summaryrefslogtreecommitdiffstats
Commit message (Collapse)AuthorAgeFilesLines
...
* awk: bug: broken redirection operators.Kaz Kylheku2023-05-231-2/+2
| | | | | | | * stdlib/awk.tl (sys:awk-redir): Fix regression from April 2018. The gensym variable introduced must be parallel bound, since it is referenced by the init expression of the other variable. This breaks all awk redirection operators.
* vm: handle cases in CALL like in GCALL.Kaz Kylheku2023-05-171-14/+58
| | | | | | | | * vm.c (vm_call): Specially handle the cases of 0 to 4 arguments, avoiding the general loop and invocation of generic_funcall. This gets us about a 1% improvement in recompiling the standard libarry (touch stdlib/*.tl; make).
* quips: reference to Super Dave Osborne.Kaz Kylheku2023-05-171-0/+1
| | | | * stdlib/quips.tl (%quips%): New entry.
* with-compile-options: reimplement using compiler-letKaz Kylheku2023-05-162-26/+64
| | | | | | | | | | | | | | | | | | | The with-compile-opts macro is rewritten such that it cad occur inside code that is being compiled, and change compiler options for individual subexpressions. It continues to work as before in scripted build steps such as when calls to (compile-file ...) are wrapped in it. However, for the time being, that now only works in interpreted code, because with this change, when a with-compile-opts form is compiled, it no longer arranges for the binding of *compile-opts* to be visible to the subforms; the binding affects the compiler's own environment. * stdlib/compiler.tl (with-compile-opts): Rewrite. * txr.1: Documented.
* New special operator: compiler-letKaz Kylheku2023-05-165-3/+87
| | | | | | | | | | | | | | | | | | | | | | * eval.c (compiler_let_s): New symbol variable. (op_let): Recognize compiler-let for sequential binding. (do_expand): Traverse and diagnose compiler-let form. (eval_init): Initialize compiler_let_s and register the interpreted version of the operator. * stdlib/compiler.tl (compiler compile): Handle compiler-let form. (compiler comp-compiler-let): New method. (no-dvbind-eval): New function. * autoload.c (compiler-set-entries): Intern the compiler-let symbol in the user package. * txr.1: Documented. * stdlib/doc-syms.tl: Updated.
* New special operator: progvKaz Kylheku2023-05-155-1/+185
| | | | | | | | | | | | | | | | | | | | | | Adding a progv operator, similar to the Common Lisp one. * eval.c (progv_s): New symbol variable. (op_progv): New static function. (do_expand): Recognize and traverse the progv form. (rt_progv): New static function: run-time support for compiled progv. (eval_init): Initialize progv_s, and register the the op_progv operator interpreting function. * stdlib/compilert (compiler compile): Handle progv operator ... (compiler comp-progv): ... via this new method. * tests/019/progv.tl: New file. * txr.1: Documented. * stdlib/doc-syms.tl: Updated.
* quips: piano-related quip.Kaz Kylheku2023-05-151-0/+1
| | | | * stdlib/quips.tl (%quips%): New entry.
* vm: bugfix: global lexicals looked up dynamically.Kaz Kylheku2023-05-153-23/+42
| | | | | | | | | | | | | | | | | | | | | | | | | The getlx and setlx VM instructions are using dynamic lookup for uncached bindings, due to using the same lookup_fun search function. They should use lookup_global_fun. That doesn't have an environment parameter though, so the type is not right. However, the VM never uses the environment parameter; it's always passing nil. We will get rid of the environment parameter in the lookup_fn callback and introduce a few wrappers. * eval.c, eval.h (lookup_global_fun, lookup_dynamic_var, lookup_dynamic_sym_lisp1): New functions. * vm.c (vm_stab_slowpath, vm_get_binding): lookup_fn argument loses environment parameter, and so we don't have to pass nil. (vm_gcall, vm_gapply): Use pass lookup_global_fun to to vm_stab. (vm_getsym, vm_getbind, vm_setsym, vm_gettab, vm_settab): lookup_fn argument loses environment parameter. (vm_execute): lookup functions replaced with the appropriate one-argument ones. GETLX and SETLX see a behavior change, due to using lookup_global_var which doesn't search the dynamic environment.
