[Vm-dev] VM Maker: VMMaker.oscog-tpr.1198.mcz
commits at source.squeak.org
commits at source.squeak.org
Wed Apr 15 21:21:47 UTC 2015
tim Rowledge uploaded a new version of VMMaker to project VM Maker:
http://source.squeak.org/VMMaker/VMMaker.oscog-tpr.1198.mcz
==================== Summary ====================
Name: VMMaker.oscog-tpr.1198
Author: tpr
Time: 15 April 2015, 2:20:30.908 pm
UUID: aac4f3ef-af49-40f6-9125-e09781490e3c
Ancestors: VMMaker.oscog-tpr.1197
Fix ARM rewriteJumpLong... after we changed register usage a little.
Change genGetActiveContextLarge:inBlock: handling of ARM SP stuff a smidge. This gets us a lot further but I'm not completely convinced it is finished with.
=============== Diff against VMMaker.oscog-tpr.1197 ===============
Item was changed:
----- Method: CogARMCompiler>>rewriteJumpLongAt:target: (in category 'inline cacheing') -----
rewriteJumpLongAt: callSiteReturnAddress target: callTargetAddress
"Rewrite a long jump instruction to jump to a different target. This variant
is used to rewrite cached primitive calls. Answer the extent of the
code change which is used to compute the range of the icache to flush."
<var: #callSiteReturnAddress type: #usqInt>
<var: #callTargetAddress type: #usqInt>
"cogit disassembleFrom: callSiteReturnAddress - 40 to: callSiteReturnAddress + 9"
+ "check that the instruction involved is actually a long jump BX ip (reg 12)
+ (CogARMCompiler new bx: 12) hex"
+ self assert: (objectMemory longAt: callSiteReturnAddress - 4) = 16rE12FFF1C .
- "check that the instruction involved is actually a long jump BX"
- self assert: (objectMemory longAt: callSiteReturnAddress - 4) = 16rE12FFF1A .
"The callTargetAddress is loaded byte by byte."
objectMemory byteAt: callSiteReturnAddress - 20 put: (callTargetAddress >> 24 bitAnd: 16rFF).
objectMemory byteAt: callSiteReturnAddress - 16 put: (callTargetAddress >> 16 bitAnd: 16rFF).
objectMemory byteAt: callSiteReturnAddress - 12 put: (callTargetAddress >> 8 bitAnd: 16rFF).
objectMemory byteAt: callSiteReturnAddress - 8 put: (callTargetAddress bitAnd: 16rFF).
self assert: (self callTargetFromReturnAddress: callSiteReturnAddress) signedIntToLong = callTargetAddress.
"self cCode: ''
inSmalltalk: [cogit disassembleFrom: callSiteReturnAddress - 20 to: callSiteReturnAddress - 1]."
^20!
Item was changed:
----- Method: CogObjectRepresentationFor32BitSpur>>genGetActiveContextLarge:inBlock: (in category 'initialization') -----
genGetActiveContextLarge: isLarge inBlock: isInBlock
"Create a trampoline to answer the active context that will
answer it if a frame is already married, and create it otherwise.
Assume numArgs is in SendNumArgsReg and ClassReg is free."
| header slotSize jumpSingle loopHead jumpNeedScavenge continuation exit |
<var: #jumpNeedScavenge type: #'AbstractInstruction *'>
<var: #continuation type: #'AbstractInstruction *'>
<var: #jumpSingle type: #'AbstractInstruction *'>
<var: #loopHead type: #'AbstractInstruction *'>
<var: #exit type: #'AbstractInstruction *'>
cogit "load the flag; stash it in both TempReg & ClassReg; do the compare (a prime candidated for use of AndCq:R:R:)"
MoveMw: FoxMethod r: FPReg R: ClassReg;
AndCq: MFMethodFlagHasContextFlag R: ClassReg R: TempReg.
jumpSingle := cogit JumpZero: 0. "jump if flag bit not set"
cogit "since the flag bit was set, get the context in the receiver reg and return"
MoveMw: FoxThisContext r: FPReg R: ReceiverResultReg;
RetN: 0.
jumpSingle jmpTarget: cogit Label.
