aboutsummaryrefslogtreecommitdiff
path: root/cpu
diff options
context:
space:
mode:
authorJose E. Marchesi <jose.marchesi@oracle.com>2019-05-23 19:03:59 +0200
committerJose E. Marchesi <jose.marchesi@oracle.com>2019-05-23 19:33:50 +0200
commitea195bb04cc5c964126aeff0f87d7161a03ca926 (patch)
tree460e3677c9b35ad26bdee3c876619b3fffa42e14 /cpu
parentfd0de36e274c8141a5dd4579cd04856dc88370da (diff)
downloadbinutils-ea195bb04cc5c964126aeff0f87d7161a03ca926.zip
binutils-ea195bb04cc5c964126aeff0f87d7161a03ca926.tar.gz
binutils-ea195bb04cc5c964126aeff0f87d7161a03ca926.tar.bz2
cpu: add eBPF cpu description
This patch adds a CPU description for the Linux kernel eBPF virtual machine, plus supporting code for disassembler and assembler. cpu/ChangeLog: 2019-05-23 Jose E. Marchesi <jose.marchesi@oracle.com> * bpf.cpu: New file. * bpf.opc: Likewise.
Diffstat (limited to 'cpu')
-rw-r--r--cpu/ChangeLog5
-rw-r--r--cpu/bpf.cpu647
-rw-r--r--cpu/bpf.opc191
3 files changed, 843 insertions, 0 deletions
diff --git a/cpu/ChangeLog b/cpu/ChangeLog
index c563860..b573c69 100644
--- a/cpu/ChangeLog
+++ b/cpu/ChangeLog
@@ -1,3 +1,8 @@
+2019-05-23 Jose E. Marchesi <jose.marchesi@oracle.com>
+
+ * bpf.cpu: New file.
+ * bpf.opc: Likewise.
+
2018-06-24 Nick Clifton <nickc@redhat.com>
2.32 branch created.
diff --git a/cpu/bpf.cpu b/cpu/bpf.cpu
new file mode 100644
index 0000000..85bac21
--- /dev/null
+++ b/cpu/bpf.cpu
@@ -0,0 +1,647 @@
+;; Linux BPF CPU description -*- Scheme -*-
+;; Copyright (C) 2019 Free Software Foundation, Inc.
+;;
+;; Contributed by Oracle Inc.
+;;
+;; This file is part of the GNU Binutils and of GDB.
+;;
+;; This program is free software; you can redistribute it and/or
+;; modify it under the terms of the GNU General Public License as
+;; published by the Free Software Foundation; either version 3 of the
+;; License, or (at your option) any later version.
+;;
+;; This program is distributed in the hope that it will be useful, but
+;; WITHOUT ANY WARRANTY; without even the implied warranty of
+;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+;; General Public License for more details.
+;;
+;; You should have received a copy of the GNU General Public License
+;; along with this program; if not, write to the Free Software
+;; Foundation, Inc., 51 Franklin Street - Fifth Floor, Boston, MA
+;; 02110-1301, USA.
+
+;; This file contains a CGEN CPU description for the Linux kernel eBPF
+;; instruction set. eBPF is documented in the linux kernel source
+;; tree. See linux/Documentation/networking/filter.txt, and also the
+;; sources in the networking subsystem, notably
+;; linux/net/core/filter.c.
+
+(include "simplify.inc")
+
+(define-arch
+ (name bpf)
+ (comment "Linux kernel BPF")
+ (insn-lsb0? #t)
+ (machs bpf)
+ (isas ebpfle ebpfbe))
+
+;;;; The ISAs
+
+;; Logically, eBPF comforms a single instruction set featuring two
+;; kind of instructions: 64-bit instructions and 128-bit instructions.
+;;
+;; The 64-bit instructions have the form:
+;;
+;; code:8 regs:8 offset:16 imm:32
+;;
+;; Whereas the 128-bit instructions (at the moment there is only one
+;; of such instructions, lddw) have the form:
+;;
+;; code:8 regs:8 offset:16 imm:32 imm:32 unused:32
+;;
+;; In both formats `regs' is itself composed by two fields:
+;;
+;; dst:4 src:4
+;;
+;; The ISA is supposed to be orthogonal to endianness: the endianness
+;; of the instruction fields follow the endianness of the host running
+;; the eBPF program, and that's all. However, this is not entirely
+;; true. The definition of an eBPF code in the Linux kernel is:
+;;
+;; struct bpf_insn {
+;; __u8 code; /* opcode */
+;; __u8 dst_reg:4; /* dest register */
+;; __u8 src_reg:4; /* source register */
+;; __s16 off; /* signed offset */
+;; __s32 imm; /* signed immediate constant */
+;; };
+;;
+;; Since the ordering of fields in C bitmaps is defined by the
+;; implementation, the impact of endianness in the encoding of eBPF
+;; instructions is effectively defined by GCC. In particular, GCC
+;; places dst_reg before src_reg in little-endian code, and the other
+;; way around in big-endian code.