* quips: new dad joke about the meta-object protocolKaz Kylheku2023-05-141-0/+1
| | | | * stdlib/quips.tl (%quips%): New entry.
* bug: symbol-value place always global.Kaz Kylheku2023-05-143-4/+41
| | | | | | | | | | | | | | | | | | | | | | | | | | | We have a problem. If v is a dynamic variable, then the form (let (v) (set (symbol-value 'v) 3)) is not behaving correctly; it's updating the top-level value of v not the rebound one. * eval.c (set_symbol_value): New static function. (eval_init): Register sys:set-symbol-value intrinsic. The top-vb variable, though no longer referenced by the symbol-value place, because existing compiled code depends on it. * stdlib/place.tl (symbol-value): Rewrite the place logic to use symbol-value to access the variable, and set-symbol-value to update it, instead of referencing sys:top-vb. (sys:get-vb): This function has to stay, because it provides run-time support for code compiled with the buggy version of the place. * tests/019/symbol-value.tl: New file.
* fix crash if built-in variable is unbound.Kaz Kylheku2023-05-121-1/+3
| | | | | | | | | | | | | | | | | | | We use lookup_var_l in many places to look up the current dynamic value of a built-in variable such as *stdout*. Those places assume that a a valid location is returned which can be subject to a deref. If the application calls makunbound to remove such a variable, that deref will crash due to a null pointer dereference. Possible repro steps are numerous, possible for many variables. One example: (makunbound '*stdout*) (put-line) * eval.c (lookukp_var_l): If the binding is not found, do not return a nulloc, but throw an error exception.
* doc: missing verb under makunbound.Kaz Kylheku2023-05-121-1/+1
| | | | | * txr.1: Add missing "removes" verb to first sentence of description of makunbound.
* compiler: spelling error in diagnostic.Kaz Kylheku2023-05-121-1/+1
| | | | | * stdlib/compiler.tl (with-compile-opts): Remove stray character from "uncrecognized".
* ffi: fix unused variable warning.Kaz Kylheku2023-05-111-2/+2
| | | | | * ffi.c (ffi_varray_null_term_get): An i variable is initialized and incremented in a for loop, but never used.
* android: restore pointer tag in gc_free_all.Kaz Kylheku2023-05-111-0/+4
| | | | | | | | | | * gc.c (gc_free_all): Just like we do in the sweep function, we must mask back the pointer tag that we removed from the heap object's pointer, before handing the pointer to the free function. Starting in Android 11, the pointer tagging is more strict. It was not enforced previously; now our logic for stripping and restoring the tags is actualy being tested.
* android: configure SDK level via clang --target.Kaz Kylheku2023-05-111-1/+1
| | | | | | | | * configure: don't set __ANDROID_API__, but instead use --target to specify a target architecture. Some of the version checking is now done using symbol attributes; the preprocessor symbol alone doesn't tell the compiler what SDK version is being targeted.
* sysif: define passwd symbol for group stuff.Kaz Kylheku2023-05-111-1/+3
| | | | | | | | | * sysif.c (sysif_init): The passwd symbol is used by both the passwd and group structure, so we need to initialize it if either HAVE_PWUID or HAVE_GRGID is set. Certain Android SDK levels have getgrgid but not getpwuid, so it's possible to end up with HAVE_GRGID but not HAVE_PWUID, in which case passwd_s ends up nil, blowing up the make_struct_type call.
* Version 286.txr-286Kaz Kylheku2023-05-076-675/+768
| | | | | | | | | | | | * RELNOTES: Updated. * configure (txr_ver): Bumped version. * stdlib/ver.tl (lib-version): Bumped. * txr.1: Bumped version and date. * txr.vim, tl.vim: Regenerated.
