| Age | Commit message (Collapse) | Author |
|---|
|
When matching tuples, the pattern matching compiler would generate
code that would fetch all elements of the tuple that will ultimately
be used, *before* testing that (for example) the first element is the
correct record tag. For example:
is_tuple Fail {x,0}
test_arity Fail {x,0} 3
get_tuple_element {x,0} 0 {x,1}
get_tuple_element {x,0} 1 {x,2}
get_tuple_element {x,0} 2 {x,3}
is_eq_exact Fail {x,1} some_tag
If {x,2} and {x,3} are not used at label Fail, we can re-arrange the
code like this:
is_tuple Fail {x,0}
test_arity Fail {x,0} 3
get_tuple_element {x,0} 0 {x,1}
is_eq_exact Fail {x,1} some_tag
get_tuple_element {x,0} 1 {x,2}
get_tuple_element {x,0} 2 {x,3}
Doing that may be beneficial in two ways.
If the branch is taken, we have eliminated the execution of two
unnecessary instructions.
Even if the branch is never or rarely taken, there is the possibility
for more optimizations following the is_eq_exact instructions.
For example, imagine that the code looks like this:
get_tuple_element {x,0} 1 {x,2}
get_tuple_element {x,0} 2 {x,3}
move {x,2} {y,0}
move {x,3} {y,1}
Assuming that {x,2} and {x,3} have no further uses in the code
that follows, that can be rewritten to:
get_tuple_element {x,0} 1 {y,0}
get_tuple_element {x,0} 2 {y,1}
When should we perform this optimization?
At the very latest, it must be done before opt_blocks/1 in
beam_block which does the elimination of unnecessary moves.
Actually, we want do the optimization before the blocks have
been established, since moving instructions out of one block
into another is cumbersome.
Therefore, we will do the optimization in a new pass that is
run before beam_block. A new pass will make debugging easier,
and beam_block already has a fair number of sub passes.
|
|
|
|
Here is an example of a move instruction that could not be optimized
away because the {x,2} register was not killed:
get_tuple_element Reg Pos {x,2}
.
.
.
move {x,2} {y,0}
put_list {x,2} nil Any
We can do the optimization if we replace all occurrences of the {x,2}
register as a source with {y,0}:
get_tuple_element Reg Pos {y,0}
.
.
.
put_list {y,0} nil Dst
|
|
The 'move' optimization was relatively clean until GC BIFs
were introduced. Instead of re-thinking the implementation,
the existing code was fixed and patched.
The current code unsuccessfully attempts to eliminate 'move'
instructions across GC BIF and allocation instructions. We can
simplify the code if we give up as soon as we encounter any
instruction that allocates.
|
|
opt_alloc/1 makes a redundant call to opt/1. It is redundant because
the opt/1 function has already been applied to the instruction
sequence prior to calling opt_alloc/1.
|
|
Add the 'da' option to create a list after the beam_a pass. Seeing
how the code looks after beam_a, but before the blocks have been
established, is sometimes useful.
For symmetry, add the 'dz' option, even though it is just a synonym
for 'S'.
|
|
* bjorn/erts/clang-opt: (27 commits)
Improve unpacking performance on x86_64
Slightly tweak the peformance for get_list
Speed up list matching
Eliminate the variable temp_bits at the top scope of process_main()
Eliminate prefetch for conditional instructions
Teach beam_makeops to pack operands for move3 and move_window
Ensure that the move_call_ext_{last,only} instructions are used
beam_makeops: Eliminate unnecessary masking when packing 3 operands
Use a cheaper tag scheme for 'd' operands
Introduce swap_temp/3 and swap/2
Introduce specialized versions of move2
Add back frequently used x(0) instructions
Rewrite the hipe_mode_switch instructions
Remove the last use of tmp_arg1
Eliminate use of tmp_arg1 and tmp_arg2 in bit syntax
Remove the i_fetch instruction
Eliminate use of i_fetch for bit syntax instructions
Eliminate the use of i_fetch for BIF instructions
Eliminate the use of i_fetch for relational operators
Eliminate the use of i_fetch in arithmetic instructions
...
|
|
When unpacking operands on 64-bit CPUs, use a smarter mask to
help the compiler optimize the code.
