This adds two optimizations using the non-zero bit mask. In some cases
involving shifts or ANDs the value can become zero, and can thus be
optimized to a move of zero. Second, useless zero-extension or an
AND with constant can be detected that would only zero bits that are
already zero.
The main advantage of this optimization is that it turns zero-extensions
into moves, thus enabling much better copy propagation (around 1% code
reduction). Here is for example a "test $0xff0000,%ecx + je" before
optimization:
mov_i64 tmp0,rcx
movi_i64 tmp1,$0xff0000
discard cc_src
and_i64 cc_dst,tmp0,tmp1
movi_i32 cc_op,$0x1c
ext32u_i64 tmp0,cc_dst
movi_i64 tmp12,$0x0
brcond_i64 tmp0,tmp12,eq,$0x0
and after (without patch on the left, with on the right):
movi_i64 tmp1,$0xff0000 movi_i64 tmp1,$0xff0000
discard cc_src discard cc_src
and_i64 cc_dst,rcx,tmp1 and_i64 cc_dst,rcx,tmp1
movi_i32 cc_op,$0x1c movi_i32 cc_op,$0x1c
ext32u_i64 tmp0,cc_dst
movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0
brcond_i64 tmp0,tmp12,eq,$0x0 brcond_i64 cc_dst,tmp12,eq,$0x0
Other similar cases: "test %eax, %eax + jne" where eax is already 32-bit
(after optimization, without patch on the left, with on the right):
discard cc_src discard cc_src
mov_i64 cc_dst,rax mov_i64 cc_dst,rax
movi_i32 cc_op,$0x1c movi_i32 cc_op,$0x1c
ext32u_i64 tmp0,cc_dst
movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0
brcond_i64 tmp0,tmp12,ne,$0x0 brcond_i64 rax,tmp12,ne,$0x0
"test $0x1, %dl + je":
movi_i64 tmp1,$0x1 movi_i64 tmp1,$0x1
discard cc_src discard cc_src
and_i64 cc_dst,rdx,tmp1 and_i64 cc_dst,rdx,tmp1
movi_i32 cc_op,$0x1a movi_i32 cc_op,$0x1a
ext8u_i64 tmp0,cc_dst
movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0
brcond_i64 tmp0,tmp12,eq,$0x0 brcond_i64 cc_dst,tmp12,eq,$0x0
In some cases TCG even outsmarts GCC. :) Here the input code has
"and $0x2,%eax + movslq %eax,%rbx + test %rbx, %rbx" and the optimizer,
thanks to copy propagation, does the following:
movi_i64 tmp12,$0x2 movi_i64 tmp12,$0x2
and_i64 rax,rax,tmp12 and_i64 rax,rax,tmp12
mov_i64 cc_dst,rax mov_i64 cc_dst,rax
ext32s_i64 tmp0,rax -> nop
mov_i64 rbx,tmp0 -> mov_i64 rbx,cc_dst
and_i64 cc_dst,rbx,rbx -> nop
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
Add a "mask" field to the tcg_temp_info struct. A bit that is zero
in "mask" will always be zero in the corresponding temporary.
Zero bits in the mask can be produced from moves of immediates,
zero-extensions, ANDs with constants, shifts; they can then be
be propagated by logical operations, shifts, sign-extensions,
negations, deposit operations, and conditional moves. Other
operations will just reset the mask to all-ones, i.e. unknown.
[rth: s/target_ulong/tcg_target_ulong/]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
The next patch will add to the TCG optimizer a field that should be
non-zero in the default case. Thus, replace the memset of the
temps array with a loop. Only the state field has to be up-to-date,
because others are not used except if the state is TCG_TEMP_COPY
or TCG_TEMP_CONST.
[rth: Extracted the loop to a function.]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
Commit 7f6f0ae5b9 added two assertions.
One of these assertions is not needed:
The pointer ts is never NULL because it is initialized with the
address of an array element.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Stefan Weil <sw@weilnetz.de>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Existing compile-time detection is spotty at best. Convert
it all to runtime detection instead.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
In dead_temp, local temps should always be marked as back to memory,
even if they have not been allocated (i.e. they are discared before
cross a basic block).
It fixes the following assertion in target-xtensa:
qemu-system-xtensa: tcg/tcg.c:1665: temp_save: Assertion `s->temps[temp].val_type == 2 || s->temps[temp].fixed_reg' failed.
Aborted
Reported-by: Max Filippov <jcmvbkbc@gmail.com>
Tested-by: Max Filippov <jcmvbkbc@gmail.com>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
The bswap16 TCG opcode assumes that the high bytes of the temp equal
to 0 before calling it. The ARM backend implementation takes this
assumption to slightly optimize the generated code.