* build: mergesort is in <stdlib.h> on MacOS.Kaz Kylheku2023-05-071-1/+1
| | | | | * configure: add mergesort to list of clashing identifiers to handle.
* doc: remove "comprised of" in two places.Kaz Kylheku2023-05-061-7/+7
| | | | | | | * txr.1: an instance of "comprised of" disappears in a wording improvement under "Dot Position in Function Calls". Another instance under doloop is replaced by consisting of. I don't have anything against it, but it bothers some people.
* hash: cache struct hash fields in locals in hash_mark.Kaz Kylheku2023-05-051-8/+11
| | | | | | * hash.c (hash_mark): Cache the table, vector and mask in local variables, so they don't have to be reloaded into registers when external functions are called.
* hash: some streamlining in weak table processing.Kaz Kylheku2023-05-053-10/+67
| | | | | | | | | | | | | | | | | | | | | | | | | | | * gc.c (mark_obj_norec): New function. Just marks an object reachable without recursing over its sub-objects. (gc_mark_norec): New function. * gc.h (gc_mark_norec): Declared. * hash.c (do_weak_tables): Cache the table, mask and vector pointer in a local variable, since these pointers are not expected to change across function calls, and can go into registers. When visiting an entry that should be reachable, we mark that entry immediately, and also use the new gc_mark_norec function to mark the chain cons cell reachable. I.e. we mark the chain backbone cons, and that cons' car field, that being the entry. Thus by the time we march through the chain, we have marked all of it. Thus, the table entries don't have to be iterated and marked any more. We use gc_mark_norec to mark the table, and explicitly mark its two special slots. The upshot of all this is that we don't have to make an extra pass over the table, and the chains, to mark things; we combine the marking with the expunging of weak values.
* defstruct: diagnostic improvement for boa args.Kaz Kylheku2023-05-051-2/+2
| | | | | | | | | | | | | | * struct.tl (defstruct): When generating the lambda that initializes slots from boa arguments, instead of we use (set (qref obj slot) val) instead of slotset. The qref macro will diagnose use of nonexistent slots.Thus warnings are produced for, say: (defstruct (point x y) nil) where x and y have not been defined, using the imperfect approach of the qref implementation, which is better than nothing.
* compiler: multiple issues in macro-parameter forms.Kaz Kylheku2023-05-051-24/+18
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | When a defmacro form is compiled, the entire form is retained as a literal in the output. This is wasteful and gives away the source code. In spite of that, errors in using the macro are incorrectly reported against defmacro, because that is the first symbol in the form. These issues arise with what arguments are passed as the first two parameters of the compiler's expand-bind-mac-params function, and what exactly it does with them. We make a tweak to that, as well as some tweaks to all the calls. * stdlib/compiler.tl (expand-bind-mac-params): There is a mix-up here in that both the ctx-form and err-form arguments are ending up in the compiled output. Let's have only the first agument, ctx-form going into the compiled output. Thus that is what is inserted into the sys:bind-mach-check call that is generated. Secondly, ctx-form should not be passed to the constructor for mac-param-parser. ctx-form is a to-be-evaluated expression which might just be a gensym; we cannot use it at compile time for error reporting. Here we must use the second argument. Thus the second argument is now used only for two purposes: copying the source code info to the output code, and for error reporting in the mac-param-parser class. This second purpose is minor, because the code has been passed through the macro expander before being compiled, which has caught all the errors. Thus the argument is changed to rlcp-form, reflecting its principal use. (comp-tree-bind, comp-tree-case): Calculate a simplified version of the tree-bind or tree-case form for error reporting and pass that as argument the ctx-form argument of expand-bind-mac-params. Just pass form as the second argument. (comp-mac-param-bind, comp-mac-env-param-bind): Just pass form as the second argument of expand-bind-mac-params.