"OK, it doesn't exist; instantiate and initialize it"
"set the hasContext flag; See CoInterpreter class>>initializeFrameIndices"
cogit
OrCq: MFMethodFlagHasContextFlag R: ClassReg;
MoveR: ClassReg Mw: FoxMethod r: FPReg.
"now get the home CogMethod into ClassReg and save for post-instantiation."
isInBlock
ifTrue:
[cogit
SubCq: 3 R: ClassReg; "-3 is -(hasContext+isBlock) flags"
MoveM16: 0 r: ClassReg R: TempReg;
SubR: TempReg R: ClassReg]
ifFalse:
[cogit SubCq: 1 R: ClassReg]. "-1 is hasContext flag"
"instantiate the context..."
slotSize := isLarge ifTrue: [LargeContextSlots] ifFalse: [SmallContextSlots].
header := objectMemory
headerForSlots: slotSize
format: objectMemory indexablePointersFormat
classIndex: ClassMethodContextCompactIndex.
self flag: #endianness.
cogit
MoveAw: objectMemory freeStartAddress R: ReceiverResultReg;
MoveCq: (self low32BitsOf: header) R: TempReg;
MoveR: TempReg Mw: 0 r: ReceiverResultReg;
MoveCq: header >> 32 R: TempReg;
MoveR: TempReg Mw: 4 r: ReceiverResultReg;
MoveR: ReceiverResultReg R: TempReg;
AddCq: (objectMemory smallObjectBytesForSlots: slotSize) R: TempReg;
MoveR: TempReg Aw: objectMemory freeStartAddress;
CmpCq: objectMemory getScavengeThreshold R: TempReg.
jumpNeedScavenge := cogit JumpAboveOrEqual: 0.
"Now initialize the fields of the context. See CoInterpreter>>marryFrame:SP:copyTemps:"
"sender gets frame pointer as a SmallInteger"
continuation :=
cogit MoveR: FPReg R: TempReg.
self genSetSmallIntegerTagsIn: TempReg.
cogit MoveR: TempReg Mw: objectMemory baseHeaderSize + (SenderIndex * objectMemory bytesPerOop) r: ReceiverResultReg.
"pc gets frame caller as a SmallInteger"
cogit MoveMw: FoxSavedFP r: FPReg R: TempReg.
self genSetSmallIntegerTagsIn: TempReg.
cogit MoveR: TempReg Mw: objectMemory baseHeaderSize + (InstructionPointerIndex * objectMemory bytesPerOop) r: ReceiverResultReg.
"Set the method field, freeing up ClassReg again, and frame's context field,"
cogit
MoveMw: (cogit offset: CogMethod of: #methodObject) r: ClassReg R: TempReg;
MoveR: TempReg Mw: objectMemory baseHeaderSize + (MethodIndex * objectMemory wordSize) r: ReceiverResultReg;
MoveR: ReceiverResultReg Mw: FoxThisContext r: FPReg.
"Now compute stack pointer; this is stackPointer (- 1 for return pc if a CISC) - framePointer - 4 (1 each for saved pc, method, context, receiver) + 1 (1-relative)"
+ "TPR note - the code here is actually doing
+ context stackPointer := ((((fp - sp) / 4) - [3|4]) + num args) asSmallInteger"
cogit
MoveR: FPReg R: TempReg;
SubR: SPReg R: TempReg;
LogicalShiftRightCq: self log2BytesPerWord R: TempReg;
SubCq: (cogit backEnd hasLinkRegister ifTrue: [3] ifFalse: [4]) R: TempReg;
AddR: SendNumArgsReg R: TempReg.
self genConvertIntegerToSmallIntegerInReg: TempReg.
cogit MoveR: TempReg Mw: objectMemory baseHeaderSize + (StackPointerIndex * objectMemory bytesPerOop) r: ReceiverResultReg.
"Set closureOrNil to either the stacked receiver or nil"
isInBlock
ifTrue:
[cogit
MoveR: SendNumArgsReg R: TempReg;
+ AddCq: 2 R: TempReg; "+2 for saved fp and saved pc" "TPR note - is this a problematic place? Maybe only 1 with LR?"
