This is a summary of discussions relative to the merge request created by (cr1901) <[email protected]> titled Enable reload in lm32 backend after testing showed minimal regressions [PR120077] since its creation.
Description: This PR closes [120077](https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120077) on BZ. I compiled `lm32-none-gcc` with newlib support, and ran the `compile.exp`, `execute.exp`, `dg.exp`, and `old-deja.exp` test suites twice; once with LRA disabled (baseline), and once with LRA enabled (lra). All 16 multilib configurations were tested. Both sets of tests were run against f4b5c2bf40d. lm32 has a binutils simulator, and while DejaGNU doesn't currently know about it in-tree, I was able to define an out-of-tree boards file that worked: 1. export `DEJAGNU=/path/to/site.exp` 2. Fill out `site.exp` with the following: ``` global boards_dir lappend boards_dir /path/to/site_exp_dir ``` 3. Fill out `lm32-sim.exp` _in the same directory as `site.exp`_ with the following: ``` # Load the generic configuration for this board. This will define a basic # set of routines needed by the tool to communicate with the board. load_generic_config "sim" # basic-sim.exp is a basic description for the standard Cygnus simulator. load_base_board_description "basic-sim" # This will look for lm32-elf-run somewhere on your PATH, or in the gcc build # dir if binutils was built in-tree. setup_sim lm32 # Makes the slash notation described in "Passing options and running multiple testsuites" # at https://gcc.gnu.org/install/test.html work. Also "does the right thing" for # [get_multilibs]. process_multilib_options "" # The compiler used to build for this board. This has *nothing* to do # with what compiler is tested if we're testing gcc. Further, this is # the default, so this line is optional for most boards. # (Does not appear to be optional for lm32-sim. Try removing it and see what happens!) set_board_info compiler "[find_gcc]" # We only support newlib on this target. # However, we include libgloss so we can find the linker scripts. # For lm32, libnosys is also required and it's not part of the default # [libgloss_link_flags]. set_board_info cflags "[newlib_include_flags] [libgloss_include_flags]" set_board_info ldflags "[newlib_link_flags] [libgloss_link_flags] -L [get_multilibs]/libgloss/libnosys" # No linker script for this board. This is the default explicitly written out. set_board_info ldscript "-Tsim.ld" # set_board_info needs_status_wrapper 1 set_board_info gcc,stack_size 16384 set_board_info noargs 1 set_board_info gdb,nosignals 1 # 16 set_board_info gcc,timeout 8 # set_board_info gdb,timeout 1 # set_board_info sim,timeout 1 # set_board_info gdb,noresults 1 ``` 4. `make check-gcc` and `make check-g++` should "just work" and pick up the new board file. The full command-lines I used were: ``` make -j5 check-gcc//lm32-sim/{,-mbarrel-shift-enabled,-mmultiply-enabled,-mdivide-enabled,-msign-extend-enabled,'-mdivide-enabled/-msign-extend-enabled','-mmultiply-enabled/-mdivide-enabled','-mmultiply-enabled/-msign-extend-enabled','-mmultiply-enabled/-mdivide-enabled/-msign-extend-enabled','-mbarrel-shift-enabled/-mmultiply-enabled','-mbarrel-shift-enabled/-mdivide-enabled','-mbarrel-shift-enabled/-msign-extend-enabled','-mbarrel-shift-enabled/-mdivide-enabled/-msign-extend-enabled','-mbarrel-shift-enabled/-mmultiply-enabled/-mdivide-enabled','-mbarrel-shift-enabled/-mmultiply-enabled/-msign-extend-enabled','-mbarrel-shift-enabled/-mmultiply-enabled/-mdivide-enabled/-msign-extend-enabled'} RUNTESTFLAGS="compile.exp execute.exp dg.exp old-deja.exp " make -j5 check-g++//lm32-sim/{,-mbarrel-shift-enabled,-mmultiply-enabled,-mdivide-enabled,-msign-extend-enabled,'-mdivide-enabled/-msign-extend-enabled','-mmultiply-enabled/-mdivide-enabled','-mmultiply-enabled/-msign-extend-enabled','-mmultiply-enabled/-mdivide-enabled/-msign-extend-enabled','-mbarrel-shift-enabled/-mmultiply-enabled','-mbarrel-shift-enabled/-mdivide-enabled','-mbarrel-shift-enabled/-msign-extend-enabled','-mbarrel-shift-enabled/-mdivide-enabled/-msign-extend-enabled','-mbarrel-shift-enabled/-mmultiply-enabled/-mdivide-enabled','-mbarrel-shift-enabled/-mmultiply-enabled/-msign-extend-enabled','-mbarrel-shift-enabled/-mmultiply-enabled/-mdivide-enabled/-msign-extend-enabled'} RUNTESTFLAGS="compile.exp execute.exp dg.exp old-deja.exp " ``` I encountered issues with log files being truncated when interleaving `check-gcc` and `check-g++`, so I ran them separately. Once I ran the testsuites with LRA disabled and enabled, I uploaded them into a git repo using [Bunsen](https://sourceware.org/bunsen/). The repo of testruns is available on [Codeberg](https://codeberg.org/cr1901/gcc-testruns), for those curious. You will want the `lm32-baseline-2026-04-30T07.50.19` and `lm32-lra-2026-04-30T14.05.53` tags. I