diff --git a/src/Makefile.am b/src/Makefile.am index f098a4d..93f8f32 100644 --- a/src/Makefile.am +++ b/src/Makefile.am @@ -72,12 +72,12 @@ pcem_SOURCES += wx-sdl2-midi.c endif if CPU_I386 -pcem_SOURCES += codegen_backend_x86.c codegen_backend_x86_ops.c codegen_backend_x86_uops.c +pcem_SOURCES += codegen_backend_x86.c codegen_backend_x86_ops.c codegen_backend_x86_ops_fpu.c codegen_backend_x86_ops_sse.c codegen_backend_x86_uops.c pcem_CFLAGS += -msse2 endif if CPU_X86_64 -pcem_SOURCES += codegen_backend_x86-64.c codegen_backend_x86-64_ops.c codegen_backend_x86-64_uops.c +pcem_SOURCES += codegen_backend_x86-64.c codegen_backend_x86-64_ops.c codegen_backend_x86-64_ops_sse.c codegen_backend_x86-64_uops.c endif if CPU_ARM diff --git a/src/codegen_backend_x86-64.c b/src/codegen_backend_x86-64.c index 01eec22..bfbf6a7 100644 --- a/src/codegen_backend_x86-64.c +++ b/src/codegen_backend_x86-64.c @@ -6,6 +6,7 @@ #include "codegen_backend.h" #include "codegen_backend_x86-64_defs.h" #include "codegen_backend_x86-64_ops.h" +#include "codegen_backend_x86-64_ops_sse.h" #include "codegen_reg.h" #include "x86.h" diff --git a/src/codegen_backend_x86-64_ops.c b/src/codegen_backend_x86-64_ops.c index 9e9928d..a4a8680 100644 --- a/src/codegen_backend_x86-64_ops.c +++ b/src/codegen_backend_x86-64_ops.c @@ -6,6 +6,7 @@ #include "codegen_backend.h" #include "codegen_backend_x86-64_defs.h" #include "codegen_backend_x86-64_ops.h" +#include "codegen_backend_x86-64_ops_helpers.h" #define RM_OP_ADD 0x00 #define RM_OP_OR 0x08 @@ -18,109 +19,6 @@ #define RM_OP_SHR 0x28 #define RM_OP_SAR 0x38 -#define JMP_LEN_BYTES 5 - -static inline void codegen_addbyte(codeblock_t *block, uint8_t val) -{ - if (block_pos >= BLOCK_MAX) - { - fatal("codegen_addbyte over! %i\n", block_pos); -// CPU_BLOCK_END(); - } - block_write_data[block_pos++] = val; -} -static inline void codegen_addbyte2(codeblock_t *block, uint8_t vala, uint8_t valb) -{ - if (block_pos > (BLOCK_MAX-2)) - { - fatal("codegen_addbyte2 over! %i\n", block_pos); - CPU_BLOCK_END(); - } - block_write_data[block_pos++] = vala; - block_write_data[block_pos++] = valb; -} -static inline void codegen_addbyte3(codeblock_t *block, uint8_t vala, uint8_t valb, uint8_t valc) -{ - if (block_pos > (BLOCK_MAX-3)) - { - fatal("codegen_addbyte3 over! %i\n", block_pos); - CPU_BLOCK_END(); - } - block_write_data[block_pos++] = vala; - block_write_data[block_pos++] = valb; - block_write_data[block_pos++] = valc; -} -static inline void codegen_addbyte4(codeblock_t *block, uint8_t vala, uint8_t valb, uint8_t valc, uint8_t vald) -{ - if (block_pos > (BLOCK_MAX-4)) - { - fatal("codegen_addbyte4 over! %i\n", block_pos); - CPU_BLOCK_END(); - } - block_write_data[block_pos++] = vala; - block_write_data[block_pos++] = valb; - block_write_data[block_pos++] = valc; - block_write_data[block_pos++] = vald; -} - -static inline void codegen_addword(codeblock_t *block, uint16_t val) -{ - if (block_pos > (BLOCK_MAX-2)) - { - fatal("codegen_addword over! %i\n", block_pos); - CPU_BLOCK_END(); - } - *(uint16_t *)&block_write_data[block_pos] = val; - block_pos += 2; -} - -static inline void codegen_addlong(codeblock_t *block, uint32_t val) -{ - if (block_pos > (BLOCK_MAX-4)) - { - fatal("codegen_addlong over! %i\n", block_pos); - CPU_BLOCK_END(); - } - *(uint32_t *)&block_write_data[block_pos] = val; - block_pos += 4; -} - -static inline void codegen_addquad(codeblock_t *block, uint64_t val) -{ - if (block_pos > (BLOCK_MAX-8)) - { - fatal("codegen_addquad over! %i\n", block_pos); - CPU_BLOCK_END(); - } - *(uint64_t *)&block_write_data[block_pos] = val; - block_pos += 8; -} - -static inline void codegen_alloc_bytes(codeblock_t *block, int size) -{ - if (block_pos > ((BLOCK_MAX - size) - JMP_LEN_BYTES)) - { - /*Current block is full. Allocate a new block*/ - struct mem_block_t *new_block = codegen_allocator_allocate(block->head_mem_block, get_block_nr(block)); - uint8_t *new_ptr = codeblock_allocator_get_ptr(new_block); - - /*Add a jump instruction to the new block*/ - codegen_addbyte(block, 0xe9); /*JMP*/ - codegen_addlong(block, (uintptr_t)new_ptr - (uintptr_t)&block_write_data[block_pos + 4]); - - /*Set write address to start of new block*/ - block_pos = 0; - block_write_data = new_ptr; - } -} - -static int is_imm8(uint32_t imm_data) -{ - if (imm_data <= 0x7f || imm_data >= 0xffffff80) - return 1; - return 0; -} - static inline void call(codeblock_t *block, uintptr_t func) { uintptr_t diff; @@ -263,17 +161,6 @@ void host_x86_ADD32_REG_REG(codeblock_t *block, int dst_reg, int src_reg) codegen_addbyte2(block, 0x01, 0xc0 | (dst_reg & 7) | ((src_reg & 7) << 3)); /*ADD dst_reg, src_reg*/ } -void host_x86_ADDPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x58, 0xc0 | src_reg | (dst_reg << 3)); /*ADDPS dst_reg, src_reg*/ -} -void host_x86_ADDSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x58, 0xc0 | src_reg | (dst_reg << 3)); -} - void host_x86_AND8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data) { if (dst_reg & 8) @@ -433,86 +320,6 @@ void host_x86_CMP32_REG_REG(codeblock_t *block, int src_reg_a, int src_reg_b) codegen_addbyte2(block, 0x39, 0xc0 | src_reg_a | (src_reg_b << 3)); /*CMP src_reg_a, src_reg_b*/ } -void host_x86_CMPPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg, int type) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x0f, 0xc2, 0xc0 | src_reg | (dst_reg << 3), type); /*CMPPS dst_reg, src_reg, type*/ -} - -void host_x86_COMISD_XREG_XREG(codeblock_t *block, int src_reg_a, int src_reg_b) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x2e, 0xc0 | src_reg_b | (src_reg_a << 3)); -} - -void host_x86_CVTDQ2PS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x5b, 0xc0 | src_reg | (dst_reg << 3)); /*CVTDQ2PS dst_reg, src_reg*/ -} -void host_x86_CVTPS2DQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x5b, 0xc0 | src_reg | (dst_reg << 3)); /*CVTPS2DQ dst_reg, src_reg*/ -} - -void host_x86_CVTSD2SI_REG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x2d, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSD2SI dst_reg, src_reg*/ -} -void host_x86_CVTSD2SI_REG64_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0xf2, 0x48, 0x0f, 0x2d); /*CVTSD2SI dst_reg, src_reg*/ - codegen_addbyte(block, 0xc0 | src_reg | (dst_reg << 3)); -} -void host_x86_CVTSD2SS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x5a, 0xc0 | src_reg | (dst_reg << 3)); -} - -void host_x86_CVTSI2SD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x2a, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSI2SD dst_reg, src_reg*/ -} -void host_x86_CVTSI2SS_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x2a, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSI2SD dst_reg, src_reg*/ -} -void host_x86_CVTSI2SD_XREG_REG64(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0xf2, 0x48, 0x0f, 0x2a); /*CVTSI2SD dst_reg, src_reg*/ - codegen_addbyte(block, 0xc0 | src_reg | (dst_reg << 3)); -} - -void host_x86_CVTSS2SD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x5a, 0xc0 | src_reg | (dst_reg << 3)); -} -void host_x86_CVTSS2SD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x5a, 0x04 | (dst_reg << 3)); /*CVTSS2SD XMMx, [base_reg + idx_reg]*/ - codegen_addbyte(block, base_reg | (idx_reg << 3)); -} - -void host_x86_DIVSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x5e, 0xc0 | src_reg | (dst_reg << 3)); -} -void host_x86_DIVSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x5e, 0xc0 | src_reg | (dst_reg << 3)); /*DIVSS dst_reg, src_reg*/ -} - void host_x86_JMP(codeblock_t *block, void *p) { jmp(block, (uintptr_t)p); @@ -655,27 +462,6 @@ void host_x86_LAHF(codeblock_t *block) codegen_addbyte(block, 0x9f); /*LAHF*/ } -void host_x86_LDMXCSR(codeblock_t *block, void *p) -{ - int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); - - if (offset >= -128 && offset < 127) - { - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x0f, 0xae, 0x50 | REG_EBP, offset); /*LDMXCSR offset[EBP]*/ - } - else if (offset < (1ull << 32)) - { - codegen_alloc_bytes(block, 7); - codegen_addbyte3(block, 0x0f, 0xae, 0x90 | REG_EBP); /*LDMXCSR offset[EBP]*/ - codegen_addlong(block, offset); - } - else - { - fatal("host_x86_LDMXCSR - out of range %p\n", p); - } -} - void host_x86_LEA_REG_IMM(codeblock_t *block, int dst_reg, int src_reg, uint32_t offset) { if (offset) @@ -709,12 +495,6 @@ void host_x86_LEA_REG_REG_SHIFT(codeblock_t *block, int dst_reg, int src_reg_a, (shift << 6) | ((src_reg_b & 7) << 3) | (src_reg_a & 7)); } -void host_x86_MAXSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x5f, 0xc0 | src_reg | (dst_reg << 3)); /*MAXSD dst_reg, src_reg*/ -} - void host_x86_MOV8_ABS_IMM(codeblock_t *block, void *p, uint32_t imm_data) { int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); @@ -1220,187 +1000,6 @@ void host_x86_MOV32_STACK_IMM(codeblock_t *block, int32_t offset, uint32_t imm_d } } -void host_x86_MOVD_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x7e, 0x04 | (src_reg << 3)); /*MOVD XMMx, [base_reg + idx_reg]*/ - codegen_addbyte(block, base_reg | (idx_reg << 3)); -} -void host_x86_MOVD_REG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x7e, 0xc0 | dst_reg | (src_reg << 3)); -} -void host_x86_MOVD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x6e, 0x04 | (dst_reg << 3)); /*MOVD XMMx, [base_reg + idx_reg]*/ - codegen_addbyte(block, base_reg | (idx_reg << 3)); -} -void host_x86_MOVD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x6e, 0xc0 | src_reg | (dst_reg << 3)); -} - -void host_x86_MOVQ_ABS_XREG(codeblock_t *block, void *p, int src_reg) -{ - int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); - - if (src_reg & 8) - fatal("host_x86_MOVQ_ABS_REG reg & 8\n"); - - if (offset >= -128 && offset < 127) - { - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x45 | (src_reg << 3)); /*MOVQ offset[EBP], src_reg*/ - codegen_addbyte(block, offset); - } - else - { - if ((uintptr_t)p >> 32) - fatal("host_x86_MOVQ_ABS_REG - out of range %p\n", p); - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x04 | (src_reg << 3)); /*MOVQ [p], src_reg*/ - codegen_addbyte(block, 0x25); - codegen_addlong(block, (uint32_t)(uintptr_t)p); - } -} -void host_x86_MOVQ_ABS_REG_REG_SHIFT_XREG(codeblock_t *block, uint32_t addr, int src_reg_a, int src_reg_b, int shift, int src_reg) -{ - if ((src_reg & 8) || (src_reg_a & 8) || (src_reg_b & 8)) - fatal("host_x86_MOVQ_ABS_REG_REG_SHIFT_REG - bad reg\n"); - - if (addr < 0x80 || addr >= 0xffffff80) - { - codegen_alloc_bytes(block, 6); - codegen_addbyte3(block, 0x66, 0x0f, 0xd6); /*MOVQ addr[src_reg_a + src_reg_b << shift], XMMx*/ - codegen_addbyte3(block, 0x44 | (src_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6), addr & 0xff); - } - else - { - codegen_alloc_bytes(block, 9); - codegen_addbyte3(block, 0x66, 0x0f, 0xd6); /*MOVQ addr[src_reg_a + src_reg_b << shift], XMMx*/ - codegen_addbyte2(block, 0x84 | (src_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6)); - codegen_addlong(block, addr); - } -} - -void host_x86_MOVQ_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x04 | (src_reg << 3)); /*MOVD XMMx, [base_reg + idx_reg]*/ - codegen_addbyte(block, base_reg | (idx_reg << 3)); -} -void host_x86_MOVQ_BASE_OFFSET_XREG(codeblock_t *block, int base_reg, int offset, int src_reg) -{ - if (offset >= -128 && offset < 127) - { - if (base_reg == REG_RSP) - { - codegen_alloc_bytes(block, 6); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x44 | (src_reg << 3)); /*MOVQ [RSP + offset], XMMx*/ - codegen_addbyte2(block, 0x24, offset); - } - else - { - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x40 | base_reg | (src_reg << 3)); /*MOVQ [base_reg + offset], XMMx*/ - codegen_addbyte(block, offset); - } - } - else - fatal("MOVQ_BASE_OFFSET_XREG - offset %i\n", offset); -} - -void host_x86_MOVQ_XREG_ABS(codeblock_t *block, int dst_reg, void *p) -{ - int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); - - if (dst_reg & 8) - fatal("host_x86_MOVQ_REG_ABS reg & 8\n"); - - if (offset >= -128 && offset < 127) - { - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x45 | (dst_reg << 3)); /*MOVQ offset[EBP], src_reg*/ - codegen_addbyte(block, offset); - } - else - { - if ((uintptr_t)p >> 32) - fatal("host_x86_MOVQ_REG_ABS - out of range %p\n", p); - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x04 | (dst_reg << 3)); /*MOVQ [p], src_reg*/ - codegen_addbyte(block, 0x25); - codegen_addlong(block, (uint32_t)(uintptr_t)p); - } -} -void host_x86_MOVQ_XREG_ABS_REG_REG_SHIFT(codeblock_t *block, int dst_reg, uint32_t addr, int src_reg_a, int src_reg_b, int shift) -{ - if ((dst_reg & 8) || (src_reg_a & 8) || (src_reg_b & 8)) - fatal("host_x86_MOVQ_REG_ABS_REG_REG_SHIFT - bad reg\n"); - - if (addr < 0x80 || addr >= 0xffffff80) - { - codegen_alloc_bytes(block, 6); - codegen_addbyte3(block, 0xf3, 0x0f, 0x7e); /*MOVQ XMMx, addr[src_reg_a + src_reg_b << shift]*/ - codegen_addbyte3(block, 0x44 | (dst_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6), addr & 0xff); - } - else - { - codegen_alloc_bytes(block, 9); - codegen_addbyte3(block, 0xf3, 0x0f, 0x7e); /*MOVQ XMMx, addr[src_reg_a + src_reg_b << shift]*/ - codegen_addbyte2(block, 0x84 | (dst_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6)); - codegen_addlong(block, addr); - } -} -void host_x86_MOVQ_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x04 | (dst_reg << 3)); /*MOVQ XMMx, [base_reg + idx_reg]*/ - codegen_addbyte(block, base_reg | (idx_reg << 3)); -} -void host_x86_MOVQ_XREG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, int offset) -{ - if (offset >= -128 && offset < 127) - { - if (base_reg == REG_ESP) - { - codegen_alloc_bytes(block, 6); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x44 | (dst_reg << 3)); /*MOVQ XMMx, [ESP + offset]*/ - codegen_addbyte2(block, 0x24, offset); - } - else - { - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x40 | base_reg | (dst_reg << 3)); /*MOVQ XMMx, [base_reg + offset]*/ - codegen_addbyte(block, offset); - } - } - else - fatal("MOVQ_REG_BASE_OFFSET - offset %i\n", offset); -} - -void host_x86_MOVQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0xc0 | src_reg | (dst_reg << 3)); /*MOVQ dst_reg, src_reg*/ -} - -void host_x86_MOVQ_REG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x48, 0x0f, 0x7e); /*MOVQ dst_reg, src_reg*/ - codegen_addbyte(block, 0xc0 | dst_reg | (src_reg << 3)); -} -void host_x86_MOVQ_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x48, 0x0f, 0x6e); /*MOVQ dst_reg, src_reg*/ - codegen_addbyte(block, 0xc0 | src_reg | (dst_reg << 3)); -} - void host_x86_MOVSX_REG_16_8(codeblock_t *block, int dst_reg, int src_reg) { codegen_alloc_bytes(block, 4); @@ -1484,28 +1083,6 @@ void host_x86_MOVZX_REG_ABS_32_8(codeblock_t *block, int dst_reg, void *p) fatal("host_x86_MOVZX_REG_ABS_32_8 - bad offset %i\n", offset); } -void host_x86_MAXPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x5f, 0xc0 | src_reg | (dst_reg << 3)); /*MAXPS dst_reg, src_reg*/ -} -void host_x86_MINPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x5d, 0xc0 | src_reg | (dst_reg << 3)); /*MINPS dst_reg, src_reg*/ -} - -void host_x86_MULPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x59, 