The same implementation is called for implementing the cross-endian
qemu_st16 opcode, where this assumption is not true anymore. One way to
fix that would be to zero the high bytes before calling it. Given the
store instruction just ignore them, it is possible to provide a slightly
more optimized version. With ARMv6+ the rev16 instruction does the work
correctly. For lower ARM versions the patch provides a version which
behaves correctly with non-zero high bytes, but fill them with junk.
Cc: Andrzej Zaborowski <balrogg@gmail.com>
Cc: Peter Maydell <peter.maydell@linaro.org>
Cc: qemu-stable@nongnu.org
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
The TCG arm backend considers likely that the offset to the TLB
entries does not exceed 12 bits for mem_index = 0. In practice this is
not true for at least the MIPS target.
The current patch fixes that by loading the bits 23-12 with a separate
instruction, and using loads with address writeback, independently of
the value of mem_idx. In total this allow a 24-bit offset, which is a
lot more than needed.
Cc: Andrzej Zaborowski <balrogg@gmail.com>
Cc: Peter Maydell <peter.maydell@linaro.org>
Cc: qemu-stable@nongnu.org
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
The operations for INDEX_op_deposit_i32 and INDEX_op_deposit_i64
are now supported and enabled by default.
Signed-off-by: Stefan Weil <sw@weilnetz.de>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
mmu access looks something like:
<check tlb>
if miss goto slow_path
<fast path>
done:
...
; end of the TB
slow_path:
<pre process>
mr r3, r27 ; move areg0 to r3
; (r3 holds the first argument for all the PPC32 ABIs)
<call mmu_helper>
b $+8
.long done
<post process>
b done
On ppc32 <call mmu_helper> is:
(SysV and Darwin)
mmu_helper is most likely not within direct branching distance from
the call site, necessitating
a. moving 32 bit offset of mmu_helper into a GPR ; 8 bytes
b. moving GPR to CTR/LR ; 4 bytes
c. (finally) branching to CTR/LR ; 4 bytes
r3 setting - 4 bytes
call - 16 bytes
dummy jump over retaddr - 4 bytes
embedded retaddr - 4 bytes
Total overhead - 28 bytes
(PowerOpen (AIX))
a. moving 32 bit offset of mmu_helper's TOC into a GPR1 ; 8 bytes
b. loading 32 bit function pointer into GPR2 ; 4 bytes
c. moving GPR2 to CTR/LR ; 4 bytes
d. loading 32 bit small area pointer into R2 ; 4 bytes
e. (finally) branching to CTR/LR ; 4 bytes
r3 setting - 4 bytes
call - 24 bytes
dummy jump over retaddr - 4 bytes
embedded retaddr - 4 bytes
Total overhead - 36 bytes
Following is done to trim the code size of slow path sections:
In tcg_target_qemu_prologue trampolines are emitted that look like this:
trampoline:
mfspr r3, LR
addi r3, 4
mtspr LR, r3 ; fixup LR to point over embedded retaddr
mr r3, r27
<jump mmu_helper> ; tail call of sorts
And slow path becomes:
slow_path:
<pre process>
<call trampoline>
.long done
<post process>
b done
call - 4 bytes (trampoline is within code gen buffer
and most likely accessible via
direct branch)
embedded retaddr - 4 bytes
Total overhead - 8 bytes
In the end the icache pressure is decreased by 20/28 bytes at the cost
of an extra jump to trampoline and adjusting LR (to skip over embedded
retaddr) once inside.
Signed-off-by: malc <av1474@comtv.ru>
Add optimized TCG qemu_ld/st generation which locates the code of TLB miss
cases at the end of a block after generating the other IRs.
Currently, this optimization supports only i386 and x86_64 hosts.
Signed-off-by: Yeongkyoon Lee <yeongkyoon.lee@samsung.com>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
Commit 9c43b68de6 do not correctly check
for dead outputs when they need to be synced to memory in case of
half-dead operations.
Fix that by applying the same pattern than for the default case.
Tested-by: Stefan Weil <sw@weilnetz.de>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
MIPS32 and later instruction sets have a multiplication instruction
directly operating on GPRs. It only produces a 32-bit result but
it is exactly what is needed by QEMU.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
When we allocate a reserved_va for the guest, the kernel will likely
choose an address well above 4G. At which point we must use a pair
of movabsq+addq to form the host address. If we have OS support,
set up a segment register to point to guest_base instead.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
The current helper flags, TCG_CALL_CONST and TCG_CALL_PURE might be
confusing and doesn't provide enough granularity for some helpers (FP
helpers for example).
This patch changes them into the following helpers flags:
- TCG_CALL_NO_READ_GLOBALS means that the helper does not read globals,
either directly or via an exception. They will not be saved to their
canonical location before calling the helper.
- TCG_CALL_NO_WRITE_GLOBALS means that the helper does not modify any
globals. They will only be saved to their canonical locations before
calling helpers, but they won't be reloaded afterwise.
