CompilerTemplateAsm-armv5te.S revision 7a0bcd0de6c4da6499a088a18d1750e51204c2a6
1/*
2 * This file was generated automatically by gen-template.py for 'armv5te'.
3 *
4 * --> DO NOT EDIT <--
5 */
6
7/* File: armv5te/header.S */
8/*
9 * Copyright (C) 2008 The Android Open Source Project
10 *
11 * Licensed under the Apache License, Version 2.0 (the "License");
12 * you may not use this file except in compliance with the License.
13 * You may obtain a copy of the License at
14 *
15 *      http://www.apache.org/licenses/LICENSE-2.0
16 *
17 * Unless required by applicable law or agreed to in writing, software
18 * distributed under the License is distributed on an "AS IS" BASIS,
19 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
20 * See the License for the specific language governing permissions and
21 * limitations under the License.
22 */
23
24#if defined(WITH_JIT)
25
26/*
27 * ARMv5 definitions and declarations.
28 */
29
30/*
31ARM EABI general notes:
32
33r0-r3 hold first 4 args to a method; they are not preserved across method calls
34r4-r8 are available for general use
35r9 is given special treatment in some situations, but not for us
36r10 (sl) seems to be generally available
37r11 (fp) is used by gcc (unless -fomit-frame-pointer is set)
38r12 (ip) is scratch -- not preserved across method calls
39r13 (sp) should be managed carefully in case a signal arrives
40r14 (lr) must be preserved
41r15 (pc) can be tinkered with directly
42
43r0 holds returns of <= 4 bytes
44r0-r1 hold returns of 8 bytes, low word in r0
45
46Callee must save/restore r4+ (except r12) if it modifies them.
47
48Stack is "full descending".  Only the arguments that don't fit in the first 4
49registers are placed on the stack.  "sp" points at the first stacked argument
50(i.e. the 5th arg).
51
52VFP: single-precision results in s0, double-precision results in d0.
53
54In the EABI, "sp" must be 64-bit aligned on entry to a function, and any
5564-bit quantities (long long, double) must be 64-bit aligned.
56*/
57
58/*
59JIT and ARM notes:
60
61The following registers have fixed assignments:
62
63  reg nick      purpose
64  r5  rFP       interpreted frame pointer, used for accessing locals and args
65  r6  rGLUE     MterpGlue pointer
66
67The following registers have fixed assignments in mterp but are scratch
68registers in compiled code
69
70  reg nick      purpose
71  r4  rPC       interpreted program counter, used for fetching instructions
72  r7  rINST     first 16-bit code unit of current instruction
73  r8  rIBASE    interpreted instruction base pointer, used for computed goto
74
75Macros are provided for common operations.  Each macro MUST emit only
76one instruction to make instruction-counting easier.  They MUST NOT alter
77unspecified registers or condition codes.
78*/
79
80/* single-purpose registers, given names for clarity */
81#define rPC     r4
82#define rFP     r5
83#define rGLUE   r6
84#define rINST   r7
85#define rIBASE  r8
86
87/*
88 * Given a frame pointer, find the stack save area.
89 *
90 * In C this is "((StackSaveArea*)(_fp) -1)".
91 */
92#define SAVEAREA_FROM_FP(_reg, _fpreg) \
93    sub     _reg, _fpreg, #sizeofStackSaveArea
94
95#define EXPORT_PC() \
96    str     rPC, [rFP, #(-sizeofStackSaveArea + offStackSaveArea_currentPc)]
97
98/*
99 * This is a #include, not a %include, because we want the C pre-processor
100 * to expand the macros into assembler assignment statements.
101 */
102#include "../../../mterp/common/asm-constants.h"
103
104
105/* File: armv5te/platform.S */
106/*
107 * ===========================================================================
108 *  CPU-version-specific defines and utility
109 * ===========================================================================
110 */
111
112/*
113 * Macro for "MOV LR,PC / LDR PC,xxx", which is not allowed pre-ARMv5.
114 * Jump to subroutine.
115 *
116 * May modify IP and LR.
117 */
118.macro  LDR_PC_LR source
119    mov     lr, pc
120    ldr     pc, \source
121.endm
122
123
124    .global dvmCompilerTemplateStart
125    .type   dvmCompilerTemplateStart, %function
126    .text
127
128dvmCompilerTemplateStart:
129
130/* ------------------------------ */
131    .balign 4
132    .global dvmCompiler_TEMPLATE_CMP_LONG
133dvmCompiler_TEMPLATE_CMP_LONG:
134/* File: armv5te/TEMPLATE_CMP_LONG.S */
135    /*
136     * Compare two 64-bit values.  Puts 0, 1, or -1 into the destination
137     * register based on the results of the comparison.
138     *
139     * We load the full values with LDM, but in practice many values could
140     * be resolved by only looking at the high word.  This could be made
141     * faster or slower by splitting the LDM into a pair of LDRs.
142     *
143     * If we just wanted to set condition flags, we could do this:
144     *  subs    ip, r0, r2
145     *  sbcs    ip, r1, r3
146     *  subeqs  ip, r0, r2
147     * Leaving { <0, 0, >0 } in ip.  However, we have to set it to a specific
148     * integer value, which we can do with 2 conditional mov/mvn instructions
149     * (set 1, set -1; if they're equal we already have 0 in ip), giving
150     * us a constant 5-cycle path plus a branch at the end to the
151     * instruction epilogue code.  The multi-compare approach below needs
152     * 2 or 3 cycles + branch if the high word doesn't match, 6 + branch
153     * in the worst case (the 64-bit values are equal).
154     */
155    /* cmp-long vAA, vBB, vCC */
156    cmp     r1, r3                      @ compare (vBB+1, vCC+1)
157    blt     .LTEMPLATE_CMP_LONG_less            @ signed compare on high part
158    bgt     .LTEMPLATE_CMP_LONG_greater
159    subs    r0, r0, r2                  @ r0<- r0 - r2
160    bxeq     lr
161    bhi     .LTEMPLATE_CMP_LONG_greater         @ unsigned compare on low part
162.LTEMPLATE_CMP_LONG_less:
163    mvn     r0, #0                      @ r0<- -1
164    bx      lr
165.LTEMPLATE_CMP_LONG_greater:
166    mov     r0, #1                      @ r0<- 1
167    bx      lr
168
169
170/* ------------------------------ */
171    .balign 4
172    .global dvmCompiler_TEMPLATE_RETURN
173dvmCompiler_TEMPLATE_RETURN:
174/* File: armv5te/TEMPLATE_RETURN.S */
175    /*
176     * Unwind a frame from the Dalvik stack for compiled OP_RETURN_XXX.
177     * If the stored value in returnAddr
178     * is non-zero, the caller is compiled by the JIT thus return to the
179     * address in the code cache following the invoke instruction. Otherwise
180     * return to the special dvmJitToInterpNoChain entry point.