+;;
+;; So, in reality, eBPF comprises two instruction sets: one for
+;; little-endian with instructions like:
+;;
+;; code:8 src:4 dst:4 offset:16 imm:32 [unused:32 imm:32]
+;;
+;; and another for big-endian with instructions like:
+;;
+;; code:8 dst:4 src:4 offset:16 imm:32 [unused:32 imm:32]
+;;
+;; where `offset' and the immediate fields are encoded in
+;; little-endian and big-endian byte-order, respectively.
+
+(define-pmacro (define-bpf-isa x-endian)
+ (define-isa
+ (name (.sym ebpf x-endian))
+ (comment "The eBPF instruction set")
+ ;; Default length to record in ifields. This is used in
+ ;; calculations involving bit numbers.
+ (default-insn-word-bitsize 64)
+ ;; Length of an unknown instruction. Used by disassembly and by the
+ ;; simulator's invalid insn handler.
+ (default-insn-bitsize 64)
+ ;; Number of bits of insn that can be initially fetched. XXX this
+ ;; should be 64 (the size of the smallest insn) but until CGEN
+ ;; gets fixed to place constant fields in their own words, we have
+ ;; to use this workaround to avoid the opcode byte to be placed at
+ ;; the wrong side of the instruction when assembling in
+ ;; big-endian.
+ (base-insn-bitsize 8)))
+
+(define-bpf-isa le)
+(define-bpf-isa be)
+
+(define-pmacro all-isas () (ISA ebpfle,ebpfbe))
+
+;;;; Hardware Hierarchy
+
+;;
+;; bpf architecture
+;; |
+;; bpfbf cpu-family
+;; |
+;; bpf machine
+;; |
+;; bpf-def model
+
+(define-cpu
+ (name bpfbf)
+ (comment "Linux kernel eBPF virtual CPU")
+ (word-bitsize 32))
+
+(define-mach
+ (name bpf)
+ (comment "Linux eBPF")
+ (cpu bpfbf)
+ (isas ebpfle ebpfbe))
+
+(define-model
+ (name bpf-def)
+ (comment "Linux eBPF default model")
+ (mach bpf)
+ (unit u-exec "execution unit" ()
+ 1 ; issue
+ 1 ; done
+ () ; state
+ () ; inputs
+ () ; outputs
+ () ; profile action (default)
+ ))
+
+;;;; Hardware Elements
+
+;; eBPF programs can access 10 general-purpose registers which are
+;; 64-bit.
+
+(define-hardware
+ (name h-gpr)
+ (comment "General Purpose Registers")
+ (attrs all-isas (MACH bpf))
+ (type register DI (16))
+ (indices keyword "%"
+ ;; XXX the frame pointer fp is read-only, so it should
+ ;; go in a different hardware.
+ (;; ABI names. Take priority when disassembling.
+ (a 0) (r1 1) (r2 2) (r3 3) (r4 4) (r5 5) (ctx 6)
+ (r7 7) (r8 8) (r9 9) (fp 10)
+ ;; Additional names recognized when assembling.
+ (r0 0) (r6 6) (r10 10))))
+
+;; The program counter. CGEN requires it, even if it is not visible
+;; to eBPF programs.
+
+(dnh h-pc "program counter" (PC PROFILE) (pc) () () ())
+
+;; A 64-bit h-sint to be used by the imm64 operand below. XXX this
+;; shouldn't be needed, as h-sint is supposed to be able to hold
+;; 64-bit values. However, in practice CGEN limits h-sint to 32 bits
+;; in 32-bit hosts. To be fixed in CGEN.
+
+(dnh h-sint64 "signed 64-bit integer" (all-isas) (immediate DI)
+ () () ())
+
+;;;; The Instruction Sets
+
+;;; Fields and Opcodes
+
+;; Convenience macro to shorten the definition of the fields below.
+(define-pmacro (dwf x-name x-comment x-attrs
+ x-word-offset x-word-length x-start x-length
+ x-mode)
+ "Define a field including its containing word."
