blob: d73e0101c2b2911c0ee8aa790bd5cbfea3aa0b91 [file] [log] [blame]
/*
* This file was generated automatically by gen-template.py for 'armv5te'.
*
* --> DO NOT EDIT <--
*/
/* File: armv5te/header.S */
/*
* Copyright (C) 2008 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#if defined(WITH_JIT)
/*
* ARMv5 definitions and declarations.
*/
/*
ARM EABI general notes:
r0-r3 hold first 4 args to a method; they are not preserved across method calls
r4-r8 are available for general use
r9 is given special treatment in some situations, but not for us
r10 (sl) seems to be generally available
r11 (fp) is used by gcc (unless -fomit-frame-pointer is set)
r12 (ip) is scratch -- not preserved across method calls
r13 (sp) should be managed carefully in case a signal arrives
r14 (lr) must be preserved
r15 (pc) can be tinkered with directly
r0 holds returns of <= 4 bytes
r0-r1 hold returns of 8 bytes, low word in r0
Callee must save/restore r4+ (except r12) if it modifies them.
Stack is "full descending". Only the arguments that don't fit in the first 4
registers are placed on the stack. "sp" points at the first stacked argument
(i.e. the 5th arg).
VFP: single-precision results in s0, double-precision results in d0.
In the EABI, "sp" must be 64-bit aligned on entry to a function, and any
64-bit quantities (long long, double) must be 64-bit aligned.
*/
/*
JIT and ARM notes:
The following registers have fixed assignments:
reg nick purpose
r5 rFP interpreted frame pointer, used for accessing locals and args
r6 rGLUE MterpGlue pointer
The following registers have fixed assignments in mterp but are scratch
registers in compiled code
reg nick purpose
r4 rPC interpreted program counter, used for fetching instructions
r7 rINST first 16-bit code unit of current instruction
r8 rIBASE interpreted instruction base pointer, used for computed goto
Macros are provided for common operations. Each macro MUST emit only
one instruction to make instruction-counting easier. They MUST NOT alter
unspecified registers or condition codes.
*/
/* single-purpose registers, given names for clarity */
#define rPC r4
#define rFP r5
#define rGLUE r6
#define rINST r7
#define rIBASE r8
/*
* Given a frame pointer, find the stack save area.
*
* In C this is "((StackSaveArea*)(_fp) -1)".
*/
#define SAVEAREA_FROM_FP(_reg, _fpreg) \
sub _reg, _fpreg, #sizeofStackSaveArea
#define EXPORT_PC() \
str rPC, [rFP, #(-sizeofStackSaveArea + offStackSaveArea_currentPc)]
/*
* This is a #include, not a %include, because we want the C pre-processor
* to expand the macros into assembler assignment statements.
*/
#include "../../../mterp/common/asm-constants.h"
/* File: armv5te/platform.S */
/*
* ===========================================================================
* CPU-version-specific defines and utility
* ===========================================================================
*/
/*
* Macro for "MOV LR,PC / LDR PC,xxx", which is not allowed pre-ARMv5.
* Jump to subroutine.
*
* May modify IP and LR.
*/
.macro LDR_PC_LR source
mov lr, pc
ldr pc, \source
.endm
.global dvmCompilerTemplateStart
.type dvmCompilerTemplateStart, %function
.text
dvmCompilerTemplateStart:
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_CMP_LONG
dvmCompiler_TEMPLATE_CMP_LONG:
/* File: armv5te/TEMPLATE_CMP_LONG.S */
/*
* Compare two 64-bit values. Puts 0, 1, or -1 into the destination
* register based on the results of the comparison.
*
* We load the full values with LDM, but in practice many values could
* be resolved by only looking at the high word. This could be made
* faster or slower by splitting the LDM into a pair of LDRs.
*
* If we just wanted to set condition flags, we could do this:
* subs ip, r0, r2
* sbcs ip, r1, r3
* subeqs ip, r0, r2
* Leaving { <0, 0, >0 } in ip. However, we have to set it to a specific
* integer value, which we can do with 2 conditional mov/mvn instructions
* (set 1, set -1; if they're equal we already have 0 in ip), giving
* us a constant 5-cycle path plus a branch at the end to the
* instruction epilogue code. The multi-compare approach below needs
* 2 or 3 cycles + branch if the high word doesn't match, 6 + branch
* in the worst case (the 64-bit values are equal).
*/
/* cmp-long vAA, vBB, vCC */
cmp r1, r3 @ compare (vBB+1, vCC+1)
blt .LTEMPLATE_CMP_LONG_less @ signed compare on high part
bgt .LTEMPLATE_CMP_LONG_greater
subs r0, r0, r2 @ r0<- r0 - r2
bxeq lr
bhi .LTEMPLATE_CMP_LONG_greater @ unsigned compare on low part
.LTEMPLATE_CMP_LONG_less:
mvn r0, #0 @ r0<- -1
bx lr
.LTEMPLATE_CMP_LONG_greater:
mov r0, #1 @ r0<- 1
bx lr
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_RETURN
dvmCompiler_TEMPLATE_RETURN:
/* File: armv5te/TEMPLATE_RETURN.S */
/*
* Unwind a frame from the Dalvik stack for compiled OP_RETURN_XXX.
* If the stored value in returnAddr
* is non-zero, the caller is compiled by the JIT thus return to the
* address in the code cache following the invoke instruction. Otherwise
* return to the special dvmJitToInterpNoChain entry point.
