This is identical for each target. So, move the initialization to
common code. Move the variable itself out of tcg_ctx and name it
cpu_env to minimize changes within targets.
This also means we can remove tcg_global_reg_new_{ptr,i32,i64},
since there are no longer global-register temps created by targets.
Backports commit 1c2adb958fc07e5b3e81ed21b801c04a15f41f4f from qemu
This enables parallel TCG code generation. However, we do not take
advantage of it yet since tb_lock is still held during tb_gen_code.
In user-mode we use a single TCG context; see the documentation
added to tcg_region_init for the rationale.
Note that targets do not need any conversion: targets initialize a
TCGContext (e.g. defining TCG globals), and after this initialization
has finished, the context is cloned by the vCPU threads, each of
them keeping a separate copy.
TCG threads claim one entry in tcg_ctxs[] by atomically increasing
n_tcg_ctxs. Do not be too annoyed by the subsequent atomic_read's
of that variable and tcg_ctxs; they are there just to play nice with
analysis tools such as thread sanitizer.
Note that we do not allocate an array of contexts (we allocate
an array of pointers instead) because when tcg_context_init
is called, we do not know yet how many contexts we'll use since
the bool behind qemu_tcg_mttcg_enabled() isn't set yet.
Previous patches folded some TCG globals into TCGContext. The non-const
globals remaining are only set at init time, i.e. before the TCG
threads are spawned. Here is a list of these set-at-init-time globals
under tcg/:
Only written by tcg_context_init:
- indirect_reg_alloc_order
- tcg_op_defs
Only written by tcg_target_init (called from tcg_context_init):
- tcg_target_available_regs
- tcg_target_call_clobber_regs
- arm: arm_arch, use_idiv_instructions
- i386: have_cmov, have_bmi1, have_bmi2, have_lzcnt,
have_movbe, have_popcnt
- mips: use_movnz_instructions, use_mips32_instructions,
use_mips32r2_instructions, got_sigill (tcg_target_detect_isa)
- ppc: have_isa_2_06, have_isa_3_00, tb_ret_addr
- s390: tb_ret_addr, s390_facilities
- sparc: qemu_ld_trampoline, qemu_st_trampoline (build_trampolines),
use_vis3_instructions
Only written by tcg_prologue_init:
- 'struct jit_code_entry one_entry'
- aarch64: tb_ret_addr
- arm: tb_ret_addr
- i386: tb_ret_addr, guest_base_flags
- ia64: tb_ret_addr
- mips: tb_ret_addr, bswap32_addr, bswap32u_addr, bswap64_addr
Backports commit 3468b59e18b179bc63c7ce934de912dfa9596122 from qemu
This is groundwork for supporting multiple TCG contexts.
The naive solution here is to split code_gen_buffer statically
among the TCG threads; this however results in poor utilization
if translation needs are different across TCG threads.
What we do here is to add an extra layer of indirection, assigning
regions that act just like pages do in virtual memory allocation.
(BTW if you are wondering about the chosen naming, I did not want
to use blocks or pages because those are already heavily used in QEMU).
We use a global lock to serialize allocations as well as statistics
reporting (we now export the size of the used code_gen_buffer with
tcg_code_size()). Note that for the allocator we could just use
a counter and atomic_inc; however, that would complicate the gathering
of tcg_code_size()-like stats. So given that the region operations are
not a fast path, a lock seems the most reasonable choice.
The effectiveness of this approach is clear after seeing some numbers.
I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark.
Note that I'm evaluating this after enabling per-thread TCG (which
is done by a subsequent commit).
* -smp 1, 1 region (entire buffer):
qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357
qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363
qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364
qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373
qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373
qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360
qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370
qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367
That is, 8 flushes.
* -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]:
qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356
qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361
qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361
qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375
qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375
qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360
qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365
qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368
Again, 8 flushes. Note how buffer utilization is not 100%, but it
is close. Smaller region sizes would yield higher utilization,
but we want region allocation to be rare (it acquires a lock), so
we do not want to go too small.