+ (define-ifield
+ (name x-name)
+ (comment x-comment)
+ (.splice attrs (.unsplice x-attrs))
+ (word-offset x-word-offset)
+ (word-length x-word-length)
+ (start x-start)
+ (length x-length)
+ (mode x-mode)))
+
+;; For arithmetic and jump instructions the 8-bit code field is
+;; subdivided in:
+;;
+;; op-code:4 op-src:1 op-class:3
+
+(dwf f-op-code "eBPF opcode code" (all-isas) 0 8 7 4 UINT)
+(dwf f-op-src "eBPF opcode source" (all-isas) 0 8 3 1 UINT)
+(dwf f-op-class "eBPF opcode instruction class" (all-isas) 0 8 2 3 UINT)
+
+(define-normal-insn-enum insn-op-code-alu "eBPF instruction codes"
+ (all-isas) OP_CODE_ f-op-code
+ (;; Codes for OP_CLASS_ALU and OP_CLASS_ALU64
+ (ADD #x0) (SUB #x1) (MUL #x2) (DIV #x3) (OR #x4) (AND #x5)
+ (LSH #x6) (RSH #x7) (NEG #x8) (MOD #x9) (XOR #xa) (MOV #xb)
+ (ARSH #xc) (END #xd)
+ ;; Codes for OP_CLASS_JMP
+ (JA #x0) (JEQ #x1) (JGT #x2) (JGE #x3) (JSET #x4)
+ (JNE #x5) (JSGT #x6) (JSGE #x7) (CALL #x8) (EXIT #x9)
+ (JLT #xa) (JLE #xb) (JSLT #xc) (JSLE #xd)))
+
+(define-normal-insn-enum insn-op-src "eBPF instruction source"
+ (all-isas) OP_SRC_ f-op-src
+ ;; X => use `src' as source operand.
+ ;; K => use `imm32' as source operand.
+ ((K #b0) (X #b1)))
+
+(define-normal-insn-enum insn-op-class "eBPF instruction class"
+ (all-isas) OP_CLASS_ f-op-class
+ ((LD #b000) (LDX #b001) (ST #b010) (STX #b011)
+ (ALU #b100) (JMP #b101) (ALU64 #b111)))
+
+;; For load/store instructions, the 8-bit code field is subdivided in:
+;;
+;; op-mode:3 op-size:2 op-class:3
+
+(dwf f-op-mode "eBPF opcode mode" (all-isas) 0 8 7 3 UINT)
+(dwf f-op-size "eBPF opcode size" (all-isas) 0 8 4 2 UINT)
+
+(define-normal-insn-enum insn-op-mode "eBPF load/store instruction modes"
+ (all-isas) OP_MODE_ f-op-mode
+ ((IMM #b000) (ABS #b001) (IND #b010) (MEM #b011)
+ ;; #b100 and #b101 are used in classic BPF only, reserved in eBPF.
+ (XADD #b110)))
+
+(define-normal-insn-enum insn-op-size "eBPF load/store instruction sizes"
+ (all-isas) OP_SIZE_ f-op-size
+ ((W #b00) ;; Word: 4 byte
+ (H #b01) ;; Half-word: 2 byte
+ (B #b10) ;; Byte: 1 byte
+ (DW #b11))) ;; Double-word: 8 byte
+
+;; The fields for the source and destination registers are a bit
+;; tricky. Due to the bizarre nibble swap between little-endian and
+;; big-endian ISAs we need to keep different variants of the fields.
+;;
+;; Note that f-regs is used in the format spec of instructions that do
+;; NOT use registers, where endianness is irrelevant i.e. f-regs is a
+;; constant 0 opcode.
+
+(dwf f-dstle "eBPF dst register field" ((ISA ebpfle)) 8 8 3 4 UINT)
+(dwf f-srcle "eBPF source register field" ((ISA ebpfle)) 8 8 7 4 UINT)
+
+(dwf f-dstbe "eBPF dst register field" ((ISA ebpfbe)) 8 8 7 4 UINT)
+(dwf f-srcbe "eBPF source register field" ((ISA ebpfbe)) 8 8 3 4 UINT)
+
+(dwf f-regs "eBPF registers field" (all-isas) 8 8 7 8 UINT)
+
+;; Finally, the fields for the immediates.
+;;
+;; The 16-bit offsets and 32-bit immediates do not present any special
+;; difficulty: we put them in their own instruction word so the
+;; byte-endianness will be properly applied.
+
+(dwf f-offset16 "eBPF offset field" (all-isas) 16 16 15 16 INT)
+(dwf f-imm32 "eBPF 32-bit immediate field" (all-isas) 32 32 31 32 INT)
+
+;; For the disjoint 64-bit signed immediate, however, we need to use a
+;; multi-ifield.
+
+(dwf f-imm64-a "eBPF 64-bit immediate a" (all-isas) 32 32 31 32 UINT)
+(dwf f-imm64-b "eBPF 64-bit immediate b" (all-isas) 64 32 31 32 UINT)
+(dwf f-imm64-c "eBPF 64-bit immediate c" (all-isas) 96 32 31 32 UINT)
+
+(define-multi-ifield
+ (name f-imm64)
+ (comment "eBPF 64-bit immediate field")
+ (attrs all-isas)
+ (mode DI)
+ (subfields f-imm64-a f-imm64-b f-imm64-c)
+ (insert (sequence ()
+ (set (ifield f-imm64-b) (const 0))
+ (set (ifield f-imm64-c) (srl (ifield f-imm64) (const 32)))
+ (set (ifield f-imm64-a) (and (ifield f-imm64) (const #xffffffff)))))
+ (extract (sequence ()
+ (set (ifield f-imm64)
+ (or (sll DI (zext DI (ifield f-imm64-c)) (const 32))
+ (zext DI (ifield f-imm64-a)))))))
+
+;;; Operands
+
+;; A couple of source and destination register operands are defined
+;; for each ISA: ebpfle and ebpfbe.
