binutils-esp32ulp allows peripheral registers to be specifiedeither as a full address (e.g. 0x3ff48000) or as a direct ULPaddress (e.g. 0x120), i.e. the address as seen from the ULP."direct" addresses are anything from 0x0 to 0x3ff (inclusive).This change ensures that 0x3ff is included in what is treatedas a direct ULP address.(Seehttps://github.com/espressif/binutils-esp32ulp/blob/249ec34/gas/config/tc-esp32ulp_esp32.c#L145for reference to how binutils treats anything larger thanDR_REG_MAX_DIRECT (0x3ff) as a full address, so everythingless *AND* equal to DR_REG_MAX_DIRECT is therefore a directULP address.)This commit contributes to being able to eventually assemble theesp32ulp_ranges.s test from binutils-esp32ulp. It addresses this line:https://github.com/espressif/binutils-esp32ulp/blob/249ec34/gas/testsuite/gas/esp32ulp/esp32/esp32ulp_ranges.s#L136

When register addresses less than or equal to 0x3ff are given to thereg_rd and reg_wr opcodes, these addresses should be used unmodifiedas the address in the machine instruction.In our implementation we split the 10 bit address field of the machineinstruction into two fields, namely "addr" for the lower 8 bits and"periph_sel" for the upper 2 bits.We already had a mechanism for determining the periph_sel part for"full" addresses (e.g. 0x3ff48000), but for direct (ULP) addresses, wealways set periph_sel to 0 instead of using the upper 2 bits from thegiven address. This commit fixes that.Note 1: In binutils-esp32ulp, they don't split the address into these 2fields but simply put the direct ULP address into a single combinedfield of 10 bits, which has the same effect. See:https://github.com/espressif/binutils-esp32ulp/blob/249ec34/gas/config/tc-esp32ulp_esp32.c#L145Note 2: In the "macro approach" in esp-idf for creating ULP code, theyalso use the split field approach (I assume our implementation ismodelled after that) and they also don't handle direct (ULP) addressescorrectly (or seemingly at all). See:https://github.com/espressif/esp-idf/blob/9d34a1c/components/ulp/include/esp32/ulp.h#L349This commit contributes to being able to eventually assemble theesp32ulp_ranges.s test from binutils-esp32ulp. It addresses this line:https://github.com/espressif/binutils-esp32ulp/blob/249ec34/gas/testsuite/gas/esp32ulp/esp32/esp32ulp_ranges.s#L136

There is no need to duplicate the "<= DR_REG_MAG_DIRECT" checkin two functions. Now the _soc_reg_to_ulp_periph_sel() functionis used only for "full addresses" (as it was some commits ago)and will return a ValueError if used with direct ULP addresses.This commit also masks values correctly when setting the addrand periph_sel fields for "direct ULP addresses", so only thebits we're interested in actually get used (rather than beingimplicitly trimmed because there aren't more bits available ina field).

Everything necessary to pass previously skipped binutils-esp32ulp tests isnow fixed. So we no longer need to skip them. (Tests using unsupportedfeatures, such as assembler macros, are still skipped.)Note 1: There is one test, esp32ulp_ranges.s, which requires symbols definedin esp32ulp_globals.s. binutils-esp32ulp joins these during the linking stagein its test scripts. Since we don't separate stages, we simply concatenate thetwo files before assembly.Note 2: binutils-esp32ulp has a bug related to how absolute symbols definedwith .set are interpreted by the JUMP instruction. If a symbol is markedglobal, the value is taken as-is, but if a symbol is not global, it's valueis divided by 4. Since py-esp32-ulp treats the symbol value the same, whetherglobal or not (the believed correct behaviour), we work around the bug in ourtest script and patch the input files to make the relevant symbols global.

ThomasWaldmann left a comment