181     */
182    SAVEAREA_FROM_FP(r0, rFP)           @ r0<- saveArea (old)
183    ldr     r10, [r0, #offStackSaveArea_prevFrame] @ r10<- saveArea->prevFrame
184    ldr     r8, [rGLUE, #offGlue_pSelfSuspendCount] @ r8<- &suspendCount
185    ldr     rPC, [r0, #offStackSaveArea_savedPc] @ rPC<- saveArea->savedPc
186#if !defined(WITH_SELF_VERIFICATION)
187    ldr     r9,  [r0, #offStackSaveArea_returnAddr] @ r9<- chaining cell ret
188#else
189    mov     r9, #0                      @ disable chaining
190#endif
191    ldr     r2, [r10, #(offStackSaveArea_method - sizeofStackSaveArea)]
192                                        @ r2<- method we're returning to
193    ldr     r3, [rGLUE, #offGlue_self]  @ r3<- glue->self
194    cmp     r2, #0                      @ break frame?
195#if !defined(WITH_SELF_VERIFICATION)
196    beq     1f                          @ bail to interpreter
197#else
198    blxeq   lr                          @ punt to interpreter and compare state
199#endif
200    ldr     r1, .LdvmJitToInterpNoChain @ defined in footer.S
201    mov     rFP, r10                    @ publish new FP
202    ldrne   r10, [r2, #offMethod_clazz] @ r10<- method->clazz
203    ldr     r8, [r8]                    @ r8<- suspendCount
204
205    str     r2, [rGLUE, #offGlue_method]@ glue->method = newSave->method
206    ldr     r0, [r10, #offClassObject_pDvmDex] @ r0<- method->clazz->pDvmDex
207    str     rFP, [r3, #offThread_curFrame] @ self->curFrame = fp
208    add     rPC, rPC, #6                @ publish new rPC (advance 6 bytes)
209    str     r0, [rGLUE, #offGlue_methodClassDex]
210    cmp     r8, #0                      @ check the suspendCount
211    movne   r9, #0                      @ clear the chaining cell address
212    str     r9, [r3, #offThread_inJitCodeCache] @ in code cache or not
213    cmp     r9, #0                      @ chaining cell exists?
214    blxne   r9                          @ jump to the chaining cell
215#if defined(EXIT_STATS)
216    mov     r0, #kCallsiteInterpreted
217#endif
218    mov     pc, r1                      @ callsite is interpreted
2191:
220    stmia   rGLUE, {rPC, rFP}           @ SAVE_PC_FP_TO_GLUE()
221    ldr     r2, .LdvmMterpStdBail       @ defined in footer.S
222    mov     r1, #0                      @ changeInterp = false
223    mov     r0, rGLUE                   @ Expecting rGLUE in r0
224    blx     r2                          @ exit the interpreter
225
226/* ------------------------------ */
227    .balign 4
228    .global dvmCompiler_TEMPLATE_INVOKE_METHOD_NO_OPT
229dvmCompiler_TEMPLATE_INVOKE_METHOD_NO_OPT:
230/* File: armv5te/TEMPLATE_INVOKE_METHOD_NO_OPT.S */
231    /*
232     * For polymorphic callsites - setup the Dalvik frame and load Dalvik PC
233     * into rPC then jump to dvmJitToInterpNoChain to dispatch the
234     * runtime-resolved callee.
235     */
236    @ r0 = methodToCall, r1 = returnCell, rPC = dalvikCallsite
237    ldrh    r7, [r0, #offMethod_registersSize]  @ r7<- methodToCall->regsSize
238    ldrh    r2, [r0, #offMethod_outsSize]  @ r2<- methodToCall->outsSize
239    ldr     r9, [rGLUE, #offGlue_interpStackEnd]    @ r9<- interpStackEnd
240    ldr     r8, [rGLUE, #offGlue_pSelfSuspendCount] @ r8<- &suspendCount
241    add     r3, r1, #1  @ Thumb addr is odd
242    SAVEAREA_FROM_FP(r1, rFP)           @ r1<- stack save area
243    sub     r1, r1, r7, lsl #2          @ r1<- newFp (old savearea - regsSize)
244    SAVEAREA_FROM_FP(r10, r1)           @ r10<- stack save area
245    sub     r10, r10, r2, lsl #2        @ r10<- bottom (newsave - outsSize)
246    ldr     r8, [r8]                    @ r3<- suspendCount (int)
247    cmp     r10, r9                     @ bottom < interpStackEnd?
248    bxlt    lr                          @ return to raise stack overflow excep.
249    @ r1 = newFP, r0 = methodToCall, r3 = returnCell, rPC = dalvikCallsite
250    ldr     r9, [r0, #offMethod_clazz]      @ r9<- method->clazz
251    ldr     r10, [r0, #offMethod_accessFlags] @ r10<- methodToCall->accessFlags
252    str     rPC, [rFP, #(offStackSaveArea_currentPc - sizeofStackSaveArea)]
253    str     rPC, [r1, #(offStackSaveArea_savedPc - sizeofStackSaveArea)]
254    ldr     rPC, [r0, #offMethod_insns]     @ rPC<- methodToCall->insns
255
256
257    @ set up newSaveArea
258    str     rFP, [r1, #(offStackSaveArea_prevFrame - sizeofStackSaveArea)]
259    str     r3, [r1, #(offStackSaveArea_returnAddr - sizeofStackSaveArea)]
260    str     r0, [r1, #(offStackSaveArea_method - sizeofStackSaveArea)]
261    cmp     r8, #0                      @ suspendCount != 0
262    bxne    lr                          @ bail to the interpreter
263    tst     r10, #ACC_NATIVE
264#if !defined(WITH_SELF_VERIFICATION)
265    bne     .LinvokeNative
266#else
267    bxne    lr                          @ bail to the interpreter
268#endif
269
270    ldr     r10, .LdvmJitToInterpNoChain
271    ldr     r3, [r9, #offClassObject_pDvmDex] @ r3<- method->clazz->pDvmDex
272    ldr     r2, [rGLUE, #offGlue_self]      @ r2<- glue->self
273
274    @ Update "glue" values for the new method
275    str     r0, [rGLUE, #offGlue_method]    @ glue->method = methodToCall
276    str     r3, [rGLUE, #offGlue_methodClassDex] @ glue->methodClassDex = ...
277    mov     rFP, r1                         @ fp = newFp
278    str     rFP, [r2, #offThread_curFrame]  @ self->curFrame = newFp
279
280    @ Start executing the callee
281#if defined(EXIT_STATS)
282    mov     r0, #kInlineCacheMiss
283#endif
284    mov     pc, r10                         @ dvmJitToInterpNoChain
285
286/* ------------------------------ */
287    .balign 4
288    .global dvmCompiler_TEMPLATE_INVOKE_METHOD_CHAIN
289dvmCompiler_TEMPLATE_INVOKE_METHOD_CHAIN:
290/* File: armv5te/TEMPLATE_INVOKE_METHOD_CHAIN.S */
291    /*
292     * For monomorphic callsite, setup the Dalvik frame and return to the
293     * Thumb code through the link register to transfer control to the callee
294     * method through a dedicated chaining cell.
