Merge pull request #2869 from BrzVlad/feature-mod-union-opt

[mono.git] / mono / arch / amd64 / amd64-codegen.h

diff --git a/mono/arch/amd64/amd64-codegen.h b/mono/arch/amd64/amd64-codegen.h
index cb7c80e1094280eedd0a0837298e01e902a0342f..26d73ce31c49214d27052cc77e2f929f8b436829 100644 (file)
--- a/mono/arch/amd64/amd64-codegen.h
+++ b/mono/arch/amd64/amd64-codegen.h
@@ -11,11 +11,27 @@
  * 
  * Copyright (C)  2000 Intel Corporation.  All rights reserved.
  * Copyright (C)  2001, 2002 Ximian, Inc.
+ * Licensed under the MIT license. See LICENSE file in the project root for full license information.
  */
 
 #ifndef AMD64_H
 #define AMD64_H
 
+// Conventions in this file:
+
+// body: implementation. other macros call this one
+// disp: displacement
+// inst: instruction
+// is_half: short if true, byte if false (then why is it named is_half...?)
+// imm: immediate
+// mem: read from (immediate-supplied address?)
+// membase: read from address in a base register plus a displacement
+// memindex: SIP addressing: (address in base register) + (displacement in index register)<<(shift)
+// reg: register, encode modR/M bits 00
+// regp: register, encode modR/M bits 11
+// size: Expected 1,2,4 or 8
+// widen: extends from 1 or 2 bytes
+
 #include <glib.h>
 
 typedef enum {
@@ -214,7 +230,7 @@ typedef union {
                        *(inst)++ = (unsigned char)0x83;        \
                        x86_reg_emit ((inst), (opc), (reg));    \
                        x86_imm_emit8 ((inst), (imm));  \
-               } else if ((reg) == X86_EAX) {  \
+               } else if ((reg) == AMD64_RAX) {        \
                        amd64_emit_rex(inst, size, 0, 0, 0); \
                        *(inst)++ = (((unsigned char)(opc)) << 3) + 5;  \
                        x86_imm_emit32 ((inst), (imm)); \
@@ -233,6 +249,21 @@ typedef union {
                x86_reg_emit ((inst), (dreg), (reg));   \
        } while (0)
 
+#define amd64_test_reg_imm_size_body(inst,reg,imm,size) \
+       do { \
+               amd64_codegen_pre(inst); \
+               amd64_emit_rex ((inst),(size),0,0,(reg)); \
+               if ((reg) == AMD64_RAX) { \
+                       *(inst)++ = (unsigned char)0xa9; \
+               } \
+               else { \
+                       *(inst)++ = (unsigned char)0xf7;        \
+                       x86_reg_emit((inst), 0, (reg)); \
+               } \
+               x86_imm_emit32((inst), (imm));  \
+               amd64_codegen_post(inst); \
+       } while (0)
+
 #if defined(__default_codegen__)
 
 #define amd64_alu_reg_imm_size(inst,opc,reg,imm,size) \
@@ -241,6 +272,9 @@ typedef union {
 #define amd64_alu_reg_reg_size(inst,opc,dreg,reg,size) \
                amd64_alu_reg_reg_size_body((inst), (opc), (dreg), (reg), (size))
 
+#define amd64_test_reg_imm_size(inst, reg, imm, size) \
+               amd64_test_reg_imm_size_body(inst, reg, imm, size)
+
 #elif defined(__native_client_codegen__)
 /* NaCl modules may not directly update RSP or RBP other than direct copies */
 /* between them. Instead the lower 4 bytes are updated and then added to R15 */
@@ -282,6 +316,8 @@ typedef union {
 
 #define amd64_alu_reg_reg(inst,opc,dreg,reg) amd64_alu_reg_reg_size ((inst),(opc),(dreg),(reg),8)
 
+#define amd64_test_reg_imm(inst,reg,imm) amd64_test_reg_imm_size(inst,reg,imm,8)
+
 #define amd64_alu_reg_membase_size(inst,opc,reg,basereg,disp,size) \
        do { \
                amd64_codegen_pre(inst);                                                  \
@@ -1029,6 +1065,7 @@ typedef union {
 #define amd64_sse_andpd_reg_membase(inst,dreg,basereg,disp) emit_sse_reg_membase ((inst),(dreg),(basereg), (disp), 0x66, 0x0f, 0x54)
 
 #define amd64_sse_movsd_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf2, 0x0f, 0x10)
+#define amd64_sse_movss_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf3, 0x0f, 0x10)
 
 #define amd64_sse_movsd_reg_membase(inst,dreg,basereg,disp) emit_sse_reg_membase ((inst), (dreg), (basereg), (disp), 0xf2, 0x0f, 0x10)
 