*/
SAVEAREA_FROM_FP(r0, rFP) @ r0<- saveArea (old)
ldr r10, [r0, #offStackSaveArea_prevFrame] @ r10<- saveArea->prevFrame
ldr r8, [rGLUE, #offGlue_pSelfSuspendCount] @ r8<- &suspendCount
ldr rPC, [r0, #offStackSaveArea_savedPc] @ rPC<- saveArea->savedPc
#if !defined(WITH_SELF_VERIFICATION)
ldr r9, [r0, #offStackSaveArea_returnAddr] @ r9<- chaining cell ret
#else
mov r9, #0 @ disable chaining
#endif
ldr r2, [r10, #(offStackSaveArea_method - sizeofStackSaveArea)]
@ r2<- method we're returning to
ldr r3, [rGLUE, #offGlue_self] @ r3<- glue->self
cmp r2, #0 @ break frame?
#if !defined(WITH_SELF_VERIFICATION)
beq 1f @ bail to interpreter
#else
blxeq lr @ punt to interpreter and compare state
#endif
ldr r1, .LdvmJitToInterpNoChain @ defined in footer.S
mov rFP, r10 @ publish new FP
ldrne r10, [r2, #offMethod_clazz] @ r10<- method->clazz
ldr r8, [r8] @ r8<- suspendCount
str r2, [rGLUE, #offGlue_method]@ glue->method = newSave->method
ldr r0, [r10, #offClassObject_pDvmDex] @ r0<- method->clazz->pDvmDex
str rFP, [r3, #offThread_curFrame] @ self->curFrame = fp
add rPC, rPC, #6 @ publish new rPC (advance 6 bytes)
str r0, [rGLUE, #offGlue_methodClassDex]
cmp r8, #0 @ check the suspendCount
movne r9, #0 @ clear the chaining cell address
str r9, [r3, #offThread_inJitCodeCache] @ in code cache or not
cmp r9, #0 @ chaining cell exists?
blxne r9 @ jump to the chaining cell
#if defined(EXIT_STATS)
mov r0, #kCallsiteInterpreted
#endif
mov pc, r1 @ callsite is interpreted
1:
stmia rGLUE, {rPC, rFP} @ SAVE_PC_FP_TO_GLUE()
ldr r2, .LdvmMterpStdBail @ defined in footer.S
mov r1, #0 @ changeInterp = false
mov r0, rGLUE @ Expecting rGLUE in r0
blx r2 @ exit the interpreter
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_INVOKE_METHOD_NO_OPT
dvmCompiler_TEMPLATE_INVOKE_METHOD_NO_OPT:
/* File: armv5te/TEMPLATE_INVOKE_METHOD_NO_OPT.S */
/*
* For polymorphic callsites - setup the Dalvik frame and load Dalvik PC
* into rPC then jump to dvmJitToInterpNoChain to dispatch the
* runtime-resolved callee.
*/
@ r0 = methodToCall, r1 = returnCell, rPC = dalvikCallsite
ldrh r7, [r0, #offMethod_registersSize] @ r7<- methodToCall->regsSize
ldrh r2, [r0, #offMethod_outsSize] @ r2<- methodToCall->outsSize
ldr r9, [rGLUE, #offGlue_interpStackEnd] @ r9<- interpStackEnd
ldr r8, [rGLUE, #offGlue_pSelfSuspendCount] @ r8<- &suspendCount
add r3, r1, #1 @ Thumb addr is odd
SAVEAREA_FROM_FP(r1, rFP) @ r1<- stack save area
sub r1, r1, r7, lsl #2 @ r1<- newFp (old savearea - regsSize)
SAVEAREA_FROM_FP(r10, r1) @ r10<- stack save area
sub r10, r10, r2, lsl #2 @ r10<- bottom (newsave - outsSize)
ldr r8, [r8] @ r8<- suspendCount (int)
cmp r10, r9 @ bottom < interpStackEnd?
bxlt lr @ return to raise stack overflow excep.
@ r1 = newFP, r0 = methodToCall, r3 = returnCell, rPC = dalvikCallsite
ldr r9, [r0, #offMethod_clazz] @ r9<- method->clazz
ldr r10, [r0, #offMethod_accessFlags] @ r10<- methodToCall->accessFlags
str rPC, [rFP, #(offStackSaveArea_currentPc - sizeofStackSaveArea)]
str rPC, [r1, #(offStackSaveArea_savedPc - sizeofStackSaveArea)]
ldr rPC, [r0, #offMethod_insns] @ rPC<- methodToCall->insns
@ set up newSaveArea
str rFP, [r1, #(offStackSaveArea_prevFrame - sizeofStackSaveArea)]
str r3, [r1, #(offStackSaveArea_returnAddr - sizeofStackSaveArea)]
str r0, [r1, #(offStackSaveArea_method - sizeofStackSaveArea)]
cmp r8, #0 @ suspendCount != 0
bxne lr @ bail to the interpreter
tst r10, #ACC_NATIVE
#if !defined(WITH_SELF_VERIFICATION)
bne .LinvokeNative
#else
bxne lr @ bail to the interpreter
#endif
ldr r10, .LdvmJitToInterpNoChain
ldr r3, [r9, #offClassObject_pDvmDex] @ r3<- method->clazz->pDvmDex
ldr r2, [rGLUE, #offGlue_self] @ r2<- glue->self
@ Update "glue" values for the new method
str r0, [rGLUE, #offGlue_method] @ glue->method = methodToCall
str r3, [rGLUE, #offGlue_methodClassDex] @ glue->methodClassDex = ...
mov rFP, r1 @ fp = newFp
str rFP, [r2, #offThread_curFrame] @ self->curFrame = newFp
@ Start executing the callee
#if defined(EXIT_STATS)
mov r0, #kInlineCacheMiss
#endif
mov pc, r10 @ dvmJitToInterpNoChain
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_INVOKE_METHOD_CHAIN
dvmCompiler_TEMPLATE_INVOKE_METHOD_CHAIN:
/* File: armv5te/TEMPLATE_INVOKE_METHOD_CHAIN.S */
/*
* For monomorphic callsite, setup the Dalvik frame and return to the
* Thumb code through the link register to transfer control to the callee
* method through a dedicated chaining cell.