295     */
296    @ r0 = methodToCall, r1 = returnCell, rPC = dalvikCallsite
297    @ methodToCall is guaranteed to be non-native
298.LinvokeChain:
299    ldrh    r7, [r0, #offMethod_registersSize]  @ r7<- methodToCall->regsSize
300    ldrh    r2, [r0, #offMethod_outsSize]  @ r2<- methodToCall->outsSize
301    ldr     r9, [rGLUE, #offGlue_interpStackEnd]    @ r9<- interpStackEnd
302    ldr     r8, [rGLUE, #offGlue_pSelfSuspendCount] @ r8<- &suspendCount
303    add     r3, r1, #1  @ Thumb addr is odd
304    SAVEAREA_FROM_FP(r1, rFP)           @ r1<- stack save area
305    sub     r1, r1, r7, lsl #2          @ r1<- newFp (old savearea - regsSize)
306    SAVEAREA_FROM_FP(r10, r1)           @ r10<- stack save area
307    add     r12, lr, #2                 @ setup the punt-to-interp address
308    sub     r10, r10, r2, lsl #2        @ r10<- bottom (newsave - outsSize)
309    ldr     r8, [r8]                    @ r3<- suspendCount (int)
310    cmp     r10, r9                     @ bottom < interpStackEnd?
311    bxlt    r12                         @ return to raise stack overflow excep.
312    @ r1 = newFP, r0 = methodToCall, r3 = returnCell, rPC = dalvikCallsite
313    ldr     r9, [r0, #offMethod_clazz]      @ r9<- method->clazz
314    str     rPC, [rFP, #(offStackSaveArea_currentPc - sizeofStackSaveArea)]
315    str     rPC, [r1, #(offStackSaveArea_savedPc - sizeofStackSaveArea)]
316    ldr     rPC, [r0, #offMethod_insns]     @ rPC<- methodToCall->insns
317
318
319    @ set up newSaveArea
320    str     rFP, [r1, #(offStackSaveArea_prevFrame - sizeofStackSaveArea)]
321    str     r3, [r1, #(offStackSaveArea_returnAddr - sizeofStackSaveArea)]
322    str     r0, [r1, #(offStackSaveArea_method - sizeofStackSaveArea)]
323    cmp     r8, #0                      @ suspendCount != 0
324    bxne    r12                         @ bail to the interpreter
325
326    ldr     r3, [r9, #offClassObject_pDvmDex] @ r3<- method->clazz->pDvmDex
327    ldr     r2, [rGLUE, #offGlue_self]      @ r2<- glue->self
328
329    @ Update "glue" values for the new method
330    str     r0, [rGLUE, #offGlue_method]    @ glue->method = methodToCall
331    str     r3, [rGLUE, #offGlue_methodClassDex] @ glue->methodClassDex = ...
332    mov     rFP, r1                         @ fp = newFp
333    str     rFP, [r2, #offThread_curFrame]  @ self->curFrame = newFp
334
335    bx      lr                              @ return to the callee-chaining cell
336
337
338
339/* ------------------------------ */
340    .balign 4
341    .global dvmCompiler_TEMPLATE_INVOKE_METHOD_PREDICTED_CHAIN
342dvmCompiler_TEMPLATE_INVOKE_METHOD_PREDICTED_CHAIN:
343/* File: armv5te/TEMPLATE_INVOKE_METHOD_PREDICTED_CHAIN.S */
344    /*
345     * For polymorphic callsite, check whether the cached class pointer matches
346     * the current one. If so setup the Dalvik frame and return to the
347     * Thumb code through the link register to transfer control to the callee
348     * method through a dedicated chaining cell.
349     *
350     * The predicted chaining cell is declared in ArmLIR.h with the
351     * following layout:
352     *
353     *  typedef struct PredictedChainingCell {
354     *      u4 branch;
355     *      const ClassObject *clazz;
356     *      const Method *method;
357     *      u4 counter;
358     *  } PredictedChainingCell;
359     *
360     * Upon returning to the callsite:
361     *    - lr  : to branch to the chaining cell
362     *    - lr+2: to punt to the interpreter
363     *    - lr+4: to fully resolve the callee and may rechain.
364     *            r3 <- class
365     *            r9 <- counter
366     */
367    @ r0 = this, r1 = returnCell, r2 = predictedChainCell, rPC = dalvikCallsite
368    ldr     r3, [r0, #offObject_clazz]  @ r3 <- this->class
369    ldr     r8, [r2, #4]    @ r8 <- predictedChainCell->clazz
370    ldr     r0, [r2, #8]    @ r0 <- predictedChainCell->method
371    ldr     r9, [r2, #12]   @ r9 <- predictedChainCell->counter
372    cmp     r3, r8          @ predicted class == actual class?
373    beq     .LinvokeChain   @ predicted chain is valid
374    ldr     r7, [r3, #offClassObject_vtable] @ r7 <- this->class->vtable
375    sub     r1, r9, #1      @ count--
376    str     r1, [r2, #12]   @ write back to PredictedChainingCell->counter
377    add     lr, lr, #4      @ return to fully-resolve landing pad
378    /*
379     * r1 <- count
380     * r2 <- &predictedChainCell
381     * r3 <- this->class
382     * r4 <- dPC
383     * r7 <- this->class->vtable
384     */
385    bx      lr
386
387/* ------------------------------ */
388    .balign 4
389    .global dvmCompiler_TEMPLATE_INVOKE_METHOD_NATIVE
390dvmCompiler_TEMPLATE_INVOKE_METHOD_NATIVE:
391/* File: armv5te/TEMPLATE_INVOKE_METHOD_NATIVE.S */
392    @ r0 = methodToCall, r1 = returnCell, rPC = dalvikCallsite
393    ldrh    r7, [r0, #offMethod_registersSize]  @ r7<- methodToCall->regsSize
394    ldr     r9, [rGLUE, #offGlue_interpStackEnd]    @ r9<- interpStackEnd
395    ldr     r8, [rGLUE, #offGlue_pSelfSuspendCount] @ r8<- &suspendCount
396    add     r3, r1, #1  @ Thumb addr is odd
397    SAVEAREA_FROM_FP(r1, rFP)           @ r1<- stack save area
398    sub     r1, r1, r7, lsl #2          @ r1<- newFp (old savearea - regsSize)
399    SAVEAREA_FROM_FP(r10, r1)           @ r10<- stack save area
400    ldr     r8, [r8]                    @ r3<- suspendCount (int)
401    cmp     r10, r9                     @ bottom < interpStackEnd?
402    bxlt    lr                          @ return to raise stack overflow excep.
