In preparation for adding more similar checks, move the VEX.L=0 check
and several X86_SPECIAL_* checks to a new field, where each bit represent
a common check on unused bits, or a restriction on the processor mode.
Likewise, many SVM intercepts can be checked during the decoding phase,
the main exception being the selective CR0 write, MSR and IOIO intercepts.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The implementation was validated with OpenSSL and with the test vectors in
https://github.com/rust-lang/stdarch/blob/master/crates/core_arch/src/x86/sha.rs.
The instructions provide a ~25% improvement on hashing a 64 MiB file:
runtime goes down from 1.8 seconds to 1.4 seconds; instruction count on
the host goes down from 5.8 billion to 4.8 billion with slightly better
IPC too. Good job Intel. ;)
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
CVTPS2PD only loads a half-register for memory, like CVTPH2PS. It can
reuse the "ph" packed half-precision size to load a half-register,
but rename it to "xh" because it is now a variation of "x" (it is not
used only for half-precision values).
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The only issue with FMA instructions is that there are _a lot_ of them (30
opcodes, each of which comes in up to 4 versions depending on VEX.W and
VEX.L; a total of 96 possibilities). However, they can be implement with
only 6 helpers, two for scalar operations and four for packed operations.
(Scalar versions do not do any merging; they only affect the bottom 32
or 64 bits of the output operand. Therefore, there is no separate XMM
and YMM of the scalar helpers).
First, we can reduce the number of helpers to one third by passing four
operands (one output and three inputs); the reordering of which operands
go to the multiply and which go to the add is done in emit.c.
Second, the different instructions also dispatch to the same softfloat
function, so the flags for float32_muladd and float64_muladd are passed
in the helper as int arguments, with a little extra complication to
handle FMADDSUB and FMSUBADD.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
F16C only consists of two instructions, which are a bit peculiar
nevertheless.
First, they access only the low half of an YMM or XMM register for the
packed-half operand; the exact size still depends on the VEX.L flag.
This is similar to the existing avx_movx flag, but not exactly because
avx_movx is hardcoded to affect operand 2. To this end I added a "ph"
format name; it's possible to reuse this approach for the VPMOVSX and
VPMOVZX instructions, though that would also require adding two more
formats for the low-quarter and low-eighth of an operand.
Second, VCVTPS2PH is somewhat weird because it *stores* the result of
the instruction into memory rather than loading it.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This adds another kind of weirdness when you thought you had seen it all:
an opcode byte that comes _after_ the address, not before. It's not
worth adding a new X86_SPECIAL_* constant for it, but it's actually
not unlike VCMP; so, forgive me for exploiting the similarity and just
deciding to dispatch to the right gen_helper_* call in a single code
generation function.
In fact, the old decoder had a bug where s->rip_offset should have
been set to 1 for 3DNow! instructions, and it's fixed now.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
There are several special cases here:
1) extending moves have different widths for the helpers vs. for the
memory loads, and the width for memory loads depends on VEX.L too.
This is represented by X86_SPECIAL_AVXExtMov.
2) some instructions, such as variable-width shifts, select the vector element
size via REX.W.
3) VSIB instructions (VGATHERxPy, VPGATHERxy) are also part of this group,
and they have (among other things) two output operands.
3) the macros for 4-operand blends (which are under 0x0f 0x3a) have to be
extended to support 2-operand blends. The 2-operand variant actually
came a few years earlier, but it is clearer to implement them in the
opposite order.
X86_TYPE_WM, introduced earlier for unaligned loads, is reused for helpers
that accept a Reg* but have a M argument.
These three-byte opcodes also include AVX new instructions, for which
the helpers were originally implemented by Paul Brook <paul@nowt.org>.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The more complicated ones here are d6-d7, e6-e7, f7. The others
are trivial.
For LDDQU, using gen_load_sse directly might corrupt the register if
the second part of the load fails. Therefore, add a custom X86_TYPE_WM
value; like X86_TYPE_W it does call gen_load(), but it also rejects a
value of 11 in the ModRM field like X86_TYPE_M.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Many SSE and AVX instructions are only valid with specific prefixes
(none, 66, F3, F2). Introduce a direct way to encode this in the
decoding table to avoid using decode groups too much.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add generic code generation that takes care of preparing operands
around calls to decode.e.gen in a table-driven manner, so that ALU
operations need not take care of that.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The new decoder is based on three principles:
- use mostly table-driven decoding, using tables derived as much as possible
from the Intel manual. Centralizing the decode the operands makes it
more homogeneous, for example all immediates are signed. All modrm
handling is in one function, and can be shared between SSE and ALU
instructions (including XMM<->GPR instructions). The SSE/AVX decoder
will also not have duplicated code between the 0F, 0F38 and 0F3A tables.
- keep the code as "non-branchy" as possible. Generally, the code for
the new decoder is more verbose, but the control flow is simpler.
Conditionals are not nested and have small bodies. All instruction
groups are resolved even before operands are decoded, and code
generation is separated as much as possible within small functions
that only handle one instruction each.
- keep address generation and (for ALU operands) memory loads and writeback
as much in common code as possible. All ALU operations for example
are implemented as T0=f(T0,T1). For non-ALU instructions,
read-modify-write memory operations are rare, but registers do not
have TCGv equivalents: therefore, the common logic sets up pointer
temporaries with the operands, while load and writeback are handled
by gvec or by helpers.
These principles make future code review and extensibility simpler, at
the cost of having a relatively large amount of code in the form of this
patch. Even EVEX should not be _too_ hard to implement (it's just a crazy
large amount of possibilities).
This patch introduces the main decoder flow, and integrates the old
decoder with the new one. The old decoder takes care of parsing
prefixes and then optionally drops to the new one. The changes to the
old decoder are minimal and allow it to be replaced incrementally with
the new one.
There is a debugging mechanism through a "LIMIT" environment variable.
In user-mode emulation, the variable is the number of instructions
decoded by the new decoder before permanently switching to the old one.
In system emulation, the variable is the highest opcode that is decoded
by the new decoder (this is less friendly, but it's the best that can
be done without requiring deterministic execution).
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
