While working with a PIC24F microcontroller, I came upon a situation where interrupts of UARTs 3 and 4 started disappearing. I reviewed my own code and even checked silicon errata for the MCU, but to no avail. Therefore, I turned my focus towards some third-party code I had incorporated in my project.
I found the problem in the USB host stack of Microchip Libraries for Applications (MLA) version 2013-06-15. The cause became apparent only with a disassembly listing, which I will use here for demonstration.
This is the original statement from the library, which clears the USB interrupt flag using a mask:
IFS5 &= 0xFFBF;
It was compiled by XC16 v1.24 at optimization level -O0 into this:
00365A 800475 MOV IFS5, W5 00365C 2FFBF4 MOV #0xFFBF, W4 00365E 628204 AND W5, W4, W4 003660 880474 MOV W4, IFS5
In this scenario, clearing the interrupt flag takes four instructions. This command sequence can unintentionally clear a flag if it is set by the hardware in-between the read and write of IFS5. The chances of this happening are not low because the USB interrupt occurs frequently, sometimes at 10,000 Hz. As it happens, the interrupt flags for UARTs 3 and 4 are present in IFS5, which solves the mystery behind this bug.
I fixed the problem by changing the original statement to the one below, which accesses the flag through a bit-field:
IFS5bits.USB1IF = 0;
This statement was compiled as:
00365A A9C08E BCLR IFS5, #6
In this case, the compiler has used an atomic bit clear instruction to clear the interrupt flag. IFS5, like other SFRs, is a bit-addressable register.
However, if I raise the compiler’s optimization level to -O1, even the original bit-mask statement is translated into a bit clear instruction.
The compiler appears to be acting inconsistently at -O0, as both the mask and bit-field C statements are logically equivalent. After searching the web as well as I could, I have found an explanation of this peculiar behavior. It is easier for the compiler to determine if a bit clear instruction can be used in case of a bit-field, as compared to a bit-mask. The use of BCLR is ultimately an optimization, albeit a very simple one.