TCG_TARGET_HAS_MEMORY_BSWAP is already unset for this backend,
which means that MO_BSWAP be handled by the middle-end and
will never be seen by the backend. Thus the indexes used with
qemu_{ld,st}_helpers will always be zero.
Tidy the comments and asserts in tcg_out_qemu_{ld,st}_direct.
It is not that we do not handle bswap "yet", but never will.
Acked-by: Alistair Francis <alistair.francis@wdc.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
The memory bswap support in the aarch64 backend merely dates from
a time when it was required. There is nothing special about the
backend support that could not have been provided by the middle-end
even prior to the introduction of the bswap flags.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Now that the middle-end can replicate the same tricks as tcg/arm
used for optimizing bswap for signed loads and for stores, do not
pretend to have these memory ops in the backend.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
By removing TCG_BSWAP_IZ we indicate that the input is
not zero-extended, and thus can remove an explicit extend.
By removing TCG_BSWAP_OZ, we allow the implementation to
leave high bits set, which will be ignored by the store.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
We can perform any required sign-extension via TCG_BSWAP_OS.
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Implement the new semantics in the fallback expansion.
Change all callers to supply the flags that keep the
semantics unchanged locally.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Notice when the input is known to be zero-extended and force
the TCG_BSWAP_IZ flag on. Honor the TCG_BSWAP_OS bit during
constant folding. Propagate the input to the output mask.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
The existing interpreter zero-extends, ignoring high bits.
Simply add a separate sign-extension opcode if required.
Ensure that the interpreter supports ext16s when bswap16 is enabled.
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Merge tcg_out_bswap32 and tcg_out_bswap32s.
Use the flags in the internal uses for loads and stores.
For mips32r2 bswap32 with zero-extension, standardize on
WSBH+ROTR+DEXT. This is the same number of insns as the
previous DSBH+DSHD+DSRL but fits in better with the flags check.
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Merge tcg_out_bswap16 and tcg_out_bswap16s. Use the flags
in the internal uses for loads and stores.
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
For INDEX_op_bswap16_i64, use 64-bit instructions so that we can
easily provide the extension to 64-bits. Drop the special case,
previously used, where the input is already zero-extended -- the
minor code size savings is not worth the complication.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
For INDEX_op_bswap32_i32, pass 0 for flags: input not zero-extended,
output does not need extension within the host 64-bit register.
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
With the use of a suitable temporary, we can use the same
algorithm when src overlaps dst. The result is the same
number of instructions either way.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
We will shortly require sari in other context;
split out both for cleanliness sake.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
We will shortly require these in other context;
make the expansion as clear as possible.
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Combine the three bswap16 routines, and differentiate via the flags.
Use the correct flags combination from the load/store routines, and
pass along the constant parameter from tcg_out_op.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Pass in the input and output size. We currently use 3 of the 5
possible combinations; the others may be used by new tcg opcodes.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Retain the current rorw bswap16 expansion for the zero-in/zero-out case.
Otherwise, perform a wider bswap plus a right-shift or extend.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This will eventually simplify front-end usage, and will allow
backends to unset TCG_TARGET_HAS_MEMORY_BSWAP without loss of
optimization.
The argument is added during expansion, not currently exposed to the
front end translators. The backends currently only support a flags
value of either TCG_BSWAP_IZ, or (TCG_BSWAP_IZ | TCG_BSWAP_OZ),
since they all require zero top bytes and leave them that way.
At the existing call sites we pass in (TCG_BSWAP_IZ | TCG_BSWAP_OZ),
except for the flags-ignored cases of a 32-bit swap of a 32-bit
value and or a 64-bit swap of a 64-bit value, where we pass 0.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
The Arm MVE VDUP implementation would like to be able to emit code to
duplicate a byte or halfword value into an i32. We have code to do
this already in tcg-op-gvec.c, so all we need to do is make the
functions global.
For consistency with other functions made available to the frontends:
* we rename to tcg_gen_dup_*
* we expose both the _i32 and _i64 forms
* we provide the #define for a _tl form
Suggested-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Message-id: 20210617121628.20116-10-peter.maydell@linaro.org
Assume that we'll have fewer temps allocated after
restarting with a fewer number of instructions.