*/
@ r0 = methodToCall, r1 = returnCell, rPC = dalvikCallsite
@ methodToCall is guaranteed to be non-native
.LinvokeChain:
ldrh r7, [r0, #offMethod_registersSize] @ r7<- methodToCall->regsSize
ldrh r2, [r0, #offMethod_outsSize] @ r2<- methodToCall->outsSize
ldr r9, [rGLUE, #offGlue_interpStackEnd] @ r9<- interpStackEnd
ldr r8, [rGLUE, #offGlue_pSelfSuspendCount] @ r8<- &suspendCount
add r3, r1, #1 @ Thumb addr is odd
SAVEAREA_FROM_FP(r1, rFP) @ r1<- stack save area
sub r1, r1, r7, lsl #2 @ r1<- newFp (old savearea - regsSize)
SAVEAREA_FROM_FP(r10, r1) @ r10<- stack save area
add r12, lr, #2 @ setup the punt-to-interp address
sub r10, r10, r2, lsl #2 @ r10<- bottom (newsave - outsSize)
ldr r8, [r8] @ r8<- suspendCount (int)
cmp r10, r9 @ bottom < interpStackEnd?
bxlt r12 @ return to raise stack overflow excep.
@ r1 = newFP, r0 = methodToCall, r3 = returnCell, rPC = dalvikCallsite
ldr r9, [r0, #offMethod_clazz] @ r9<- method->clazz
str rPC, [rFP, #(offStackSaveArea_currentPc - sizeofStackSaveArea)]
str rPC, [r1, #(offStackSaveArea_savedPc - sizeofStackSaveArea)]
ldr rPC, [r0, #offMethod_insns] @ rPC<- methodToCall->insns
@ set up newSaveArea
str rFP, [r1, #(offStackSaveArea_prevFrame - sizeofStackSaveArea)]
str r3, [r1, #(offStackSaveArea_returnAddr - sizeofStackSaveArea)]
str r0, [r1, #(offStackSaveArea_method - sizeofStackSaveArea)]
cmp r8, #0 @ suspendCount != 0
bxne r12 @ bail to the interpreter
ldr r3, [r9, #offClassObject_pDvmDex] @ r3<- method->clazz->pDvmDex
ldr r2, [rGLUE, #offGlue_self] @ r2<- glue->self
@ Update "glue" values for the new method
str r0, [rGLUE, #offGlue_method] @ glue->method = methodToCall
str r3, [rGLUE, #offGlue_methodClassDex] @ glue->methodClassDex = ...
mov rFP, r1 @ fp = newFp
str rFP, [r2, #offThread_curFrame] @ self->curFrame = newFp
bx lr @ return to the callee-chaining cell
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_INVOKE_METHOD_PREDICTED_CHAIN
dvmCompiler_TEMPLATE_INVOKE_METHOD_PREDICTED_CHAIN:
/* File: armv5te/TEMPLATE_INVOKE_METHOD_PREDICTED_CHAIN.S */
/*
* For polymorphic callsite, check whether the cached class pointer matches
* the current one. If so setup the Dalvik frame and return to the
* Thumb code through the link register to transfer control to the callee
* method through a dedicated chaining cell.
*
* The predicted chaining cell is declared in ArmLIR.h with the
* following layout:
*
* typedef struct PredictedChainingCell {
* u4 branch;
* const ClassObject *clazz;
* const Method *method;
* u4 counter;
* } PredictedChainingCell;
*
* Upon returning to the callsite:
* - lr : to branch to the chaining cell
* - lr+2: to punt to the interpreter
* - lr+4: to fully resolve the callee and may rechain.
* r3 <- class
* r9 <- counter
*/
@ r0 = this, r1 = returnCell, r2 = predictedChainCell, rPC = dalvikCallsite
ldr r3, [r0, #offObject_clazz] @ r3 <- this->class
ldr r8, [r2, #4] @ r8 <- predictedChainCell->clazz
ldr r0, [r2, #8] @ r0 <- predictedChainCell->method
ldr r9, [r2, #12] @ r9 <- predictedChainCell->counter
cmp r3, r8 @ predicted class == actual class?
beq .LinvokeChain @ predicted chain is valid
ldr r7, [r3, #offClassObject_vtable] @ r7 <- this->class->vtable
sub r1, r9, #1 @ count--
str r1, [r2, #12] @ write back to PredictedChainingCell->counter
add lr, lr, #4 @ return to fully-resolve landing pad
/*
* r1 <- count
* r2 <- &predictedChainCell
* r3 <- this->class
* r4 <- dPC
* r7 <- this->class->vtable
*/
bx lr
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_INVOKE_METHOD_NATIVE
dvmCompiler_TEMPLATE_INVOKE_METHOD_NATIVE:
/* File: armv5te/TEMPLATE_INVOKE_METHOD_NATIVE.S */
@ r0 = methodToCall, r1 = returnCell, rPC = dalvikCallsite
ldrh r7, [r0, #offMethod_registersSize] @ r7<- methodToCall->regsSize
ldr r9, [rGLUE, #offGlue_interpStackEnd] @ r9<- interpStackEnd
ldr r8, [rGLUE, #offGlue_pSelfSuspendCount] @ r8<- &suspendCount
add r3, r1, #1 @ Thumb addr is odd
SAVEAREA_FROM_FP(r1, rFP) @ r1<- stack save area
sub r1, r1, r7, lsl #2 @ r1<- newFp (old savearea - regsSize)
SAVEAREA_FROM_FP(r10, r1) @ r10<- stack save area
ldr r8, [r8] @ r3<- suspendCount (int)
cmp r10, r9 @ bottom < interpStackEnd?
bxlt lr @ return to raise stack overflow excep.