0xc0 | src_reg | (dst_reg << 3)); /*MULPS dst_reg, src_reg*/ -} -void host_x86_MULSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x59, 0xc0 | src_reg | (dst_reg << 3)); -} - void host_x86_NOP(codeblock_t *block) { codegen_alloc_bytes(block, 1); @@ -1599,260 +1176,6 @@ void host_x86_OR32_REG_REG(codeblock_t *block, int dst_reg, int src_reg) codegen_addbyte2(block, 0x09, 0xc0 | (dst_reg & 7) | ((src_reg & 7) << 3)); /*OR dst_reg, src_reg*/ } -void host_x86_PACKSSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x63, 0xc0 | src_reg | (dst_reg << 3)); /*PACKSSWB dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ - codegen_addbyte(block, 0x88); -} -void host_x86_PACKSSDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x6b, 0xc0 | src_reg | (dst_reg << 3)); /*PACKSSDW dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ - codegen_addbyte(block, 0x88); -} -void host_x86_PACKUSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x67, 0xc0 | src_reg | (dst_reg << 3)); /*PACKUSWB dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ - codegen_addbyte(block, 0x88); -} - -void host_x86_PADDB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xfc, 0xc0 | src_reg | (dst_reg << 3)); /*PADDB dst_reg, src_reg*/ -} -void host_x86_PADDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xfd, 0xc0 | src_reg | (dst_reg << 3)); /*PADDW dst_reg, src_reg*/ -} -void host_x86_PADDD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xfe, 0xc0 | src_reg | (dst_reg << 3)); /*PADDD dst_reg, src_reg*/ -} -void host_x86_PADDSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xec, 0xc0 | src_reg | (dst_reg << 3)); /*PADDSB dst_reg, src_reg*/ -} -void host_x86_PADDSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xed, 0xc0 | src_reg | (dst_reg << 3)); /*PADDSW dst_reg, src_reg*/ -} -void host_x86_PADDUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xdc, 0xc0 | src_reg | (dst_reg << 3)); /*PADDUSB dst_reg, src_reg*/ -} -void host_x86_PADDUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xdd, 0xc0 | src_reg | (dst_reg << 3)); /*PADDUSW dst_reg, src_reg*/ -} - -void host_x86_PAND_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xdb, 0xc0 | src_reg | (dst_reg << 3)); /*PAND dst_reg, src_reg*/ -} -void host_x86_PANDN_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xdf, 0xc0 | src_reg | (dst_reg << 3)); /*PANDN dst_reg, src_reg*/ -} -void host_x86_POR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xeb, 0xc0 | src_reg | (dst_reg << 3)); /*POR dst_reg, src_reg*/ -} -void host_x86_PXOR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xef, 0xc0 | src_reg | (dst_reg << 3)); /*PXOR dst_reg, src_reg*/ -} - -void host_x86_PCMPEQB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x74, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQB dst_reg, src_reg*/ -} -void host_x86_PCMPEQW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x75, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQW dst_reg, src_reg*/ -} -void host_x86_PCMPEQD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x76, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQD dst_reg, src_reg*/ -} -void host_x86_PCMPGTB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x64, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTB dst_reg, src_reg*/ -} -void host_x86_PCMPGTW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x65, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTW dst_reg, src_reg*/ -} -void host_x86_PCMPGTD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x66, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTD dst_reg, src_reg*/ -} - -void host_x86_PMADDWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xf5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ -} -void host_x86_PMULHW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xe5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ -} -void host_x86_PMULLW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xd5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ -} - -void host_x86_PSLLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x30 | dst_reg); /*PSLLW dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSLLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x30 | dst_reg); /*PSLLD dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSLLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x30 | dst_reg); /*PSLLD dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRAW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x20 | dst_reg); /*PSRAW dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRAD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x20 | dst_reg); /*PSRAD dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRAQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x20 | dst_reg); /*PSRAD dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x10 | dst_reg); /*PSRLW dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x10 | dst_reg); /*PSRLD dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x10 | dst_reg); /*PSRLD dst_reg, imm*/ - codegen_addbyte(block, shift); -} - -void host_x86_PSUBB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xf8, 0xc0 | src_reg | (dst_reg << 3)); /*PADDB dst_reg, src_reg*/ -} -void host_x86_PSUBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xf9, 0xc0 | src_reg | (dst_reg << 3)); /*PADDW dst_reg, src_reg*/ -} -void host_x86_PSUBD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xfa, 0xc0 | src_reg | (dst_reg << 3)); /*PADDD dst_reg, src_reg*/ -} -void host_x86_PSUBSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xe8, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBSB dst_reg, src_reg*/ -} -void host_x86_PSUBSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xe9, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBSW dst_reg, src_reg*/ -} -void host_x86_PSUBUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xd8, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBUSB dst_reg, src_reg*/ -} -void host_x86_PSUBUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xd9, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBUSW dst_reg, src_reg*/ -} - -void host_x86_PUNPCKHBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x60, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLBW dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ - codegen_addbyte(block, 0xee); -} -void host_x86_PUNPCKHWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x61, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLWD dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ - codegen_addbyte(block, 0xee); -} -void host_x86_PUNPCKHDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x62, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLDQ dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ - codegen_addbyte(block, 0xee); -} -void host_x86_PUNPCKLBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x60, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLBW dst_reg, src_reg*/ -} -void host_x86_PUNPCKLWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x61, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLWD dst_reg, src_reg*/ -} -void host_x86_PUNPCKLDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x62, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLDQ dst_reg, src_reg*/ -} - void host_x86_POP(codeblock_t *block, int dst_reg) { if (dst_reg & 8) @@ -2035,17 +1358,6 @@ void host_x86_SHR32_IMM(codeblock_t *block, int dst_reg, int shift) codegen_addbyte3(block, 0xc1, 0xc0 | RM_OP_SHR | dst_reg, shift); /*SHR dst_reg, shift*/ } -void host_x86_SQRTSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x51, 0xc0 | src_reg | (dst_reg << 3)); /*SQRTSD dst_reg, src_reg*/ -} -void host_x86_SQRTSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x51, 0xc0 | src_reg | (dst_reg << 3)); /*SQRTSS dst_reg, src_reg*/ -} - void host_x86_SUB8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data) { if (dst_reg & 8) @@ -2146,17 +1458,6 @@ void host_x86_SUB32_REG_REG(codeblock_t *block, int dst_reg, int src_reg) codegen_addbyte2(block, 0x29, 0xc0 | (dst_reg & 7) | ((src_reg & 7) << 3)); /*SUB dst_reg, src_reg*/ } -void host_x86_SUBPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x5c, 0xc0 | src_reg | (dst_reg << 3)); /*SUBPS dst_reg, src_reg*/ -} -void host_x86_SUBSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x5c, 0xc0 | src_reg | (dst_reg << 3)); -} - #define MODRM_MOD_REG(rm, reg) (0xc0 | reg | (rm << 3)) void host_x86_TEST8_REG(codeblock_t *block, int src_host_reg, int dst_host_reg) @@ -2195,12 +1496,6 @@ void host_x86_TEST32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data) } } -void host_x86_UNPCKLPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x14, 0xc0 | src_reg | (dst_reg << 3)); -} - void host_x86_XOR8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data) { if (dst_reg & 8) diff --git a/src/codegen_backend_x86-64_ops.h b/src/codegen_backend_x86-64_ops.h index aa7e0c7..eaee044 100644 --- a/src/codegen_backend_x86-64_ops.h +++ b/src/codegen_backend_x86-64_ops.h @@ -7,9 +7,6 @@ void host_x86_ADD8_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_ADD16_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_ADD32_REG_REG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_ADDPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_ADDSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_AND8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data); void host_x86_AND16_REG_IMM(codeblock_t *block, int dst_reg, uint16_t imm_data); void host_x86_AND32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data); @@ -28,28 +25,6 @@ void host_x86_CMP8_REG_REG(codeblock_t *block, int src_reg_a, int src_reg_b); void host_x86_CMP16_REG_REG(codeblock_t *block, int src_reg_a, int src_reg_b); void host_x86_CMP32_REG_REG(codeblock_t *block, int src_reg_a, int src_reg_b); -#define CMPPS_EQ 0 -#define CMPPS_NLT 5 -#define CMPPS_NLE 6 -void host_x86_CMPPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg, int type); - -void host_x86_COMISD_XREG_XREG(codeblock_t *block, int src_reg_a, int src_reg_b); - -void host_x86_CVTDQ2PS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_CVTPS2DQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_CVTSD2SI_REG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_CVTSD2SI_REG64_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_CVTSD2SS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_CVTSI2SD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_CVTSI2SD_XREG_REG64(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_CVTSI2SS_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_CVTSS2SD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_CVTSS2SD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); - -void host_x86_DIVSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_DIVSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_JMP(codeblock_t *block, void *p); void host_x86_JNZ(codeblock_t *block, void *p); @@ -76,14 +51,10 @@ uint32_t *host_x86_JZ_long(codeblock_t *block); void host_x86_LAHF(codeblock_t *block); -void host_x86_LDMXCSR(codeblock_t *block, void *p); - void host_x86_LEA_REG_IMM(codeblock_t *block, int dst_reg, int src_reg, uint32_t offset); void host_x86_LEA_REG_REG(codeblock_t *block, int dst_reg, int src_reg_a, int src_reg_b); void host_x86_LEA_REG_REG_SHIFT(codeblock_t *block, int dst_reg, int src_reg_a, int src_reg_b, int shift); -void host_x86_MAXSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_MOV8_ABS_IMM(codeblock_t *block, void *p, uint32_t imm_data); void host_x86_MOV32_ABS_IMM(codeblock_t *block, void *p, uint32_t imm_data); @@ -128,24 +99,6 @@ void host_x86_MOV32_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_MOV32_STACK_IMM(codeblock_t *block, int32_t offset, uint32_t imm_data); -void host_x86_MOVD_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg); -void host_x86_MOVD_REG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_MOVD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); -void host_x86_MOVD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_MOVQ_ABS_XREG(codeblock_t *block, void *p, int src_reg); -void host_x86_MOVQ_ABS_REG_REG_SHIFT_XREG(codeblock_t *block, uint32_t addr, int src_reg_a, int src_reg_b, int shift, int src_reg); -void host_x86_MOVQ_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg); -void host_x86_MOVQ_BASE_OFFSET_XREG(codeblock_t *block, int base_reg, int offset, int src_reg); -void host_x86_MOVQ_XREG_ABS(codeblock_t *block, int dst_reg, void *p); -void host_x86_MOVQ_XREG_ABS_REG_REG_SHIFT(codeblock_t *block, int dst_reg, uint32_t addr, int src_reg_a, int src_reg_b, int shift); -void host_x86_MOVQ_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); -void host_x86_MOVQ_XREG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, int offset); - -void host_x86_MOVQ_REG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_MOVQ_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_MOVQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_MOVSX_REG_16_8(codeblock_t *block, int dst_reg, int src_reg); void host_x86_MOVSX_REG_32_8(codeblock_t *block, int dst_reg, int src_reg); void host_x86_MOVSX_REG_32_16(codeblock_t *block, int dst_reg, int src_reg); @@ -159,13 +112,6 @@ void host_x86_MOVZX_REG_32_16(codeblock_t *block, int dst_reg, int src_reg); void host_x86_MOVZX_REG_ABS_32_8(codeblock_t *block, int dst_reg, void *p); -void host_x86_MAXPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_MINPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_MULPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_MULSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_MULSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_NOP(codeblock_t *block); void host_x86_OR8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data); @@ -176,59 +122,6 @@ void host_x86_OR8_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_OR16_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_OR32_REG_REG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PACKSSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PACKSSDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PACKUSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PADDB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PAND_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PANDN_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_POR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PXOR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PCMPEQB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PCMPEQW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PCMPEQD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PCMPGTB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PCMPGTW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PCMPGTD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PMADDWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PMULHW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PMULLW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PSLLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSLLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSLLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRAW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRAD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRAQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); - -void