1/* 2 * Copyright (C) 2016 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17/* 18 Art assembly interpreter notes: 19 20 First validate assembly code by implementing ExecuteXXXImpl() style body (doesn't 21 handle invoke, allows higher-level code to create frame & shadow frame. 22 23 Once that's working, support direct entry code & eliminate shadow frame (and 24 excess locals allocation. 25 26 Some (hopefully) temporary ugliness. We'll treat rFP as pointing to the 27 base of the vreg array within the shadow frame. Access the other fields, 28 dex_pc_, method_ and number_of_vregs_ via negative offsets. For now, we'll continue 29 the shadow frame mechanism of double-storing object references - via rFP & 30 number_of_vregs_. 31 32 */ 33 34/* 35x86 ABI general notes: 36 37Caller save set: 38 eax, edx, ecx, st(0)-st(7) 39Callee save set: 40 ebx, esi, edi, ebp 41Return regs: 42 32-bit in eax 43 64-bit in edx:eax (low-order 32 in eax) 44 fp on top of fp stack st(0) 45 46Parameters passed on stack, pushed right-to-left. On entry to target, first 47parm is at 4(%esp). Traditional entry code is: 48 49functEntry: 50 push %ebp # save old frame pointer 51 mov %ebp,%esp # establish new frame pointer 52 sub FrameSize,%esp # Allocate storage for spill, locals & outs 53 54Once past the prologue, arguments are referenced at ((argno + 2)*4)(%ebp) 55 56Stack must be 16-byte aligned to support SSE in native code. 57 58If we're not doing variable stack allocation (alloca), the frame pointer can be 59eliminated and all arg references adjusted to be esp relative. 60*/ 61 62/* 63Mterp and x86 notes: 64 65Some key interpreter variables will be assigned to registers. 66 67 nick reg purpose 68 rPC esi interpreted program counter, used for fetching instructions 69 rFP edi interpreted frame pointer, used for accessing locals and args 70 rINSTw bx first 16-bit code of current instruction 71 rINSTbl bl opcode portion of instruction word 72 rINSTbh bh high byte of inst word, usually contains src/tgt reg names 73 rIBASE edx base of instruction handler table 74 rREFS ebp base of object references in shadow frame. 75 76Notes: 77 o High order 16 bits of ebx must be zero on entry to handler 78 o rPC, rFP, rINSTw/rINSTbl valid on handler entry and exit 79 o eax and ecx are scratch, rINSTw/ebx sometimes scratch 80 81Macros are provided for common operations. Each macro MUST emit only 82one instruction to make instruction-counting easier. They MUST NOT alter 83unspecified registers or condition codes. 84*/ 85 86/* 87 * This is a #include, not a %include, because we want the C pre-processor 88 * to expand the macros into assembler assignment statements. 89 */ 90#include "asm_support.h" 91 92/* 93 * Handle mac compiler specific 94 */ 95#if defined(__APPLE__) 96 #define MACRO_LITERAL(value) $$(value) 97 #define FUNCTION_TYPE(name) 98 #define SIZE(start,end) 99 // Mac OS' symbols have an _ prefix. 100 #define SYMBOL(name) _ ## name 101#else 102 #define MACRO_LITERAL(value) $$value 103 #define FUNCTION_TYPE(name) .type name, @function 104 #define SIZE(start,end) .size start, .