asm: add support for the may_goto insn and JCOND opcode

The `may_goto` instruction is a bit special and relatively new, from
March 2024, added by 011832b97b31 ("bpf: Introduce may_goto
instruction")[^1]. It uses the `BPF_JCOND` opcode `0xe0` with the source
register being `BPF_MAY_GOTO = 0`. Note that this instruction is not
documented in the latest instruction set kernel documentation[^2].

As I understand, the goal of this instruction is to be injected in loops
(also potentially by LLVM) to offload the verifier responsibility of
loop verification to a runtime check. The runtime check being
decrementing a counter so far.

[^1]: https://lore.kernel.org/bpf/20240306031929.42666-2-alexei.starovoitov@gmail.com
[^2]: https://www.kernel.org/doc/html/v6.17/bpf/standardization/instruction-set.html#jump-instructions

Co-Authored-By: Mahe Tardy <mahe.tardy@gmail.com>
Signed-off-by: Dylan Reimerink <dylan.reimerink@isovalent.com>

Author: mtardy

asm/instruction.go

@@ -396,9 +396,9 @@ func (ins Instruction) Format(f fmt.State, c rune) {
 		}
 
 	case cls.IsJump():
-		fmt.Fprintf(f, "%v ", op)
 		switch jop := op.JumpOp(); jop {
 		case Call:
+			fmt.Fprintf(f, "%v ", op)
 			switch ins.Src {
 			case PseudoCall:
 				// bpf-to-bpf call
@@ -411,13 +411,23 @@ func (ins Instruction) Format(f fmt.State, c rune) {
 			}
 
 		case Ja:
+			fmt.Fprintf(f, "%v ", op)
 			if ins.OpCode.Class() == Jump32Class {
 				fmt.Fprintf(f, "imm: %d", ins.Constant)
 			} else {
 				fmt.Fprintf(f, "off: %d", ins.Offset)
 			}
 
+		case JCOND:
+			switch ins.Src {
+			case PseudoMayGoto:
+				fmt.Fprintf(f, "JCond may_goto off: %d", ins.Offset)
+			default:
+				fmt.Fprintf(f, "%v", op)
+			}
+
 		default:
+			fmt.Fprintf(f, "%v ", op)
 			fmt.Fprintf(f, "dst: %s off: %d ", ins.Dst, ins.Offset)
 			if op.Source() == ImmSource {
 				fmt.Fprintf(f, "imm: %d", ins.Constant)

asm/instruction_test.go

@@ -393,6 +393,8 @@ func TestISAv4(t *testing.T) {
 		0xcb, 0x21, 0x00, 0x00, 0x10, 0x01, 0x00, 0x00, // store_release((u16 *)(r1 + 0x0), w2)
 		0xc3, 0x21, 0x00, 0x00, 0x10, 0x01, 0x00, 0x00, // store_release((u32 *)(r1 + 0x0), w2)
 		0xdb, 0x21, 0x00, 0x00, 0x10, 0x01, 0x00, 0x00, // store_release((u64 *)(r1 + 0x0), r2)
+
+		0xe5, 0x00, 0x1e, 0x00, 0x00, 0x00, 0x00, 0x00, // may_goto +30
 	}
 
 	insns, err := AppendInstructions(nil, bytes.NewReader(rawInsns), binary.LittleEndian, platform.Linux)
@@ -427,6 +429,7 @@ func TestISAv4(t *testing.T) {
 		"StXAtomicStRelH dst: r1 src: r2 off: 0",
 		"StXAtomicStRelW dst: r1 src: r2 off: 0",
 		"StXAtomicStRelDW dst: r1 src: r2 off: 0",
+		"JCond may_goto off: 30",
 	}
 
 	for i, ins := range insns {

asm/jump.go

@@ -44,6 +44,8 @@ const (
 	JSLT JumpOp = 0xc0
 	// JSLE jumps by offset if signed r <= signed imm
 	JSLE JumpOp = 0xd0
+	// JCOND is a conditional pseudo jump to encode the may_goto instruction
+	JCOND JumpOp = 0xe0
 )
 
 // Return emits an exit instruction.

asm/jump_string.go

@@ -44,6 +44,7 @@ func TestGetSetJumpOp(t *testing.T) {
 		JLE,
 		JSLT,
 		JSLE,
+		JCOND,
 	} {
 		test(Jump32Class, op, true)
 		test(JumpClass, op, true)

asm/opcode_test.go

@@ -40,6 +40,7 @@ const (
 	PseudoCall      = R1 // BPF_PSEUDO_CALL
 	PseudoFunc      = R4 // BPF_PSEUDO_FUNC
 	PseudoKfuncCall = R2 // BPF_PSEUDO_KFUNC_CALL
+	PseudoMayGoto   = R0 // BPF_MAY_GOTO
 )
 
 func (r Register) String() string {