+
+(dno dstle "destination register" ((ISA ebpfle)) h-gpr f-dstle)
+(dno srcle "source register" ((ISA ebpfle)) h-gpr f-srcle)
+
+(dno dstbe "destination register" ((ISA ebpfbe)) h-gpr f-dstbe)
+(dno srcbe "source register" ((ISA ebpfbe)) h-gpr f-srcbe)
+
+;; Jump instructions have a 16-bit PC-relative address.
+;; CALL instructions have a 32-bit PC-relative address.
+
+(dno disp16 "16-bit PC-relative address" (all-isas PCREL-ADDR) h-sint
+ f-offset16)
+(dno disp32 "32-bit PC-relative address" (all-isas PCREL-ADDR) h-sint
+ f-imm32)
+
+;; Immediate operands in eBPF are signed, and we want the disassembler
+;; to print negative values in a sane way. Therefore we use the macro
+;; below to register a printer, which is itself defined as a C
+;; function in bpf.opc.
+
+;; define-normal-signed-immediate-operand
+(define-pmacro (dnsio x-name x-comment x-attrs x-type x-index)
+ (define-operand
+ (name x-name)
+ (comment x-comment)
+ (.splice attrs (.unsplice x-attrs))
+ (type x-type)
+ (index x-index)
+ (handlers (print "immediate"))))
+
+(dnsio imm32 "32-bit immediate" (all-isas) h-sint f-imm32)
+(dnsio offset16 "16-bit offset" (all-isas) h-sint f-offset16)
+
+;; The 64-bit immediate cannot use the default
+;; cgen_parse_signed_integer, because it assumes operands are at much
+;; 32-bit wide. Use our own.
+
+(define-operand
+ (name imm64)
+ (comment "64-bit immediate")
+ (attrs all-isas)
+ (type h-sint64)
+ (index f-imm64)
+ (handlers (parse "imm64") (print "immediate")))
+
+;; The endle/endbe instructions take an operand to specify the word
+;; width in endianness conversions. We use both a parser and printer,
+;; which are defined as C functions in bpf.opc.
+
+(define-operand
+ (name endsize)
+ (comment "endianness size immediate: 16, 32 or 64")
+ (attrs all-isas)
+ (type h-uint)
+ (index f-imm32)
+ (handlers (parse "endsize") (print "endsize")))
+
+;;; ALU instructions
+
+;; For each opcode in insn-op-code-alu representing and integer
+;; arithmetic instruction (ADD, SUB, etc) we define a bunch of
+;; instruction variants:
+;;
+;; ADD[32]{i,r}le for the little-endian ISA
+;; ADD[32]{i,r}be for the big-endian ISA
+;;
+;; The `i' variants perform `src OP dst -> dst' operations.
+;; The `r' variants perform `dst OP imm32 -> dst' operations.
+;;
+;; The variants with 32 in their name are of ALU class. Otherwise
+;; they are ALU64 class.