- TCG_CALL_NO_SIDE_EFFECTS means that the call to the function is
removed if the return value is not used.
It provides convenience flags, to avoid helper definitions longer than
80 characters. It also provides compatibility flags, and updates the
documentation.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Operations with side effects (in practice qemu_ld/st ops), only need to
synchronize globals to make sure the CPU state is consistent in case of
exception.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Mapping a memory address using a global and accessing it through
ld/st operations is currently broken. As it doesn't make any sense
to do that performance wise, let's forbid that.
Update the TCG documentation, and remove partial support for that.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Some branch related ops are marked with TCG_OPF_SIDE_EFFECTS, some other
not. In practice they don't need to, as they are all marked with
TCG_OPF_BB_END, which is handled specifically in all the code.
The call op is marked as TCG_OPF_SIDE_EFFECTS, which might be not true
as there is are specific flags (TCG_CALL_CONST and TCG_CALL_PURE) for
specifying that. On the other hand it always clobber arguments, so mark
it as such even if the call op is handled in a different code path.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
The liveness analysis ensures that globals and temps are at the correct
state at a basic block end or with an op with side effects. Avoid
looping on all temps, this can be time consuming on targets with a lot
of globals. Keep an assert in debug mode.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Start with local temps in TEMP_VAL_MEM state, to make possible a later
check that all the temps are correctly saved back to memory.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Always mark dead input arguments as dead, even if the op is at the basic
block end. This will allow to check that all temps are correctly saved.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Now that the liveness analysis provides more information, rewrite
tcg_reg_alloc_mov(). This changes the behaviour about propagating
constants and memory accesses. We now take the assumption that once
a value is loaded into a register (from memory or from a constant),
it's better to keep it there than to reload it later. This assumption
is now always almost correct given that we are now sure the
corresponding temp is going to be used later (otherwise it would have
been synchronized and marked as dead already). The assumption is wrong
if one of the op after clobbers some registers including the one
of the holding the temp (this can be avoided by allocating clobbered
registers last, which is what most TCG target do), or in case of lack
of available register.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Now that the liveness analysis might mark some output temps as dead, call
temp_dead() if needed.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Rework the liveness analysis by tracking temps that need to go back to
memory in addition to dead temps tracking. This allows to mark output
arguments as "need sync", and to synchronize them back to memory as soon
as they are not written anymore. This way even arguments mapping to
globals can be marked as "dead", avoiding moves to a new register when
input and outputs are aliased.
In addition it means that registers are freed as soon as temps are not
used anymore, instead of waiting for a basic block end or an op with side
effects. This reduces register spilling especially on CPUs with few
registers, and spread the mov over all the TB, increasing the
performances on in-order CPUs.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Synchronize an output argument when requested by the liveness analysis.
This is needed so that the temp can be declared dead later.
For that, add a new op_sync_args table in which each bit tells if the
corresponding output argument needs to be synchronized with the memory.
Pass it to the tcg_reg_alloc_* functions, and honor this bit. We need to
synchronize the argument before marking it as dead, and we have to make
sure all the infos about the temp are correctly filled.
At the same time change some types from unsigned int to uint16_t when
passing op_dead_args.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Add a new function temp_sync() to synchronize the canonical location
of a temp with the value in the corresponding register, but without
freeing the associated register. Rewrite temp_save() to call
temp_sync() followed by temp_dead().
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Add a new function tcg_reg_sync() to synchronize the canonical location
of a temp with the value in the associated register, but without freeing
it. Rewrite tcg_reg_free() to first call tcg_reg_sync() and then to free
the register.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
A lot of code is duplicated to mark a temporary as dead. Replace it
by temp_dead(), which in addition marks the temp as saved in memory
for globals and local temps, instead of doing this a posteriori in
temp_save().
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
On x86_64, remove the constraint on the third argument register which
is not needed:
- For loads the helper arguments are env, addr, mem_idx. The addr
value should not be in the two first argument registers as they are
used in tcg_out_tlb_load().
- For stores the helper arguments are env, addr, data, mem_idx.
The addr and data values should not be in the two first argument
registers as they are used in tcg_out_tlb_load(). The data value
should also not be in the two first argument registers, but could
be in the third argument register in which case it would be already
loaded at the right location.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Now that CONFIG_TCG_PASS_AREG0 has been removed, it's easier to get
an optimal code for the load/store functions.
First swap the two registers used in tcg_out_tlb_load() so that the
address end-up in the second register instead of the first one. Adjust
tcg_out_qemu_ld() and tcg_out_qemu_st() to respectively call
tcg_out_qemu_ld_direct() and tcg_out_qemu_st_direct() with the correct
registers. Then replace the register shifting by direct load of the
arguments.
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