host_x86_PSUBB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PUNPCKHBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PUNPCKHWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PUNPCKHDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PUNPCKLBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PUNPCKLWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PUNPCKLDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_POP(codeblock_t *block, int src_reg); void host_x86_PUSH(codeblock_t *block, int src_reg); @@ -259,9 +152,6 @@ void host_x86_SHR8_IMM(codeblock_t *block, int dst_reg, int shift); void host_x86_SHR16_IMM(codeblock_t *block, int dst_reg, int shift); void host_x86_SHR32_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_SQRTSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_SQRTSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_SUB8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data); void host_x86_SUB16_REG_IMM(codeblock_t *block, int dst_reg, uint16_t imm_data); void host_x86_SUB32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data); @@ -271,16 +161,11 @@ void host_x86_SUB8_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_SUB16_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_SUB32_REG_REG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_SUBPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_SUBSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_TEST8_REG(codeblock_t *block, int src_host_reg, int dst_host_reg); void host_x86_TEST16_REG(codeblock_t *block, int src_host_reg, int dst_host_reg); void host_x86_TEST32_REG(codeblock_t *block, int src_reg, int dst_reg); void host_x86_TEST32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data); -void host_x86_UNPCKLPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_XOR8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data); void host_x86_XOR16_REG_IMM(codeblock_t *block, int dst_reg, uint16_t imm_data); void host_x86_XOR32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data); diff --git a/src/codegen_backend_x86-64_ops_helpers.h b/src/codegen_backend_x86-64_ops_helpers.h new file mode 100644 index 0000000..4e63bbd --- /dev/null +++ b/src/codegen_backend_x86-64_ops_helpers.h @@ -0,0 +1,102 @@ +#define JMP_LEN_BYTES 5 + +static inline void codegen_addbyte(codeblock_t *block, uint8_t val) +{ + if (block_pos >= BLOCK_MAX) + { + fatal("codegen_addbyte over! %i\n", block_pos); +// CPU_BLOCK_END(); + } + block_write_data[block_pos++] = val; +} +static inline void codegen_addbyte2(codeblock_t *block, uint8_t vala, uint8_t valb) +{ + if (block_pos > (BLOCK_MAX-2)) + { + fatal("codegen_addbyte2 over! %i\n", block_pos); + CPU_BLOCK_END(); + } + block_write_data[block_pos++] = vala; + block_write_data[block_pos++] = valb; +} +static inline void codegen_addbyte3(codeblock_t *block, uint8_t vala, uint8_t valb, uint8_t valc) +{ + if (block_pos > (BLOCK_MAX-3)) + { + fatal("codegen_addbyte3 over! %i\n", block_pos); + CPU_BLOCK_END(); + } + block_write_data[block_pos++] = vala; + block_write_data[block_pos++] = valb; + block_write_data[block_pos++] = valc; +} +static inline void codegen_addbyte4(codeblock_t *block, uint8_t vala, uint8_t valb, uint8_t valc, uint8_t vald) +{ + if (block_pos > (BLOCK_MAX-4)) + { + fatal("codegen_addbyte4 over! %i\n", block_pos); + CPU_BLOCK_END(); + } + block_write_data[block_pos++] = vala; + block_write_data[block_pos++] = valb; + block_write_data[block_pos++] = valc; + block_write_data[block_pos++] = vald; +} + +static inline void codegen_addword(codeblock_t *block, uint16_t val) +{ + if (block_pos > (BLOCK_MAX-2)) + { + fatal("codegen_addword over! %i\n", block_pos); + CPU_BLOCK_END(); + } + *(uint16_t *)&block_write_data[block_pos] = val; + block_pos += 2; +} + +static inline void codegen_addlong(codeblock_t *block, uint32_t val) +{ + if (block_pos > (BLOCK_MAX-4)) + { + fatal("codegen_addlong over! %i\n", block_pos); + CPU_BLOCK_END(); + } + *(uint32_t *)&block_write_data[block_pos] = val; + block_pos += 4; +} + +static inline void codegen_addquad(codeblock_t *block, uint64_t val) +{ + if (block_pos > (BLOCK_MAX-8)) + { + fatal("codegen_addquad over! %i\n", block_pos); + CPU_BLOCK_END(); + } + *(uint64_t *)&block_write_data[block_pos] = val; + block_pos += 8; +} + +static inline void codegen_alloc_bytes(codeblock_t *block, int size) +{ + if (block_pos > ((BLOCK_MAX - size) - JMP_LEN_BYTES)) + { + /*Current block is full. Allocate a new block*/ + struct mem_block_t *new_block = codegen_allocator_allocate(block->head_mem_block, get_block_nr(block)); + uint8_t *new_ptr = codeblock_allocator_get_ptr(new_block); + + /*Add a jump instruction to the new block*/ + codegen_addbyte(block, 0xe9); /*JMP*/ + codegen_addlong(block, (uintptr_t)new_ptr - (uintptr_t)&block_write_data[block_pos + 4]); + + /*Set write address to start of new block*/ + block_pos = 0; + block_write_data = new_ptr; + } +} + +static inline int is_imm8(uint32_t imm_data) +{ + if (imm_data <= 0x7f || imm_data >= 0xffffff80) + return 1; + return 0; +} diff --git a/src/codegen_backend_x86-64_ops_sse.c b/src/codegen_backend_x86-64_ops_sse.c new file mode 100644 index 0000000..cdff2c6 --- /dev/null +++ b/src/codegen_backend_x86-64_ops_sse.c @@ -0,0 +1,614 @@ +#ifdef __amd64__ + +#include "ibm.h" +#include "codegen.h" +#include "codegen_allocator.h" +#include "codegen_backend.h" +#include "codegen_backend_x86-64_defs.h" +#include "codegen_backend_x86-64_ops_sse.h" +#include "codegen_backend_x86-64_ops_helpers.h" + +void host_x86_ADDPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x58, 0xc0 | src_reg | (dst_reg << 3)); /*ADDPS dst_reg, src_reg*/ +} +void host_x86_ADDSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x58, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_CMPPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg, int type) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x0f, 0xc2, 0xc0 | src_reg | (dst_reg << 3), type); /*CMPPS dst_reg, src_reg, type*/ +} + +void host_x86_COMISD_XREG_XREG(codeblock_t *block, int src_reg_a, int src_reg_b) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x2e, 0xc0 | src_reg_b | (src_reg_a << 3)); +} + +void host_x86_CVTDQ2PS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x5b, 0xc0 | src_reg | (dst_reg << 3)); /*CVTDQ2PS dst_reg, src_reg*/ +} +void host_x86_CVTPS2DQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x5b, 0xc0 | src_reg | (dst_reg << 3)); /*CVTPS2DQ dst_reg, src_reg*/ +} + +void host_x86_CVTSD2SI_REG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x2d, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSD2SI dst_reg, src_reg*/ +} +void host_x86_CVTSD2SI_REG64_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0xf2, 0x48, 0x0f, 0x2d); /*CVTSD2SI dst_reg, src_reg*/ + codegen_addbyte(block, 0xc0 | src_reg | (dst_reg << 3)); +} +void host_x86_CVTSD2SS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x5a, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_CVTSI2SD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x2a, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSI2SD dst_reg, src_reg*/ +} +void host_x86_CVTSI2SS_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x2a, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSI2SD dst_reg, src_reg*/ +} +void host_x86_CVTSI2SD_XREG_REG64(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0xf2, 0x48, 0x0f, 0x2a); /*CVTSI2SD dst_reg, src_reg*/ + codegen_addbyte(block, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_CVTSS2SD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x5a, 0xc0 | src_reg | (dst_reg << 3)); +} +void host_x86_CVTSS2SD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x5a, 0x04 | (dst_reg << 3)); /*CVTSS2SD XMMx, [base_reg + idx_reg]*/ + codegen_addbyte(block, base_reg | (idx_reg << 3)); +} + +void host_x86_DIVSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x5e, 0xc0 | src_reg | (dst_reg << 3)); +} +void host_x86_DIVSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x5e, 0xc0 | src_reg | (dst_reg << 3)); /*DIVSS dst_reg, src_reg*/ +} + +void host_x86_LDMXCSR(codeblock_t *block, void *p) +{ + int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); + + if (offset >= -128 && offset < 127) + { + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x0f, 0xae, 0x50 | REG_EBP, offset); /*LDMXCSR offset[EBP]*/ + } + else if (offset < (1ull << 32)) + { + codegen_alloc_bytes(block, 7); + codegen_addbyte3(block, 0x0f, 0xae, 0x90 | REG_EBP); /*LDMXCSR offset[EBP]*/ + codegen_addlong(block, offset); + } + else + { + fatal("host_x86_LDMXCSR - out of range %p\n", p); + } +} + +void host_x86_MAXSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x5f, 0xc0 | src_reg | (dst_reg << 3)); /*MAXSD dst_reg, src_reg*/ +} + +void host_x86_MOVD_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x7e, 0x04 | (src_reg << 3)); /*MOVD XMMx, [base_reg + idx_reg]*/ + codegen_addbyte(block, base_reg | (idx_reg << 3)); +} +void host_x86_MOVD_REG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x7e, 0xc0 | dst_reg | (src_reg << 3)); +} +void host_x86_MOVD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x6e, 0x04 | (dst_reg << 3)); /*MOVD XMMx, [base_reg + idx_reg]*/ + codegen_addbyte(block, base_reg | (idx_reg << 3)); +} +void host_x86_MOVD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x6e, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_MOVQ_ABS_XREG(codeblock_t *block, void *p, int src_reg) +{ + int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); + + if (src_reg & 8) + fatal("host_x86_MOVQ_ABS_REG reg & 8\n"); + + if (offset >= -128 && offset < 127) + { + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x45 | (src_reg << 3)); /*MOVQ offset[EBP], src_reg*/ + codegen_addbyte(block, offset); + } + else + { + if ((uintptr_t)p >> 32) + fatal("host_x86_MOVQ_ABS_REG - out of range %p\n", p); + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x04 | (src_reg << 3)); /*MOVQ [p], src_reg*/ + codegen_addbyte(block, 0x25); + codegen_addlong(block, (uint32_t)(uintptr_t)p); + } +} +void host_x86_MOVQ_ABS_REG_REG_SHIFT_XREG(codeblock_t *block, uint32_t addr, int src_reg_a, int src_reg_b, int shift, int src_reg) +{ + if ((src_reg & 8) || (src_reg_a & 8) || (src_reg_b & 8)) + fatal("host_x86_MOVQ_ABS_REG_REG_SHIFT_REG - bad reg\n"); + + if (addr < 0x80 || addr >= 0xffffff80) + { + codegen_alloc_bytes(block, 6); + codegen_addbyte3(block, 0x66, 0x0f, 0xd6); /*MOVQ addr[src_reg_a + src_reg_b << shift], XMMx*/ + codegen_addbyte3(block, 0x44 | (src_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6), addr & 0xff); + } + else + { + codegen_alloc_bytes(block, 9); + codegen_addbyte3(block, 0x66, 0x0f, 0xd6); /*MOVQ addr[src_reg_a + src_reg_b << shift], XMMx*/ + codegen_addbyte2(block, 0x84 | (src_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6)); + codegen_addlong(block, addr); + } +} + +void host_x86_MOVQ_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x04 | (src_reg << 3)); /*MOVD XMMx, [base_reg + idx_reg]*/ + codegen_addbyte(block, base_reg | (idx_reg << 3)); +} +void host_x86_MOVQ_BASE_OFFSET_XREG(codeblock_t *block, int base_reg, int offset, int src_reg) +{ + if (offset >= -128 && offset < 127) + { + if (base_reg == REG_RSP) + { + codegen_alloc_bytes(block, 6); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x44 | (src_reg << 3)); /*MOVQ [RSP + offset], XMMx*/ + codegen_addbyte2(block, 0x24, offset); + } + else + { + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x40 | base_reg | (src_reg << 3)); /*MOVQ [base_reg + offset], XMMx*/ + codegen_addbyte(block, offset); + } + } + else + fatal("MOVQ_BASE_OFFSET_XREG - offset %i\n", offset); +} + +void host_x86_MOVQ_XREG_ABS(codeblock_t *block, int dst_reg, void *p) +{ + int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); + + if (dst_reg & 8) + fatal("host_x86_MOVQ_REG_ABS reg & 8\n"); + + if (offset >= -128 && offset < 127) + { + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x45 | (dst_reg << 3)); /*MOVQ offset[EBP], src_reg*/ + codegen_addbyte(block, offset); + } + else + { + if ((uintptr_t)p >> 32) + fatal("host_x86_MOVQ_REG_ABS - out of range %p\n", p); + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x04 | (dst_reg << 3)); /*MOVQ [p], src_reg*/ + codegen_addbyte(block, 0x25); + codegen_addlong(block, (uint32_t)(uintptr_t)p); + } +} +void host_x86_MOVQ_XREG_ABS_REG_REG_SHIFT(codeblock_t *block, int dst_reg, uint32_t addr, int src_reg_a, int src_reg_b, int shift) +{ + if ((dst_reg & 8) || (src_reg_a & 8) || (src_reg_b & 8)) + fatal("host_x86_MOVQ_REG_ABS_REG_REG_SHIFT - bad reg\n"); + + if (addr < 0x80 || addr >= 0xffffff80) + { + codegen_alloc_bytes(block, 6); + codegen_addbyte3(block, 0xf3, 0x0f, 0x7e); /*MOVQ XMMx, addr[src_reg_a + src_reg_b << shift]*/ + codegen_addbyte3(block, 0x44 | (dst_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6), addr & 0xff); + } + else + { + codegen_alloc_bytes(block, 9); + codegen_addbyte3(block, 0xf3, 0x0f, 0x7e); /*MOVQ XMMx, addr[src_reg_a + src_reg_b << shift]*/ + codegen_addbyte2(block, 0x84 | (dst_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6)); + codegen_addlong(block, addr); + } +} +void host_x86_MOVQ_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x04 | (dst_reg << 3)); /*MOVQ XMMx, [base_reg + idx_reg]*/ + codegen_addbyte(block, base_reg | (idx_reg << 3)); +} +void host_x86_MOVQ_XREG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, int offset) +{ + if (offset >= -128 && offset < 127) + { + if (base_reg == REG_ESP) + { + codegen_alloc_bytes(block, 6); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x44 | (dst_reg << 3)); /*MOVQ XMMx, [ESP + offset]*/ + codegen_addbyte2(block, 0x24, offset); + } + else + { + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x40 | base_reg | (dst_reg << 3)); /*MOVQ XMMx, [base_reg + offset]*/ + codegen_addbyte(block, offset); + } + } + else + fatal("MOVQ_REG_BASE_OFFSET - offset %i\n", offset); +} + +void host_x86_MOVQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0xc0 | src_reg | (dst_reg << 3)); /*MOVQ dst_reg, src_reg*/ +} + +void host_x86_MOVQ_REG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x48, 0x0f, 0x7e); /*MOVQ dst_reg, src_reg*/ + codegen_addbyte(block, 0xc0 | dst_reg | (src_reg << 3)); +} +void host_x86_MOVQ_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x48, 0x0f, 0x6e); /*MOVQ dst_reg, src_reg*/ + codegen_addbyte(block, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_MAXPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x5f, 0xc0 | src_reg | (dst_reg << 3)); /*MAXPS dst_reg, src_reg*/ +} +void host_x86_MINPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x5d, 0xc0 | src_reg | (dst_reg << 3)); /*MINPS dst_reg, src_reg*/ +} + +void host_x86_MULPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x59, 0xc0 | src_reg | (dst_reg << 3)); /*MULPS dst_reg, src_reg*/ +} +void host_x86_MULSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x59, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_PACKSSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x63, 