+
+(define-pmacro (define-alu-insn-un x-basename x-suffix x-op-class x-op-code x-endian)
+ (dni (.sym x-basename x-suffix x-endian)
+ (.str x-basename x-suffix)
+ ((ISA (.sym ebpf x-endian)))
+ (.str x-basename x-suffix " $dst" x-endian)
+ (+ (f-imm32 0) (f-offset16 0) ((.sym f-src x-endian) 0) (.sym dst x-endian)
+ x-op-class OP_SRC_X x-op-code) () ()))
+
+(define-pmacro (define-alu-insn-bin x-basename x-suffix x-op-class x-op-code x-endian)
+ (begin
+ (dni (.sym x-basename x-suffix "i" x-endian)
+ (.str x-basename x-suffix " immediate")
+ ((ISA (.sym ebpf x-endian)))
+ (.str x-basename x-suffix " $dst" x-endian ",$imm32")
+ (+ imm32 (f-offset16 0) ((.sym f-src x-endian) 0) (.sym dst x-endian)
+ x-op-class OP_SRC_K x-op-code) () ())
+ (dni (.sym x-basename x-suffix "r" x-endian)
+ (.str x-basename x-suffix " register")
+ ((ISA (.sym ebpf x-endian)))
+ (.str x-basename x-suffix " $dst" x-endian ",$src" x-endian)
+ (+ (f-imm32 0) (f-offset16 0) (.sym src x-endian) (.sym dst x-endian)
+ x-op-class OP_SRC_X x-op-code) () ())))
+
+(define-pmacro (daiu x-basename x-op-code x-endian)
+ (begin
+ (define-alu-insn-un x-basename "" OP_CLASS_ALU64 x-op-code x-endian)
+ (define-alu-insn-un x-basename "32" OP_CLASS_ALU x-op-code x-endian)))
+
+(define-pmacro (daib x-basename x-op-code x-endian)
+ (begin
+ (define-alu-insn-bin x-basename "" OP_CLASS_ALU64 x-op-code x-endian)
+ (define-alu-insn-bin x-basename "32" OP_CLASS_ALU x-op-code x-endian)))
+
+(define-pmacro (define-alu-instructions x-endian)
+ (begin
+ (daib add OP_CODE_ADD x-endian)
+ (daib sub OP_CODE_SUB x-endian)
+ (daib mul OP_CODE_MUL x-endian)
+ (daib div OP_CODE_DIV x-endian)
+ (daib or OP_CODE_OR x-endian)
+ (daib and OP_CODE_AND x-endian)
+ (daib lsh OP_CODE_LSH x-endian)
+ (daib rsh OP_CODE_RSH x-endian)
+ (daib mod OP_CODE_MOD x-endian)
+ (daib xor OP_CODE_XOR x-endian)
+ (daib mov OP_CODE_MOV x-endian)
+ (daib arsh OP_CODE_ARSH x-endian)
+ (daiu neg OP_CODE_NEG x-endian)))
+
+(define-alu-instructions le)
+(define-alu-instructions be)
+
+;;; Endianness conversion instructions
+
+;; The endianness conversion instructions come in several variants:
+;;
+;; END{le,be}le for the little-endian ISA
+;; END{le,be}be for the big-endian ISA
+;;
+;; Please do not be confused by the repeated `be' and `le' here. Each
+;; ISA has both endle and endbe instructions. It is the disposition
+;; of the source and destination register fields that change between
+;; ISAs, not the semantics of the instructions themselves (see section
+;; "The ISAs" above in this very file.)
+
+(define-pmacro (define-endian-insn x-suffix x-op-src x-endian)
+ (dni (.sym "end" x-suffix x-endian)
+ (.str "end" x-suffix " register")
+ ((ISA (.sym ebpf x-endian)))
+ (.str "end" x-suffix " $dst" x-endian ",$endsize")
+ (+ (f-offset16 0) ((.sym f-src x-endian) 0) (.sym dst x-endian) endsize
+ OP_CLASS_ALU x-op-src OP_CODE_END) () ()))
+
+(define-endian-insn "le" OP_SRC_K le)
+(define-endian-insn "be" OP_SRC_X le)
+(define-endian-insn "le" OP_SRC_K be)
+(define-endian-insn "be" OP_SRC_X be)
+
+;;; Load/Store instructions
+
+;; The lddw instruction takes a 64-bit immediate as an operand. Since
+;; this instruction also takes a `dst' operand, we need to define a
+;; variant for each ISA:
+;;
+;; LDDWle for the little-endian ISA
+;; LDDWbe for the big-endian ISA
+
+(define-pmacro (define-lddw x-endian)
+ (dni (.sym lddw x-endian)
+ (.str "lddw" x-endian)
+ ((ISA (.sym ebpf x-endian)))
+ (.str "lddw $dst" x-endian ",$imm64")
+ (+ imm64 (f-offset16 0) ((.sym f-src x-endian) 0)
+ (.sym dst x-endian)
+ OP_CLASS_LD OP_SIZE_DW OP_MODE_IMM) () ()))