-end 105 #define SYMBOL(name) name 106#endif 107 108.macro PUSH _reg 109 pushl \_reg 110 .cfi_adjust_cfa_offset 4 111 .cfi_rel_offset \_reg, 0 112.endm 113 114.macro POP _reg 115 popl \_reg 116 .cfi_adjust_cfa_offset -4 117 .cfi_restore \_reg 118.endm 119 120/* 121 * Instead of holding a pointer to the shadow frame, we keep rFP at the base of the vregs. So, 122 * to access other shadow frame fields, we need to use a backwards offset. Define those here. 123 */ 124#define OFF_FP(a) (a - SHADOWFRAME_VREGS_OFFSET) 125#define OFF_FP_NUMBER_OF_VREGS OFF_FP(SHADOWFRAME_NUMBER_OF_VREGS_OFFSET) 126#define OFF_FP_DEX_PC OFF_FP(SHADOWFRAME_DEX_PC_OFFSET) 127#define OFF_FP_LINK OFF_FP(SHADOWFRAME_LINK_OFFSET) 128#define OFF_FP_METHOD OFF_FP(SHADOWFRAME_METHOD_OFFSET) 129#define OFF_FP_RESULT_REGISTER OFF_FP(SHADOWFRAME_RESULT_REGISTER_OFFSET) 130#define OFF_FP_DEX_PC_PTR OFF_FP(SHADOWFRAME_DEX_PC_PTR_OFFSET) 131#define OFF_FP_CODE_ITEM OFF_FP(SHADOWFRAME_CODE_ITEM_OFFSET) 132#define OFF_FP_COUNTDOWN_OFFSET OFF_FP(SHADOWFRAME_HOTNESS_COUNTDOWN_OFFSET) 133#define OFF_FP_SHADOWFRAME OFF_FP(0) 134 135/* Frame size must be 16-byte aligned. 136 * Remember about 4 bytes for return address + 4 * 4 for spills 137 */ 138#define FRAME_SIZE 28 139 140/* Frame diagram while executing ExecuteMterpImpl, high to low addresses */ 141#define IN_ARG3 (FRAME_SIZE + 16 + 16) 142#define IN_ARG2 (FRAME_SIZE + 16 + 12) 143#define IN_ARG1 (FRAME_SIZE + 16 + 8) 144#define IN_ARG0 (FRAME_SIZE + 16 + 4) 145/* Spill offsets relative to %esp */ 146#define LOCAL0 (FRAME_SIZE - 4) 147#define LOCAL1 (FRAME_SIZE - 8) 148#define LOCAL2 (FRAME_SIZE - 12) 149/* Out Arg offsets, relative to %esp */ 150#define OUT_ARG3 ( 12) 151#define OUT_ARG2 ( 8) 152#define OUT_ARG1 ( 4) 153#define OUT_ARG0 ( 0) /* <- ExecuteMterpImpl esp + 0 */ 154 155/* During bringup, we'll use the shadow frame model instead of rFP */ 156/* single-purpose registers, given names for clarity */ 157#define rSELF IN_ARG0(%esp) 158#define rPC %esi 159#define rFP %edi 160#define rINST %ebx 161#define rINSTw %bx 162#define rINSTbh %bh 163#define rINSTbl %bl 164#define rIBASE %edx 165#define rREFS %ebp 166#define rPROFILE OFF_FP_COUNTDOWN_OFFSET(rFP) 167 168#define MTERP_LOGGING 0 169 170/* 171 * "export" the PC to dex_pc field in the shadow frame, f/b/o future exception objects. Must 172 * be done *before* something throws. 173 * 174 * It's okay to do this more than once. 175 * 176 * NOTE: the fast interpreter keeps track of dex pc as a direct pointer to the mapped 177 * dex byte codes. However, the rest of the runtime expects dex pc to be an instruction 178 * offset into the code_items_[] array. For effiency, we will "export" the 179 * current dex pc as a direct pointer using the EXPORT_PC macro, and rely on GetDexPC 180 * to convert to a dex pc when needed. 181 */ 182.macro EXPORT_PC 183 movl rPC, OFF_FP_DEX_PC_PTR(rFP) 184.endm 185 186/* 187 * Refresh handler table. 