@ r1 = newFP, r0 = methodToCall, r3 = returnCell, rPC = dalvikCallsite
str rPC, [rFP, #(offStackSaveArea_currentPc - sizeofStackSaveArea)]
str rPC, [r1, #(offStackSaveArea_savedPc - sizeofStackSaveArea)]
ldr rPC, [r0, #offMethod_insns] @ rPC<- methodToCall->insns
@ set up newSaveArea
str rFP, [r1, #(offStackSaveArea_prevFrame - sizeofStackSaveArea)]
str r3, [r1, #(offStackSaveArea_returnAddr - sizeofStackSaveArea)]
ldr r3, [rGLUE, #offGlue_self] @ r3<- glue->self
str r0, [r1, #(offStackSaveArea_method - sizeofStackSaveArea)]
cmp r8, #0 @ suspendCount != 0
ldr r8, [r0, #offMethod_nativeFunc] @ r8<- method->nativeFunc
#if !defined(WITH_SELF_VERIFICATION)
bxne lr @ bail to the interpreter
#else
bx lr @ bail to interpreter unconditionally
#endif
@ go ahead and transfer control to the native code
ldr r9, [r3, #offThread_jniLocal_topCookie] @ r9<- thread->localRef->...
mov r2, #0
str r1, [r3, #offThread_curFrame] @ self->curFrame = newFp
str r2, [r3, #offThread_inJitCodeCache] @ not in the jit code cache
str r9, [r1, #(offStackSaveArea_localRefCookie - sizeofStackSaveArea)]
@ newFp->localRefCookie=top
mov r9, r3 @ r9<- glue->self (preserve)
SAVEAREA_FROM_FP(r10, r1) @ r10<- new stack save area
mov r2, r0 @ r2<- methodToCall
mov r0, r1 @ r0<- newFP
add r1, rGLUE, #offGlue_retval @ r1<- &retval
blx r8 @ off to the native code
@ native return; r9=self, r10=newSaveArea
@ equivalent to dvmPopJniLocals
ldr r2, [r10, #offStackSaveArea_returnAddr] @ r2 = chaining cell ret
ldr r0, [r10, #offStackSaveArea_localRefCookie] @ r0<- saved->top
ldr r1, [r9, #offThread_exception] @ check for exception
str rFP, [r9, #offThread_curFrame] @ self->curFrame = fp
cmp r1, #0 @ null?
str r0, [r9, #offThread_jniLocal_topCookie] @ new top <- old top
ldr r0, [rFP, #(offStackSaveArea_currentPc - sizeofStackSaveArea)]
@ r0 = dalvikCallsitePC
bne .LhandleException @ no, handle exception
str r2, [r9, #offThread_inJitCodeCache] @ set the mode properly
cmp r2, #0 @ return chaining cell still exists?
bxne r2 @ yes - go ahead
@ continue executing the next instruction through the interpreter
ldr r1, .LdvmJitToInterpNoChain @ defined in footer.S
add rPC, r0, #6 @ reconstruct new rPC (advance 6 bytes)
#if defined(EXIT_STATS)
mov r0, #kCallsiteInterpreted
#endif
mov pc, r1
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_CMPG_DOUBLE
dvmCompiler_TEMPLATE_CMPG_DOUBLE:
/* File: armv5te/TEMPLATE_CMPG_DOUBLE.S */
/* File: armv5te/TEMPLATE_CMPL_DOUBLE.S */
/*
* For the JIT: incoming arguments in r0-r1, r2-r3
* result in r0
*
* Compare two floating-point values. Puts 0, 1, or -1 into the
* destination register based on the results of the comparison.
*
* Provide a "naninst" instruction that puts 1 or -1 into r1 depending
* on what value we'd like to return when one of the operands is NaN.
*
* See OP_CMPL_FLOAT for an explanation.
*
* For: cmpl-double, cmpg-double
*/
/* op vAA, vBB, vCC */
push {r0-r3} @ save operands
mov r11, lr @ save return address
LDR_PC_LR ".L__aeabi_cdcmple" @ PIC way of "bl __aeabi_cdcmple"
bhi .LTEMPLATE_CMPG_DOUBLE_gt_or_nan @ C set and Z clear, disambiguate
mvncc r0, #0 @ (less than) r1<- -1
moveq r0, #0 @ (equal) r1<- 0, trumps less than
add sp, #16 @ drop unused operands
bx r11
@ Test for NaN with a second comparison. EABI forbids testing bit
@ patterns, and we can't represent 0x7fc00000 in immediate form, so
@ make the library call.
.LTEMPLATE_CMPG_DOUBLE_gt_or_nan:
pop {r2-r3} @ restore operands in reverse order
pop {r0-r1} @ restore operands in reverse order
LDR_PC_LR ".L__aeabi_cdcmple" @ r0<- Z set if eq, C clear if <
movcc r0, #1 @ (greater than) r1<- 1
bxcc r11
mov r0, #1 @ r1<- 1 or -1 for NaN
bx r11
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_CMPL_DOUBLE
dvmCompiler_TEMPLATE_CMPL_DOUBLE:
/* File: armv5te/TEMPLATE_CMPL_DOUBLE.S */
/*
* For the JIT: incoming arguments in r0-r1, r2-r3
* result in r0
*
* Compare two floating-point values. Puts 0, 1, or -1 into the
* destination register based on the results of the comparison.
