Object-Relative Addressing 83
;; R4 contains the referencing
;; object’s virtual address
ld4 R1, [R4 + offset] ;; load 32-bit object offset and
;; sign-extend it into R1
;; fast decompression path
add R2, R4, R1 ;; compute 64-bit address
tst R1, 1 ;; test least significant bit (LSB)
bre L2 ;; jump to L2 in case non-zero
L1: ... ;; referenced object’s virtual
;; address is in R2 here
...
L2: ;; slow decompression path
mask R1 ;; compute LAT index by masking R1
ld8 R2, [R5 + R1] ;; load 64-bit address from LAT
;; R5 contains LAT address and
;; R1 contains LAT index
jmp L1
Fig.3. Low-level pseudocode for decompressing 32-bit object references: the if-then decompres-
sion approach
the 32-bit offset plus the offset of the object field in the referenced object could then be
encoded in a single addressing mode.
In our experimental setup using a PowerPC setup, we were not able to implement
these optimizations because the PowerPC ISA does not provide predication, nor does
it support the ‘base plus index plus offset’ addressing mode. Instead, we consider two
implementations to pointer decompression that are generally applicable across different
ISAs. These two decompression implementations have different performance trade-offs
which we discuss now and which we will experimentally evaluate in section 5.
If-then pointer decompression. The if-then implementation is shown in Figure 3.
The assembler code generated for decompressing 32-bit object references optimizes the
corresponding high-level pseudocode by optimizing for the most common case, namely
the fast path. We (speculatively) compute the virtual address of the referenced object by
adding the 32-bit offset with the referencing object’s virtual address. In case the least
significant bit of the 32-bit offset is zero, we then continue fetching and executing along
the fall-through path. Only in case the least significant bit of the 32-bit offset is set, we
jump to the slow path. The slow path selects a number of bits from the 32-bit offset that
will serve as index into the LAT. The slow path then indexes the LAT which reads the
64-bit virtual address of the referenced object.
Patched pointer decompression. Patched pointer decompression optimizes the com-
mon case even further by assuming that the fast path is always taken. This results in
the code shown in Figure 4. In other words, the 32-bit offset is added to the referencing
object’s virtual address to obtain the referenced object’s virtual address. This avoids the
conditional branch as needed in the if-then decompression implementation. In case the
referenced object may not be reachable using a 32-bit offset, the decompression code
剩余21页未读，继续阅读
weixin_38688906
- 粉丝: 12
- 资源: 904
我的内容管理
展开
最新资源
- 中国微型数字传声器：技术革新与市场前景
- 智能安防：基于Hi3515的嵌入式云台控制系统设计
- 手机电量低时辐射真增千倍？解析手机使用谣言
- 56F803型DSP驱动的高精度大功率超声波电源控制策略研究
- ARM与GPRS结合的远程监测系统设计
- GPS与RFID技术结合的智能巡检系统设计
- CPLD驱动的低功耗爆炸场温度测试系统设计
- 基于FPGA的智能驱动控制系统：可扩展设计与工业网络协议
- 基于ATmega128和CH374的嵌入式USB接口设计
- 基于AT89C52的温度补偿超声波测距仪：高精度设计与应用
- MSP430F448单片机在交流数字电压表中的应用
- 提升变频器应用效率的12项实用技巧
- STM32F103在数字电镀电源并联均流系统中的应用
- PSpice仿真下的升压开关电源设计：拓扑分析与CCM稳定性提升
- 轻巧高效：MSP430主导的低成本无线传感器网络节点设计
- FPGA在EDA/PLD中实现LVDS接口的应用解析
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