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This function should have been updated for vector types
when they were introduced.
Fixes: d2fd745fe8
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/367
Cc: qemu-stable@nongnu.org
Tested-by: Stefan Weil <sw@weilnetz.de>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
We should not be aligning the offset in temp_allocate_frame,
because the odd offset produces an aligned address in the end.
Instead, pass the logical offset into tcg_set_frame and add
the stack bias last.
Cc: qemu-stable@nongnu.org
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Wrap guest memory operations for tci like we do for cpu_ld*_data.
We cannot actually use the cpu_ldst.h interface without duplicating
the memory trace operations performed within, which will already
have been expanded into the tcg opcode stream.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
These macros are only used in one place. By expanding,
we get to apply some common-subexpression elimination
and create some local variables.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This reverts commit dc09f047ed.
For tcg, tracepoints are expanded inline in tcg opcodes.
Using a helper which generates a second tracepoint is incorrect.
For system mode, the extraction and re-packing of MemOp and mmu_idx
lost the alignment information from MemOp. So we were no longer
raising alignment exceptions for !TARGET_ALIGNED_ONLY guests.
This can be seen in tests/tcg/xtensa/test_load_store.S.
For user mode, we must update to the new signature of g2h() so that
the revert compiles. We can leave set_helper_retaddr for later.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
We can share this code between 32-bit and 64-bit loads and stores.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
We already had the 32-bit versions for a 32-bit host; expand this
to 64-bit hosts as well. The 64-bit opcodes are new.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
We already had mulu2_i32 for a 32-bit host; expand this to 64-bit
hosts as well. The muls2_i32 and the 64-bit opcodes are new.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
These were already present in tcg-target.c.inc,
but not in the interpreter.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
When this opcode is not available in the backend, tcg middle-end
will expand this as a series of 5 opcodes. So implementing this
saves bytecode space.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This operation is critical to staying within the interpretation
loop longer, which avoids the overhead of setup and teardown for
many TBs.
The check in tcg_prologue_init is disabled because TCI does
want to use NULL to indicate exit, as opposed to branching to
a real epilogue.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This removes all of the problems with unaligned accesses
to the bytecode stream.
With an 8-bit opcode at the bottom, we have 24 bits remaining,
which are generally split into 6 4-bit slots. This fits well
with the maximum length opcodes, e.g. INDEX_op_add2_i32, which
have 6 register operands.
We have, in previous patches, rearranged things such that there
are no operations with a label which have more than one other
operand. Which leaves us with a 20-bit field in which to encode
a label, giving us a maximum TB size of 512k -- easily large.
Change the INDEX_op_tci_movi_{i32,i64} opcodes to tci_mov[il].
The former puts the immediate in the upper 20 bits of the insn,
like we do for the label displacement. The later uses a label
to reference an entry in the constant pool. Thus, in the worst
case we still have a single memory reference for any constant,
but now the constants are out-of-line of the bytecode and can
be shared between different moves saving space.
Change INDEX_op_call to use a label to reference a pair of
pointers in the constant pool. This removes the only slightly
dodgy link with the layout of struct TCGHelperInfo.
The re-encode cannot be done in pieces.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Inline it into its one caller, tci_write_reg64.
Drop the asserts that are redundant with tcg_read_r.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
The encoding planned for tci does not have enough room for
brcond2, with 4 registers and a condition as input as well
as the label. Resolve the condition into TCG_REG_TMP, and
relax brcond to one register plus a label, considering the
condition to always be reg != 0.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
We're about to adjust the offset range on host memory ops,
and the format of branches. Both will require a temporary.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This requires adjusting where arguments are stored.
Place them on the stack at left-aligned positions.
Adjust the stack frame to be at entirely positive offsets.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
As the only call-clobbered regs for TCI, these should
receive the least priority.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
The current setting is much too pessimistic. Indicating only
the one or two registers that are actually assigned after a
call should avoid unnecessary movement between the register
array and the stack array.
Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