188 */ 189.macro REFRESH_IBASE 190 movl rSELF, rIBASE 191 movl THREAD_CURRENT_IBASE_OFFSET(rIBASE), rIBASE 192.endm 193 194/* 195 * Refresh handler table. 196 * IBase handles uses the caller save register so we must restore it after each call. 197 * Also it is used as a result of some 64-bit operations (like imul) and we should 198 * restore it in such cases also. 199 * 200 * TODO: Consider spilling the IBase instead of restoring it from Thread structure. 201 */ 202.macro RESTORE_IBASE 203 movl rSELF, rIBASE 204 movl THREAD_CURRENT_IBASE_OFFSET(rIBASE), rIBASE 205.endm 206 207/* 208 * If rSELF is already loaded then we can use it from known reg. 209 */ 210.macro RESTORE_IBASE_FROM_SELF _reg 211 movl THREAD_CURRENT_IBASE_OFFSET(\_reg), rIBASE 212.endm 213 214/* 215 * Refresh rINST. 216 * At enter to handler rINST does not contain the opcode number. 217 * However some utilities require the full value, so this macro 218 * restores the opcode number. 219 */ 220.macro REFRESH_INST _opnum 221 movb rINSTbl, rINSTbh 222 movb MACRO_LITERAL(\_opnum), rINSTbl 223.endm 224 225/* 226 * Fetch the next instruction from rPC into rINSTw. Does not advance rPC. 227 */ 228.macro FETCH_INST 229 movzwl (rPC), rINST 230.endm 231 232/* 233 * Remove opcode from rINST, compute the address of handler and jump to it. 234 */ 235.macro GOTO_NEXT 236 movzx rINSTbl,%eax 237 movzbl rINSTbh,rINST 238 shll MACRO_LITERAL(${handler_size_bits}), %eax 239 addl rIBASE, %eax 240 jmp *%eax 241.endm 242 243/* 244 * Advance rPC by instruction count. 245 */ 246.macro ADVANCE_PC _count 247 leal 2*\_count(rPC), rPC 248.endm 249 250/* 251 * Advance rPC by instruction count, fetch instruction and jump to handler. 252 */ 253.macro ADVANCE_PC_FETCH_AND_GOTO_NEXT _count 254 ADVANCE_PC \_count 255 FETCH_INST 256 GOTO_NEXT 257.endm 258 259/* 260 * Get/set the 32-bit value from a Dalvik register. 261 */ 262#define VREG_ADDRESS(_vreg) (rFP,_vreg,4) 263#define VREG_HIGH_ADDRESS(_vreg) 4(rFP,_vreg,4) 264#define VREG_REF_ADDRESS(_vreg) (rREFS,_vreg,4) 265#define VREG_REF_HIGH_ADDRESS(_vreg) 4(rREFS,_vreg,4) 266 267.macro GET_VREG _reg _vreg 268 movl (rFP,\_vreg,4), \_reg 269.endm 270 271/* Read wide value to xmm. */ 272.macro GET_WIDE_FP_VREG _reg _vreg 273 movq (rFP,\_vreg,4), \_reg 274.endm 275 276.macro SET_VREG _reg _vreg 277 movl \_reg, (rFP,\_vreg,4) 278 movl MACRO_LITERAL(0), (rREFS,\_vreg,4) 279.endm 280 281/* Write wide value from xmm. xmm is clobbered. */ 282.macro SET_WIDE_FP_VREG _reg _vreg 283 movq \_reg, (rFP,\_vreg,4) 284 pxor \_reg, \_reg 285 movq \_reg, (rREFS,\_vreg,4) 286.endm 287 288.macro SET_VREG_OBJECT _reg _vreg 289 movl \_reg, (rFP,\_vreg,4) 290 movl \_reg, (rREFS,\_vreg,4) 291.endm 292 293.macro GET_VREG_HIGH _reg _vreg 294 movl 4(rFP,\_vreg,4), \_reg 295.endm 296 297.macro SET_VREG_HIGH _reg _vreg 298 movl \_reg, 4(rFP,\_vreg,4) 299 movl MACRO_LITERAL(0), 4(rREFS,\_vreg,4) 300.endm 301 302.macro CLEAR_REF _vreg 303 movl MACRO_LITERAL(0), (rREFS,\_vreg,4) 304.endm 305 306.macro CLEAR_WIDE_REF _vreg 307 movl MACRO_LITERAL(0), (rREFS,\_vreg,4) 308 movl MACRO_LITERAL(0), 4(rREFS,\_vreg,4) 309.endm 310