* compiler: bugfix: wrong condition in late-peephole.Kaz Kylheku2023-05-041-1/+1
| | | | | | | | | | | * stdlib/optimize.tl (basic-blocks late-peephole): The test whether lab2 is used is bogus, and will never be true. The correct test is simply whether the block has two or more rlinks. This makes no difference in the standard library images. When the bug appears, the manifestation would be that a needed label is deleted, resulting in an exception from the assembler.
* compiler: liveness bug involving closures.Kaz Kylheku2023-05-041-3/+17
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | 2022-09-13 commit 6e354e1c2d5d64d18f527d52db75e344a9223d95, subject "compiler: bugfixes in dead code elimination", introduced a problem. By allowing the closure body blocks to be included in the links of the previous basic block that ends in the close instruction, it caused liveness info to flow out out of close blocks into the close instruction, which is wrong. Thus registers used inside a closure, which are entirely private, wrongly appear live outside of the closure, interfering with optimizations like eliminating dead registers. We can't simply roll back the commit because the bug it fixes will reappear. The fix is to pair the next field with a prev field, and maintain them; don't rely on the rlinks to point to the previous block. * stdlib/optimize.tl (basic-block): New slot, prev. (back-block join-block): As we delete the next block, we must update that block's next block's prev link. (basic-blocks link-graph): Build the prev links. Fix the bug in handling the close instruction: do not list the close body code among the links, only the branch target of the close. (basic-blocks do-peephole-block): In a few cases in which we set the bl.next to nil, we also set the bl.next.prev to nil, if bl.next exists. (basic-blocks elim-dead-clode): Reset the bl.prev of every block also. (basic-block check-bypass-empty): Here, we no longer depend on rlinks containing the previous block; the prev gives it to us. So we move that fixup out of the link, and also fix up the next blocks prev pointer.
* sort: missing vector coverage in tests.Kaz Kylheku2023-05-031-4/+18
| | | | | | | * tests/012/sort.tl: The larger input tests are testing only vectors, thus covering neither quicksort nor array binary merge. Cases added.
* sort: optimizations eliding keyfun and access.Kaz Kylheku2023-05-031-15/+45
| | | | | | | | * lib.c (quicksort): Avoid calls to keyfun when it's known to be identity, (mergesort): Likewise. Also, avoid redundant accesses to the vector when merging, for that index which has not moved between iterations.
* sort: move tests into tests/012.Kaz Kylheku2023-05-021-0/+0
| | | | | | | | * tests/010/sort.tl: File moved to tests/012. The reason is that the tests 010 run with the --gc-debug torture tests. That test case runs way too long under that test because of the testing of many permutations and whatnot.
* gc: use single allocation for prot_array.Kaz Kylheku2023-05-024-27/+21
| | | | | | | | | | | | | | | | | | | | | | | | | | | * gc.c (prot_array): Add self pointer; arr member becomes flexible array. (prot_array_mark): We now check the handle itself for null, because the whole thing is freed. (prot_array_free): Function removed. (prot_array_ops): Wire cobj_destroy_free_op in place of prot_array_free. This fixes a memory leak because prot_array_free was not freeing the handle, only the array. (gc_prot_array_alloc): Fix to allocate everything in one swoop and store the self-pointer in the named member rather than arr[-1]. The self argument is not required; we drop it. The size argument cannot be anywhere near INT_PTR_MAX, because such an array wouldn't fit into virtual memory, so it is always safe to add a small value to the size. (prot_array_free): Obtain the self-pointer, and free the handle, replacing it with a null pointer. * gc.h (gc_prot_array_alloc): Declaration updated. * lib.c (ssort_vec): Don't pass self to gc_prot_array_alloc. * lib.h (container): New macro.
