如何优化EDS SRAM复制

你好！我的CAN中断处理程序功能太慢，我必须优化它。CAN缓冲区在DMA RAM中，因为它应该是这样，所以DMA不会锁定SRAM。问题是：编译器生成一个很长的代码，我在其中将DMA ram（它是_u eds_space的一部分）复制到普通数据空间。（PIC24EP512GP806，Windows7，MpalabX 3.30，XC16 1.26）。本评论的最后是关键部分的拆卸：20个asm指令长循环，循环14次，用于将7个单词从ram复制到ram。那太疯狂了。我用XC16 1.26编译。优化级别1加上我试着添加附加的moptimize-page设置选项，也尝试了XC16 1.25。我试过各种变通方法，但是生成的代码长度相应地保持不变。我确信这可以用更短的代码完成，但我不想编写内联程序集。我能用C代码缩短这个拷贝吗？在这个例子中，我展示了CAN发送函数，但是我对CAN接收函数.typedef结构{unsigned int DmaBuf[7]；}DMA_BUF；...unsigned int CAN_SendMsg(...)有同样的问题。，u eds_unsigned int*pDmaBuff)//u eds_unsigned int*pDmaBuff：指向DMA读取的CAN发送缓冲区，{volatile DMA_BUF DmaBufOnStack;//我在堆栈上的正常SRAM中构建消息，因此我可以单独处理_u eds_问题...(构建消息)...373: *(u eds_BufOnStack;)。DMA_BUF*)pDmaBuff = DmaBufpDmaBufpDmaBufOnStack;00A692EB0000CLRW000A000000000000000000A694 370013BRA0XA6A6B694646464646464646464646464646464646464646464646464646464646464646464646464646464646464646464646BRLSRW0，#15，W500500A698W500A6988888888888888780200MOVW0 W0，#15，W500500A69W500A698W500A698W500A69878020008W500A6987802000000000000000000000000042 BSET SR，#100A6A4 498285 ADDC W3，W5，W500A6A6 A90042BCLR SR，000A6A8AE2042BTSS SR，000A6A8AE2042BTSS SR，100A6AAA80042BSETSR，100A6A6A6AAAAAAAAAAA800A800A800AAA80040042BSETSR，100A6AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA6，[W4]00A6B8 8801A1MOV W1，DSWPAG00A6BA E80000 INC W0，W000A6BC 500FEE亚W0，α0xE，[W15] 900A6BE 39 FFEBB文胸NC，0xA696…}



      以下为原文

    Hi!
My CAN interrupt handler function is too slow, and I have to optimize it. The CAN buffer is in the DMA ram, as it should be, so the DMA won't lock the SRAM ever. The issue is: the compiler generates a very long code where I copy from the DMA ram (which is part of the __eds__ space) to the normal data space. (PIC24EP512GP806, windows7, MpalabX 3.30, XC16 1.26 ).
The end of this comment is the disassembly for the critical part:
20 asm instruction long loop, looped 14 times for copying 7 words from ram to ram. That's crazy.
I compile with XC16 1.26. Optimization level1 plus I tried adding additionally -moptimize-page-setting option and I tried XC16 1.25 also.
I tried various workarounds, but the generated code length relevantly remained the same.
I am sure this can be done with shorter code, but I don't want to write in-line assembly. Can I make this copy shorter with C only code?
In this example I show the CAN send function, but I have the same issue with the CAN receive function.


typedef struct
{
unsigned int DmaBuf[ 7 ];
}DMA_BUF;
...
unsigned int CAN_SendMsg( ... , __eds__ unsigned int *pDmaBuff )
//__eds__ unsigned int *pDmaBuff: points to the CAN sending buffer, which the DMA reads from.
{
volatile DMA_BUF DmaBufOnStack; //I build up the message in normal SRAM on the stack, so I can handle the __eds__ issue separately
...
(building the message)
...
373: *(__eds__ DMA_BUF * )pDmaBuff = DmaBufOnStack;
00A692 EB0000 CLR W0
00A694 370013 BRA 0xA6BC
00A696 DE02CF LSR W0, #15, W5
00A698 780200 MOV W0, W4
00A69A A1F004 BCLR W4, #15
00A69C D00082 SL W2, W1
00A69E D00304 SL W4, W6
00A6A0 408206 ADD W1, W6, W4
00A6A2 A82042 BSET SR, #1
00A6A4 498285 ADDC W3, W5, W5
00A6A6 A90042 BCLR SR, #0
00A6A8 AE2042 BTSS SR, #1
00A6AA A80042 BSET SR, #0
00A6AC D38204 RRC W4, W4
00A6AE 5000EE SUB W0, #0xE, W1
00A6B0 78C36F MOV.B [W15+W1], W6
00A6B2 8001A1 MOV DSWPAG, W1
00A6B4 8801A5 MOV W5, DSWPAG
00A6B6 784A06 MOV.B W6, [W4]
00A6B8 8801A1 MOV W1, DSWPAG
00A6BA E80000 INC W0, W0
00A6BC 500FEE SUB W0, #0xE, [W15]
00A6BE 39FFEB BRA NC, 0xA696
...
}