- AddCq: 2 R: TempReg; "+2 for saved fp and saved pc"
MoveXwr: TempReg R: FPReg R: TempReg]
ifFalse:
[cogit MoveCw: objectMemory nilObject R: TempReg].
cogit MoveR: TempReg Mw: objectMemory baseHeaderSize + (ClosureIndex * objectMemory bytesPerOop) r: ReceiverResultReg.
"Set the receiver"
cogit
MoveMw: FoxMFReceiver r: FPReg R: TempReg;
MoveR: TempReg Mw: objectMemory baseHeaderSize + (ReceiverIndex * objectMemory bytesPerOop) r: ReceiverResultReg.
"Now copy the arguments. This is tricky because of the shortage of registers,. ClassReg ranges
from 1 to numArgs (SendNumArgsReg), and from ReceiverIndex + 1 to ReceiverIndex + numArgs.
1 to: numArgs do:
[:i|
temp := longAt(FPReg + ((SendNumArgs - i + 2) * BytesPerWord)). +2 for saved pc and savedfp
longAtput(FPReg + FoxMFReceiver + (i * BytesPerWord), temp)]"
+ "TPR note: this is a prime candidate for passing off to the backed to do at least faintly optimal code"
cogit MoveCq: 1 R: ClassReg.
loopHead := cogit CmpR: SendNumArgsReg R: ClassReg.
exit := cogit JumpGreater: 0.
cogit
MoveR: SendNumArgsReg R: TempReg;
SubR: ClassReg R: TempReg;
+ AddCq: (cogit backEnd hasLinkRegister ifTrue: [1] ifFalse: [2]) R: TempReg; "+2 for saved fp and saved pc" "TPR note another LR problem place?"
- AddCq: 2 R: TempReg; "+2 for saved fp and saved pc"
MoveXwr: TempReg R: FPReg R: TempReg;
AddCq: ReceiverIndex + (objectMemory baseHeaderSize / objectMemory wordSize) R: ClassReg; "Now convert ClassReg from frame index to context index"
MoveR: TempReg Xwr: ClassReg R: ReceiverResultReg;
SubCq: ReceiverIndex + (objectMemory baseHeaderSize / objectMemory wordSize) - 1 R: ClassReg; "convert back adding 1 ;-)"
Jump: loopHead.
exit jmpTarget: cogit Label.
"Finally copy the temps.
ClassReg := FPReg + FoxMFReceiver.
SendNumArgsReg := SendNumArgsReg+ReceiverIndex.
[ClassReg := ClassReg - 4.
backEnd hasLinkRegister
ifTrue: [ClassReg > SPReg]
ifFalse: [ClassReg >= SPReg]] whileTrue:
[receiver[SendNumArgsReg] := *ClassReg.
SendNumArgsReg := SendNumArgsReg + 1]]"
cogit
MoveR: FPReg R: ClassReg;
AddCq: FoxMFReceiver R: ClassReg;
AddCq: ReceiverIndex + 1 + (objectMemory baseHeaderSize / objectMemory wordSize) R: SendNumArgsReg.
loopHead :=
cogit SubCq: objectMemory wordSize R: ClassReg.
cogit CmpR: SPReg R: ClassReg.
+ exit := cogit JumpBelow: 0.
- exit := cogit backEnd hasLinkRegister
- ifTrue: [cogit JumpBelowOrEqual: 0]
- ifFalse: [cogit JumpBelow: 0].
cogit
MoveMw: 0 r: ClassReg R: TempReg;
MoveR: TempReg Xwr: SendNumArgsReg R: ReceiverResultReg;
AddCq: 1 R: SendNumArgsReg;
Jump: loopHead.
exit jmpTarget: cogit Label.
cogit RetN: 0.
jumpNeedScavenge jmpTarget:
(cogit backEnd saveAndRestoreLinkRegAround: [cogit CallRT: ceScheduleScavengeTrampoline]). "We need to push the LR here for ARM, and pop it back after the callRT:"
cogit Jump: continuation.
^0!
More information about the Vm-dev
mailing list