*
* Provide a "naninst" instruction that puts 1 or -1 into r1 depending
* on what value we'd like to return when one of the operands is NaN.
*
* See OP_CMPL_FLOAT for an explanation.
*
* For: cmpl-double, cmpg-double
*/
/* op vAA, vBB, vCC */
push {r0-r3} @ save operands
mov r11, lr @ save return address
LDR_PC_LR ".L__aeabi_cdcmple" @ PIC way of "bl __aeabi_cdcmple"
bhi .LTEMPLATE_CMPL_DOUBLE_gt_or_nan @ C set and Z clear, disambiguate
mvncc r0, #0 @ (less than) r1<- -1
moveq r0, #0 @ (equal) r1<- 0, trumps less than
add sp, #16 @ drop unused operands
bx r11
@ Test for NaN with a second comparison. EABI forbids testing bit
@ patterns, and we can't represent 0x7fc00000 in immediate form, so
@ make the library call.
.LTEMPLATE_CMPL_DOUBLE_gt_or_nan:
pop {r2-r3} @ restore operands in reverse order
pop {r0-r1} @ restore operands in reverse order
LDR_PC_LR ".L__aeabi_cdcmple" @ r0<- Z set if eq, C clear if <
movcc r0, #1 @ (greater than) r1<- 1
bxcc r11
mvn r0, #0 @ r1<- 1 or -1 for NaN
bx r11
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_CMPG_FLOAT
dvmCompiler_TEMPLATE_CMPG_FLOAT:
/* File: armv5te/TEMPLATE_CMPG_FLOAT.S */
/* File: armv5te/TEMPLATE_CMPL_FLOAT.S */
/*
* For the JIT: incoming arguments in r0-r1, r2-r3
* result in r0
*
* Compare two floating-point values. Puts 0, 1, or -1 into the
* destination register based on the results of the comparison.
*
* Provide a "naninst" instruction that puts 1 or -1 into r1 depending
* on what value we'd like to return when one of the operands is NaN.
*
* The operation we're implementing is:
* if (x == y)
* return 0;
* else if (x < y)
* return -1;
* else if (x > y)
* return 1;
* else
* return {-1,1}; // one or both operands was NaN
*
* The straightforward implementation requires 3 calls to functions
* that return a result in r0. We can do it with two calls if our
* EABI library supports __aeabi_cfcmple (only one if we want to check
* for NaN directly):
* check x <= y
* if <, return -1
* if ==, return 0
* check y <= x
* if <, return 1
* return {-1,1}
*
* for: cmpl-float, cmpg-float
*/
/* op vAA, vBB, vCC */
mov r9, r0 @ Save copies - we may need to redo
mov r10, r1
mov r11, lr @ save return address
LDR_PC_LR ".L__aeabi_cfcmple" @ cmp <=: C clear if <, Z set if eq
bhi .LTEMPLATE_CMPG_FLOAT_gt_or_nan @ C set and Z clear, disambiguate
mvncc r0, #0 @ (less than) r0<- -1
moveq r0, #0 @ (equal) r0<- 0, trumps less than
bx r11
@ Test for NaN with a second comparison. EABI forbids testing bit
@ patterns, and we can't represent 0x7fc00000 in immediate form, so
@ make the library call.
.LTEMPLATE_CMPG_FLOAT_gt_or_nan:
mov r0, r10 @ restore in reverse order
mov r1, r9
LDR_PC_LR ".L__aeabi_cfcmple" @ r0<- Z set if eq, C clear if <
movcc r0, #1 @ (greater than) r1<- 1
bxcc r11
mov r0, #1 @ r1<- 1 or -1 for NaN
bx r11
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_CMPL_FLOAT
dvmCompiler_TEMPLATE_CMPL_FLOAT:
/* File: armv5te/TEMPLATE_CMPL_FLOAT.S */
/*
* For the JIT: incoming arguments in r0-r1, r2-r3
* result in r0
*
* Compare two floating-point values. Puts 0, 1, or -1 into the
* destination register based on the results of the comparison.
*
* Provide a "naninst" instruction that puts 1 or -1 into r1 depending
* on what value we'd like to return when one of the operands is NaN.
*
* The operation we're implementing is:
* if (x == y)
* return 0;
* else if (x < y)
* return -1;
* else if (x > y)
* return 1;
* else
* return {-1,1}; // one or both operands was NaN
*
* The straightforward implementation requires 3 calls to functions
* that return a result in r0. We can do it with two calls if our
* EABI library supports __aeabi_cfcmple (only one if we want to check
* for NaN directly):
* check x <= y
* if <, return -1
* if ==, return 0
* check y <= x
* if <, return 1
* return {-1,1}
*
* for: cmpl-float, cmpg-float
*/
/* op vAA, vBB, vCC */
mov r9, r0 @ Save copies - we may need to redo
mov r10, r1
mov r11, lr @ save return address
LDR_PC_LR ".L__aeabi_cfcmple" @ cmp <=: C clear if <, Z set if eq
bhi .LTEMPLATE_CMPL_FLOAT_gt_or_nan @ C set and Z clear, disambiguate
mvncc r0, #0 @ (less than) r0<- -1
moveq r0, #0 @ (equal) r0<- 0, trumps less than
bx r11
@ Test for NaN with a second comparison. EABI forbids testing bit
@ patterns, and we can't represent 0x7fc00000 in immediate form, so
@ make the library call.
.LTEMPLATE_CMPL_FLOAT_gt_or_nan:
mov r0, r10 @ restore in reverse order
mov r1, r9
LDR_PC_LR ".L__aeabi_cfcmple" @ r0<- Z set if eq, C clear if <
movcc r0, #1 @ (greater than) r1<- 1
bxcc r11
mvn r0, #0 @ r1<- 1 or -1 for NaN
bx r11
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_MUL_LONG
dvmCompiler_TEMPLATE_MUL_LONG:
/* File: armv5te/TEMPLATE_MUL_LONG.S */
/*
* Signed 64-bit integer multiply.