It turns out that on x86_64, if we use the mask 0xFFFFUL (selecting
the 16 least significant bits), the compiler can combine a move and
a mask operation into the single insruction 'movzwl', which will
eliminate one instruction.
|
|
Fetch the head and tail parts to temporary variables before
writing them to their destinations. That should allow the CPU to
perform the moves in parallel, which might improve performance.
|
|
The combination is_non_empty_list followed by get_list is extremly
common (but not in estone_SUITE, which is why it has not been noticed
before). Therefore it is worthwile to introduce a combined
instruction.
|
|
|
|
Not pre-fetching in conditional instructions (instructions that use
-fail_action) seems to improve performance slightly.
The reason for that is that conditional instructions may jump to the
failure label, wasting the pre-fetched instruction. Another reason
is that that many conditional instructions do a function call, and
the register pressure is therefore high. Avoiding the pre-fetch
may reduce the register pressure and pontentially result in more
efficient code.
|
|
* maint:
ssh: be more generous about disconnect expects
ssh: add disjunction to ssh_trpt_test_lib:match
|
|
* hans/ssh/no_common_algs/OTP-11531:
ssh: be more generous about disconnect expects
ssh: add disjunction to ssh_trpt_test_lib:match
|
|
|
|
|
|
It is currently only possible to pack up to 4 operands. However,
the move_window4 instrucion has 5 operands and move_window5 and
move3 instrucations have 6 operands.
Teach beam_makeops to pack instructions with 5 or 6 operands.
Also rewrite the move_window instructions in beam_emu.c to macros
to allow their operands to get packed.
|
|
Update transformations to ensure that the move_call_ext_last
and move_call_ext_last are used.
|
|
When packing 3 operands into one word, there would be an unnecessary
mask operation when extracting the last operand.
|
|
Since 'd' operands can only either an X register or an Y register,
we only need a single bit to distinguish them. Furthermore, we can
pre-multiply the register number with the word size to speed up
address calculation.
|
|
Sequences of three move instructionst that effectively swap the
contents of two registers are fairly common. We can replace them
with a swap_temp/3 instruction. The third operand is the temporary
register to be used for swapping, since the temporary register
may actually be used.
If swap_temp/3 instruction is followed by a call, the temporary
register will often (but not always) be killed by the call. If
it is killed, we can replace the swap_temp/3 instruction with a
slightly cheaper swap/2 instruction.
|
|
Currently, move2/2 does the two moves sequentially to ensure
that the instruction will always work correctly.
We can do better than that. If the two move instructions have
any registers in common, we can introduce simpler and slightly
more efficient instructions to handle those cases:
move_shift/3
move_dup/3
For the remaining cases when the the move instructions
have no common registers, the move2/4 instruction can perform
the moves in parallel which is probably slightly more efficient.
For clarity's sake, we will remain the instruction to move2_par/4.
|
|
|
|
The 'cmd' variable that were shared by several hipe_mode_switch
instructions would cause clang to produce sub-optimal code,
probably because it considered the instructions as part of of
loop that needed to be optimized.
What would was that 'cmd' would be assigned to the ESI register
(lower 32 bits of the RSI register). It would use ESI for other
purposes in instructions, but at the end of every instruction
it would set ESI to 1 just in case the next instruction happened
to be hipe_trap_return. This can be seen clearly if this commit
is omitted and the define HIPE_MODE_SWITCH_CMD_RETURN in
hipe/hipe_mode_switch.h is changed from 1 to some other number
such as 42. You will see that 42 is assigned to ESI at the end
of every instruction.
Eliminate this problem by elimininating the shared 'cmd' variable.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The i_fetch instruction fetches two operands and places them
in the tmp_arg1 and tmp_arg2 variables. The next instruction
(such as i_plus) does not have to handle different types of
operands, but can get get them simply from the tmp_arg*
variables. Thus, i_fetch was introduced as a way to temper
a potentail combinatorial explosion.
Unfortunately, clang will generate terrible code because of
the tmp_arg1 and tmp_arg2 variables being live across multiple
instructions. Note that Clang has no way to predict the control
flow from one instruction to another. Clang must assume that
any instruction can jump to any other instruction. Somehow GCC
manages to cope with this situation much better.