403    @ r1 = newFP, r0 = methodToCall, r3 = returnCell, rPC = dalvikCallsite
404    str     rPC, [rFP, #(offStackSaveArea_currentPc - sizeofStackSaveArea)]
405    str     rPC, [r1, #(offStackSaveArea_savedPc - sizeofStackSaveArea)]
406    ldr     rPC, [r0, #offMethod_insns]     @ rPC<- methodToCall->insns
407
408
409    @ set up newSaveArea
410    str     rFP, [r1, #(offStackSaveArea_prevFrame - sizeofStackSaveArea)]
411    str     r3, [r1, #(offStackSaveArea_returnAddr - sizeofStackSaveArea)]
412    ldr     r3, [rGLUE, #offGlue_self]      @ r3<- glue->self
413    str     r0, [r1, #(offStackSaveArea_method - sizeofStackSaveArea)]
414    cmp     r8, #0                      @ suspendCount != 0
415    ldr     r8, [r0, #offMethod_nativeFunc] @ r8<- method->nativeFunc
416#if !defined(WITH_SELF_VERIFICATION)
417    bxne    lr                          @ bail to the interpreter
418#else
419    bx      lr                          @ bail to interpreter unconditionally
420#endif
421
422    @ go ahead and transfer control to the native code
423    ldr     r9, [r3, #offThread_jniLocal_topCookie] @ r9<- thread->localRef->...
424    mov     r2, #0
425    str     r1, [r3, #offThread_curFrame]   @ self->curFrame = newFp
426    str     r2, [r3, #offThread_inJitCodeCache] @ not in the jit code cache
427    str     r9, [r1, #(offStackSaveArea_localRefCookie - sizeofStackSaveArea)]
428                                        @ newFp->localRefCookie=top
429    mov     r9, r3                      @ r9<- glue->self (preserve)
430    SAVEAREA_FROM_FP(r10, r1)           @ r10<- new stack save area
431
432    mov     r2, r0                      @ r2<- methodToCall
433    mov     r0, r1                      @ r0<- newFP
434    add     r1, rGLUE, #offGlue_retval  @ r1<- &retval
435
436    blx     r8                          @ off to the native code
437
438    @ native return; r9=self, r10=newSaveArea
439    @ equivalent to dvmPopJniLocals
440    ldr     r2, [r10, #offStackSaveArea_returnAddr] @ r2 = chaining cell ret
441    ldr     r0, [r10, #offStackSaveArea_localRefCookie] @ r0<- saved->top
442    ldr     r1, [r9, #offThread_exception] @ check for exception
443    str     rFP, [r9, #offThread_curFrame]  @ self->curFrame = fp
444    cmp     r1, #0                      @ null?
445    str     r0, [r9, #offThread_jniLocal_topCookie] @ new top <- old top
446    ldr     r0, [rFP, #(offStackSaveArea_currentPc - sizeofStackSaveArea)]
447
448    @ r0 = dalvikCallsitePC
449    bne     .LhandleException           @ no, handle exception
450
451    str     r2, [r9, #offThread_inJitCodeCache] @ set the mode properly
452    cmp     r2, #0                      @ return chaining cell still exists?
453    bxne    r2                          @ yes - go ahead
454
455    @ continue executing the next instruction through the interpreter
456    ldr     r1, .LdvmJitToInterpNoChain @ defined in footer.S
457    add     rPC, r0, #6                 @ reconstruct new rPC (advance 6 bytes)
458#if defined(EXIT_STATS)
459    mov     r0, #kCallsiteInterpreted
460#endif
461    mov     pc, r1
462
463
464
465
466/* ------------------------------ */
467    .balign 4
468    .global dvmCompiler_TEMPLATE_CMPG_DOUBLE
469dvmCompiler_TEMPLATE_CMPG_DOUBLE:
470/* File: armv5te/TEMPLATE_CMPG_DOUBLE.S */
471/* File: armv5te/TEMPLATE_CMPL_DOUBLE.S */
472    /*
473     * For the JIT: incoming arguments in r0-r1, r2-r3
474     *              result in r0
475     *
476     * Compare two floating-point values.  Puts 0, 1, or -1 into the
477     * destination register based on the results of the comparison.
478     *
479     * Provide a "naninst" instruction that puts 1 or -1 into r1 depending
480     * on what value we'd like to return when one of the operands is NaN.
481     *
482     * See OP_CMPL_FLOAT for an explanation.
483     *
484     * For: cmpl-double, cmpg-double
485     */
486    /* op vAA, vBB, vCC */
487    push    {r0-r3}                     @ save operands
488    mov     r11, lr                     @ save return address
489    LDR_PC_LR ".L__aeabi_cdcmple"       @ PIC way of "bl __aeabi_cdcmple"
490    bhi     .LTEMPLATE_CMPG_DOUBLE_gt_or_nan       @ C set and Z clear, disambiguate
491    mvncc   r0, #0                      @ (less than) r1<- -1
492    moveq   r0, #0                      @ (equal) r1<- 0, trumps less than
493    add     sp, #16                     @ drop unused operands
494    bx      r11
495
496    @ Test for NaN with a second comparison.  EABI forbids testing bit
497    @ patterns, and we can't represent 0x7fc00000 in immediate form, so
498    @ make the library call.
499.LTEMPLATE_CMPG_DOUBLE_gt_or_nan:
500    pop     {r2-r3}                     @ restore operands in reverse order
501    pop     {r0-r1}                     @ restore operands in reverse order
502    LDR_PC_LR ".L__aeabi_cdcmple"       @ r0<- Z set if eq, C clear if <
503    movcc   r0, #1                      @ (greater than) r1<- 1
504    bxcc    r11
505    mov     r0, #1                            @ r1<- 1 or -1 for NaN
506    bx      r11
507
508
509
510/* ------------------------------ */
511    .balign 4
512    .global dvmCompiler_TEMPLATE_CMPL_DOUBLE
513dvmCompiler_TEMPLATE_CMPL_DOUBLE:
514/* File: armv5te/TEMPLATE_CMPL_DOUBLE.S */
515    /*
516     * For the JIT: incoming arguments in r0-r1, r2-r3
517     *              result in r0
518     *
519     * Compare two floating-point values.  Puts 0, 1, or -1 into the
520     * destination register based on the results of the comparison.
521     *
522     * Provide a "naninst" instruction that puts 1 or -1 into r1 depending
523     * on what value we'd like to return when one of the operands is NaN.
524     *
525     * See OP_CMPL_FLOAT for an explanation.