*
* For JIT: op1 in r0/r1, op2 in r2/r3, return in r0/r1
*
* Consider WXxYZ (r1r0 x r3r2) with a long multiply:
* WX
* x YZ
* --------
* ZW ZX
* YW YX
*
* The low word of the result holds ZX, the high word holds
* (ZW+YX) + (the high overflow from ZX). YW doesn't matter because
* it doesn't fit in the low 64 bits.
*
* Unlike most ARM math operations, multiply instructions have
* restrictions on using the same register more than once (Rd and Rm
* cannot be the same).
*/
/* mul-long vAA, vBB, vCC */
mul ip, r2, r1 @ ip<- ZxW
umull r9, r10, r2, r0 @ r9/r10 <- ZxX
mla r2, r0, r3, ip @ r2<- YxX + (ZxW)
add r10, r2, r10 @ r10<- r10 + low(ZxW + (YxX))
mov r0,r9
mov r1,r10
bx lr
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_SHL_LONG
dvmCompiler_TEMPLATE_SHL_LONG:
/* File: armv5te/TEMPLATE_SHL_LONG.S */
/*
* Long integer shift. This is different from the generic 32/64-bit
* binary operations because vAA/vBB are 64-bit but vCC (the shift
* distance) is 32-bit. Also, Dalvik requires us to ignore all but the low
* 6 bits.
*/
/* shl-long vAA, vBB, vCC */
and r2, r2, #63 @ r2<- r2 & 0x3f
mov r1, r1, asl r2 @ r1<- r1 << r2
rsb r3, r2, #32 @ r3<- 32 - r2
orr r1, r1, r0, lsr r3 @ r1<- r1 | (r0 << (32-r2))
subs ip, r2, #32 @ ip<- r2 - 32
movpl r1, r0, asl ip @ if r2 >= 32, r1<- r0 << (r2-32)
mov r0, r0, asl r2 @ r0<- r0 << r2
bx lr
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_SHR_LONG
dvmCompiler_TEMPLATE_SHR_LONG:
/* File: armv5te/TEMPLATE_SHR_LONG.S */
/*
* Long integer shift. This is different from the generic 32/64-bit
* binary operations because vAA/vBB are 64-bit but vCC (the shift
* distance) is 32-bit. Also, Dalvik requires us to ignore all but the low
* 6 bits.
*/
/* shr-long vAA, vBB, vCC */
and r2, r2, #63 @ r0<- r0 & 0x3f
mov r0, r0, lsr r2 @ r0<- r2 >> r2
rsb r3, r2, #32 @ r3<- 32 - r2
orr r0, r0, r1, asl r3 @ r0<- r0 | (r1 << (32-r2))
subs ip, r2, #32 @ ip<- r2 - 32
movpl r0, r1, asr ip @ if r2 >= 32, r0<-r1 >> (r2-32)
mov r1, r1, asr r2 @ r1<- r1 >> r2
bx lr
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_USHR_LONG
dvmCompiler_TEMPLATE_USHR_LONG:
/* File: armv5te/TEMPLATE_USHR_LONG.S */
/*
* Long integer shift. This is different from the generic 32/64-bit
* binary operations because vAA/vBB are 64-bit but vCC (the shift
* distance) is 32-bit. Also, Dalvik requires us to ignore all but the low
* 6 bits.
*/
/* ushr-long vAA, vBB, vCC */
and r2, r2, #63 @ r0<- r0 & 0x3f
mov r0, r0, lsr r2 @ r0<- r2 >> r2
rsb r3, r2, #32 @ r3<- 32 - r2
orr r0, r0, r1, asl r3 @ r0<- r0 | (r1 << (32-r2))
subs ip, r2, #32 @ ip<- r2 - 32
movpl r0, r1, lsr ip @ if r2 >= 32, r0<-r1 >>> (r2-32)
mov r1, r1, lsr r2 @ r1<- r1 >>> r2
bx lr
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_THROW_EXCEPTION_COMMON
dvmCompiler_TEMPLATE_THROW_EXCEPTION_COMMON:
/* File: armv5te/TEMPLATE_THROW_EXCEPTION_COMMON.S */
/*
* Throw an exception from JIT'ed code.
* On entry:
* r0 Dalvik PC that raises the exception
*/
b .LhandleException
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_SAVE_STATE
dvmCompiler_TEMPLATE_SAVE_STATE:
/* File: armv5te/TEMPLATE_SAVE_STATE.S */
/*
* This handler performs a register save for selfVerification mode.
* On entry:
* Top of stack + 4: r7 value to save
* Top of stack + 0: r0 value to save
* r0 - offset from rGLUE to the beginning of the heapArgSpace record
* r7 - the value of regMap
*
* The handler must save regMap, r0-r12 and then return with r0-r12
* with their original values (note that this means r0 and r7 must take
* the values on the stack - not the ones in those registers on entry.
* Finally, the two registers previously pushed must be popped.
*/
add r0, r0, rGLUE @ pointer to heapArgSpace
stmia r0!, {r7} @ save regMap
ldr r7, [r13, #0] @ recover r0 value
stmia r0!, {r7} @ save r0
ldr r7, [r13, #4] @ recover r7 value
stmia r0!, {r1-r12}
pop {r0, r7} @ recover r0, r7
bx lr
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_RESTORE_STATE
dvmCompiler_TEMPLATE_RESTORE_STATE:
/* File: armv5te/TEMPLATE_RESTORE_STATE.S */
/*
* This handler restores state following a selfVerification memory access.
