xref: /openbmc/linux/arch/loongarch/net/bpf_jit.h (revision 519b58bb)
1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /*
3  * BPF JIT compiler for LoongArch
4  *
5  * Copyright (C) 2022 Loongson Technology Corporation Limited
6  */
7 #include <linux/bpf.h>
8 #include <linux/filter.h>
9 #include <asm/cacheflush.h>
10 #include <asm/inst.h>
11 
12 struct jit_ctx {
13 	const struct bpf_prog *prog;
14 	unsigned int idx;
15 	unsigned int flags;
16 	unsigned int epilogue_offset;
17 	u32 *offset;
18 	union loongarch_instruction *image;
19 	u32 stack_size;
20 };
21 
22 struct jit_data {
23 	struct bpf_binary_header *header;
24 	u8 *image;
25 	struct jit_ctx ctx;
26 };
27 
28 #define emit_insn(ctx, func, ...)						\
29 do {										\
30 	if (ctx->image != NULL) {						\
31 		union loongarch_instruction *insn = &ctx->image[ctx->idx];	\
32 		emit_##func(insn, ##__VA_ARGS__);				\
33 	}									\
34 	ctx->idx++;								\
35 } while (0)
36 
37 #define is_signed_imm12(val)	signed_imm_check(val, 12)
38 #define is_signed_imm14(val)	signed_imm_check(val, 14)
39 #define is_signed_imm16(val)	signed_imm_check(val, 16)
40 #define is_signed_imm26(val)	signed_imm_check(val, 26)
41 #define is_signed_imm32(val)	signed_imm_check(val, 32)
42 #define is_signed_imm52(val)	signed_imm_check(val, 52)
43 #define is_unsigned_imm12(val)	unsigned_imm_check(val, 12)
44 
45 static inline int bpf2la_offset(int bpf_insn, int off, const struct jit_ctx *ctx)
46 {
47 	/* BPF JMP offset is relative to the next instruction */
48 	bpf_insn++;
49 	/*
50 	 * Whereas LoongArch branch instructions encode the offset
51 	 * from the branch itself, so we must subtract 1 from the
52 	 * instruction offset.
53 	 */
54 	return (ctx->offset[bpf_insn + off] - (ctx->offset[bpf_insn] - 1));
55 }
56 
57 static inline int epilogue_offset(const struct jit_ctx *ctx)
58 {
59 	int from = ctx->idx;
60 	int to = ctx->epilogue_offset;
61 
62 	return (to - from);
63 }
64 
65 /* Zero-extend 32 bits into 64 bits */
66 static inline void emit_zext_32(struct jit_ctx *ctx, enum loongarch_gpr reg, bool is32)
67 {
68 	if (!is32)
69 		return;
70 
71 	emit_insn(ctx, lu32id, reg, 0);
72 }
73 
74 /* Signed-extend 32 bits into 64 bits */
75 static inline void emit_sext_32(struct jit_ctx *ctx, enum loongarch_gpr reg, bool is32)
76 {
77 	if (!is32)
78 		return;
79 
80 	emit_insn(ctx, addiw, reg, reg, 0);
81 }
82 
83 static inline void move_imm(struct jit_ctx *ctx, enum loongarch_gpr rd, long imm, bool is32)
84 {
85 	long imm_11_0, imm_31_12, imm_51_32, imm_63_52, imm_51_0, imm_51_31;
86 
87 	/* or rd, $zero, $zero */
88 	if (imm == 0) {
89 		emit_insn(ctx, or, rd, LOONGARCH_GPR_ZERO, LOONGARCH_GPR_ZERO);
90 		return;
91 	}
92 
93 	/* addiw rd, $zero, imm_11_0 */
94 	if (is_signed_imm12(imm)) {