@@ -1039,14 +1076,17 @@ typedef union {
 #define amd64_sse_movss_reg_membase(inst,dreg,basereg,disp) emit_sse_reg_membase ((inst), (dreg), (basereg), (disp), 0xf3, 0x0f, 0x10)
 
 #define amd64_sse_comisd_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst),(dreg),(reg),0x66,0x0f,0x2f)
+#define amd64_sse_comiss_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst),(dreg),(reg),0x67,0x0f,0x2f)
 
 #define amd64_sse_comisd_reg_membase(inst,dreg,basereg,disp) emit_sse_reg_membase ((inst), (dreg), (basereg), (disp), 0x66, 0x0f, 0x2f)
 
 #define amd64_sse_ucomisd_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst),(dreg),(reg),0x66,0x0f,0x2e)
 
 #define amd64_sse_cvtsd2si_reg_reg(inst,dreg,reg) emit_sse_reg_reg_size ((inst), (dreg), (reg), 0xf2, 0x0f, 0x2d, 8)
+#define amd64_sse_cvtss2si_reg_reg(inst,dreg,reg) emit_sse_reg_reg_size ((inst), (dreg), (reg), 0xf3, 0x0f, 0x2d, 8)
 
 #define amd64_sse_cvttsd2si_reg_reg_size(inst,dreg,reg,size) emit_sse_reg_reg_size ((inst), (dreg), (reg), 0xf2, 0x0f, 0x2c, (size))
+#define amd64_sse_cvtss2si_reg_reg_size(inst,dreg,reg,size) emit_sse_reg_reg_size ((inst), (dreg), (reg), 0xf3, 0x0f, 0x2c, (size))
 
 #define amd64_sse_cvttsd2si_reg_reg(inst,dreg,reg) amd64_sse_cvttsd2si_reg_reg_size ((inst), (dreg), (reg), 8)
 
@@ -1054,17 +1094,25 @@ typedef union {
 
 #define amd64_sse_cvtsi2sd_reg_reg(inst,dreg,reg) amd64_sse_cvtsi2sd_reg_reg_size ((inst), (dreg), (reg), 8)
 
+#define amd64_sse_cvtsi2ss_reg_reg_size(inst,dreg,reg,size) emit_sse_reg_reg_size ((inst), (dreg), (reg), 0xf3, 0x0f, 0x2a, (size))
+
+#define amd64_sse_cvtsi2ss_reg_reg(inst,dreg,reg) amd64_sse_cvtsi2ss_reg_reg_size ((inst), (dreg), (reg), 8)
+
 #define amd64_sse_cvtsd2ss_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf2, 0x0f, 0x5a)
 
 #define amd64_sse_cvtss2sd_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf3, 0x0f, 0x5a)
 
 #define amd64_sse_addsd_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf2, 0x0f, 0x58)
+#define amd64_sse_addss_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf3, 0x0f, 0x58)
 
 #define amd64_sse_subsd_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf2, 0x0f, 0x5c)
+#define amd64_sse_subss_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf3, 0x0f, 0x5c)
 
 #define amd64_sse_mulsd_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf2, 0x0f, 0x59)
+#define amd64_sse_mulss_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf3, 0x0f, 0x59)
 
 #define amd64_sse_divsd_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf2, 0x0f, 0x5e)
+#define amd64_sse_divss_reg_reg(inst,dreg,reg) emit_sse_reg_reg ((inst), (dreg), (reg), 0xf3, 0x0f, 0x5e)
 
 #define amd64_sse_sqrtsd_reg_reg(inst,dreg,reg) emit_sse_reg_reg((inst), (dreg), (reg), 0xf2, 0x0f, 0x51)
 
@@ -1353,6 +1401,23 @@ typedef union {
 #define amd64_sse_psllq_reg_reg(inst, dreg, reg) emit_sse_reg_reg((inst), (dreg), (reg), 0x66, 0x0f, 0xf3)
 
 
+#define amd64_sse_cvtdq2pd_reg_reg(inst, dreg, reg) emit_sse_reg_reg((inst), (dreg), (reg), 0xF3, 0x0F, 0xE6)
+
+#define amd64_sse_cvtdq2ps_reg_reg(inst, dreg, reg) emit_sse_reg_reg_op2((inst), (dreg), (reg), 0x0F, 0x5B)
+
+#define amd64_sse_cvtpd2dq_reg_reg(inst, dreg, reg) emit_sse_reg_reg((inst), (dreg), (reg), 0xF2, 0x0F, 0xE6)
+
+#define amd64_sse_cvtpd2ps_reg_reg(inst, dreg, reg) emit_sse_reg_reg((inst), (dreg), (reg), 0x66, 0x0F, 0x5A)
+
+#define amd64_sse_cvtps2dq_reg_reg(inst, dreg, reg) emit_sse_reg_reg((inst), (dreg), (reg), 0x66, 0x0F, 0x5B)
+
+#define amd64_sse_cvtps2pd_reg_reg(inst, dreg, reg) emit_sse_reg_reg_op2((inst), (dreg), (reg), 0x0F, 0x5A)
+
+#define amd64_sse_cvttpd2dq_reg_reg(inst, dreg, reg) emit_sse_reg_reg((inst), (dreg), (reg), 0x66, 0x0F, 0xE6)
+
+#define amd64_sse_cvttps2dq_reg_reg(inst, dreg, reg) emit_sse_reg_reg((inst), (dreg), (reg), 0xF3, 0x0F, 0x5B)
+
+
 #define amd64_movd_xreg_reg_size(inst,dreg,sreg,size) emit_sse_reg_reg_size((inst), (dreg), (sreg), 0x66, 0x0f, 0x6e, (size))
 