* sort: support stable sorting via ssort and snsort.Kaz Kylheku2023-05-028-9/+211
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | For array-like objecgts, these objects use an array-based merge sort, using an auxiliary array equal in size to the original array. To provide the auxiliary array, a new kind of very simple vector-like object is introduced into the gc module: protected array. This looks like a raw dynamic C array of val type, returned as a val *. Under the hood, there is a heap object there, which makes the array traversable by the garbage collector. The whole point of this exercise is to make the new mergesort function safe even if the caller-supplied functions misbehave in such a way that the auxiliary array holds the only references to heap objects. * gc.c (struct prot_array): New struct, (prot_array_cls): New static variable. (gc_late_init): Register COBJ class, retaining in prot_array_cls. (prot_array_mark, prot_array_free): New static functions. (prot_array_ops): New static structure. (prot_array_alloc, prot_array_free): New functions. * gc.h (prot_array_alloc, prot_array_free): Declared. * lib.c (mergesort, ssort_vec): New static function. (snsort, ssort): New functions. * lib.h (snsort, ssort): Declared. * tests/010/sort.tl: Cover ssort. * txr.1: Documented. * stdlib/doc-syms.tl: Updated.
* sort: correct name in error reporting.Kaz Kylheku2023-05-021-4/+3
| | | | | | * lib.c (sort_vec): Take self argument instead of assuming that we are sort; this can be called by nsort. (nsort, sort): Pass self to sort_vec.
* sort: larger test.Kaz Kylheku2023-05-011-0/+8
| | | | | | | | * tests/010/sort.tl: Add some test cases of larger list. The exhaustive permutation tests are good but only go up to a relatively short size, where the median-of-three doesn't even kick in. We also cover choosing an alternative less function.
* sort: replace Lomuto partitioning with HoareKaz Kylheku2023-05-012-30/+50
| | | | | | | | | | | | | | | | | | I'm seeing numbers aobut the same performance on a sorted vector of integers, and 21% faster on vector of N random integers in the range [0, N). Also, this original algorithm handles well the case of an array consisting of a repeated value. The code we are replacing degrates to quadratic time. * lib.c (med_of_three, middle_pivot): We don't use the return value, so don't calculate and return one. (quicksort): Revise to Hoare: scanning from both ends of the array, exchanging elements. * tests/010/sort.tl: New file. We test sort with lists and vectors from length zero to eight, all permutations.
* hash: new function, hash-props.Kaz Kylheku2023-05-015-0/+51
| | | | | | | | | | | | | | | | We don't have a function in the hash table module which can create a populated hash table in one step without requiring the caller to create auxiliary lists. This new function fills that gap, albeit with some limitations. * hash.c (hash_props): New function. (hash_init): Register hash-props intrinsic. * tests/010/hash.tl: New tests. * txr.1: Documented. * stdlib/doc-syms.tl: Updated.
* time: improved convenience in time functions.Kaz Kylheku2023-04-303-29/+80
| | | | | | | | | | | | | | | | * time.c (time_str_local, time_str_utc): New static functions. (time_fields_local, time_fields_utc, time_struct_local, time_struct_utc): Time argument becomes optional, defaulted to current time. (time_init): Use time_s symbol instead of interning twice. Register new time-str-local and time-str-utc intrinsics. Fix registration of functions that take optional args. * txr.1: New functions documented; optional arguments documented; existing documentation revised. * stdlib/doc-syms.tl: Updated.
* match: ^#S() and ^#H(()) patterns must workKaz Kylheku2023-04-296-592/+549
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Quasiquote patterns not containing unquotes are not working, because the parser transforms them into quoted objects. For instance ^#S(time) becomes the form (quote #S(time)) and not the form (sys:qquote (sys:struct-lit time)). The pattern matching compiler doesn't treat quote specially, only sys:qquote. * parser.y (unquotes_occur): Function removed. (vector, hash, struct, tree, json_vals, json_pairs): Remove use of unquotes_occur. Thus vector, hash, struct, tree and JSON syntax occurring within a backquote will be turned into a special literal whether or not it contains unquotes. * lib.c (obj_print_impl): Do not print the form (sys:hash-lit) as #Hnil, but #H(). * stdlib/match.tl (transform-qquote): Add a case which will handle ^#H(), as if it were ^H(()). Bugfix in the ^H(() ...) case. The use of @(coll) means it fails to match the empty syntax when no key/value pairs are specified, whereas @(all) respects vacuous truth. * test/011/patmatch.tl: A few tests. * y.tab.shipped, y.tab.h.shipped: Updated.