95 		emit_insn(ctx, addiw, rd, LOONGARCH_GPR_ZERO, imm);
96 		goto zext;
97 	}
98 
99 	/* ori rd, $zero, imm_11_0 */
100 	if (is_unsigned_imm12(imm)) {
101 		emit_insn(ctx, ori, rd, LOONGARCH_GPR_ZERO, imm);
102 		goto zext;
103 	}
104 
105 	/* lu52id rd, $zero, imm_63_52 */
106 	imm_63_52 = (imm >> 52) & 0xfff;
107 	imm_51_0 = imm & 0xfffffffffffff;
108 	if (imm_63_52 != 0 && imm_51_0 == 0) {
109 		emit_insn(ctx, lu52id, rd, LOONGARCH_GPR_ZERO, imm_63_52);
110 		return;
111 	}
112 
113 	/* lu12iw rd, imm_31_12 */
114 	imm_31_12 = (imm >> 12) & 0xfffff;
115 	emit_insn(ctx, lu12iw, rd, imm_31_12);
116 
117 	/* ori rd, rd, imm_11_0 */
118 	imm_11_0 = imm & 0xfff;
119 	if (imm_11_0 != 0)
120 		emit_insn(ctx, ori, rd, rd, imm_11_0);
121 
122 	if (!is_signed_imm32(imm)) {
123 		if (imm_51_0 != 0) {
124 			/*
125 			 * If bit[51:31] is all 0 or all 1,
126 			 * it means bit[51:32] is sign extended by lu12iw,
127 			 * no need to call lu32id to do a new filled operation.
128 			 */
129 			imm_51_31 = (imm >> 31) & 0x1fffff;
130 			if (imm_51_31 != 0 || imm_51_31 != 0x1fffff) {
131 				/* lu32id rd, imm_51_32 */
132 				imm_51_32 = (imm >> 32) & 0xfffff;
133 				emit_insn(ctx, lu32id, rd, imm_51_32);
134 			}
135 		}
136 
137 		/* lu52id rd, rd, imm_63_52 */
138 		if (!is_signed_imm52(imm))
139 			emit_insn(ctx, lu52id, rd, rd, imm_63_52);
140 	}
141 
142 zext:
143 	emit_zext_32(ctx, rd, is32);
144 }
145 
146 static inline void move_reg(struct jit_ctx *ctx, enum loongarch_gpr rd,
147 			    enum loongarch_gpr rj)
148 {
149 	emit_insn(ctx, or, rd, rj, LOONGARCH_GPR_ZERO);
150 }
151 
152 static inline int invert_jmp_cond(u8 cond)
153 {
154 	switch (cond) {
155 	case BPF_JEQ:
156 		return BPF_JNE;
157 	case BPF_JNE:
158 	case BPF_JSET:
159 		return BPF_JEQ;
160 	case BPF_JGT:
161 		return BPF_JLE;
162 	case BPF_JGE:
163 		return BPF_JLT;
164 	case BPF_JLT:
165 		return BPF_JGE;
166 	case BPF_JLE:
167 		return BPF_JGT;
168 	case BPF_JSGT:
169 		return BPF_JSLE;
170 	case BPF_JSGE:
171 		return BPF_JSLT;
172 	case BPF_JSLT:
173 		return BPF_JSGE;
174 	case BPF_JSLE:
175 		return BPF_JSGT;
176 	}
177 	return -1;
178 }
179 
180 static inline void cond_jmp_offset(struct jit_ctx *ctx, u8 cond, enum loongarch_gpr rj,
181 				   enum loongarch_gpr rd, int jmp_offset)
182 {
183 	switch (cond) {
184 	case BPF_JEQ:
185 		/* PC += jmp_offset if rj == rd */
186 		emit_insn(ctx, beq, rj, rd, jmp_offset);
187 		return;
188 	case BPF_JNE:
189 	case BPF_JSET:
190 		/* PC += jmp_offset if rj != rd */
191 		emit_insn(ctx, bne, rj, rd, jmp_offset);
192 		return;
193 	case BPF_JGT:
194 		/* PC += jmp_offset if rj > rd (unsigned) */
195 		emit_insn(ctx, bltu, rd, rj, jmp_offset);
196 		return;
197 	case BPF_JLT:
198 		/* PC += jmp_offset if rj < rd (unsigned) */