then used Bunsen's `r-diff-testruns` command output to analyze the differences at a file level (i.e. I'm summarizing the various optimization level and multilib configs): |Test file |Without LRA |With LRA | |:------------------------------------------|:-----------------------------|:-----------------------------| | c-c++-common/missing-close-symbol.c | all PASS | transient error | | c-c++-common/Wbidi-chars-ranges.c | all PASS | transient error | | gcc.c-torture/compile/asmgoto-2.c | all UNSUPPORTED | all PASS | | gcc.c-torture/compile/asmgoto-5.c | all UNSUPPORTED | all PASS | | gcc.c-torture/compile/asmgoto-6.c | all UNSUPPORTED | all PASS | | gcc.c-torture/compile/limits-externdecl.c | all FAIL | all PASS | | gcc.c-torture/compile/pr98096.c | all UNSUPPORTED | all PASS | | gcc.c-torture/execute/built-in-setjmp.c | all FAIL | all PASS | | gcc.dg/asm-hard-reg-error-4.c | all UNSUPPORTED | all PASS | | gcc.dg/asm-hard-reg-error-5.c | all UNSUPPORTED | all PASS | | gcc.dg/asm-hard-reg-error-6.c | all UNSUPPORTED | all PASS | | gcc.dg/c11-atomic-2.c | all FAIL | all PASS | | gcc.dg/missing-symbol-3.c | all PASS | transient error | | gcc.dg/pr100590.c | all UNSUPPORTED | all PASS | | gcc.dg/pr107385.c | all UNSUPPORTED | all PASS | | gcc.dg/pr108095.c | all UNSUPPORTED | all PASS | | gcc.dg/pr110079.c | all UNSUPPORTED | all PASS | | gcc.dg/pr43058.c | all FAIL | all PASS | | gcc.dg/pr48141.c | all FAIL | all PASS | | gcc.dg/pr70281.c | all PASS | transient error | | gcc.dg/pr97459-2.c | all FAIL | all PASS | | gcc.dg/pr97459-4.c | all FAIL | all PASS | | gcc.dg/pr97459-5.c | all FAIL | all PASS | | gcc.dg/pr97459-6.c | all FAIL | all PASS | | gcc.dg/pr97954.c | all UNSUPPORTED | all PASS | | gcc.dg/struct-by-value-1.c | UNSUPPORTED/FAIL mix | all PASS | | g++.dg/eh/ctor2.C | PASS/FAIL mix | disjoint PASS/FAIL mix | | g++.dg/inherit/covariant4.C | PASS/FAIL mix | PASS/FAIL mix | | g++.dg/ipa/pr71207.C | PASS/FAIL mix | PASS/FAIL mix | | g++.dg/tree-ssa/pr96979.C | all FAIL | all PASS | | g++.dg/warn/uninit-pr105562.C | all FAIL | all PASS | | g++.dg/warn/Warray-bounds-pr122824.C | all FAIL | all PASS | | g++.old-deja/g++.eh/catch5p.C | PASS/FAIL mix | disjoint PASS/FAIL mix | | g++.old-deja/g++.eh/catch6.C | PASS/FAIL mix | disjoint PASS/FAIL mix | | g++.old-deja/g++.eh/catch6p.C | PASS/FAIL mix | disjoint PASS/FAIL mix | | g++.old-deja/g++.mike/align1.C | PASS/FAIL mix | disjoint PASS/FAIL mix | | g++.old-deja/g++.mike/dyncast1.C | PASS/FAIL mix | disjoint PASS/FAIL mix | | g++.old-deja/g++.mike/p658.C | all UNSUPPORTED | all UNRESOLVED/FAIL mix | | g++.old-deja/g++.other/dyncast4.C | PASS/FAIL mix | disjoint PASS/FAIL mix | | g++.old-deja/g++.robertl/eh990323-2.C | PASS/FAIL mix | disjoint PASS/FAIL mix | Notes: * "Transient errors" mean "the test failed , but I could not duplicate." Race condition in dejagnu? Bit flip? Who knows :)! All of the failures seem to be excess-error related, like some columns of the error output got stripped: ``` FAIL: c-c++-common/Wbidi-chars-ranges.c -Wc++-compat (test for excess errors) Excess errors: <U+2066>if (isAdmin)<U+2069> <U+2066> begin admins only */ ~~~~~~~~ ^ | | | end of bidirectional context U+2066 (LEFT-TO-RIGHT ISOLATE) ``` * For `g++.dg/inherit/covariant4.C` are `g++.dg/ipa/pr71207.C`, the test results don't change, but the failure reason varies between `core: 4 byte misaligned read to address` and "the test timed out". I wasn't sure if this is noteworthy with the 8 second timeout in my board file, so I included them just in case. Based on these results, I don't believe that the lm32 backend really depends on differences between reload and LRA. Specifically: * Many tests that fail with LRA disabled, pass with LRA enabled. * Disregarding transient errors, _there is no test which was previously passing in all multilib configurations that started failing when LRA is enabled._ * Tests that started failing when LRA is enabled seem to all be part of tests which were already flaky before LRA was enabled (most of `g++.old-deja`, `g++.dg/inherit/covariant4.C`, and `g++.dg/ipa/pr71207.C`?). For these reasons, I'm offering this PR and test results as reason to just "flip the LRA switch on" for lm32. I don't use the backend myself as much as I did 10 years ago; everybody loves RISC-V nowadays. but the target is unchanging, and the backend is simple enough for a single person to study that I think it's worth keeping around as gcc's version of [Cpu0](https://jonathan2251.github.io/lbd/llvmstructure.html). The full and up to date discussion can be found at https://forge.sourceware.org/gcc/gcc-TEST/pulls/153 The merge request has been closed without being merged directly on the forge repository. On 2026-05-11 19:46:03+00:00, Drea Pinski (pinskia) wrote: https://gcc.gnu.org/pipermail/gcc-cvs/2026-May/453720.html