526     *
527     * For: cmpl-double, cmpg-double
528     */
529    /* op vAA, vBB, vCC */
530    push    {r0-r3}                     @ save operands
531    mov     r11, lr                     @ save return address
532    LDR_PC_LR ".L__aeabi_cdcmple"       @ PIC way of "bl __aeabi_cdcmple"
533    bhi     .LTEMPLATE_CMPL_DOUBLE_gt_or_nan       @ C set and Z clear, disambiguate
534    mvncc   r0, #0                      @ (less than) r1<- -1
535    moveq   r0, #0                      @ (equal) r1<- 0, trumps less than
536    add     sp, #16                     @ drop unused operands
537    bx      r11
538
539    @ Test for NaN with a second comparison.  EABI forbids testing bit
540    @ patterns, and we can't represent 0x7fc00000 in immediate form, so
541    @ make the library call.
542.LTEMPLATE_CMPL_DOUBLE_gt_or_nan:
543    pop     {r2-r3}                     @ restore operands in reverse order
544    pop     {r0-r1}                     @ restore operands in reverse order
545    LDR_PC_LR ".L__aeabi_cdcmple"       @ r0<- Z set if eq, C clear if <
546    movcc   r0, #1                      @ (greater than) r1<- 1
547    bxcc    r11
548    mvn     r0, #0                            @ r1<- 1 or -1 for NaN
549    bx      r11
550
551
552/* ------------------------------ */
553    .balign 4
554    .global dvmCompiler_TEMPLATE_CMPG_FLOAT
555dvmCompiler_TEMPLATE_CMPG_FLOAT:
556/* File: armv5te/TEMPLATE_CMPG_FLOAT.S */
557/* File: armv5te/TEMPLATE_CMPL_FLOAT.S */
558    /*
559     * For the JIT: incoming arguments in r0-r1, r2-r3
560     *              result in r0
561     *
562     * Compare two floating-point values.  Puts 0, 1, or -1 into the
563     * destination register based on the results of the comparison.
564     *
565     * Provide a "naninst" instruction that puts 1 or -1 into r1 depending
566     * on what value we'd like to return when one of the operands is NaN.
567     *
568     * The operation we're implementing is:
569     *   if (x == y)
570     *     return 0;
571     *   else if (x < y)
572     *     return -1;
573     *   else if (x > y)
574     *     return 1;
575     *   else
576     *     return {-1,1};  // one or both operands was NaN
577     *
578     * The straightforward implementation requires 3 calls to functions
579     * that return a result in r0.  We can do it with two calls if our
580     * EABI library supports __aeabi_cfcmple (only one if we want to check
581     * for NaN directly):
582     *   check x <= y
583     *     if <, return -1
584     *     if ==, return 0
585     *   check y <= x
586     *     if <, return 1
587     *   return {-1,1}
588     *
589     * for: cmpl-float, cmpg-float
590     */
591    /* op vAA, vBB, vCC */
592    mov     r9, r0                      @ Save copies - we may need to redo
593    mov     r10, r1
594    mov     r11, lr                     @ save return address
595    LDR_PC_LR ".L__aeabi_cfcmple"       @ cmp <=: C clear if <, Z set if eq
596    bhi     .LTEMPLATE_CMPG_FLOAT_gt_or_nan       @ C set and Z clear, disambiguate
597    mvncc   r0, #0                      @ (less than) r0<- -1
598    moveq   r0, #0                      @ (equal) r0<- 0, trumps less than
599    bx      r11
600    @ Test for NaN with a second comparison.  EABI forbids testing bit
601    @ patterns, and we can't represent 0x7fc00000 in immediate form, so
602    @ make the library call.
603.LTEMPLATE_CMPG_FLOAT_gt_or_nan:
604    mov     r0, r10                     @ restore in reverse order
605    mov     r1, r9
606    LDR_PC_LR ".L__aeabi_cfcmple"       @ r0<- Z set if eq, C clear if <
607    movcc   r0, #1                      @ (greater than) r1<- 1
608    bxcc    r11
609    mov     r0, #1                            @ r1<- 1 or -1 for NaN
610    bx      r11
611
612
613
614
615/* ------------------------------ */
616    .balign 4
617    .global dvmCompiler_TEMPLATE_CMPL_FLOAT
618dvmCompiler_TEMPLATE_CMPL_FLOAT:
619/* File: armv5te/TEMPLATE_CMPL_FLOAT.S */
620    /*
621     * For the JIT: incoming arguments in r0-r1, r2-r3
622     *              result in r0
623     *
624     * Compare two floating-point values.  Puts 0, 1, or -1 into the
625     * destination register based on the results of the comparison.
626     *
627     * Provide a "naninst" instruction that puts 1 or -1 into r1 depending
628     * on what value we'd like to return when one of the operands is NaN.
629     *
630     * The operation we're implementing is:
631     *   if (x == y)
632     *     return 0;
633     *   else if (x < y)
634     *     return -1;
635     *   else if (x > y)
636     *     return 1;
637     *   else
638     *     return {-1,1};  // one or both operands was NaN
639     *
640     * The straightforward implementation requires 3 calls to functions
641     * that return a result in r0.  We can do it with two calls if our
642     * EABI library supports __aeabi_cfcmple (only one if we want to check
643     * for NaN directly):
644     *   check x <= y
645     *     if <, return -1
646     *     if ==, return 0
647     *   check y <= x
648     *     if <, return 1
649     *   return {-1,1}
650     *
651     * for: cmpl-float, cmpg-float
652     */
653    /* op vAA, vBB, vCC */
654    mov     r9, r0                      @ Save copies - we may need to redo
655    mov     r10, r1
656    mov     r11, lr                     @ save return address
657    LDR_PC_LR ".L__aeabi_cfcmple"       @ cmp <=: C clear if <, Z set if eq
658    bhi     .LTEMPLATE_CMPL_FLOAT_gt_or_nan       @ C set and Z clear, disambiguate
659    mvncc   r0, #0                      @ (less than) r0<- -1
660    moveq   r0, #0                      @ (equal) r0<- 0, trumps less than
661    bx      r11
662    @ Test for NaN with a second comparison.  EABI forbids testing bit
663    @ patterns, and we can't represent 0x7fc00000 in immediate form, so
664    @ make the library call.
665.LTEMPLATE_CMPL_FLOAT_gt_or_nan:
666    mov     r0, r10                     @ restore in reverse order
667    mov     r1, r9
668    LDR_PC_LR ".L__aeabi_cfcmple"       @ r0<- Z set if eq, C clear if <
669    movcc   r0, #1                      @ (greater than) r1<- 1
670    bxcc    r11
671    mvn     r0, #0                            @ r1<- 1 or -1 for NaN
672    bx      r11
673
674
675
676/* ------------------------------ */
677    .balign 4
678    .global dvmCompiler_TEMPLATE_MUL_LONG
679dvmCompiler_TEMPLATE_MUL_LONG:
680/* File: armv5te/TEMPLATE_MUL_LONG.S */
681    /*
682     * Signed 64-bit integer multiply.