 #define amd64_movd_reg_xreg_size(inst,dreg,sreg,size) emit_sse_reg_reg_size((inst), (sreg), (dreg), 0x66, 0x0f, 0x7e, (size))
@@ -1420,7 +1485,7 @@ typedef union {
 #define amd64_alu_reg8_reg8_size(inst,opc,dreg,reg,is_dreg_h,is_reg_h,size) do { amd64_codegen_pre(inst); amd64_emit_rex ((inst),(size),(dreg),0,(reg)); x86_alu_reg8_reg8((inst),(opc),((dreg)&0x7),((reg)&0x7),(is_dreg_h),(is_reg_h)); amd64_codegen_post(inst); } while (0)
 #define amd64_alu_reg_mem_size(inst,opc,reg,mem,size) do { amd64_codegen_pre(inst); amd64_emit_rex ((inst),(size),0,0,(reg)); x86_alu_reg_mem((inst),(opc),((reg)&0x7),(mem)); amd64_codegen_post(inst); } while (0)
 //#define amd64_alu_reg_membase_size(inst,opc,reg,basereg,disp,size) do { amd64_codegen_pre(inst); amd64_emit_rex ((inst),(size),(reg),0,(basereg)); x86_alu_reg_membase((inst),(opc),((reg)&0x7),((basereg)&0x7),(disp)); amd64_codegen_post(inst); } while (0)
-#define amd64_test_reg_imm_size(inst,reg,imm,size) do { amd64_codegen_pre(inst); amd64_emit_rex ((inst),(size),0,0,(reg)); x86_test_reg_imm((inst),((reg)&0x7),(imm)); amd64_codegen_post(inst); } while (0)
+//#define amd64_test_reg_imm_size(inst,reg,imm,size) do { amd64_codegen_pre(inst); amd64_emit_rex ((inst),(size),0,0,(reg)); x86_test_reg_imm((inst),((reg)&0x7),(imm)); amd64_codegen_post(inst); } while (0)
 #define amd64_test_mem_imm_size(inst,mem,imm,size) do { amd64_codegen_pre(inst); amd64_emit_rex ((inst),(size),0,0,0); x86_test_mem_imm((inst),(mem),(imm)); amd64_codegen_post(inst); } while (0)
 #define amd64_test_membase_imm_size(inst,basereg,disp,imm,size) do { amd64_codegen_pre(inst); amd64_emit_rex ((inst),(size),0,0,(basereg)); x86_test_membase_imm((inst),((basereg)&0x7),(disp),(imm)); amd64_codegen_post(inst); } while (0)
 #define amd64_test_reg_reg_size(inst,dreg,reg,size) do { amd64_codegen_pre(inst); amd64_emit_rex ((inst),(size),(dreg),0,(reg)); x86_test_reg_reg((inst),((dreg)&0x7),((reg)&0x7)); amd64_codegen_post(inst); } while (0)
@@ -1665,7 +1730,7 @@ typedef union {
 #define amd64_alu_reg8_reg8(inst,opc,dreg,reg,is_dreg_h,is_reg_h) amd64_alu_reg8_reg8_size(inst,opc,dreg,reg,is_dreg_h,is_reg_h,8)
 #define amd64_alu_reg_mem(inst,opc,reg,mem) amd64_alu_reg_mem_size(inst,opc,reg,mem,8)
 #define amd64_alu_reg_membase(inst,opc,reg,basereg,disp) amd64_alu_reg_membase_size(inst,opc,reg,basereg,disp,8)
-#define amd64_test_reg_imm(inst,reg,imm) amd64_test_reg_imm_size(inst,reg,imm,8)
+//#define amd64_test_reg_imm(inst,reg,imm) amd64_test_reg_imm_size(inst,reg,imm,8)
 #define amd64_test_mem_imm(inst,mem,imm) amd64_test_mem_imm_size(inst,mem,imm,8)
 #define amd64_test_membase_imm(inst,basereg,disp,imm) amd64_test_membase_imm_size(inst,basereg,disp,imm,8)
 #define amd64_test_reg_reg(inst,dreg,reg) amd64_test_reg_reg_size(inst,dreg,reg,8)