* On entry:
* r0 - offset from rGLUE to the 1st element of the coreRegs save array.
*/
add r0, r0, rGLUE @ pointer to heapArgSpace.coreRegs[0]
ldmia r0, {r0-r12}
bx lr
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_STRING_COMPARETO
dvmCompiler_TEMPLATE_STRING_COMPARETO:
/* File: armv5te/TEMPLATE_STRING_COMPARETO.S */
/*
* String's compareTo.
*
* Requires r0/r1 to have been previously checked for null. Will
* return negative if this's string is < comp, 0 if they are the
* same and positive if >.
*
* IMPORTANT NOTE:
*
* This code relies on hard-coded offsets for string objects, and must be
* kept in sync with definitions in UtfString.h. See asm-constants.h
*
* On entry:
* r0: this object pointer
* r1: comp object pointer
*
*/
mov r2, r0 @ this to r2, opening up r0 for return value
subs r0, r2, r1 @ Same?
bxeq lr
ldr r4, [r2, #STRING_FIELDOFF_OFFSET]
ldr r9, [r1, #STRING_FIELDOFF_OFFSET]
ldr r7, [r2, #STRING_FIELDOFF_COUNT]
ldr r10, [r1, #STRING_FIELDOFF_COUNT]
ldr r2, [r2, #STRING_FIELDOFF_VALUE]
ldr r1, [r1, #STRING_FIELDOFF_VALUE]
/*
* At this point, we have:
* value: r2/r1
* offset: r4/r9
* count: r7/r10
* We're going to compute
* r11 <- countDiff
* r10 <- minCount
*/
subs r11, r7, r10
movls r10, r7
/* Now, build pointers to the string data */
add r2, r2, r4, lsl #1
add r1, r1, r9, lsl #1
/*
* Note: data pointers point to previous element so we can use pre-index
* mode with base writeback.
*/
add r2, #16-2 @ offset to contents[-1]
add r1, #16-2 @ offset to contents[-1]
/*
* At this point we have:
* r2: *this string data
* r1: *comp string data
* r10: iteration count for comparison
* r11: value to return if the first part of the string is equal
* r0: reserved for result
* r3, r4, r7, r8, r9, r12 available for loading string data
*/
subs r10, #2
blt do_remainder2
/*
* Unroll the first two checks so we can quickly catch early mismatch
* on long strings (but preserve incoming alignment)
*/
ldrh r3, [r2, #2]!
ldrh r4, [r1, #2]!
ldrh r7, [r2, #2]!
ldrh r8, [r1, #2]!
subs r0, r3, r4
subeqs r0, r7, r8
bxne lr
cmp r10, #28
bgt do_memcmp16
subs r10, #3
blt do_remainder
loopback_triple:
ldrh r3, [r2, #2]!
ldrh r4, [r1, #2]!
ldrh r7, [r2, #2]!
ldrh r8, [r1, #2]!
ldrh r9, [r2, #2]!
ldrh r12,[r1, #2]!
subs r0, r3, r4
subeqs r0, r7, r8
subeqs r0, r9, r12
bxne lr
subs r10, #3
bge loopback_triple
do_remainder:
adds r10, #3
beq returnDiff
loopback_single:
ldrh r3, [r2, #2]!
ldrh r4, [r1, #2]!
subs r0, r3, r4
bxne lr
subs r10, #1
bne loopback_single
returnDiff:
mov r0, r11
bx lr
do_remainder2:
adds r10, #2
bne loopback_single
mov r0, r11
bx lr
/* Long string case */
do_memcmp16:
mov r4, lr
ldr lr, .Lmemcmp16
mov r7, r11
add r0, r2, #2
add r1, r1, #2
mov r2, r10
blx lr
cmp r0, #0
bxne r4
mov r0, r7
bx r4
.Lmemcmp16:
.word __memcmp16
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_STRING_INDEXOF
dvmCompiler_TEMPLATE_STRING_INDEXOF:
/* File: armv5te/TEMPLATE_STRING_INDEXOF.S */
/*
* String's indexOf.
*
* Requires r0 to have been previously checked for null. Will
* return index of match of r1 in r0.
*
* IMPORTANT NOTE:
*
* This code relies on hard-coded offsets for string objects, and must be
* kept in sync wth definitions in UtfString.h See asm-constants.h
*
* On entry:
* r0: string object pointer
* r1: char to match
* r2: Starting offset in string data
*/
ldr r7, [r0, #STRING_FIELDOFF_OFFSET]
ldr r8, [r0, #STRING_FIELDOFF_COUNT]
ldr r0, [r0, #STRING_FIELDOFF_VALUE]
/*
* At this point, we have:
* r0: object pointer
* r1: char to match
* r2: starting offset
* r7: offset
* r8: string length
*/
/* Build pointer to start of string data */
add r0, #16
add r0, r0, r7, lsl #1
/* Save a copy of starting data in r7 */
mov r7, r0
/* Clamp start to [0..count] */
cmp r2, #0
movlt r2, #0
cmp r2, r8
movgt r2, r8
/* Build pointer to start of data to compare and pre-bias */
add r0, r0, r2, lsl #1
sub r0, #2
/* Compute iteration count */
sub r8, r2
/*
* At this point we have:
* r0: start of data to test
* r1: chat to compare
* r8: iteration count
* r7: original start of string
* r3, r4, r9, r10, r11, r12 available for loading string data
*/
subs r8, #4
blt indexof_remainder
indexof_loop4:
ldrh r3, [r0, #2]!
ldrh r4, [r0, #2]!
ldrh r10, [r0, #2]!