* -smp 8, static partitioning of 8 regions (10 MB per region):
qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354
qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370
qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365
qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377
qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358
qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367
qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364
qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358
qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362
qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372
qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374
qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376
qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374
qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372
qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359
qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362
qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368
qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378
qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367
qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364
That is, 20 flushes. Note how a static partitioning approach uses
the code buffer poorly, leading to many unnecessary flushes.
Backports commit e8feb96fcc6c16eab8923332e86ff4ef0e2ac276 from qemu
Groundwork for supporting multiple TCG contexts.
Note that having n_tcg_ctxs is unnecessary. However, it is
convenient to have it, since it will simplify iterating over the
array: we'll have just a for loop instead of having to iterate
over a NULL-terminated array (which would require n+1 elems)
or having to check with ifdef's for usermode/softmmu.
Backports commit df2cce2968069526553d82331ce9817eaca6b03a from qemu
Opcodes are added for scalar and vector shifts, but considering the
varied semantics of these do not expose them to the front ends. Do
go ahead and provide them in case they are needed for backend expansion.
Backports commit d0ec97967f940bbc11dced83422b39c224127f1e from qemu
We had two fields specific to INDEX_op_call. Rename these and
add some macros so that the fields may be reused for other opcodes.
Backports commit cd9090aa9dbba30db8aec9a2fc103aaf1ab0f5a7 from qemu
With no fixed array allocation, we can't overflow a buffer.
This will be important as optimizations related to host vectors
may expand the number of ops used.
Use QTAILQ to link the ops together.
Backports commit 15fa08f8451babc88d733bd411d4c94976f9d0f8 from qemu
Rather than have separate code only used for guest_base,
rely on a recent change to handle constant pool entries.
Backports commit ba2c747992f8c315c2fbddba196ce9137430d61d from qemu
Both ARMv6 and AArch64 currently may drop complex guest_base values
into the constant pool. But generic code wasn't expecting that, and
the pool is not emitted. Correct that.
Backports commit 5b38ee31616d1532c3c3a6dc644a9160d608ed2f from qemu
Using the offset of a temporary, relative to TCGContext, rather than
its index means that we don't use 0. That leaves offset 0 free for
a NULL representation without having to leave index 0 unused.
Backports commit e89b28a63501c0ad6d2501fe851d0c5202055e70 from qemu
The GET and MAKE functions weren't really specific enough.
We now have a full complement of functions that convert exactly
between temporaries, arguments, tcgv pointers, and indices.
The target/sparc change is also a bug fix, which would have affected
a host that defines TCG_TARGET_HAS_extr[lh]_i64_i32, i.e. MIPS64.
Backports commit dc41aa7d34989b552efe712ffe184236216f960b from qemu
Transform TCGv_* to an "argument" or a temporary.
For now, an argument is simply the temporary index.
Backports commit ae8b75dc6ec808378487064922f25f1e7ea7a9be from qemu
Copy s->nb_globals or s->nb_temps to a local variable for the purposes
of iteration. This should allow the compiler to use low-overhead
looping constructs on some hosts.
Backports commit ac3b88911ebc6fc841f28898ee8aed40839debe2 from qemu
Rather than have a separate buffer of 10*max_ops entries,
give each opcode 10 entries. The result is actually a bit
smaller and should have slightly more cache locality.
Backports commit 75e8b9b7aa0b95a761b9add7e2f09248b101a392 from qemu
In preparation for adding tc.size to be able to keep track of
TB's using the binary search tree implementation from glib.
Backports commit e7e168f41364c6e83d0f75fc1b3ce7f9c41ccf76 from qemu
A new shared header tcg-pool.inc.c adds new_pool_label,
for registering a tcg_target_ulong to be emitted after
the generated code, plus relocation data to install a
pointer to the data.
A new pointer is added to the TCGContext, so that we
dump the constant pool as data, not code.
Backports commit 57a269469dbf70013dab3a176e1735636010a772 from qemu
Dispense with TCGBackendData, as it has never been used for more than
holding a single pointer. Use a define in the cpu/tcg-target.h to
signal requirement for TCGLabelQemuLdst, so that we can drop the no-op
tcg-be-null.h stubs. Rename tcg-be-ldst.h to tcg-ldst.inc.c.