0xc0 | src_reg | (dst_reg << 3)); /*PACKSSWB dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ + codegen_addbyte(block, 0x88); +} +void host_x86_PACKSSDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x6b, 0xc0 | src_reg | (dst_reg << 3)); /*PACKSSDW dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ + codegen_addbyte(block, 0x88); +} +void host_x86_PACKUSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x67, 0xc0 | src_reg | (dst_reg << 3)); /*PACKUSWB dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ + codegen_addbyte(block, 0x88); +} + +void host_x86_PADDB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xfc, 0xc0 | src_reg | (dst_reg << 3)); /*PADDB dst_reg, src_reg*/ +} +void host_x86_PADDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xfd, 0xc0 | src_reg | (dst_reg << 3)); /*PADDW dst_reg, src_reg*/ +} +void host_x86_PADDD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xfe, 0xc0 | src_reg | (dst_reg << 3)); /*PADDD dst_reg, src_reg*/ +} +void host_x86_PADDSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xec, 0xc0 | src_reg | (dst_reg << 3)); /*PADDSB dst_reg, src_reg*/ +} +void host_x86_PADDSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xed, 0xc0 | src_reg | (dst_reg << 3)); /*PADDSW dst_reg, src_reg*/ +} +void host_x86_PADDUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xdc, 0xc0 | src_reg | (dst_reg << 3)); /*PADDUSB dst_reg, src_reg*/ +} +void host_x86_PADDUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xdd, 0xc0 | src_reg | (dst_reg << 3)); /*PADDUSW dst_reg, src_reg*/ +} + +void host_x86_PAND_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xdb, 0xc0 | src_reg | (dst_reg << 3)); /*PAND dst_reg, src_reg*/ +} +void host_x86_PANDN_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xdf, 0xc0 | src_reg | (dst_reg << 3)); /*PANDN dst_reg, src_reg*/ +} +void host_x86_POR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xeb, 0xc0 | src_reg | (dst_reg << 3)); /*POR dst_reg, src_reg*/ +} +void host_x86_PXOR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xef, 0xc0 | src_reg | (dst_reg << 3)); /*PXOR dst_reg, src_reg*/ +} + +void host_x86_PCMPEQB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x74, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQB dst_reg, src_reg*/ +} +void host_x86_PCMPEQW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x75, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQW dst_reg, src_reg*/ +} +void host_x86_PCMPEQD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x76, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQD dst_reg, src_reg*/ +} +void host_x86_PCMPGTB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x64, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTB dst_reg, src_reg*/ +} +void host_x86_PCMPGTW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x65, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTW dst_reg, src_reg*/ +} +void host_x86_PCMPGTD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x66, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTD dst_reg, src_reg*/ +} + +void host_x86_PMADDWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xf5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ +} +void host_x86_PMULHW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xe5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ +} +void host_x86_PMULLW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xd5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ +} + +void host_x86_PSLLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x30 | dst_reg); /*PSLLW dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSLLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x30 | dst_reg); /*PSLLD dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSLLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x30 | dst_reg); /*PSLLD dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRAW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x20 | dst_reg); /*PSRAW dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRAD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x20 | dst_reg); /*PSRAD dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRAQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x20 | dst_reg); /*PSRAD dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x10 | dst_reg); /*PSRLW dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x10 | dst_reg); /*PSRLD dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x10 | dst_reg); /*PSRLD dst_reg, imm*/ + codegen_addbyte(block, shift); +} + +void host_x86_PSUBB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xf8, 0xc0 | src_reg | (dst_reg << 3)); /*PADDB dst_reg, src_reg*/ +} +void host_x86_PSUBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xf9, 0xc0 | src_reg | (dst_reg << 3)); /*PADDW dst_reg, src_reg*/ +} +void host_x86_PSUBD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xfa, 0xc0 | src_reg | (dst_reg << 3)); /*PADDD dst_reg, src_reg*/ +} +void host_x86_PSUBSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xe8, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBSB dst_reg, src_reg*/ +} +void host_x86_PSUBSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xe9, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBSW dst_reg, src_reg*/ +} +void host_x86_PSUBUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xd8, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBUSB dst_reg, src_reg*/ +} +void host_x86_PSUBUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xd9, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBUSW dst_reg, src_reg*/ +} + +void host_x86_PUNPCKHBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x60, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLBW dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ + codegen_addbyte(block, 0xee); +} +void host_x86_PUNPCKHWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x61, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLWD dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ + codegen_addbyte(block, 0xee); +} +void host_x86_PUNPCKHDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x62, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLDQ dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ + codegen_addbyte(block, 0xee); +} +void host_x86_PUNPCKLBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x60, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLBW dst_reg, src_reg*/ +} +void host_x86_PUNPCKLWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x61, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLWD dst_reg, src_reg*/ +} +void host_x86_PUNPCKLDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x62, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLDQ dst_reg, src_reg*/ +} + +void host_x86_SQRTSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x51, 0xc0 | src_reg | (dst_reg << 3)); /*SQRTSD dst_reg, src_reg*/ +} +void host_x86_SQRTSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x51, 0xc0 | src_reg | (dst_reg << 3)); /*SQRTSS dst_reg, src_reg*/ +} + +void host_x86_SUBPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x5c, 0xc0 | src_reg | (dst_reg << 3)); /*SUBPS dst_reg, src_reg*/ +} +void host_x86_SUBSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x5c, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_UNPCKLPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x14, 0xc0 | src_reg | (dst_reg << 3)); +} + +#endif diff --git a/src/codegen_backend_x86-64_ops_sse.h b/src/codegen_backend_x86-64_ops_sse.h new file mode 100644 index 0000000..39e39f4 --- /dev/null +++ b/src/codegen_backend_x86-64_ops_sse.h @@ -0,0 +1,114 @@ +void host_x86_ADDPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_ADDSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +#define CMPPS_EQ 0 +#define CMPPS_NLT 5 +#define CMPPS_NLE 6 +void host_x86_CMPPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg, int type); + +void host_x86_COMISD_XREG_XREG(codeblock_t *block, int src_reg_a, int src_reg_b); + +void host_x86_CVTDQ2PS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_CVTPS2DQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_CVTSD2SI_REG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_CVTSD2SI_REG64_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_CVTSD2SS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_CVTSI2SD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_CVTSI2SD_XREG_REG64(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_CVTSI2SS_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_CVTSS2SD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_CVTSS2SD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); + +void host_x86_DIVSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_DIVSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_LDMXCSR(codeblock_t *block, void *p); + +void host_x86_MAXSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_MOVD_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg); +void host_x86_MOVD_REG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_MOVD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); +void host_x86_MOVD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_MOVQ_ABS_XREG(codeblock_t *block, void *p, int src_reg); +void host_x86_MOVQ_ABS_REG_REG_SHIFT_XREG(codeblock_t *block, uint32_t addr, int src_reg_a, int src_reg_b, int shift, int src_reg); +void host_x86_MOVQ_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg); +void host_x86_MOVQ_BASE_OFFSET_XREG(codeblock_t *block, int base_reg, int offset, int src_reg); +void host_x86_MOVQ_XREG_ABS(codeblock_t *block, int dst_reg, void *p); +void host_x86_MOVQ_XREG_ABS_REG_REG_SHIFT(codeblock_t *block, int dst_reg, uint32_t addr, int src_reg_a, int src_reg_b, int shift); +void host_x86_MOVQ_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); +void host_x86_MOVQ_XREG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, int offset); + +void host_x86_MOVQ_REG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_MOVQ_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_MOVQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_MAXPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_MINPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_MULPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_MULSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_MULSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PACKSSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PACKSSDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PACKUSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PADDB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PAND_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PANDN_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_POR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PXOR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PCMPEQB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PCMPEQW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PCMPEQD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PCMPGTB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PCMPGTW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PCMPGTD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PMADDWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PMULHW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PMULLW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PSLLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSLLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSLLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRAW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRAD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRAQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); + +void host_x86_PSUBB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PUNPCKHBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PUNPCKHWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PUNPCKHDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PUNPCKLBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PUNPCKLWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PUNPCKLDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_SQRTSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_SQRTSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_SUBPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_SUBSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_UNPCKLPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); diff --git a/src/codegen_backend_x86-64_uops.c b/src/codegen_backend_x86-64_uops.c index e17ad0e..aaabbbc 100644 --- a/src/codegen_backend_x86-64_uops.c +++ b/src/codegen_backend_x86-64_uops.c @@ -8,6 +8,7 @@ #include "codegen_backend.h" #include "codegen_backend_x86-64_defs.h" #include "codegen_backend_x86-64_ops.h" +#include "codegen_backend_x86-64_ops_sse.h" #include "codegen_ir_defs.h" #define STACK_ARG0 (0) diff --git a/src/codegen_backend_x86.c b/src/codegen_backend_x86.c index 7beffc5..7aef0e3 100644 --- a/src/codegen_backend_x86.c +++ b/src/codegen_backend_x86.c @@ -7,6 +7,7 @@ #include "codegen_backend.h" #include "codegen_backend_x86_defs.h" #include "codegen_backend_x86_ops.h" +#include "codegen_backend_x86_ops_sse.h" #include "codegen_reg.h" #include "x86.h" diff --git a/src/codegen_backend_x86_ops.c b/src/codegen_backend_x86_ops.c index 47f7493..ce64b33 100644 --- a/src/codegen_backend_x86_ops.c +++ b/src/codegen_backend_x86_ops.c @@ -6,6 +6,7 @@ #include "codegen_backend.h" #include "codegen_backend_x86_defs.h" #include "codegen_backend_x86_ops.h" +#include "codegen_backend_x86_ops_helpers.h" #define RM_OP_ADD 0x00 #define RM_OP_OR 0x08 @@ -18,91 +19,6 @@ #define RM_OP_SHR 0x28 #define RM_OP_SAR 0x38 -#define JMP_LEN_BYTES 5 - -static inline void codegen_addbyte(codeblock_t *block, uint8_t val) -{ - if (block_pos >= BLOCK_MAX) - fatal("codegen_addbyte over! %i\n", block_pos); - block_write_data[block_pos++] = val; -} -static inline void codegen_addbyte2(codeblock_t *block, uint8_t vala, uint8_t valb) -{ - if (block_pos > (BLOCK_MAX-2)) - fatal("codegen_addbyte2 over! %i\n", block_pos); - block_write_data[block_pos++] = vala; - block_write_data[block_pos++] = valb; -} -static inline void codegen_addbyte3(codeblock_t *block, uint8_t vala, uint8_t valb, uint8_t valc) -{ - if (block_pos > (BLOCK_MAX-3)) - fatal("codegen_addbyte3 over! %i\n", block_pos); - block_write_data[block_pos++] = vala; - block_write_data[block_pos++] = valb; - block_write_data[block_pos++] = valc; -} -static inline void codegen_addbyte4(codeblock_t *block, uint8_t vala, uint8_t valb, uint8_t valc, uint8_t vald) -{ - if (block_pos > (BLOCK_MAX-4)) - fatal("codegen_addbyte4 over! %i\n", block_pos); - block_write_data[block_pos++] = vala; - block_write_data[block_pos++] = valb; - block_write_data[block_pos++] = valc; - block_write_data[block_pos++] = vald; -} - -static inline void codegen_addword(codeblock_t *block, uint16_t val) -{ - if (block_pos > (BLOCK_MAX-2)) - fatal("codegen_addword over! %i\n", block_pos); - *(uint16_t *)&block_write_data[block_pos] = val; - block_pos += 2; -} - -static inline void codegen_addlong(codeblock_t *block, uint32_t val) -{ - if (block_pos > (BLOCK_MAX-4)) - fatal("codegen_addlong over! %i\n", block_pos); - *(uint32_t *)&block_write_data[block_pos] = val; - block_pos += 4; -} - -static inline void codegen_addquad(codeblock_t *block, uint64_t val) -{ - if (block_pos > (BLOCK_MAX-8)) - fatal("codegen_addquad over! %i\n", block_pos); - *(uint64_t *)&block_write_data[block_pos] = val; - block_pos += 8; -} - -static void codegen_allocate_new_block(codeblock_t *block) -{ - /*Current block is full. Allocate a new block*/ - struct mem_block_t *new_block = codegen_allocator_allocate(block->head_mem_block, get_block_nr(block)); - uint8_t *new_ptr = codeblock_allocator_get_ptr(new_block); - - /*Add a jump instruction to the new block*/ - codegen_addbyte(block, 0xe9); /*JMP*/ - codegen_addlong(block, (uintptr_t)new_ptr - (uintptr_t)&block_write_data[block_pos + 4]); - - /*Set write address to start of new block*/ - block_pos = 0; - block_write_data = new_ptr; -} - -static inline void codegen_alloc_bytes(codeblock_t *block, int size) -{ - if (block_pos > ((BLOCK_MAX - size) - JMP_LEN_BYTES)) - codegen_allocate_new_block(block); -} - -static int is_imm8(uint32_t imm_data) -{ - if (imm_data <= 0x7f || imm_data >= 0xffffff80) - return 1; - return 0; -} - void host_x86_ADD32_REG_ABS(codeblock_t *block, int dst_reg, void *p) { int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); @@ -191,17 +107,6 @@ void host_x86_ADD32_REG_REG(codeblock_t *block, int dst_reg, int src_reg) codegen_addbyte2(block, 0x01, 0xc0 | dst_reg | (src_reg << 3)); /*ADD dst_reg, src_reg*/ } -void host_x86_ADDPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x58, 0xc0 | src_reg | (dst_reg << 3)); /*ADDPS dst_reg, src_reg*/ -} -void host_x86_ADDSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x58, 0xc0 | src_reg | (dst_reg << 3)); -} - void host_x86_AND8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data) { if (dst_reg == REG_EAX) @@ -336,74 +241,6 @@ void host_x86_CMP32_REG_REG(codeblock_t *block, int src_reg_a, int src_reg_b) codegen_addbyte2(block, 0x39, 0xc0 | src_reg_a | (src_reg_b << 3)); /*CMP src_reg_a, src_reg_b*/ } -void host_x86_CMPPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg, int type) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x0f, 0xc2, 0xc0 | src_reg | (dst_reg << 3), type); /*CMPPS dst_reg, src_reg, type*/ -} - -void host_x86_COMISD_XREG_XREG(codeblock_t *block, int src_reg_a, int src_reg_b) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x2e, 0xc0 | src_reg_b | (src_reg_a << 3)); -} - -void host_x86_CVTDQ2PS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x5b, 0xc0 | src_reg | (dst_reg << 3)); /*CVTDQ2PS dst_reg, src_reg*/ -} -void host_x86_CVTPS2DQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x5b, 0xc0 | src_reg | (dst_reg << 3)); /*CVTPS2DQ dst_reg, src_reg*/ -} - -void host_x86_CVTSD2SI_REG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x2d, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSD2SI dst_reg, src_reg*/ -} -void host_x86_CVTSD2SS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x5a, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSD2SS dst_reg, src_reg*/ -} - -void host_x86_CVTSI2SD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x2a, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSI2SD dst_reg, src_reg*/ -} -void host_x86_CVTSI2SS_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x2a, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSI2SD dst_reg, src_reg*/ -} - -void host_x86_CVTSS2SD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x5a, 0xc0 | src_reg | (dst_reg << 3)); -} -void host_x86_CVTSS2SD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0xf3, 0x0f, 0x5a, 0x04 | (dst_reg << 3)); /*CVTSS2SD XMMx, [base_reg + idx_reg]*/ - codegen_addbyte(block, base_reg | (idx_reg << 3)); -} - -void host_x86_DIVSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x5e, 0xc0 | src_reg | (dst_reg << 3)); /*DIVSD dst_reg, src_reg*/ -} -void host_x86_DIVSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x5e, 0xc0 | src_reg | (dst_reg << 3)); /*DIVSS dst_reg, src_reg*/ -} - void host_x86_INC32_ABS(codeblock_t *block, void *p) { codegen_alloc_bytes(block, 6); @@ -562,23 +399,6 @@ void host_x86_LAHF(codeblock_t *block) codegen_addbyte(block, 0x9f); /*LAHF*/ } -void host_x86_LDMXCSR(codeblock_t *block, void *p) -{ - int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); - - if (offset >= -128 && offset < 127) - { - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x0f, 0xae, 0x50 | REG_EBP, offset); /*LDMXCSR offset[EBP]*/ - } - else - { - codegen_alloc_bytes(block, 7); - codegen_addbyte3(block, 0x0f, 0xae, 0x15); /*LDMXCSR [p]*/ - codegen_addlong(block, (uint32_t)p); - } -} - void host_x86_LEA_REG_IMM(codeblock_t *block, int dst_reg, int src_reg, uint32_t offset) { if (offset) @@ -605,12 +425,6 @@ void host_x86_LEA_REG_REG_SHIFT(codeblock_t *block, int dst_reg, int src_reg_a, codegen_addbyte3(block, 0x8d, 0x04 | (dst_reg << 3), (shift << 6) | (src_reg_b << 3) | src_reg_a); /*LEA dst_reg, [src_reg_a + src_reg_b * (1 << shift)]*/ } -void host_x86_MAXSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x5f, 0xc0 | src_reg | (dst_reg << 3)); /*MAXSD dst_reg, src_reg*/ -} - void host_x86_MOV8_ABS_IMM(codeblock_t *block, void *p, uint32_t imm_data) { int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); @@ -958,177 +772,6 @@ void host_x86_MOV32_STACK_IMM(codeblock_t *block, int32_t offset, uint32_t imm_d } } -void host_x86_MOVD_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x7e, 0x04 | (src_reg << 3)); /*MOVD XMMx, [base_reg + idx_reg]*/ - codegen_addbyte(block, base_reg | (idx_reg << 3)); -} -void host_x86_MOVD_REG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x7e, 0xc0 | dst_reg | (src_reg << 3)); -} -void host_x86_MOVD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x6e, 0x04 | (dst_reg << 3)); /*MOVD XMMx, [base_reg + idx_reg]*/ - codegen_addbyte(block, base_reg | (idx_reg << 3)); -} -void host_x86_MOVD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x6e, 0xc0 | src_reg | (dst_reg << 3)); -} - -void host_x86_MOVQ_ABS_XREG(codeblock_t *block, void *p, int src_reg) -{ - int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); - - if (offset >= -128 && offset < 127) - { - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x45 | (src_reg << 3)); /*MOVQ offset[EBP], src_reg*/ - codegen_addbyte(block, offset); - } - else - { - codegen_alloc_bytes(block, 8); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x05 | (src_reg << 3)); /*MOVQ [p], src_reg*/ - codegen_addlong(block, (uint32_t)p); - } -} -void host_x86_MOVQ_ABS_REG_REG_SHIFT_XREG(codeblock_t *block, uint32_t addr, int src_reg_a, int src_reg_b, int shift, int src_reg) -{ - if (addr < 0x80 || addr >= 0xffffff80) - { - codegen_alloc_bytes(block, 6); - codegen_addbyte3(block, 0x66, 0x0f, 0xd6); /*MOVQ addr[src_reg_a + src_reg_b << shift], XMMx*/ - codegen_addbyte3(block, 0x44 | (src_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6), addr & 0xff); - } - else - { - codegen_alloc_bytes(block, 9); - codegen_addbyte3(block, 0x66, 0x0f, 0xd6); /*MOVQ addr[src_reg_a + src_reg_b << shift], XMMx*/ - codegen_addbyte2(block, 0x84 | (src_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6)); - codegen_addlong(block, addr); - } -} -void host_x86_MOVQ_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x04 | (src_reg << 3)); /*MOVQ XMMx, [base_reg + idx_reg]*/ - codegen_addbyte(block, base_reg | (idx_reg << 3)); -} -void host_x86_MOVQ_BASE_OFFSET_XREG(codeblock_t *block, int base_reg, int offset, int src_reg) -{ - if (offset >= -128 && offset < 127) - { - if (base_reg == REG_ESP) - { - codegen_alloc_bytes(block, 6); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x44 | (src_reg << 3)); /*MOVQ [ESP + offset], XMMx*/ - codegen_addbyte2(block, 0x24, offset); - } - else - { - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x40 | base_reg | (src_reg << 3)); /*MOVQ [base_reg + offset], XMMx*/ - codegen_addbyte(block, offset); - } - } - else - fatal("MOVQ_BASE_OFFSET_XREG - offset %i\n", offset); -} -void host_x86_MOVQ_STACK_OFFSET_XREG(codeblock_t *block, int offset, int src_reg) -{ - if (!offset) - { - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x04 | (src_reg << 3)); /*MOVQ [ESP], src_reg*/ - codegen_addbyte(block, 0x24); - } - else if (offset >= -80 || offset < 0x80) - { - codegen_alloc_bytes(block, 6); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x44 | (src_reg << 3)); /*MOVQ offset[ESP], src_reg*/ - codegen_addbyte2(block, 0x24, offset & 0xff); - } - else - { - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x84 | (src_reg << 3)); /*MOVQ offset[ESP], src_reg*/ - codegen_addbyte(block, 0x24); - codegen_addlong(block, offset); - } - -} - -void host_x86_MOVQ_XREG_ABS(codeblock_t *block, int dst_reg, void *p) -{ - int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); - - if (offset >= -128 && offset < 127) - { - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x45 | (dst_reg << 3)); /*MOVQ offset[EBP], src_reg*/ - codegen_addbyte(block, offset); - } - else - { - codegen_alloc_bytes(block, 8); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x05 | (dst_reg << 3)); /*MOVQ [p], src_reg*/ - codegen_addlong(block, (uint32_t)p); - } -} -void host_x86_MOVQ_XREG_ABS_REG_REG_SHIFT(codeblock_t *block, int dst_reg, uint32_t addr, int src_reg_a, int src_reg_b, int shift) -{ - if (addr < 0x80 || addr >= 0xffffff80) - { - codegen_alloc_bytes(block, 6); - codegen_addbyte3(block, 0xf3, 0x0f, 0x7e); /*MOVQ XMMx, addr[src_reg_a + src_reg_b << shift]*/ - codegen_addbyte3(block, 0x44 | (dst_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6), addr & 0xff); - } - else - { - codegen_alloc_bytes(block, 9); - codegen_addbyte3(block, 0xf3, 0x0f, 0x7e); /*MOVQ XMMx, addr[src_reg_a + src_reg_b << shift]*/ - codegen_addbyte2(block, 0x84 | (dst_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6)); - codegen_addlong(block, addr); - } -} -void host_x86_MOVQ_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x04 | (dst_reg << 3)); /*MOVQ XMMx, [base_reg + idx_reg]*/ - codegen_addbyte(block, base_reg | (idx_reg << 3)); -} -void host_x86_MOVQ_XREG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, int offset) -{ - if (offset >= -128 && offset < 127) - { - if (base_reg == REG_ESP) - { - codegen_alloc_bytes(block, 6); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x44 | (dst_reg << 3)); /*MOVQ XMMx, [ESP + offset]*/ - codegen_addbyte2(block, 0x24, offset); - } - else - { - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x40 | base_reg | (dst_reg << 3)); /*MOVQ XMMx, [base_reg + offset]*/ - codegen_addbyte(block, offset); - } - } - else - fatal("MOVQ_REG_BASE_OFFSET - offset %i\n", offset); -} -void host_x86_MOVQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0xc0 | src_reg | (dst_reg << 3)); /*MOVQ dst_reg, src_reg*/ -} - void host_x86_MOVSX_REG_16_8(codeblock_t *block, int dst_reg, int src_reg) { codegen_alloc_bytes(block, 4); @@ -1205,28 +848,6 @@ void host_x86_MOVZX_BASE_INDEX_32_16(codeblock_t *block, int dst_reg, int base_r codegen_addbyte4(block, 0x0f, 0xb7, 0x04 | (dst_reg << 3), base_reg | (idx_reg << 3)); /*MOVZX dst_reg, W[base_reg + idx_reg]*/ } -void host_x86_MAXPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x5f, 0xc0 | src_reg | (dst_reg << 3)); /*MAXPS dst_reg, src_reg*/ -} -void host_x86_MINPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x5d, 0xc0 | src_reg | (dst_reg << 3)); /*MINPS dst_reg, src_reg*/ -} - -void host_x86_MULPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x59, 0xc0 | src_reg | (dst_reg << 3)); /*MULPS dst_reg, src_reg*/ -} -void host_x86_MULSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x59, 0xc0 | src_reg | (dst_reg << 3)); -} - void host_x86_OR8_REG_REG(codeblock_t *block, int dst_reg, int src_reg) { codegen_alloc_bytes(block, 2); @@ -1297,260 +918,6 @@ void host_x86_OR32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data) } } -void host_x86_PACKSSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x63, 0xc0 | src_reg | (dst_reg << 3)); /*PACKSSWB dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ - codegen_addbyte(block, 0x88); -} -void host_x86_PACKSSDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x6b, 0xc0 | src_reg | (dst_reg << 3)); /*PACKSSDW dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ - codegen_addbyte(block, 0x88); -} -void host_x86_PACKUSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x67, 0xc0 | src_reg | (dst_reg << 3)); /*PACKUSWB dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ - codegen_addbyte(block, 0x88); -} - -void host_x86_PADDB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xfc, 0xc0 | src_reg | (dst_reg << 3)); /*PADDB dst_reg, src_reg*/ -} -void host_x86_PADDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xfd, 0xc0 | src_reg | (dst_reg << 3)); /*PADDW dst_reg, src_reg*/ -} -void host_x86_PADDD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xfe, 0xc0 | src_reg | (dst_reg << 3)); /*PADDD dst_reg, src_reg*/ -} -void host_x86_PADDSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xec, 0xc0 | src_reg | (dst_reg << 3)); /*PADDSB dst_reg, src_reg*/ -} -void host_x86_PADDSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xed, 0xc0 | src_reg | (dst_reg << 3)); /*PADDSW dst_reg, src_reg*/ -} -void host_x86_PADDUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xdc, 0xc0 | src_reg | (dst_reg << 3)); /*PADDUSB dst_reg, src_reg*/ -} -void host_x86_PADDUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xdd, 0xc0 | src_reg | (dst_reg << 3)); /*PADDUSW dst_reg, src_reg*/ -} - -void host_x86_PAND_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xdb, 0xc0 | src_reg | (dst_reg << 3)); /*PAND dst_reg, src_reg*/ -} -void host_x86_PANDN_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xdf, 0xc0 | src_reg | (dst_reg << 3)); /*PANDN dst_reg, src_reg*/ -} -void host_x86_POR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xeb, 0xc0 | src_reg | (dst_reg << 3)); /*POR dst_reg, src_reg*/ -} -void host_x86_PXOR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xef, 0xc0 | src_reg | (dst_reg << 3)); /*PXOR dst_reg, src_reg*/ -} - -void host_x86_PCMPEQB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x74, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQB dst_reg, src_reg*/ -} -void host_x86_PCMPEQW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x75, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQW dst_reg, src_reg*/ -} -void host_x86_PCMPEQD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x76, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQD dst_reg, src_reg*/ -} -void host_x86_PCMPGTB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x64, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTB dst_reg, src_reg*/ -} -void host_x86_PCMPGTW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x65, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTW dst_reg, src_reg*/ -} -void host_x86_PCMPGTD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x66, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTD dst_reg, src_reg*/ -} - -void host_x86_PMADDWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xf5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ -} -void host_x86_PMULHW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xe5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ -} -void host_x86_PMULLW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xd5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ -} - -void host_x86_PSLLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x30 | dst_reg); /*PSLLW dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSLLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x30 | dst_reg); /*PSLLD dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSLLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x30 | dst_reg); /*PSLLD dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRAW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x20 | dst_reg); /*PSRAW dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRAD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x20 | dst_reg); /*PSRAD dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRAQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x20 | dst_reg); /*PSRAD dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x10 | dst_reg); /*PSRLW dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x10 | dst_reg); /*PSRLD dst_reg, imm*/ - codegen_addbyte(block, shift); -} -void host_x86_PSRLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) -{ - codegen_alloc_bytes(block, 5); - codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x10 | dst_reg); /*PSRLD dst_reg, imm*/ - codegen_addbyte(block, shift); -} - -void host_x86_PSUBB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xf8, 0xc0 | src_reg | (dst_reg << 3)); /*PADDB dst_reg, src_reg*/ -} -void host_x86_PSUBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xf9, 0xc0 | src_reg | (dst_reg << 3)); /*PADDW dst_reg, src_reg*/ -} -void host_x86_PSUBD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xfa, 0xc0 | src_reg | (dst_reg << 3)); /*PADDD dst_reg, src_reg*/ -} -void host_x86_PSUBSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xe8, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBSB dst_reg, src_reg*/ -} -void host_x86_PSUBSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xe9, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBSW dst_reg, src_reg*/ -} -void host_x86_PSUBUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xd8, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBUSB dst_reg, src_reg*/ -} -void host_x86_PSUBUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0xd9, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBUSW dst_reg, src_reg*/ -} - -void host_x86_PUNPCKHBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x60, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLBW dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ - codegen_addbyte(block, 0xee); -} -void host_x86_PUNPCKHWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x61, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLWD dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ - codegen_addbyte(block, 0xee); -} -void host_x86_PUNPCKHDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 9); - codegen_addbyte4(block, 0x66, 0x0f, 0x62, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLDQ dst_reg, src_reg*/ - codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ - codegen_addbyte(block, 0xee); -} -void host_x86_PUNPCKLBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x60, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLBW dst_reg, src_reg*/ -} -void host_x86_PUNPCKLWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x61, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLWD dst_reg, src_reg*/ -} -void host_x86_PUNPCKLDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0x66, 0x0f, 0x62, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLDQ dst_reg, src_reg*/ -} - void host_x86_POP(codeblock_t *block, int src_reg) { codegen_alloc_bytes(block, 1); @@ -1667,17 +1034,6 @@ void host_x86_SHR32_IMM(codeblock_t *block, int dst_reg, int shift) codegen_addbyte3(block, 0xc1, 0xc0 | RM_OP_SHR | dst_reg, shift); /*SHR dst_reg, shift*/ } -void host_x86_SQRTSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x51, 0xc0 | src_reg | (dst_reg << 3)); /*SQRTSD dst_reg, src_reg*/ -} -void host_x86_SQRTSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf3, 0x0f, 0x51, 0xc0 | src_reg | (dst_reg << 3)); /*SQRTSS dst_reg, src_reg*/ -} - void host_x86_SUB8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data) { if (dst_reg == REG_EAX) @@ -1748,17 +1104,6 @@ void host_x86_SUB32_REG_REG(codeblock_t *block, int dst_reg, int src_reg) codegen_addbyte2(block, 0x29, 0xc0 | dst_reg | (src_reg << 3)); /*SUB dst_reg, src_reg*/ } -void host_x86_SUBPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x5c, 0xc0 | src_reg | (dst_reg << 3)); /*SUBPS dst_reg, src_reg*/ -} -void host_x86_SUBSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 4); - codegen_addbyte4(block, 0xf2, 0x0f, 0x5c, 0xc0 | src_reg | (dst_reg << 3)); -} - void host_x86_TEST8_REG(codeblock_t *block, int src_host_reg, int dst_host_reg) { codegen_alloc_bytes(block, 2); @@ -1790,12 +1135,6 @@ void host_x86_TEST32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data) } } -void host_x86_UNPCKLPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) -{ - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0x0f, 0x14, 0xc0 | src_reg | (dst_reg << 3)); -} - void host_x86_XOR8_REG_REG(codeblock_t *block, int dst_reg, int src_reg) { codegen_alloc_bytes(block, 2); @@ -1866,73 +1205,4 @@ void host_x86_XOR32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data) } } -void host_x87_FILDq_BASE(codeblock_t *block, int base_reg) -{ - if (base_reg == REG_ESP) - { - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0xdf, 0x2c, 0x24); /*FILDq [ESP]*/ - } - else - { - codegen_alloc_bytes(block, 2); - codegen_addbyte2(block, 0xdf, 0x28 | base_reg); /*FILDq [base_reg]*/ - } -} -void host_x87_FISTPq_BASE(codeblock_t *block, int base_reg) -{ - if (base_reg == REG_ESP) - { - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0xdf, 0x3c, 0x24); /*FISTPq [ESP]*/ - } - else - { - codegen_alloc_bytes(block, 2); - codegen_addbyte2(block, 0xdf, 0x38 | base_reg); /*FISTPq [base_reg]*/ - } -} -void host_x87_FLDCW(codeblock_t *block, void *p) -{ - int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); - - if (offset >= -128 && offset < 127) - { - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0xd9, 0x68 | REG_EBP, offset); /*FLDCW offset[EBP]*/ - } - else - { - codegen_alloc_bytes(block, 6); - codegen_addbyte2(block, 0xd9, 0x2d); /*FLDCW [p]*/ - codegen_addlong(block, (uint32_t)p); - } -} -void host_x87_FLDd_BASE(codeblock_t *block, int base_reg) -{ - if (base_reg == REG_ESP) - { - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0xdd, 0x04, 0x24); /*FILDq [ESP]*/ - } - else - { - codegen_alloc_bytes(block, 2); - codegen_addbyte2(block, 0xdd, 0x08 | base_reg); /*FILDq [base_reg]*/ - } -} -void host_x87_FSTPd_BASE(codeblock_t *block, int base_reg) -{ - if (base_reg == REG_ESP) - { - codegen_alloc_bytes(block, 3); - codegen_addbyte3(block, 0xdd, 0x1c, 0x24); /*FILDq [ESP]*/ - } - else - { - codegen_alloc_bytes(block, 2); - codegen_addbyte2(block, 0xdd, 0x18 | base_reg); /*FILDq [base_reg]*/ - } -} - #endif diff --git a/src/codegen_backend_x86_ops.h b/src/codegen_backend_x86_ops.h index db7b37c..08b7e9f 100644 --- a/src/codegen_backend_x86_ops.h +++ b/src/codegen_backend_x86_ops.h @@ -8,9 +8,6 @@ void host_x86_ADD8_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_ADD16_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_ADD32_REG_REG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_ADDPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_ADDSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_AND8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data); void host_x86_AND16_REG_IMM(codeblock_t *block, int dst_reg, uint16_t imm_data); void host_x86_AND32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data); @@ -28,28 +25,6 @@ void host_x86_CMP8_REG_REG(codeblock_t *block, int src_reg_a, int src_reg_b); void host_x86_CMP16_REG_REG(codeblock_t *block, int src_reg_a, int src_reg_b); void host_x86_CMP32_REG_REG(codeblock_t *block, int src_reg_a, int src_reg_b); -#define CMPPS_EQ 0 -#define CMPPS_NLT 5 -#define CMPPS_NLE 6 -void host_x86_CMPPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg, int type); - -void host_x86_COMISD_XREG_XREG(codeblock_t *block, int src_reg_a, int src_reg_b); - -void host_x86_CVTDQ2PS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_CVTPS2DQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_CVTSD2SI_REG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_CVTSD2SS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_CVTSI2SD_XREG_REG(codeblock_t *block, int dest_reg, int src_reg); -void host_x86_CVTSI2SS_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_CVTSS2SD_XREG_XREG(codeblock_t *block, int dest_reg, int src_reg); -void host_x86_CVTSS2SD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); - -void host_x86_DIVSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_DIVSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_INC32_ABS(codeblock_t *block, void *p); void host_x86_JMP(codeblock_t *block, void *p); @@ -80,14 +55,10 @@ uint32_t *host_x86_JZ_long(codeblock_t *block); void host_x86_LAHF(codeblock_t *block); -void host_x86_LDMXCSR(codeblock_t *block, void *p); - void host_x86_LEA_REG_IMM(codeblock_t *block, int dst_reg, int src_reg_a, uint32_t offset); void host_x86_LEA_REG_REG(codeblock_t *block, int dst_reg, int src_reg_a, int src_reg_b); void host_x86_LEA_REG_REG_SHIFT(codeblock_t *block, int dst_reg, int src_reg_a, int src_reg_b, int shift); -void host_x86_MAXSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_MOV8_ABS_IMM(codeblock_t *block, void *p, uint32_t imm_data); void host_x86_MOV32_ABS_IMM(codeblock_t *block, void *p, uint32_t imm_data); @@ -129,23 +100,6 @@ void host_x86_MOV32_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_MOV32_STACK_IMM(codeblock_t *block, int32_t offset, uint32_t imm_data); -void host_x86_MOVD_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg); -void host_x86_MOVD_REG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_MOVD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); -void host_x86_MOVD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_MOVQ_ABS_XREG(codeblock_t *block, void *p, int src_reg); -void host_x86_MOVQ_ABS_REG_REG_SHIFT_XREG(codeblock_t *block, uint32_t addr, int src_reg_a, int src_reg_b, int shift, int src_reg); -void host_x86_MOVQ_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg); -void host_x86_MOVQ_BASE_OFFSET_XREG(codeblock_t *block, int base_reg, int offset, int src_reg); -void host_x86_MOVQ_STACK_OFFSET_XREG(codeblock_t *block, int offset, int src_reg); - -void host_x86_MOVQ_XREG_ABS(codeblock_t *block, int dst_reg, void *p); -void host_x86_MOVQ_XREG_ABS_REG_REG_SHIFT(codeblock_t *block, int dst_reg, uint32_t addr, int src_reg_a, int src_reg_b, int shift); -void host_x86_MOVQ_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); -void host_x86_MOVQ_XREG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, int offset); -void host_x86_MOVQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_MOVSX_REG_16_8(codeblock_t *block, int dst_reg, int src_reg); void host_x86_MOVSX_REG_32_8(codeblock_t *block, int dst_reg, int src_reg); void host_x86_MOVSX_REG_32_16(codeblock_t *block, int dst_reg, int src_reg); @@ -159,12 +113,6 @@ void host_x86_MOVZX_REG_ABS_32_8(codeblock_t *block, int dst_reg, void *p); void host_x86_MOVZX_BASE_INDEX_32_8(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); void host_x86_MOVZX_BASE_INDEX_32_16(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); -void host_x86_MAXPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_MINPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_MULPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_MULSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_OR8_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_OR16_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_OR32_REG_REG(codeblock_t *block, int dst_reg, int src_reg); @@ -173,59 +121,6 @@ void host_x86_OR8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data); void host_x86_OR16_REG_IMM(codeblock_t *block, int dst_reg, uint16_t imm_data); void host_x86_OR32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data); -void host_x86_PACKSSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PACKSSDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PACKUSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PADDB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PADDUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PAND_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PANDN_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_POR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PXOR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PCMPEQB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PCMPEQW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PCMPEQD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PCMPGTB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PCMPGTW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PCMPGTD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PMADDWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PMULHW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PMULLW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PSLLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSLLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSLLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRAW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRAD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRAQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_PSRLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); - -void host_x86_PSUBB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PSUBUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - -void host_x86_PUNPCKHBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PUNPCKHWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PUNPCKHDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PUNPCKLBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PUNPCKLWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_PUNPCKLDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_POP(codeblock_t *block, int src_reg); void host_x86_PUSH(codeblock_t *block, int src_reg); @@ -256,9 +151,6 @@ void host_x86_SHR8_IMM(codeblock_t *block, int dst_reg, int shift); void host_x86_SHR16_IMM(codeblock_t *block, int dst_reg, int shift); void host_x86_SHR32_IMM(codeblock_t *block, int dst_reg, int shift); -void host_x86_SQRTSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_SQRTSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_SUB8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data); void host_x86_SUB16_REG_IMM(codeblock_t *block, int