683     *
684     * For JIT: op1 in r0/r1, op2 in r2/r3, return in r0/r1
685     *
686     * Consider WXxYZ (r1r0 x r3r2) with a long multiply:
687     *        WX
688     *      x YZ
689     *  --------
690     *     ZW ZX
691     *  YW YX
692     *
693     * The low word of the result holds ZX, the high word holds
694     * (ZW+YX) + (the high overflow from ZX).  YW doesn't matter because
695     * it doesn't fit in the low 64 bits.
696     *
697     * Unlike most ARM math operations, multiply instructions have
698     * restrictions on using the same register more than once (Rd and Rm
699     * cannot be the same).
700     */
701    /* mul-long vAA, vBB, vCC */
702    mul     ip, r2, r1                  @  ip<- ZxW
703    umull   r9, r10, r2, r0             @  r9/r10 <- ZxX
704    mla     r2, r0, r3, ip              @  r2<- YxX + (ZxW)
705    add     r10, r2, r10                @  r10<- r10 + low(ZxW + (YxX))
706    mov     r0,r9
707    mov     r1,r10
708    bx      lr
709
710/* ------------------------------ */
711    .balign 4
712    .global dvmCompiler_TEMPLATE_SHL_LONG
713dvmCompiler_TEMPLATE_SHL_LONG:
714/* File: armv5te/TEMPLATE_SHL_LONG.S */
715    /*
716     * Long integer shift.  This is different from the generic 32/64-bit
717     * binary operations because vAA/vBB are 64-bit but vCC (the shift
718     * distance) is 32-bit.  Also, Dalvik requires us to ignore all but the low
719     * 6 bits.
720     */
721    /* shl-long vAA, vBB, vCC */
722    and     r2, r2, #63                 @ r2<- r2 & 0x3f
723    mov     r1, r1, asl r2              @  r1<- r1 << r2
724    rsb     r3, r2, #32                 @  r3<- 32 - r2
725    orr     r1, r1, r0, lsr r3          @  r1<- r1 | (r0 << (32-r2))
726    subs    ip, r2, #32                 @  ip<- r2 - 32
727    movpl   r1, r0, asl ip              @  if r2 >= 32, r1<- r0 << (r2-32)
728    mov     r0, r0, asl r2              @  r0<- r0 << r2
729    bx      lr
730
731/* ------------------------------ */
732    .balign 4
733    .global dvmCompiler_TEMPLATE_SHR_LONG
734dvmCompiler_TEMPLATE_SHR_LONG:
735/* File: armv5te/TEMPLATE_SHR_LONG.S */
736    /*
737     * Long integer shift.  This is different from the generic 32/64-bit
738     * binary operations because vAA/vBB are 64-bit but vCC (the shift
739     * distance) is 32-bit.  Also, Dalvik requires us to ignore all but the low
740     * 6 bits.
741     */
742    /* shr-long vAA, vBB, vCC */
743    and     r2, r2, #63                 @ r0<- r0 & 0x3f
744    mov     r0, r0, lsr r2              @  r0<- r2 >> r2
745    rsb     r3, r2, #32                 @  r3<- 32 - r2
746    orr     r0, r0, r1, asl r3          @  r0<- r0 | (r1 << (32-r2))
747    subs    ip, r2, #32                 @  ip<- r2 - 32
748    movpl   r0, r1, asr ip              @  if r2 >= 32, r0<-r1 >> (r2-32)
749    mov     r1, r1, asr r2              @  r1<- r1 >> r2
750    bx      lr
751
752
753/* ------------------------------ */
754    .balign 4
755    .global dvmCompiler_TEMPLATE_USHR_LONG
756dvmCompiler_TEMPLATE_USHR_LONG:
757/* File: armv5te/TEMPLATE_USHR_LONG.S */
758    /*
759     * Long integer shift.  This is different from the generic 32/64-bit
760     * binary operations because vAA/vBB are 64-bit but vCC (the shift
761     * distance) is 32-bit.  Also, Dalvik requires us to ignore all but the low
762     * 6 bits.
763     */
764    /* ushr-long vAA, vBB, vCC */
765    and     r2, r2, #63                 @ r0<- r0 & 0x3f
766    mov     r0, r0, lsr r2              @  r0<- r2 >> r2
767    rsb     r3, r2, #32                 @  r3<- 32 - r2
768    orr     r0, r0, r1, asl r3          @  r0<- r0 | (r1 << (32-r2))
769    subs    ip, r2, #32                 @  ip<- r2 - 32
770    movpl   r0, r1, lsr ip              @  if r2 >= 32, r0<-r1 >>> (r2-32)
771    mov     r1, r1, lsr r2              @  r1<- r1 >>> r2
772    bx      lr
773
774
775/* ------------------------------ */
776    .balign 4
777    .global dvmCompiler_TEMPLATE_THROW_EXCEPTION_COMMON
778dvmCompiler_TEMPLATE_THROW_EXCEPTION_COMMON:
779/* File: armv5te/TEMPLATE_THROW_EXCEPTION_COMMON.S */
780    /*
781     * Throw an exception from JIT'ed code.
782     * On entry:
783     *    r0    Dalvik PC that raises the exception
784     */
785    b       .LhandleException
786
787/* ------------------------------ */
788    .balign 4
789    .global dvmCompiler_TEMPLATE_SAVE_STATE
790dvmCompiler_TEMPLATE_SAVE_STATE:
791/* File: armv5te/TEMPLATE_SAVE_STATE.S */
792    /*
793     * This handler performs a register save for selfVerification mode.
794     * On entry:
795     *    Top of stack + 4: r7 value to save
796     *    Top of stack + 0: r0 value to save
797     *    r0 - offset from rGLUE to the beginning of the heapArgSpace record
798     *    r7 - the value of regMap
799     *
800     * The handler must save regMap, r0-r12 and then return with r0-r12
801     * with their original values (note that this means r0 and r7 must take
802     * the values on the stack - not the ones in those registers on entry.
803     * Finally, the two registers previously pushed must be popped.
804     */
805    add     r0, r0, rGLUE               @ pointer to heapArgSpace
806    stmia   r0!, {r7}                   @ save regMap
807    ldr     r7, [r13, #0]               @ recover r0 value
808    stmia   r0!, {r7}                   @ save r0
809    ldr     r7, [r13, #4]               @ recover r7 value
810    stmia   r0!, {r1-r12}
811    pop     {r0, r7}                    @ recover r0, r7
812    bx      lr
813
814/* ------------------------------ */
815    .balign 4
816    .global dvmCompiler_TEMPLATE_RESTORE_STATE
817dvmCompiler_TEMPLATE_RESTORE_STATE:
818/* File: armv5te/TEMPLATE_RESTORE_STATE.S */
819    /*
820     * This handler restores state following a selfVerification memory access.
821     * On entry:
822     *    r0 - offset from rGLUE to the 1st element of the coreRegs save array.