+
+(define-lddw le)
+(define-lddw be)
+
+;; The absolute/indirect load instructions are non-generic loads
+;; designed to be used in socket filters. They come in several
+;; variants:
+;;
+;; LD{abs,ind}{w,h,b,dw}le for the little-endian ISA
+;; LD{abs,ind}[w,h,b,dw}be for the big-endian ISA
+
+(define-pmacro (dlsi x-basename x-suffix x-class x-size x-mode x-endian)
+ (dni (.sym x-basename x-suffix x-endian)
+ (.str x-basename x-suffix)
+ ((ISA (.sym ebpf x-endian)))
+ (.str x-basename x-suffix " $dst" x-endian ",$src" x-endian ",$imm32")
+ (+ imm32 (f-offset16 0) (.sym src x-endian) (.sym dst x-endian)
+ (.sym OP_CLASS_ x-class) (.sym OP_SIZE_ x-size)
+ (.sym OP_MODE_ x-mode)) () ()))
+
+(define-pmacro (define-ldabsind x-endian)
+ (begin
+ (dlsi "ldabs" "w" LD W ABS x-endian)
+ (dlsi "ldabs" "h" LD H ABS x-endian)
+ (dlsi "ldabs" "b" LD B ABS x-endian)
+ (dlsi "ldabs" "dw" LD DW ABS x-endian)
+
+ (dlsi "ldind" "w" LD W IND x-endian)
+ (dlsi "ldind" "h" LD H IND x-endian)
+ (dlsi "ldind" "b" LD B IND x-endian)
+ (dlsi "ldind" "dw" LD DW IND x-endian)))
+
+(define-ldabsind le)
+(define-ldabsind be)
+
+;; Generic load and store instructions are provided for several word
+;; sizes. They come in several variants:
+;;
+;; LDX{b,h,w,dw}le, STX{b,h,w,dw}le for the little-endian ISA
+;;
+;; LDX{b,h,w,dw}be, STX{b,h,w,dw}be for the big-endian ISA
+;;
+;; Loads operate on [$SRC+-OFFSET] -> $DST
+;; Stores operate on $SRC -> [$DST+-OFFSET]
+
+(define-pmacro (dxli x-basename x-suffix x-size x-endian)
+ (dni (.sym x-basename x-suffix x-endian)
+ (.str x-basename x-suffix)
+ ((ISA (.sym ebpf x-endian)))
+ (.str x-basename x-suffix " $dst" x-endian ",[$src" x-endian "+$offset16]")
+ (+ (f-imm32 0) offset16 (.sym src x-endian) (.sym dst x-endian)
+ OP_CLASS_LDX (.sym OP_SIZE_ x-size) OP_MODE_MEM)
+ () ()))
+
+(define-pmacro (dxsi x-basename x-suffix x-size x-endian)
+ (dni (.sym x-basename x-suffix x-endian)
+ (.str x-basename x-suffix)
+ ((ISA (.sym ebpf x-endian)))
+ (.str x-basename x-suffix " [$dst" x-endian "+$offset16],$src" x-endian)
+ (+ (f-imm32 0) offset16 (.sym src x-endian) (.sym dst x-endian)
+ OP_CLASS_STX (.sym OP_SIZE_ x-size) OP_MODE_MEM)
+ () ()))
+
+(define-pmacro (define-ldstx-insns x-endian)
+ (begin
+ (dxli "ldx" "w" W x-endian)
+ (dxli "ldx" "h" H x-endian)
+ (dxli "ldx" "b" B x-endian)
+ (dxli "ldx" "dw" DW x-endian)
+
+ (dxsi "stx" "w" W x-endian)
+ (dxsi "stx" "h" H x-endian)
+ (dxsi "stx" "b" B x-endian)
+ (dxsi "stx" "dw" DW x-endian)))
+
+(define-ldstx-insns le)
+(define-ldstx-insns be)
+
+;; Generic store instructions of the form IMM32 -> [$DST+OFFSET] are
+;; provided in several variants:
+;;
+;; ST{b,h,w,dw}le for the little-endian ISA
+;; ST{b,h,w,dw}be for the big-endian ISA
+
+(define-pmacro (dsti x-suffix x-size x-endian)
+ (dni (.sym "st" x-suffix x-endian)
+ (.str "st" x-suffix)
+ ((ISA (.sym ebpf x-endian)))
+ (.str "st" x-suffix " [$dst" x-endian "+$offset16],$imm32")
+ (+ imm32 offset16 ((.sym f-src x-endian) 0) (.sym dst x-endian)
+ OP_CLASS_ST (.sym OP_SIZE_ x-size) OP_MODE_MEM) () ()))
+
+(define-pmacro (define-st-insns x-endian)
+ (begin
+ (dsti "b" B x-endian)
+ (dsti "h" H x-endian)
+ (dsti "w" W x-endian)
+ (dsti "dw" DW x-endian)))
+
+(define-st-insns le)
+(define-st-insns be)
+
+;;; Jump instructions
+
+;; Compare-and-jump instructions, on the other hand, make use of
+;; registers. Therefore, we need to define several variants in both
+;; ISAs:
+;;
+;; J{eq,gt,ge,lt,le,set,ne,sgt,sge,slt,sle}{i,r}le for the
+;; little-endian ISA.
+;; J{eq,gt,ge,lt,le,set,ne.sgt,sge,slt,sle}{i,r}be for the
+;; big-endian ISA.
