Fast-path execution through memcpy/memset Injection about rv32emu

[jserv]

strcpy, memset and memcpy built with newlib are disassembled below:

000461f8 <strcpy>:
   461f8:       00050793                mv      a5,a0
   461fc:       0005c703                lbu     a4,0(a1)
   46200:       00178793                addi    a5,a5,1
   46204:       00158593                addi    a1,a1,1
   46208:       fee78fa3                sb      a4,-1(a5)
   4620c:       fe0718e3                bnez    a4,461fc <strcpy+0x4>
   46210:       00008067                ret

00044cf8 <memcpy>:
   44cf8:       00050313                mv      t1,a0
   44cfc:       00060e63                beqz    a2,44d18 <memcpy+0x20>
   44d00:       00058383                lb      t2,0(a1)
   44d04:       00730023                sb      t2,0(t1)
   44d08:       fff60613                addi    a2,a2,-1
   44d0c:       00130313                addi    t1,t1,1
   44d10:       00158593                addi    a1,a1,1
   44d14:       fe0616e3                bnez    a2,44d00 <memcpy+0x8>
   44d18:       00008067                ret

00044d1c <memset>:
   44d1c:       00050313                mv      t1,a0
   44d20:       00060a63                beqz    a2,44d34 <memset+0x18>
   44d24:       00b30023                sb      a1,0(t1)
   44d28:       fff60613                addi    a2,a2,-1
   44d2c:       00130313                addi    t1,t1,1
   44d30:       fe061ae3                bnez    a2,44d24 <memset+0x8>
   44d34:       00008067                ret

We may detect and replace with macro fusion instructions instead.

[jserv]

In lightrec, its optimizer.c would replace MIPS memset instruction sequence with host counterpart.

static const u32 memset_code[] = {
	0x10a00006,	// beqz		a1, 2f
	0x24a2ffff,	// addiu	v0,a1,-1
	0x2403ffff,	// li		v1,-1
	0xac800000,	// 1: sw	zero,0(a0)
	0x2442ffff,	// addiu	v0,v0,-1
	0x1443fffd,	// bne		v0,v1, 1b
	0x24840004,	// addiu	a0,a0,4
	0x03e00008,	// 2: jr	ra
	0x00000000,	// nop
};

static int lightrec_replace_memset(struct lightrec_state *state, struct block *block)
{
	unsigned int i;
	union code c;

	for (i = 0; i < block->nb_ops; i++) {
		c = block->opcode_list[i].c;

		if (c.opcode != memset_code[i])
			return 0;

		if (i == ARRAY_SIZE(memset_code) - 1) {
			/* success! */
			pr_debug("Block at PC 0x%x is a memset\n", block->pc);
			block_set_flags(block,
					BLOCK_IS_MEMSET | BLOCK_NEVER_COMPILE);

			/* Return non-zero to skip other optimizers. */
			return 1;
		}
	}

	return 0;
}

[jserv]

Close via #196