823     */
824    add     r0, r0, rGLUE               @ pointer to heapArgSpace.coreRegs[0]
825    ldmia   r0, {r0-r12}
826    bx      lr
827
828/* ------------------------------ */
829    .balign 4
830    .global dvmCompiler_TEMPLATE_STRING_COMPARETO
831dvmCompiler_TEMPLATE_STRING_COMPARETO:
832/* File: armv5te/TEMPLATE_STRING_COMPARETO.S */
833    /*
834     * String's compareTo.
835     *
836     * Requires r0/r1 to have been previously checked for null.  Will
837     * return negative if this's string is < comp, 0 if they are the
838     * same and positive if >.
839     *
840     * IMPORTANT NOTE:
841     *
842     * This code relies on hard-coded offsets for string objects, and must be
843     * kept in sync with definitions in UtfString.h.  See asm-constants.h
844     *
845     * On entry:
846     *    r0:   this object pointer
847     *    r1:   comp object pointer
848     *
849     */
850
851    mov    r2, r0         @ this to r2, opening up r0 for return value
852    subs   r0, r2, r1     @ Same?
853    bxeq   lr
854
855    ldr    r4, [r2, #STRING_FIELDOFF_OFFSET]
856    ldr    r9, [r1, #STRING_FIELDOFF_OFFSET]
857    ldr    r7, [r2, #STRING_FIELDOFF_COUNT]
858    ldr    r10, [r1, #STRING_FIELDOFF_COUNT]
859    ldr    r2, [r2, #STRING_FIELDOFF_VALUE]
860    ldr    r1, [r1, #STRING_FIELDOFF_VALUE]
861
862    /*
863     * At this point, we have:
864     *    value:  r2/r1
865     *    offset: r4/r9
866     *    count:  r7/r10
867     * We're going to compute
868     *    r11 <- countDiff
869     *    r10 <- minCount
870     */
871     subs  r11, r7, r10
872     movls r10, r7
873
874     /* Now, build pointers to the string data */
875     add   r2, r2, r4, lsl #1
876     add   r1, r1, r9, lsl #1
877     /*
878      * Note: data pointers point to previous element so we can use pre-index
879      * mode with base writeback.
880      */
881     add   r2, #16-2   @ offset to contents[-1]
882     add   r1, #16-2   @ offset to contents[-1]
883
884     /*
885      * At this point we have:
886      *   r2: *this string data
887      *   r1: *comp string data
888      *   r10: iteration count for comparison
889      *   r11: value to return if the first part of the string is equal
890      *   r0: reserved for result
891      *   r3, r4, r7, r8, r9, r12 available for loading string data
892      */
893
894    subs  r10, #2
895    blt   do_remainder2
896
897      /*
898       * Unroll the first two checks so we can quickly catch early mismatch
899       * on long strings (but preserve incoming alignment)
900       */
901
902    ldrh  r3, [r2, #2]!
903    ldrh  r4, [r1, #2]!
904    ldrh  r7, [r2, #2]!
905    ldrh  r8, [r1, #2]!
906    subs  r0, r3, r4
907    subeqs  r0, r7, r8
908    bxne  lr
909    cmp   r10, #28
910    bgt   do_memcmp16
911    subs  r10, #3
912    blt   do_remainder
913
914loopback_triple:
915    ldrh  r3, [r2, #2]!
916    ldrh  r4, [r1, #2]!
917    ldrh  r7, [r2, #2]!
918    ldrh  r8, [r1, #2]!
919    ldrh  r9, [r2, #2]!
920    ldrh  r12,[r1, #2]!
921    subs  r0, r3, r4
922    subeqs  r0, r7, r8
923    subeqs  r0, r9, r12
924    bxne  lr
925    subs  r10, #3
926    bge   loopback_triple
927
928do_remainder:
929    adds  r10, #3
930    beq   returnDiff
931
932loopback_single:
933    ldrh  r3, [r2, #2]!
934    ldrh  r4, [r1, #2]!
935    subs  r0, r3, r4
936    bxne  lr
937    subs  r10, #1
938    bne     loopback_single
939
940returnDiff:
941    mov   r0, r11
942    bx    lr
943
944do_remainder2:
945    adds  r10, #2
946    bne   loopback_single
947    mov   r0, r11
948    bx    lr
949
950    /* Long string case */
951do_memcmp16:
952    mov   r4, lr
953    ldr   lr, .Lmemcmp16
954    mov   r7, r11
955    add   r0, r2, #2
956    add   r1, r1, #2
957    mov   r2, r10
958    blx   lr
959    cmp   r0, #0
960    bxne  r4
961    mov   r0, r7
962    bx    r4
963
964.Lmemcmp16:
965    .word __memcmp16
966
967
968/* ------------------------------ */
969    .balign 4
970    .global dvmCompiler_TEMPLATE_STRING_INDEXOF
971dvmCompiler_TEMPLATE_STRING_INDEXOF:
972/* File: armv5te/TEMPLATE_STRING_INDEXOF.S */
973    /*
974     * String's indexOf.
975     *
976     * Requires r0 to have been previously checked for null.  Will
977     * return index of match of r1 in r0.
978     *
979     * IMPORTANT NOTE:
980     *
981     * This code relies on hard-coded offsets for string objects, and must be
982     * kept in sync wth definitions in UtfString.h  See asm-constants.h
983     *
984     * On entry:
985     *    r0:   string object pointer
986     *    r1:   char to match
987     *    r2:   Starting offset in string data
988     */
989
990    ldr    r7, [r0, #STRING_FIELDOFF_OFFSET]
991    ldr    r8, [r0, #STRING_FIELDOFF_COUNT]
992    ldr    r0, [r0, #STRING_FIELDOFF_VALUE]
993
994    /*
995     * At this point, we have:
996     *    r0: object pointer
997     *    r1: char to match
998     *    r2: starting offset
999     *    r7: offset
1000     *    r8: string length
1001     */
1002
1003     /* Build pointer to start of string data */
1004     add   r0, #16
1005     add   r0, r0, r7, lsl #1
1006
1007     /* Save a copy of starting data in r7 */
1008     mov   r7, r0
1009
1010     /* Clamp start to [0..count] */
1011     cmp   r2, #0
1012     movlt r2, #0
1013     cmp   r2, r8
1014     movgt r2, r8
1015
1016     /* Build pointer to start of data to compare and pre-bias */
1017     add   r0, r0, r2, lsl #1
1018     sub   r0, #2
1019
1020     /* Compute iteration count */
1021     sub   r8, r2
1022
1023     /*
1024      * At this point we have:
1025      *   r0: start of data to test
1026      *   r1: chat to compare
1027      *   r8: iteration count
1028      *   r7: original start of string
1029      *   r3, r4, r9, r10, r11, r12 available for loading string data
1030      */
1031
1032    subs  r8, #4
1033    blt   indexof_remainder
1034
1035indexof_loop4:
1036    ldrh  r3, [r0, #2]!
1037    ldrh  r4, [r0, #2]!
