]> git.proxmox.com Git - mirror_ubuntu-disco-kernel.git/commitdiff
bpf, x64: save several bytes by using mov over movabsq when possible
authorDaniel Borkmann <daniel@iogearbox.net>
Sat, 24 Feb 2018 00:07:59 +0000 (01:07 +0100)
committerAlexei Starovoitov <ast@kernel.org>
Sat, 24 Feb 2018 06:50:00 +0000 (22:50 -0800)
While analyzing some of the more complex BPF programs from Cilium,
I found that LLVM generally prefers to emit LD_IMM64 instead of MOV32
BPF instructions for loading unsigned 32-bit immediates into a
register. Given we cannot change the current/stable LLVM versions
that are already out there, lets optimize this case such that the
JIT prefers to emit 'mov %eax, imm32' over 'movabsq %rax, imm64'
whenever suitable in order to reduce the image size by 4-5 bytes per
such load in the typical case, reducing image size on some of the
bigger programs by up to 4%. emit_mov_imm32() and emit_mov_imm64()
have been added as helpers.

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
arch/x86/net/bpf_jit_comp.c

index 4bc36bd1b97a425e399cdb4dfaaa688172fc769f..f3e5cd8c1e6899ca8546441ab3053d1e40de1cbf 100644 (file)
@@ -60,7 +60,12 @@ static bool is_imm8(int value)
 
 static bool is_simm32(s64 value)
 {
-       return value == (s64) (s32) value;
+       return value == (s64)(s32)value;
+}
+
+static bool is_uimm32(u64 value)
+{
+       return value == (u64)(u32)value;
 }
 
 /* mov dst, src */
@@ -355,6 +360,68 @@ static void emit_load_skb_data_hlen(u8 **pprog)
        *pprog = prog;
 }
 
+static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
+                          u32 dst_reg, const u32 imm32)
+{
+       u8 *prog = *pprog;
+       u8 b1, b2, b3;
+       int cnt = 0;
+
+       /* optimization: if imm32 is positive, use 'mov %eax, imm32'
+        * (which zero-extends imm32) to save 2 bytes.
+        */
+       if (sign_propagate && (s32)imm32 < 0) {
+               /* 'mov %rax, imm32' sign extends imm32 */
+               b1 = add_1mod(0x48, dst_reg);
+               b2 = 0xC7;
+               b3 = 0xC0;
+               EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
+               goto done;
+       }
+
+       /* optimization: if imm32 is zero, use 'xor %eax, %eax'
+        * to save 3 bytes.
+        */
+       if (imm32 == 0) {
+               if (is_ereg(dst_reg))
+                       EMIT1(add_2mod(0x40, dst_reg, dst_reg));
+               b2 = 0x31; /* xor */
+               b3 = 0xC0;
+               EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
+               goto done;
+       }
+
+       /* mov %eax, imm32 */
+       if (is_ereg(dst_reg))
+               EMIT1(add_1mod(0x40, dst_reg));
+       EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
+done:
+       *pprog = prog;
+}
+
+static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
+                          const u32 imm32_hi, const u32 imm32_lo)
+{
+       u8 *prog = *pprog;
+       int cnt = 0;
+
+       if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
+               /* For emitting plain u32, where sign bit must not be
+                * propagated LLVM tends to load imm64 over mov32
+                * directly, so save couple of bytes by just doing
+                * 'mov %eax, imm32' instead.
+                */
+               emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
+       } else {
+               /* movabsq %rax, imm64 */
+               EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
+               EMIT(imm32_lo, 4);
+               EMIT(imm32_hi, 4);
+       }
+
+       *pprog = prog;
+}
+
 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
                  int oldproglen, struct jit_context *ctx)
 {
@@ -377,7 +444,7 @@ static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
                const s32 imm32 = insn->imm;
                u32 dst_reg = insn->dst_reg;
                u32 src_reg = insn->src_reg;
-               u8 b1 = 0, b2 = 0, b3 = 0;
+               u8 b2 = 0, b3 = 0;
                s64 jmp_offset;
                u8 jmp_cond;
                bool reload_skb_data;
@@ -485,58 +552,13 @@ static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
                        break;
 
                case BPF_ALU64 | BPF_MOV | BPF_K:
-                       /* optimization: if imm32 is positive,
-                        * use 'mov eax, imm32' (which zero-extends imm32)
-                        * to save 2 bytes
-                        */
-                       if (imm32 < 0) {
-                               /* 'mov rax, imm32' sign extends imm32 */
-                               b1 = add_1mod(0x48, dst_reg);
-                               b2 = 0xC7;
-                               b3 = 0xC0;
-                               EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
-                               break;
-                       }
-
                case BPF_ALU | BPF_MOV | BPF_K:
-                       /* optimization: if imm32 is zero, use 'xor <dst>,<dst>'
-                        * to save 3 bytes.
-                        */
-                       if (imm32 == 0) {
-                               if (is_ereg(dst_reg))
-                                       EMIT1(add_2mod(0x40, dst_reg, dst_reg));
-                               b2 = 0x31; /* xor */
-                               b3 = 0xC0;
-                               EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
-                               break;
-                       }
-
-                       /* mov %eax, imm32 */
-                       if (is_ereg(dst_reg))
-                               EMIT1(add_1mod(0x40, dst_reg));
-                       EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
+                       emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
+                                      dst_reg, imm32);
                        break;
 
                case BPF_LD | BPF_IMM | BPF_DW:
-                       /* optimization: if imm64 is zero, use 'xor <dst>,<dst>'
-                        * to save 7 bytes.
-                        */
-                       if (insn[0].imm == 0 && insn[1].imm == 0) {
-                               b1 = add_2mod(0x48, dst_reg, dst_reg);
-                               b2 = 0x31; /* xor */
-                               b3 = 0xC0;
-                               EMIT3(b1, b2, add_2reg(b3, dst_reg, dst_reg));
-
-                               insn++;
-                               i++;
-                               break;
-                       }
-
-                       /* movabsq %rax, imm64 */
-                       EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
-                       EMIT(insn[0].imm, 4);
-                       EMIT(insn[1].imm, 4);
-
+                       emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
                        insn++;
                        i++;
                        break;
@@ -604,7 +626,8 @@ static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
                                EMIT_mov(BPF_REG_0, src_reg);
                        else
                                /* mov rax, imm32 */
-                               EMIT3_off32(0x48, 0xC7, 0xC0, imm32);
+                               emit_mov_imm32(&prog, true,
+                                              BPF_REG_0, imm32);
 
                        if (BPF_CLASS(insn->code) == BPF_ALU64)
                                EMIT1(add_1mod(0x48, AUX_REG));