+
+(define-pmacro (dcji x-cond x-op-code x-endian)
+ (begin
+ (dni (.sym j x-cond i x-endian)
+ (.str j x-cond "i")
+ ((ISA (.sym ebpf x-endian)))
+ (.str "j" x-cond " $dst" x-endian ",$imm32,$disp16")
+ (+ imm32 disp16 ((.sym f-src x-endian) 0) (.sym dst x-endian)
+ OP_CLASS_JMP OP_SRC_K (.sym OP_CODE_ x-op-code)) () ())
+ (dni (.sym j x-cond r x-endian)
+ (.str j x-cond "r")
+ ((ISA (.sym ebpf x-endian)))
+ (.str "j" x-cond " $dst" x-endian ",$src" x-endian ",$disp16")
+ (+ (f-imm32 0) disp16 (.sym src x-endian) (.sym dst x-endian)
+ OP_CLASS_JMP OP_SRC_X (.sym OP_CODE_ x-op-code)) () ())))
+
+(define-pmacro (define-condjump-insns x-endian)
+ (begin
+ (dcji "eq" JEQ x-endian)
+ (dcji "gt" JGT x-endian)
+ (dcji "ge" JGE x-endian)
+ (dcji "lt" JLT x-endian)
+ (dcji "le" JLE x-endian)
+ (dcji "set" JSET x-endian)
+ (dcji "ne" JNE x-endian)
+ (dcji "sgt" JSGT x-endian)
+ (dcji "sge" JSGE x-endian)
+ (dcji "slt" JSLT x-endian)
+ (dcji "sle" JSLE x-endian)))
+
+(define-condjump-insns le)
+(define-condjump-insns be)
+
+;; The jump-always, `call' and `exit' instructions dont make use of
+;; either source nor destination registers, so only one variant per
+;; instruction is defined.
+
+(dni ja "ja" (all-isas) "ja $disp16"
+ (+ (f-imm32 0) disp16 (f-regs 0)
+ OP_CLASS_JMP OP_SRC_K OP_CODE_JA) () ())
+
+(dni call "call" (all-isas) "call $disp32"
+ (+ disp32 (f-offset16 0) (f-regs 0)
+ OP_CLASS_JMP OP_SRC_K OP_CODE_CALL) () ())
+
+(dni "exit" "exit" (all-isas) "exit"
+ (+ (f-imm32 0) (f-offset16 0) (f-regs 0)
+ OP_CLASS_JMP (f-op-src 0) OP_CODE_EXIT) () ())
+
+;;; Atomic instructions
+
+;; The atomic exchange-and-add instructions come in two flavors: one
+;; for swapping 64-bit quantities and another for 32-bit quantities.
+
+(define-pmacro (define-atomic-insns x-endian)
+ (begin
+ (dni (.str "xadddw" x-endian)
+ "xadddw"
+ ((ISA (.sym ebpf x-endian)))
+ (.str "xadddw [$dst" x-endian "+$offset16],$src" x-endian)
+ (+ (f-imm32 0) (.sym src x-endian) (.sym dst x-endian)
+ offset16 OP_MODE_XADD OP_SIZE_DW OP_CLASS_STX) () ())
+ (dni (.str "xaddw" x-endian)
+ "xaddw"
+ ((ISA (.sym ebpf x-endian)))
+ (.str "xaddw [$dst" x-endian "+$offset16],$src" x-endian)
+ (+ (f-imm32 0) (.sym src x-endian) (.sym dst x-endian)
+ offset16 OP_MODE_XADD OP_SIZE_W OP_CLASS_STX) () ())))
+
+(define-atomic-insns le)
+(define-atomic-insns be)
diff --git a/cpu/bpf.opc b/cpu/bpf.opc
new file mode 100644
index 0000000..e70ee04
--- /dev/null
+++ b/cpu/bpf.opc
@@ -0,0 +1,191 @@
+/* EBPF opcode support. -*- c -*-
+
+ Copyright (C) 2019 Free Software Foundation, Inc.
+
+ Contributed by Oracle, Inc.
+
+ This file is part of the GNU Binutils and of GDB.
+
+ This program is free software; you can redistribute it and/or modify
+ it under the terms of the GNU General Public License as published by
+ the Free Software Foundation; either version 3 of the License, or
+ (at your option) any later version.
+
+ This program is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ GNU General Public License for more details.
+
+ You should have received a copy of the GNU General Public License
+ along with this program; if not, write to the Free Software
+ Foundation, Inc., 51 Franklin Street - Fifth Floor, Boston,
+ MA 02110-1301, USA. */
+
+/*
+ Each section is delimited with start and end markers.