Backports commit 659ef5cbb893872d25e9d95191cc23b16546c8a1 from qemu
Allocating an arbitrarily-sized array of tbs results in either
(a) a lot of memory wasted or (b) unnecessary flushes of the code
cache when we run out of TB structs in the array.
An obvious solution would be to just malloc a TB struct when needed,
and keep the TB array as an array of pointers (recall that tb_find_pc()
needs the TB array to run in O(log n)).
Perhaps a better solution, which is implemented in this patch, is to
allocate TB's right before the translated code they describe. This
results in some memory waste due to padding to have code and TBs in
separate cache lines--for instance, I measured 4.7% of padding in the
used portion of code_gen_buffer when booting aarch64 Linux on a
host with 64-byte cache lines. However, it can allow for optimizations
in some host architectures, since TCG backends could safely assume that
the TB and the corresponding translated code are very close to each
other in memory. See this message by rth for a detailed explanation:
https://lists.gnu.org/archive/html/qemu-devel/2017-03/msg05172.html
Subject: Re: GSoC 2017 Proposal: TCG performance enhancements
Backports commit 6e3b2bfd6af488a896f7936e99ef160f8f37e6f2 from qemu
Instead of exporting goto_ptr directly to TCG frontends, export
tcg_gen_lookup_and_goto_ptr(), which calls goto_ptr with the pointer
returned by the lookup_tb_ptr() helper. This is the only use case
we have for goto_ptr and lookup_tb_ptr, so having this function is
very convenient. Furthermore, it trivially allows us to avoid calling
the lookup helper if goto_ptr is not implemented by the backend.
Backports commit cedbcb01529cb6cf9a2289cdbebbc63f6149fc18 from qemu
This allows an output operand to match an input operand
only when the input operand needs a register.
Backports commit 17280ff4a5f264e01e55ae514ee6d3586f9577b2 from qemu
This will let us choose how to interpret a given constraint
depending on whether the opcode is 32- or 64-bit. Which will
let us share more constraint combinations between opcodes.
At the same time, change the interface to return the advanced
pointer instead of passing it in/out by reference.
Backports commit 069ea736b50b75fdec99c9b8cc603b97bd98419e from qemu
This will allow the target to tailor the constraints to the
auto-detected ISA extensions.
Backports commit f69d277ece43c42c7ab0144c2ff05ba740f6706b from qemu
This comes from free from unifying tcg_reg_alloc_mov and
tcg_reg_alloc_movi's handling of TEMP_VAL_CONST. It triggers
often on moves to cc_dst, such as the following translation
of "sub $0x3c,%esp":
before: after:
subl $0x3c,%ebp subl $0x3c,%ebp
movl %ebp,0x10(%r14) movl %ebp,0x10(%r14)
movl $0x3c,%ebx movl $0x3c,0x2c(%r14)
movl %ebx,0x2c(%r14)
Backports commit 0fe4fca4e1a5e06a270127dd80bb753d4dda61c6 from qemu
Rather than rely on recursion during the middle of register allocation,
lower indirect registers to loads and stores off the indirect base into
plain temps.
For an x86_64 host, with sufficient registers, this results in identical
code, modulo the actual register assignments.
For an i686 host, with insufficient registers, this means that temps can
be (temporarily) spilled to the stack in order to satisfy an allocation.
This as opposed to the possibility of not being able to spill, to allocate
a register for the indirect base, in order to perform a spill.
Backports commit 5a18407f55ade924aa6397c9a043a9ffd59645fe from qemu
We only need two bits per temporary. Fold the two bytes into one,
and reduce the memory and cachelines required during compilation.
Backports commit c70fbf0a9938baf3b4f843355a77c17a7e945b98 from qemu
Reduce the size of other bitfields to make room.
This reduces the cache footprint of compilation.
Backports commit bee158cb4dde35c41632a3a129c869f14a32f8f0 from qemu
Instead of using -1 as end of chain, use 0, and link through the 0
entry as a fully circular double-linked list.
Backports commit dcb8e75870e2de199db853697f8839cb603beefe from qemu
This reduces both memory usage and per-insn cacheline usage
during code generation.
Backports commit a1b3c48d2b23d6eaeb4529d3e1183d2648731bf8 from qemu
