8_type)); | |||
| assert(jl_typetagis(a, jl_array_uint8_type)); | |||
There was a problem hiding this comment.
| assert(jl_typetagis(a, jl_array_uint8_type)); | |
| assert(jl_typeis(a, jl_array_uint8_type)); |
There was a problem hiding this comment.
jl_typetagis is always faster than jl_typeis, which is kept as a legacy alias, if you can guarantee that the type is not sometimes represented as a tag.
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. |
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. |
|
@nanosoldier |
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. |
|
Fascinating. I didn’t expect this to improve performance by that much, but seems consistent in the two runs. |
|
I wonder if removing the load makes it codegen a bit better. |
|
Okay, it turns out our codegen for this is just pretty bad: But this PR, as a side-effect, simplified that slightly, which let LLVM mostly optimize it away in this case: Vs. if we replace that with |
|
The fix for that conflicts slightly with our change here, because we need to edit some of the lines that need to be moved, so I can do that as a followup after merging this, if it is ready to go, or before if this ends up needing to be rebased anyways. |
vtjnash
removed
the
needs nanosoldier run
|
Is |
|
|
The value here is two-fold. One, we speed up sysimg load times slightly, since we can entirely skip relocations for these values and avoid faulting in some writable pages. This should let us (later) move more content into the ConstData section, such as String objects. Secondly, some type-manipulation native code becomes simpler, since it is based on small consecutive constants instead of requiring extra pointer loads to check for pointer identity. This is similar to the existing small-union optimization, but for type tags, and only for specific ones. This makes use of the fact that the zero page (about 4096 byes) cannot be allocated in the usual C memory model, which lets us define up to 255 of these special types, after taking into account the 4 gc bits. As a current implementation limitation for performance, these cannot be parametric types. Note there are probably more places in Base that can be updated to use `jl_typetagof` instead of `jl_typeof`, where we know if it is a type or tag. For example, this optimize most of builtins.c file, except for the `jl_object_id` function.
String is simpler and smaller (since we do not yet have a Buffer{UInt8}
type, which might be almost as simple) making it easier to manipulate.
Combined with the previous changes, this lets us eliminate many more
relocations from the image (previously every Array->data), which is
quite a few megabytes of array header objects simply gone.
MASTER:
sysimg size breakdown:
sys data: 79207516
isbits data: 58603452
symbols: 463610
tags list: 1078555
reloc list: 4491653
gvar list: 118848
fptr list: 214000
.text 11109176
.data 144184704
BEFORE (moving existing Strings):
sysimg size breakdown:
sys data: 75146604
isbits data: 62679532
symbols: 463411
tags list: 1015525
reloc list: 4481370
gvar list: 118688
fptr list: 213952
.text 11129192
.data 144126152
AFTER (making CodeInfo into Strings):
sys data: 71348124
isbits data: 63789244
symbols: 463411
tags list: 933984
reloc list: 4398377
gvar list: 118688
fptr list: 213904
.text 11132968
.data 141272824
oscardssmith
added
types and dispatch
- Remove JL_NORETURN - Account for changes in #49556
- Remove JL_NORETURN - Account for changes in #49556
This allows us to avoid needing relocations for some builtin objects (esp. Strings) when loading the system image, saving many objects from being faulted into memory. Takes some of the same code from #48432, but in a different direction so it is applicable to GC, codegen, and pkgimages too.
Closes #48432