dst_reg, uint16_t imm_data); void host_x86_SUB32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data); @@ -267,17 +159,12 @@ void host_x86_SUB8_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_SUB16_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_SUB32_REG_REG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_SUBPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); -void host_x86_SUBSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_TEST8_REG(codeblock_t *block, int src_host_reg, int dst_host_reg); void host_x86_TEST16_REG(codeblock_t *block, int src_host_reg, int dst_host_reg); void host_x86_TEST32_REG(codeblock_t *block, int src_host_reg, int dst_host_reg); void host_x86_TEST32_REG_IMM(codeblock_t *block, int src_host_reg, uint32_t imm_data); -void host_x86_UNPCKLPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); - void host_x86_XOR8_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_XOR16_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_XOR32_REG_REG(codeblock_t *block, int dst_reg, int src_reg); @@ -285,9 +172,3 @@ void host_x86_XOR32_REG_REG(codeblock_t *block, int dst_reg, int src_reg); void host_x86_XOR8_REG_IMM(codeblock_t *block, int dst_reg, uint8_t imm_data); void host_x86_XOR16_REG_IMM(codeblock_t *block, int dst_reg, uint16_t imm_data); void host_x86_XOR32_REG_IMM(codeblock_t *block, int dst_reg, uint32_t imm_data); - -void host_x87_FILDq_BASE(codeblock_t *block, int base_reg); -void host_x87_FISTPq_BASE(codeblock_t *block, int base_reg); -void host_x87_FLDCW(codeblock_t *block, void *p); -void host_x87_FLDd_BASE(codeblock_t *block, int base_reg); -void host_x87_FSTPd_BASE(codeblock_t *block, int base_reg); diff --git a/src/codegen_backend_x86_ops_fpu.c b/src/codegen_backend_x86_ops_fpu.c new file mode 100644 index 0000000..1c3f990 --- /dev/null +++ b/src/codegen_backend_x86_ops_fpu.c @@ -0,0 +1,80 @@ +#if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined WIN32 || defined _WIN32 || defined _WIN32 + +#include "ibm.h" +#include "codegen.h" +#include "codegen_allocator.h" +#include "codegen_backend.h" +#include "codegen_backend_x86_defs.h" +#include "codegen_backend_x86_ops_fpu.h" +#include "codegen_backend_x86_ops_helpers.h" + +void host_x87_FILDq_BASE(codeblock_t *block, int base_reg) +{ + if (base_reg == REG_ESP) + { + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0xdf, 0x2c, 0x24); /*FILDq [ESP]*/ + } + else + { + codegen_alloc_bytes(block, 2); + codegen_addbyte2(block, 0xdf, 0x28 | base_reg); /*FILDq [base_reg]*/ + } +} +void host_x87_FISTPq_BASE(codeblock_t *block, int base_reg) +{ + if (base_reg == REG_ESP) + { + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0xdf, 0x3c, 0x24); /*FISTPq [ESP]*/ + } + else + { + codegen_alloc_bytes(block, 2); + codegen_addbyte2(block, 0xdf, 0x38 | base_reg); /*FISTPq [base_reg]*/ + } +} +void host_x87_FLDCW(codeblock_t *block, void *p) +{ + int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); + + if (offset >= -128 && offset < 127) + { + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0xd9, 0x68 | REG_EBP, offset); /*FLDCW offset[EBP]*/ + } + else + { + codegen_alloc_bytes(block, 6); + codegen_addbyte2(block, 0xd9, 0x2d); /*FLDCW [p]*/ + codegen_addlong(block, (uint32_t)p); + } +} +void host_x87_FLDd_BASE(codeblock_t *block, int base_reg) +{ + if (base_reg == REG_ESP) + { + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0xdd, 0x04, 0x24); /*FILDq [ESP]*/ + } + else + { + codegen_alloc_bytes(block, 2); + codegen_addbyte2(block, 0xdd, 0x08 | base_reg); /*FILDq [base_reg]*/ + } +} +void host_x87_FSTPd_BASE(codeblock_t *block, int base_reg) +{ + if (base_reg == REG_ESP) + { + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0xdd, 0x1c, 0x24); /*FILDq [ESP]*/ + } + else + { + codegen_alloc_bytes(block, 2); + codegen_addbyte2(block, 0xdd, 0x18 | base_reg); /*FILDq [base_reg]*/ + } +} + +#endif diff --git a/src/codegen_backend_x86_ops_fpu.h b/src/codegen_backend_x86_ops_fpu.h new file mode 100644 index 0000000..3692715 --- /dev/null +++ b/src/codegen_backend_x86_ops_fpu.h @@ -0,0 +1,5 @@ +void host_x87_FILDq_BASE(codeblock_t *block, int base_reg); +void host_x87_FISTPq_BASE(codeblock_t *block, int base_reg); +void host_x87_FLDCW(codeblock_t *block, void *p); +void host_x87_FLDd_BASE(codeblock_t *block, int base_reg); +void host_x87_FSTPd_BASE(codeblock_t *block, int base_reg); diff --git a/src/codegen_backend_x86_ops_helpers.h b/src/codegen_backend_x86_ops_helpers.h new file mode 100644 index 0000000..9c100ec --- /dev/null +++ b/src/codegen_backend_x86_ops_helpers.h @@ -0,0 +1,84 @@ +#define JMP_LEN_BYTES 5 + +static inline void codegen_addbyte(codeblock_t *block, uint8_t val) +{ + if (block_pos >= BLOCK_MAX) + fatal("codegen_addbyte over! %i\n", block_pos); + block_write_data[block_pos++] = val; +} +static inline void codegen_addbyte2(codeblock_t *block, uint8_t vala, uint8_t valb) +{ + if (block_pos > (BLOCK_MAX-2)) + fatal("codegen_addbyte2 over! %i\n", block_pos); + block_write_data[block_pos++] = vala; + block_write_data[block_pos++] = valb; +} +static inline void codegen_addbyte3(codeblock_t *block, uint8_t vala, uint8_t valb, uint8_t valc) +{ + if (block_pos > (BLOCK_MAX-3)) + fatal("codegen_addbyte3 over! %i\n", block_pos); + block_write_data[block_pos++] = vala; + block_write_data[block_pos++] = valb; + block_write_data[block_pos++] = valc; +} +static inline void codegen_addbyte4(codeblock_t *block, uint8_t vala, uint8_t valb, uint8_t valc, uint8_t vald) +{ + if (block_pos > (BLOCK_MAX-4)) + fatal("codegen_addbyte4 over! %i\n", block_pos); + block_write_data[block_pos++] = vala; + block_write_data[block_pos++] = valb; + block_write_data[block_pos++] = valc; + block_write_data[block_pos++] = vald; +} + +static inline void codegen_addword(codeblock_t *block, uint16_t val) +{ + if (block_pos > (BLOCK_MAX-2)) + fatal("codegen_addword over! %i\n", block_pos); + *(uint16_t *)&block_write_data[block_pos] = val; + block_pos += 2; +} + +static inline void codegen_addlong(codeblock_t *block, uint32_t val) +{ + if (block_pos > (BLOCK_MAX-4)) + fatal("codegen_addlong over! %i\n", block_pos); + *(uint32_t *)&block_write_data[block_pos] = val; + block_pos += 4; +} + +static inline void codegen_addquad(codeblock_t *block, uint64_t val) +{ + if (block_pos > (BLOCK_MAX-8)) + fatal("codegen_addquad over! %i\n", block_pos); + *(uint64_t *)&block_write_data[block_pos] = val; + block_pos += 8; +} + +static void codegen_allocate_new_block(codeblock_t *block) +{ + /*Current block is full. Allocate a new block*/ + struct mem_block_t *new_block = codegen_allocator_allocate(block->head_mem_block, get_block_nr(block)); + uint8_t *new_ptr = codeblock_allocator_get_ptr(new_block); + + /*Add a jump instruction to the new block*/ + codegen_addbyte(block, 0xe9); /*JMP*/ + codegen_addlong(block, (uintptr_t)new_ptr - (uintptr_t)&block_write_data[block_pos + 4]); + + /*Set write address to start of new block*/ + block_pos = 0; + block_write_data = new_ptr; +} + +static inline void codegen_alloc_bytes(codeblock_t *block, int size) +{ + if (block_pos > ((BLOCK_MAX - size) - JMP_LEN_BYTES)) + codegen_allocate_new_block(block); +} + +static inline int is_imm8(uint32_t imm_data) +{ + if (imm_data <= 0x7f || imm_data >= 0xffffff80) + return 1; + return 0; +} diff --git a/src/codegen_backend_x86_ops_sse.c b/src/codegen_backend_x86_ops_sse.c new file mode 100644 index 0000000..dc20d34 --- /dev/null +++ b/src/codegen_backend_x86_ops_sse.c @@ -0,0 +1,588 @@ +#if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined WIN32 || defined _WIN32 || defined _WIN32 + +#include "ibm.h" +#include "codegen.h" +#include "codegen_allocator.h" +#include "codegen_backend.h" +#include "codegen_backend_x86_defs.h" +#include "codegen_backend_x86_ops_sse.h" +#include "codegen_backend_x86_ops_helpers.h" + +void host_x86_ADDPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x58, 0xc0 | src_reg | (dst_reg << 3)); /*ADDPS dst_reg, src_reg*/ +} +void host_x86_ADDSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x58, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_CMPPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg, int type) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x0f, 0xc2, 0xc0 | src_reg | (dst_reg << 3), type); /*CMPPS dst_reg, src_reg, type*/ +} + +void host_x86_COMISD_XREG_XREG(codeblock_t *block, int src_reg_a, int src_reg_b) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x2e, 0xc0 | src_reg_b | (src_reg_a << 3)); +} + +void host_x86_CVTDQ2PS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x5b, 0xc0 | src_reg | (dst_reg << 3)); /*CVTDQ2PS dst_reg, src_reg*/ +} +void host_x86_CVTPS2DQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x5b, 0xc0 | src_reg | (dst_reg << 3)); /*CVTPS2DQ dst_reg, src_reg*/ +} + +void host_x86_CVTSD2SI_REG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x2d, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSD2SI dst_reg, src_reg*/ +} +void host_x86_CVTSD2SS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x5a, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSD2SS dst_reg, src_reg*/ +} + +void host_x86_CVTSI2SD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x2a, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSI2SD dst_reg, src_reg*/ +} +void host_x86_CVTSI2SS_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x2a, 0xc0 | src_reg | (dst_reg << 3)); /*CVTSI2SD dst_reg, src_reg*/ +} + +void host_x86_CVTSS2SD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x5a, 0xc0 | src_reg | (dst_reg << 3)); +} +void host_x86_CVTSS2SD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0xf3, 0x0f, 0x5a, 0x04 | (dst_reg << 3)); /*CVTSS2SD XMMx, [base_reg + idx_reg]*/ + codegen_addbyte(block, base_reg | (idx_reg << 3)); +} + +void host_x86_DIVSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x5e, 0xc0 | src_reg | (dst_reg << 3)); /*DIVSD dst_reg, src_reg*/ +} +void host_x86_DIVSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x5e, 0xc0 | src_reg | (dst_reg << 3)); /*DIVSS dst_reg, src_reg*/ +} + +void host_x86_LDMXCSR(codeblock_t *block, void *p) +{ + int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); + + if (offset >= -128 && offset < 127) + { + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x0f, 0xae, 0x50 | REG_EBP, offset); /*LDMXCSR offset[EBP]*/ + } + else + { + codegen_alloc_bytes(block, 7); + codegen_addbyte3(block, 0x0f, 0xae, 0x15); /*LDMXCSR [p]*/ + codegen_addlong(block, (uint32_t)p); + } +} + +void host_x86_MAXSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x5f, 0xc0 | src_reg | (dst_reg << 3)); /*MAXSD dst_reg, src_reg*/ +} + +void host_x86_MOVD_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x7e, 0x04 | (src_reg << 3)); /*MOVD XMMx, [base_reg + idx_reg]*/ + codegen_addbyte(block, base_reg | (idx_reg << 3)); +} +void host_x86_MOVD_REG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x7e, 0xc0 | dst_reg | (src_reg << 3)); +} +void host_x86_MOVD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x6e, 0x04 | (dst_reg << 3)); /*MOVD XMMx, [base_reg + idx_reg]*/ + codegen_addbyte(block, base_reg | (idx_reg << 3)); +} +void host_x86_MOVD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x6e, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_MOVQ_ABS_XREG(codeblock_t *block, void *p, int src_reg) +{ + int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); + + if (offset >= -128 && offset < 127) + { + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x45 | (src_reg << 3)); /*MOVQ offset[EBP], src_reg*/ + codegen_addbyte(block, offset); + } + else + { + codegen_alloc_bytes(block, 8); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x05 | (src_reg << 3)); /*MOVQ [p], src_reg*/ + codegen_addlong(block, (uint32_t)p); + } +} +void host_x86_MOVQ_ABS_REG_REG_SHIFT_XREG(codeblock_t *block, uint32_t addr, int src_reg_a, int src_reg_b, int shift, int src_reg) +{ + if (addr < 0x80 || addr >= 0xffffff80) + { + codegen_alloc_bytes(block, 6); + codegen_addbyte3(block, 0x66, 0x0f, 0xd6); /*MOVQ addr[src_reg_a + src_reg_b << shift], XMMx*/ + codegen_addbyte3(block, 0x44 | (src_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6), addr & 0xff); + } + else + { + codegen_alloc_bytes(block, 9); + codegen_addbyte3(block, 0x66, 0x0f, 0xd6); /*MOVQ addr[src_reg_a + src_reg_b << shift], XMMx*/ + codegen_addbyte2(block, 0x84 | (src_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6)); + codegen_addlong(block, addr); + } +} +void host_x86_MOVQ_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x04 | (src_reg << 3)); /*MOVQ XMMx, [base_reg + idx_reg]*/ + codegen_addbyte(block, base_reg | (idx_reg << 3)); +} +void host_x86_MOVQ_BASE_OFFSET_XREG(codeblock_t *block, int base_reg, int offset, int src_reg) +{ + if (offset >= -128 && offset < 127) + { + if (base_reg == REG_ESP) + { + codegen_alloc_bytes(block, 6); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x44 | (src_reg << 3)); /*MOVQ [ESP + offset], XMMx*/ + codegen_addbyte2(block, 0x24, offset); + } + else + { + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x40 | base_reg | (src_reg << 3)); /*MOVQ [base_reg + offset], XMMx*/ + codegen_addbyte(block, offset); + } + } + else + fatal("MOVQ_BASE_OFFSET_XREG - offset %i\n", offset); +} +void host_x86_MOVQ_STACK_OFFSET_XREG(codeblock_t *block, int offset, int src_reg) +{ + if (!offset) + { + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x04 | (src_reg << 3)); /*MOVQ [ESP], src_reg*/ + codegen_addbyte(block, 0x24); + } + else if (offset >= -80 || offset < 0x80) + { + codegen_alloc_bytes(block, 6); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x44 | (src_reg << 3)); /*MOVQ offset[ESP], src_reg*/ + codegen_addbyte2(block, 0x24, offset & 0xff); + } + else + { + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0xd6, 0x84 | (src_reg << 3)); /*MOVQ offset[ESP], src_reg*/ + codegen_addbyte(block, 0x24); + codegen_addlong(block, offset); + } + +} + +void host_x86_MOVQ_XREG_ABS(codeblock_t *block, int dst_reg, void *p) +{ + int offset = (uintptr_t)p - (((uintptr_t)&cpu_state) + 128); + + if (offset >= -128 && offset < 127) + { + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x45 | (dst_reg << 3)); /*MOVQ offset[EBP], src_reg*/ + codegen_addbyte(block, offset); + } + else + { + codegen_alloc_bytes(block, 8); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x05 | (dst_reg << 3)); /*MOVQ [p], src_reg*/ + codegen_addlong(block, (uint32_t)p); + } +} +void host_x86_MOVQ_XREG_ABS_REG_REG_SHIFT(codeblock_t *block, int dst_reg, uint32_t addr, int src_reg_a, int src_reg_b, int shift) +{ + if (addr < 0x80 || addr >= 0xffffff80) + { + codegen_alloc_bytes(block, 6); + codegen_addbyte3(block, 0xf3, 0x0f, 0x7e); /*MOVQ XMMx, addr[src_reg_a + src_reg_b << shift]*/ + codegen_addbyte3(block, 0x44 | (dst_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6), addr & 0xff); + } + else + { + codegen_alloc_bytes(block, 9); + codegen_addbyte3(block, 0xf3, 0x0f, 0x7e); /*MOVQ XMMx, addr[src_reg_a + src_reg_b << shift]*/ + codegen_addbyte2(block, 0x84 | (dst_reg << 3), src_reg_a | (src_reg_b << 3) | (shift << 6)); + codegen_addlong(block, addr); + } +} +void