1038    ldrh  r10, [r0, #2]!
1039    ldrh  r11, [r0, #2]!
1040    cmp   r3, r1
1041    beq   match_0
1042    cmp   r4, r1
1043    beq   match_1
1044    cmp   r10, r1
1045    beq   match_2
1046    cmp   r11, r1
1047    beq   match_3
1048    subs  r8, #4
1049    bge   indexof_loop4
1050
1051indexof_remainder:
1052    adds    r8, #4
1053    beq     indexof_nomatch
1054
1055indexof_loop1:
1056    ldrh  r3, [r0, #2]!
1057    cmp   r3, r1
1058    beq   match_3
1059    subs  r8, #1
1060    bne   indexof_loop1
1061
1062indexof_nomatch:
1063    mov   r0, #-1
1064    bx    lr
1065
1066match_0:
1067    sub   r0, #6
1068    sub   r0, r7
1069    asr   r0, r0, #1
1070    bx    lr
1071match_1:
1072    sub   r0, #4
1073    sub   r0, r7
1074    asr   r0, r0, #1
1075    bx    lr
1076match_2:
1077    sub   r0, #2
1078    sub   r0, r7
1079    asr   r0, r0, #1
1080    bx    lr
1081match_3:
1082    sub   r0, r7
1083    asr   r0, r0, #1
1084    bx    lr
1085
1086
1087/* ------------------------------ */
1088    .balign 4
1089    .global dvmCompiler_TEMPLATE_INTERPRET
1090dvmCompiler_TEMPLATE_INTERPRET:
1091/* File: armv5te/TEMPLATE_INTERPRET.S */
1092    /*
1093     * This handler transfers control to the interpeter without performing
1094     * any lookups.  It may be called either as part of a normal chaining
1095     * operation, or from the transition code in header.S.  We distinquish
1096     * the two cases by looking at the link register.  If called from a
1097     * translation chain, it will point to the chaining Dalvik PC + 1.
1098     * On entry:
1099     *    lr - if NULL:
1100     *        r1 - the Dalvik PC to begin interpretation.
1101     *    else
1102     *        [lr, #-1] contains Dalvik PC to begin interpretation
1103     *    rGLUE - pointer to interpState
1104     *    rFP - Dalvik frame pointer
1105     */
1106    cmp     lr, #0
1107    ldrne   r1,[lr, #-1]
1108    ldr     r2, .LinterpPunt
1109    mov     r0, r1                       @ set Dalvik PC
1110    bx      r2
1111    @ doesn't return
1112
1113.LinterpPunt:
1114    .word   dvmJitToInterpPunt
1115
1116    .size   dvmCompilerTemplateStart, .-dvmCompilerTemplateStart
1117/* File: armv5te/footer.S */
1118/*
1119 * ===========================================================================
1120 *  Common subroutines and data
1121 * ===========================================================================
1122 */
1123
1124    .text
1125    .align  2
1126.LinvokeNative:
1127    @ Prep for the native call
1128    @ r1 = newFP, r0 = methodToCall
1129    ldr     r3, [rGLUE, #offGlue_self]      @ r3<- glue->self
1130    mov     r2, #0
1131    ldr     r9, [r3, #offThread_jniLocal_topCookie] @ r9<- thread->localRef->...
1132    str     r2, [r3, #offThread_inJitCodeCache] @ not in jit code cache
1133    str     r1, [r3, #offThread_curFrame]   @ self->curFrame = newFp
1134    str     r9, [r1, #(offStackSaveArea_localRefCookie - sizeofStackSaveArea)]
1135                                        @ newFp->localRefCookie=top
1136    mov     r9, r3                      @ r9<- glue->self (preserve)
1137    SAVEAREA_FROM_FP(r10, r1)           @ r10<- new stack save area
1138
1139    mov     r2, r0                      @ r2<- methodToCall
1140    mov     r0, r1                      @ r0<- newFP
1141    add     r1, rGLUE, #offGlue_retval  @ r1<- &retval
1142
1143    LDR_PC_LR "[r2, #offMethod_nativeFunc]"
1144
1145    @ native return; r9=self, r10=newSaveArea
1146    @ equivalent to dvmPopJniLocals
1147    ldr     r2, [r10, #offStackSaveArea_returnAddr] @ r2 = chaining cell ret
1148    ldr     r0, [r10, #offStackSaveArea_localRefCookie] @ r0<- saved->top
1149    ldr     r1, [r9, #offThread_exception] @ check for exception
1150    str     rFP, [r9, #offThread_curFrame]  @ self->curFrame = fp
1151    cmp     r1, #0                      @ null?
1152    str     r0, [r9, #offThread_jniLocal_topCookie] @ new top <- old top
1153    ldr     r0, [r10, #offStackSaveArea_savedPc] @ reload rPC
1154
1155    @ r0 = dalvikCallsitePC
1156    bne     .LhandleException           @ no, handle exception
1157
1158    str     r2, [r9, #offThread_inJitCodeCache] @ set the new mode
1159    cmp     r2, #0                      @ return chaining cell still exists?
1160    bxne    r2                          @ yes - go ahead
1161
1162    @ continue executing the next instruction through the interpreter
1163    ldr     r1, .LdvmJitToInterpNoChain @ defined in footer.S
1164    add     rPC, r0, #6                 @ reconstruct new rPC (advance 6 bytes)
1165#if defined(EXIT_STATS)
1166    mov     r0, #kCallsiteInterpreted
1167#endif
1168    mov     pc, r1
1169
1170/*
1171 * On entry:
1172 * r0  Faulting Dalvik PC
1173 */
1174.LhandleException:
1175    ldr     r3, [rGLUE, #offGlue_self]  @ r3<- glue->self
1176    mov     r2, #0
1177    str     r2, [r3, #offThread_inJitCodeCache] @ in interpreter land
1178    ldr     r1, .LdvmMterpCommonExceptionThrown @ PIC way of getting &func
1179    ldr     rIBASE, .LdvmAsmInstructionStart    @ same as above
1180    mov     rPC, r0                 @ reload the faulting Dalvik address
1181    mov     pc, r1                  @ branch to dvmMterpCommonExceptionThrown
1182
1183    .align  2
1184.LdvmAsmInstructionStart:
1185    .word   dvmAsmInstructionStart
1186.LdvmJitToInterpNoChain:
1187    .word   dvmJitToInterpNoChain
1188.LdvmMterpStdBail:
1189    .word   dvmMterpStdBail
1190.LdvmMterpCommonExceptionThrown:
1191    .word   dvmMterpCommonExceptionThrown
1192.L__aeabi_cdcmple:
1193    .word   __aeabi_cdcmple
1194.L__aeabi_cfcmple:
1195    .word   __aeabi_cfcmple
1196
1197    .global dmvCompilerTemplateEnd
1198dmvCompilerTemplateEnd:
1199
1200#endif /* WITH_JIT */
1201
1202