+
+ <arch>-opc.h additions use: "-- opc.h"
+ <arch>-opc.c additions use: "-- opc.c"
+ <arch>-asm.c additions use: "-- asm.c"
+ <arch>-dis.c additions use: "-- dis.c"
+ <arch>-ibd.h additions use: "-- ibd.h". */
+
+/* -- opc.h */
+
+#undef CGEN_DIS_HASH_SIZE
+#define CGEN_DIS_HASH_SIZE 1
+
+#undef CGEN_DIS_HASH
+#define CGEN_DIS_HASH(buffer, value) 0
+
+/* Allows reason codes to be output when assembler errors occur. */
+#define CGEN_VERBOSE_ASSEMBLER_ERRORS
+
+#define CGEN_VALIDATE_INSN_SUPPORTED
+extern int bpf_cgen_insn_supported (CGEN_CPU_DESC, const CGEN_INSN *);
+
+
+/* -- opc.c */
+
+/* -- asm.c */
+
+/* Parse a signed 64-bit immediate. */
+
+static const char *
+parse_imm64 (CGEN_CPU_DESC cd,
+ const char **strp,
+ int opindex,
+ int64_t *valuep)
+{
+ bfd_vma value;
+ enum cgen_parse_operand_result result;
+ const char *errmsg;
+
+ errmsg = (* cd->parse_operand_fn)
+ (cd, CGEN_PARSE_OPERAND_INTEGER, strp, opindex, BFD_RELOC_NONE,
+ &result, &value);
+ if (!errmsg)
+ *valuep = value;
+
+ return errmsg;
+}
+
+/* Endianness size operands are integer immediates whose values can be
+ 16, 32 or 64. */
+
+static const char *
+parse_endsize (CGEN_CPU_DESC cd,
+ const char **strp,
+ int opindex,
+ unsigned long *valuep)
+{
+ const char *errmsg;
+
+ errmsg = cgen_parse_unsigned_integer (cd, strp, opindex, valuep);
+ if (errmsg)
+ return errmsg;
+
+ switch (*valuep)
+ {
+ case 16:
+ case 32:
+ case 64:
+ break;
+ default:
+ return _("expected 16, 32 or 64 in");
+ }
+
+ return NULL;
+}
+
+/* Special check to ensure that the right instruction variant is used
+ for the given endianness induced by the ISA selected in the CPU.
+ See bpf.cpu for a discussion on how eBPF is really two instruction
+ sets. */
+
+int
+bpf_cgen_insn_supported (CGEN_CPU_DESC cd, const CGEN_INSN *insn)
+{
+ CGEN_BITSET isas = CGEN_INSN_BITSET_ATTR_VALUE (insn, CGEN_INSN_ISA);
+
+ return cgen_bitset_intersect_p (&isas, cd->isas);
+}
+
+
+/* -- dis.c */
+
+/* We need to customize the disassembler a bit:
+ - Use 8 bytes per line by default.
+*/
+
+#define CGEN_PRINT_INSN bpf_print_insn
+
+static int
+bpf_print_insn (CGEN_CPU_DESC cd, bfd_vma pc, disassemble_info *info)
+{
+ bfd_byte buf[CGEN_MAX_INSN_SIZE];
+ int buflen;
+ int status;
+
+ info->bytes_per_chunk = 1;
+ info->bytes_per_line = 8;
+
+ /* Attempt to read the base part of the insn. */
+ buflen = cd->base_insn_bitsize / 8;
+ status = (*info->read_memory_func) (pc, buf, buflen, info);
+
+ /* Try again with the minimum part, if min < base. */
+ if (status != 0 && (cd->min_insn_bitsize < cd->base_insn_bitsize))
+ {
+ buflen = cd->min_insn_bitsize / 8;
+ status = (*info->read_memory_func) (pc, buf, buflen, info);
+ }
+
+ if (status != 0)
+ {
+ (*info->memory_error_func) (status, pc, info);
+ return -1;
+ }
+
+ return print_insn (cd, pc, info, buf, buflen);
+}
+
+/* Signed immediates should be printed in hexadecimal. */
+
+static void
+print_immediate (CGEN_CPU_DESC cd ATTRIBUTE_UNUSED,
+ void *dis_info,
+ int64_t value,
+ unsigned int attrs ATTRIBUTE_UNUSED,
+ bfd_vma pc ATTRIBUTE_UNUSED,
+ int length ATTRIBUTE_UNUSED)
+{
+ disassemble_info *info = (disassemble_info *) dis_info;
+
+ if (value <= 9)
+ (*info->fprintf_func) (info->stream, "%" PRId64, value);
+ else
+ (*info->fprintf_func) (info->stream, "%#" PRIx64, value);
+
+ /* This is to avoid -Wunused-function for print_normal. */
+ if (0)
+ print_normal (cd, dis_info, value, attrs, pc, length);
+}
+
+/* Endianness bit sizes should be printed in decimal. */
+
+static void
+print_endsize (CGEN_CPU_DESC cd ATTRIBUTE_UNUSED,
+ void *dis_info,
+ unsigned long value,
+ unsigned int attrs ATTRIBUTE_UNUSED,
+ bfd_vma pc ATTRIBUTE_UNUSED,
+ int length ATTRIBUTE_UNUSED)
+{
+ disassemble_info *info = (disassemble_info *) dis_info;
+ (*info->fprintf_func) (info->stream, "%lu", value);
+}
+
+
+/* -- */
+