host_x86_MOVQ_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x04 | (dst_reg << 3)); /*MOVQ XMMx, [base_reg + idx_reg]*/ + codegen_addbyte(block, base_reg | (idx_reg << 3)); +} +void host_x86_MOVQ_XREG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, int offset) +{ + if (offset >= -128 && offset < 127) + { + if (base_reg == REG_ESP) + { + codegen_alloc_bytes(block, 6); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x44 | (dst_reg << 3)); /*MOVQ XMMx, [ESP + offset]*/ + codegen_addbyte2(block, 0x24, offset); + } + else + { + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0x40 | base_reg | (dst_reg << 3)); /*MOVQ XMMx, [base_reg + offset]*/ + codegen_addbyte(block, offset); + } + } + else + fatal("MOVQ_REG_BASE_OFFSET - offset %i\n", offset); +} +void host_x86_MOVQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x7e, 0xc0 | src_reg | (dst_reg << 3)); /*MOVQ dst_reg, src_reg*/ +} + +void host_x86_MAXPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x5f, 0xc0 | src_reg | (dst_reg << 3)); /*MAXPS dst_reg, src_reg*/ +} +void host_x86_MINPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x5d, 0xc0 | src_reg | (dst_reg << 3)); /*MINPS dst_reg, src_reg*/ +} + +void host_x86_MULPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x59, 0xc0 | src_reg | (dst_reg << 3)); /*MULPS dst_reg, src_reg*/ +} +void host_x86_MULSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x59, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_PACKSSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x63, 0xc0 | src_reg | (dst_reg << 3)); /*PACKSSWB dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ + codegen_addbyte(block, 0x88); +} +void host_x86_PACKSSDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x6b, 0xc0 | src_reg | (dst_reg << 3)); /*PACKSSDW dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ + codegen_addbyte(block, 0x88); +} +void host_x86_PACKUSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x67, 0xc0 | src_reg | (dst_reg << 3)); /*PACKUSWB dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0x88 (move bits 64-95 to 32-63)*/ + codegen_addbyte(block, 0x88); +} + +void host_x86_PADDB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xfc, 0xc0 | src_reg | (dst_reg << 3)); /*PADDB dst_reg, src_reg*/ +} +void host_x86_PADDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xfd, 0xc0 | src_reg | (dst_reg << 3)); /*PADDW dst_reg, src_reg*/ +} +void host_x86_PADDD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xfe, 0xc0 | src_reg | (dst_reg << 3)); /*PADDD dst_reg, src_reg*/ +} +void host_x86_PADDSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xec, 0xc0 | src_reg | (dst_reg << 3)); /*PADDSB dst_reg, src_reg*/ +} +void host_x86_PADDSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xed, 0xc0 | src_reg | (dst_reg << 3)); /*PADDSW dst_reg, src_reg*/ +} +void host_x86_PADDUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xdc, 0xc0 | src_reg | (dst_reg << 3)); /*PADDUSB dst_reg, src_reg*/ +} +void host_x86_PADDUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xdd, 0xc0 | src_reg | (dst_reg << 3)); /*PADDUSW dst_reg, src_reg*/ +} + +void host_x86_PAND_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xdb, 0xc0 | src_reg | (dst_reg << 3)); /*PAND dst_reg, src_reg*/ +} +void host_x86_PANDN_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xdf, 0xc0 | src_reg | (dst_reg << 3)); /*PANDN dst_reg, src_reg*/ +} +void host_x86_POR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xeb, 0xc0 | src_reg | (dst_reg << 3)); /*POR dst_reg, src_reg*/ +} +void host_x86_PXOR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xef, 0xc0 | src_reg | (dst_reg << 3)); /*PXOR dst_reg, src_reg*/ +} + +void host_x86_PCMPEQB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x74, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQB dst_reg, src_reg*/ +} +void host_x86_PCMPEQW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x75, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQW dst_reg, src_reg*/ +} +void host_x86_PCMPEQD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x76, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPEQD dst_reg, src_reg*/ +} +void host_x86_PCMPGTB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x64, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTB dst_reg, src_reg*/ +} +void host_x86_PCMPGTW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x65, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTW dst_reg, src_reg*/ +} +void host_x86_PCMPGTD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x66, 0xc0 | src_reg | (dst_reg << 3)); /*PCMPGTD dst_reg, src_reg*/ +} + +void host_x86_PMADDWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xf5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ +} +void host_x86_PMULHW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xe5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ +} +void host_x86_PMULLW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xd5, 0xc0 | src_reg | (dst_reg << 3)); /*PMULLW dst_reg, src_reg*/ +} + +void host_x86_PSLLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x30 | dst_reg); /*PSLLW dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSLLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x30 | dst_reg); /*PSLLD dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSLLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x30 | dst_reg); /*PSLLD dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRAW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x20 | dst_reg); /*PSRAW dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRAD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x20 | dst_reg); /*PSRAD dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRAQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x20 | dst_reg); /*PSRAD dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x71, 0xc0 | 0x10 | dst_reg); /*PSRLW dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x72, 0xc0 | 0x10 | dst_reg); /*PSRLD dst_reg, imm*/ + codegen_addbyte(block, shift); +} +void host_x86_PSRLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift) +{ + codegen_alloc_bytes(block, 5); + codegen_addbyte4(block, 0x66, 0x0f, 0x73, 0xc0 | 0x10 | dst_reg); /*PSRLD dst_reg, imm*/ + codegen_addbyte(block, shift); +} + +void host_x86_PSUBB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xf8, 0xc0 | src_reg | (dst_reg << 3)); /*PADDB dst_reg, src_reg*/ +} +void host_x86_PSUBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xf9, 0xc0 | src_reg | (dst_reg << 3)); /*PADDW dst_reg, src_reg*/ +} +void host_x86_PSUBD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xfa, 0xc0 | src_reg | (dst_reg << 3)); /*PADDD dst_reg, src_reg*/ +} +void host_x86_PSUBSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xe8, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBSB dst_reg, src_reg*/ +} +void host_x86_PSUBSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xe9, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBSW dst_reg, src_reg*/ +} +void host_x86_PSUBUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xd8, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBUSB dst_reg, src_reg*/ +} +void host_x86_PSUBUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0xd9, 0xc0 | src_reg | (dst_reg << 3)); /*PSUBUSW dst_reg, src_reg*/ +} + +void host_x86_PUNPCKHBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x60, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLBW dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ + codegen_addbyte(block, 0xee); +} +void host_x86_PUNPCKHWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x61, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLWD dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ + codegen_addbyte(block, 0xee); +} +void host_x86_PUNPCKHDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 9); + codegen_addbyte4(block, 0x66, 0x0f, 0x62, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLDQ dst_reg, src_reg*/ + codegen_addbyte4(block, 0x66, 0x0f, 0x70, 0xc0 | dst_reg | (dst_reg << 3)); /*PSHUFD dst_reg, dst_reg, 0xee (move top 64-bits to low 64-bits)*/ + codegen_addbyte(block, 0xee); +} +void host_x86_PUNPCKLBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x60, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLBW dst_reg, src_reg*/ +} +void host_x86_PUNPCKLWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x61, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLWD dst_reg, src_reg*/ +} +void host_x86_PUNPCKLDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0x66, 0x0f, 0x62, 0xc0 | src_reg | (dst_reg << 3)); /*PUNPCKLDQ dst_reg, src_reg*/ +} + +void host_x86_SQRTSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x51, 0xc0 | src_reg | (dst_reg << 3)); /*SQRTSD dst_reg, src_reg*/ +} +void host_x86_SQRTSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf3, 0x0f, 0x51, 0xc0 | src_reg | (dst_reg << 3)); /*SQRTSS dst_reg, src_reg*/ +} + +void host_x86_SUBPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x5c, 0xc0 | src_reg | (dst_reg << 3)); /*SUBPS dst_reg, src_reg*/ +} +void host_x86_SUBSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 4); + codegen_addbyte4(block, 0xf2, 0x0f, 0x5c, 0xc0 | src_reg | (dst_reg << 3)); +} + +void host_x86_UNPCKLPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_alloc_bytes(block, 3); + codegen_addbyte3(block, 0x0f, 0x14, 0xc0 | src_reg | (dst_reg << 3)); +} + +#endif diff --git a/src/codegen_backend_x86_ops_sse.h b/src/codegen_backend_x86_ops_sse.h new file mode 100644 index 0000000..596f361 --- /dev/null +++ b/src/codegen_backend_x86_ops_sse.h @@ -0,0 +1,112 @@ +void host_x86_ADDPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_ADDSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +#define CMPPS_EQ 0 +#define CMPPS_NLT 5 +#define CMPPS_NLE 6 +void host_x86_CMPPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg, int type); + +void host_x86_COMISD_XREG_XREG(codeblock_t *block, int src_reg_a, int src_reg_b); + +void host_x86_CVTDQ2PS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_CVTPS2DQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_CVTSD2SI_REG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_CVTSD2SS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_CVTSI2SD_XREG_REG(codeblock_t *block, int dest_reg, int src_reg); +void host_x86_CVTSI2SS_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_CVTSS2SD_XREG_XREG(codeblock_t *block, int dest_reg, int src_reg); +void host_x86_CVTSS2SD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); + +void host_x86_DIVSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_DIVSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_LDMXCSR(codeblock_t *block, void *p); + +void host_x86_MAXSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_MOVD_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg); +void host_x86_MOVD_REG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_MOVD_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); +void host_x86_MOVD_XREG_REG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_MOVQ_ABS_XREG(codeblock_t *block, void *p, int src_reg); +void host_x86_MOVQ_ABS_REG_REG_SHIFT_XREG(codeblock_t *block, uint32_t addr, int src_reg_a, int src_reg_b, int shift, int src_reg); +void host_x86_MOVQ_BASE_INDEX_XREG(codeblock_t *block, int base_reg, int idx_reg, int src_reg); +void host_x86_MOVQ_BASE_OFFSET_XREG(codeblock_t *block, int base_reg, int offset, int src_reg); +void host_x86_MOVQ_STACK_OFFSET_XREG(codeblock_t *block, int offset, int src_reg); + +void host_x86_MOVQ_XREG_ABS(codeblock_t *block, int dst_reg, void *p); +void host_x86_MOVQ_XREG_ABS_REG_REG_SHIFT(codeblock_t *block, int dst_reg, uint32_t addr, int src_reg_a, int src_reg_b, int shift); +void host_x86_MOVQ_XREG_BASE_INDEX(codeblock_t *block, int dst_reg, int base_reg, int idx_reg); +void host_x86_MOVQ_XREG_BASE_OFFSET(codeblock_t *block, int dst_reg, int base_reg, int offset); +void host_x86_MOVQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_MAXPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_MINPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_MULPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_MULSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PACKSSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PACKSSDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PACKUSWB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PADDB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PADDUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PAND_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PANDN_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_POR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PXOR_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PCMPEQB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PCMPEQW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PCMPEQD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PCMPGTB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PCMPGTW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PCMPGTD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PMADDWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PMULHW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PMULLW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PSLLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSLLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSLLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRAW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRAD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRAQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRLW_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRLD_XREG_IMM(codeblock_t *block, int dst_reg, int shift); +void host_x86_PSRLQ_XREG_IMM(codeblock_t *block, int dst_reg, int shift); + +void host_x86_PSUBB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBUSB_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PSUBUSW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_PUNPCKHBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PUNPCKHWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PUNPCKHDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PUNPCKLBW_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PUNPCKLWD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_PUNPCKLDQ_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_SQRTSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_SQRTSS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_SUBPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); +void host_x86_SUBSD_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); + +void host_x86_UNPCKLPS_XREG_XREG(codeblock_t *block, int dst_reg, int src_reg); diff --git a/src/codegen_backend_x86_uops.c b/src/codegen_backend_x86_uops.c index 55bcc13..b05d4c1 100644 --- a/src/codegen_backend_x86_uops.c +++ b/src/codegen_backend_x86_uops.c @@ -9,6 +9,8 @@ #include "codegen_backend.h" #include "codegen_backend_x86_defs.h" #include "codegen_backend_x86_ops.h" +#include "codegen_backend_x86_ops_fpu.h" +#include "codegen_backend_x86_ops_sse.h" #include "codegen_ir_defs.h" /*void codegen_debug()