ldrh r11, [r0, #2]!
cmp r3, r1
beq match_0
cmp r4, r1
beq match_1
cmp r10, r1
beq match_2
cmp r11, r1
beq match_3
subs r8, #4
bge indexof_loop4
indexof_remainder:
adds r8, #4
beq indexof_nomatch
indexof_loop1:
ldrh r3, [r0, #2]!
cmp r3, r1
beq match_3
subs r8, #1
bne indexof_loop1
indexof_nomatch:
mov r0, #-1
bx lr
match_0:
sub r0, #6
sub r0, r7
asr r0, r0, #1
bx lr
match_1:
sub r0, #4
sub r0, r7
asr r0, r0, #1
bx lr
match_2:
sub r0, #2
sub r0, r7
asr r0, r0, #1
bx lr
match_3:
sub r0, r7
asr r0, r0, #1
bx lr
/* ------------------------------ */
.balign 4
.global dvmCompiler_TEMPLATE_INTERPRET
dvmCompiler_TEMPLATE_INTERPRET:
/* File: armv5te/TEMPLATE_INTERPRET.S */
/*
* This handler transfers control to the interpeter without performing
* any lookups. It may be called either as part of a normal chaining
* operation, or from the transition code in header.S. We distinquish
* the two cases by looking at the link register. If called from a
* translation chain, it will point to the chaining Dalvik PC + 1.
* On entry:
* lr - if NULL:
* r1 - the Dalvik PC to begin interpretation.
* else
* [lr, #-1] contains Dalvik PC to begin interpretation
* rGLUE - pointer to interpState
* rFP - Dalvik frame pointer
*/
cmp lr, #0
ldrne r1,[lr, #-1]
ldr r2, .LinterpPunt
mov r0, r1 @ set Dalvik PC
bx r2
@ doesn't return
.LinterpPunt:
.word dvmJitToInterpPunt
.size dvmCompilerTemplateStart, .-dvmCompilerTemplateStart
/* File: armv5te/footer.S */
/*
* ===========================================================================
* Common subroutines and data
* ===========================================================================
*/
.text
.align 2
.LinvokeNative:
@ Prep for the native call
@ r1 = newFP, r0 = methodToCall
ldr r3, [rGLUE, #offGlue_self] @ r3<- glue->self
mov r2, #0
ldr r9, [r3, #offThread_jniLocal_topCookie] @ r9<- thread->localRef->...
str r2, [r3, #offThread_inJitCodeCache] @ not in jit code cache
str r1, [r3, #offThread_curFrame] @ self->curFrame = newFp
str r9, [r1, #(offStackSaveArea_localRefCookie - sizeofStackSaveArea)]
@ newFp->localRefCookie=top
mov r9, r3 @ r9<- glue->self (preserve)
SAVEAREA_FROM_FP(r10, r1) @ r10<- new stack save area
mov r2, r0 @ r2<- methodToCall
mov r0, r1 @ r0<- newFP
add r1, rGLUE, #offGlue_retval @ r1<- &retval
LDR_PC_LR "[r2, #offMethod_nativeFunc]"
@ Refresh Jit's on/off status
ldr r3, [rGLUE, #offGlue_ppJitProfTable]
@ native return; r9=self, r10=newSaveArea
@ equivalent to dvmPopJniLocals
ldr r2, [r10, #offStackSaveArea_returnAddr] @ r2 = chaining cell ret
ldr r0, [r10, #offStackSaveArea_localRefCookie] @ r0<- saved->top
ldr r1, [r9, #offThread_exception] @ check for exception
ldr r3, [r3] @ r1 <- pointer to Jit profile table
str rFP, [r9, #offThread_curFrame] @ self->curFrame = fp
cmp r1, #0 @ null?
str r0, [r9, #offThread_jniLocal_topCookie] @ new top <- old top
ldr r0, [r10, #offStackSaveArea_savedPc] @ reload rPC
str r3, [rGLUE, #offGlue_pJitProfTable] @ cache current JitProfTable
@ r0 = dalvikCallsitePC
bne .LhandleException @ no, handle exception
str r2, [r9, #offThread_inJitCodeCache] @ set the new mode
cmp r2, #0 @ return chaining cell still exists?
bxne r2 @ yes - go ahead
@ continue executing the next instruction through the interpreter
ldr r1, .LdvmJitToInterpNoChain @ defined in footer.S
add rPC, r0, #6 @ reconstruct new rPC (advance 6 bytes)
#if defined(EXIT_STATS)
mov r0, #kCallsiteInterpreted
#endif
mov pc, r1
/*
* On entry:
* r0 Faulting Dalvik PC
*/
.LhandleException:
ldr r3, [rGLUE, #offGlue_self] @ r3<- glue->self
mov r2, #0
str r2, [r3, #offThread_inJitCodeCache] @ in interpreter land
ldr r1, .LdvmMterpCommonExceptionThrown @ PIC way of getting &func
ldr rIBASE, .LdvmAsmInstructionStart @ same as above
mov rPC, r0 @ reload the faulting Dalvik address
mov pc, r1 @ branch to dvmMterpCommonExceptionThrown
.align 2
.LdvmAsmInstructionStart:
.word dvmAsmInstructionStart
.LdvmJitToInterpNoChain:
.word dvmJitToInterpNoChain
.LdvmMterpStdBail:
.word dvmMterpStdBail
.LdvmMterpCommonExceptionThrown:
.word dvmMterpCommonExceptionThrown
.L__aeabi_cdcmple:
.word __aeabi_cdcmple
.L__aeabi_cfcmple:
.word __aeabi_cfcmple
.global dmvCompilerTemplateEnd
dmvCompilerTemplateEnd:
#endif /* WITH_JIT */