Therefore, to improve the quality of the code generated by clang, we
must eliminate all uses of the tmp_arg1 and tmp_arg2 variables. This
commit eliminates the use of i_fetch in combination with the
arithmetic and logical instructions.
While we are touching the code for the bsr and bsl instructions,
also move the tmp_big[] array from top scope of process main into
the block that encloses the bsr and bsl instructions.
|
|
The 'r' type is now mandatory. That means in order to handle
both of the following instructions:
move x(0) y(7)
move x(1) y(7)
we would need to define two specific operations in ops.tab:
move r y
move x y
We want to make 'r' operands optional. That is, if we have
only this specific instruction:
move x y
it will match both of the following instructions:
move x(0) y(7)
move x(1) y(7)
Make 'r' optional allows us to save code space when we don't
want to make handling of x(0) a special case, but we can still
use 'r' to optimize commonly used instructions.
|
|
Consider the try_case_end instruction:
try_case_end s
The 's' operand type means that the operand can either be a
literal of one of the types atom, integer, or empty list, or
a register. That worked well before R12. In R12 additional
types of literals where introduced. Because of way the
overloading was done, an 's' operand cannot handle the
new types of literals. Therefore, code such as the following
is necessary in ops.tab to avoid giving an 's' operand a
literal:
try_case_end Literal=q => move Literal x | try_case_end x
While this work, it is error-prone in that it is easy to
forget to add that kind of rule. It would also be complicated
in case we wanted to introduce a new kind of addition operator
such as:
i_plus jssd
Since there are two 's' operands, two scratch registers and
two 'move' instructions would be needed.
Therefore, we'll need to find a smarter way to find tag
register operands. We will overload the pid and port tags
for X and Y register, respectively. That works because pids
and port are immediate values (fit in one word), and there
are no literals for pids and ports.
|
|
|
|
As part of improving code generation for clang, we want to
eliminate the special variable that stores the content of X
register zero most of the time. In a future, that will allow us
to eliminate the special case of handling r(0) for most
instructions, thus reducing the code size and allow other
simplifcations.
Therefore, in this commit, eliminate the variable that is used
to store r(0) and make r(0) as synonym for x(0). I have chosen
to keep the r(0) define to keep the size of the diff managable.
|
|
The purpose of this series of commits is to improve code generation
for the Clang compiler.
As a first step we want to change the meaning of 'x' in a
transformation such as:
operation Literal=q => move Literal x | operation x
Currently, a plain 'x' means reg[0] or x(0), which is the first
element in the X register array. That element is distinct from
r(0) which is a variable in process_main(). Therefore, since r(0)
and x(0) are currently distinct it is fine to use x(0) as a
scratch register.
However, in the next commit we will eliminate the separate variable
for storing the contents of X register zero (thus, x(0) and r(0)
will point to the same location in the X register array). Therefore,
we must use another scratch register in transformation. Redefine
a plain 'x' in a transformation to mean x(1023). Also define
SCRATCH_X_REG so that we can refer to the register by name from
C code.
|
|
|
|
Consider an hypothetical instruction:
do_something x x c
The loader would crash if we tried to load an instance of the
instruction with the last operand referencing a literal:
{do_something,{x,0},{x,1},{literal,{a,b,c}}}
Teach beam_makeops to turn off packing for such unsafe instructions.
|
|
* maint:
Add a smoke test of erts_debug:df/1
Correct disassembly of the i_get_map_elements instruction
|
|
* bjorn/erts/beam_debug:
Add a smoke test of erts_debug:df/1
Correct disassembly of the i_get_map_elements instruction
|
|
|
|
* egil/fix-configure-pthread_getname/OTP-12887:
erts: Fix configure pthread_getname
|
|
|
|
|
|
* maint:
ssh: testcases for no common algorithms in key exchange
|
|
* hans/ssh/no_common_algs/OTP-11531:
ssh: testcases for no common algorithms in key exchange
|
|
* maint:
ssh: Initial ssh_tprt_test_lib.erl and ssh_protocol_SUITE
|
|
|
|
|
|
* ia/ssl/tune-tests:
ssl: Exclude broken OpenSSL version from ECC test
ssl: Tune timeouts
|