199 		emit_insn(ctx, bltu, rj, rd, jmp_offset);
200 		return;
201 	case BPF_JGE:
202 		/* PC += jmp_offset if rj >= rd (unsigned) */
203 		emit_insn(ctx, bgeu, rj, rd, jmp_offset);
204 		return;
205 	case BPF_JLE:
206 		/* PC += jmp_offset if rj <= rd (unsigned) */
207 		emit_insn(ctx, bgeu, rd, rj, jmp_offset);
208 		return;
209 	case BPF_JSGT:
210 		/* PC += jmp_offset if rj > rd (signed) */
211 		emit_insn(ctx, blt, rd, rj, jmp_offset);
212 		return;
213 	case BPF_JSLT:
214 		/* PC += jmp_offset if rj < rd (signed) */
215 		emit_insn(ctx, blt, rj, rd, jmp_offset);
216 		return;
217 	case BPF_JSGE:
218 		/* PC += jmp_offset if rj >= rd (signed) */
219 		emit_insn(ctx, bge, rj, rd, jmp_offset);
220 		return;
221 	case BPF_JSLE:
222 		/* PC += jmp_offset if rj <= rd (signed) */
223 		emit_insn(ctx, bge, rd, rj, jmp_offset);
224 		return;
225 	}
226 }
227 
228 static inline void cond_jmp_offs26(struct jit_ctx *ctx, u8 cond, enum loongarch_gpr rj,
229 				   enum loongarch_gpr rd, int jmp_offset)
230 {
231 	cond = invert_jmp_cond(cond);
232 	cond_jmp_offset(ctx, cond, rj, rd, 2);
233 	emit_insn(ctx, b, jmp_offset);
234 }
235 
236 static inline void uncond_jmp_offs26(struct jit_ctx *ctx, int jmp_offset)
237 {
238 	emit_insn(ctx, b, jmp_offset);
239 }
240 
241 static inline int emit_cond_jmp(struct jit_ctx *ctx, u8 cond, enum loongarch_gpr rj,
242 				enum loongarch_gpr rd, int jmp_offset)
243 {
244 	/*
245 	 * A large PC-relative jump offset may overflow the immediate field of
246 	 * the native conditional branch instruction, triggering a conversion
247 	 * to use an absolute jump instead, this jump sequence is particularly
248 	 * nasty. For now, use cond_jmp_offs26() directly to keep it simple.
249 	 * In the future, maybe we can add support for far branching, the branch
250 	 * relaxation requires more than two passes to converge, the code seems
251 	 * too complex to understand, not quite sure whether it is necessary and
252 	 * worth the extra pain. Anyway, just leave it as it is to enhance code
253 	 * readability now.
254 	 */
255 	if (is_signed_imm26(jmp_offset)) {
256 		cond_jmp_offs26(ctx, cond, rj, rd, jmp_offset);
257 		return 0;
258 	}
259 
260 	return -EINVAL;
261 }
262 
263 static inline int emit_uncond_jmp(struct jit_ctx *ctx, int jmp_offset)
264 {
265 	if (is_signed_imm26(jmp_offset)) {
266 		uncond_jmp_offs26(ctx, jmp_offset);
267 		return 0;
268 	}
269 
270 	return -EINVAL;
271 }
272 
273 static inline int emit_tailcall_jmp(struct jit_ctx *ctx, u8 cond, enum loongarch_gpr rj,
274 				    enum loongarch_gpr rd, int jmp_offset)
275 {
276 	if (is_signed_imm16(jmp_offset)) {
277 		cond_jmp_offset(ctx, cond, rj, rd, jmp_offset);
278 		return 0;
279 	}
280 
281 	return -EINVAL;
282 }
283