* compiler: simplify live-info defined set semantics.Kaz Kylheku2023-04-181-6/+4
| | | | | | | | * stdlib/optimize.tl (basic-blocks local-liveness): Just store the mask of defined registers into each live-info. Do not propagate the defined mask from the next instruction backwards. The way the defined mask is used in calc-liveness, this makes no difference, and is simpler and faster.
* compiler: bugfix: lingering funarg eval order issue.Kaz Kylheku2023-04-171-3/+1
| | | | | | | | * stdlib/compiler.tl (compiler comp-call-impl): We can no longer free the temporary registers as-we-go based on whether the argument expression frag uses them as the output register frag. Let's just put them all into the aoregs list to be freed afterward.
* compiler: propagate t-reg rename into end insn.Kaz Kylheku2023-04-171-1/+1
| | | | | | | | * stdlib/optimize.tl (basic-blocks rename): When we stop the renaming due to an end instruction and the src being a v-reg, we can still do the rename in that end instruction itself. If the v-reg becomes invalid, that doesn't happen until after the instruction.
* compiler: rewrite t-regs through defining instruction.Kaz Kylheku2023-04-171-6/+24
| | | | | | | | | | | | | * stdlib/optimize.tl (subst-preserve): Rename list param to insn for clarity. (careful-subst-preserve): New function. This is like subst-preserve, but used only for instructions that have destination registers. It performs a rewrite such that those destination positions are avoided. (basic-blocks rename): When the instruction has src or dst as a target, don't just stop before that insn. Do the substitution in the source operands using careful-subst-preserve.
* compiler: move peephole pattern and remove condition.Kaz Kylheku2023-04-171-15/+11
| | | | | | | | | | | | | | | | | * stdlib/optimize.tl (basic-blocks do-peephole-block): Remove the local function only-locally-used-treg. This is unnecessary because the optimization is valid even if the treg is used in downstream basic blocks. It was necessary previously in the old version of this optimization in which we deleted the first instruction which sets the treg's value. We are now depending on it being identified as a dead register. Also, moving the rule to the end. The reason is that there are cases when the pattern matches, but it returns insns. That causes the rewrite macro to march down to the next instruction, skipping other patterns. This could be bad, unless the pattern is the last one tried before the @else fallback.
* compiler: better handling for mutated locals in funargs.Kaz Kylheku2023-04-171-20/+49
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Instead of the conservative strategy in compiler comp-var of loading variables into t-registers, and relying on optimization to remove them, let's just go back to the old way: variables are just registers. For function calls, we can detect mutated variables and generate the conservative code. * stdlib/compiler.tl (frag): New slots vbin and alt-oreg. When a variable access is compiled, the binding is recorded in vbin, and the desired output register in alt-oreg. (simplify-var-spy): New struct type, used for detecting mutated lexical variables when we compile a function argument list. (compiler comp-var): Revert to the old compilation strategy for lexicals: the code fragment is empty, and the output register is just the v-reg. However, we record the variable binding and remember the caller's desired register in the new frag fields. (compiler comp-setq): Also revert the strategy here. Here we get our frag from a recursive compilation, so we just annotate it. (compiler comp-call-impl): Use the simplify-var-spy to obtain a list of the lexical variables that were mutated. This is used for rewriting the frags, if necessary. (handle-mutated-var-args): New function. If the mutated-vars list is non-empty, it rewrites the frag list. Every element in the frag which is a compiled reference to a lexical variable which is mutated over the evaluation of the arg list is substituted with a conservative frag which loads the variable into a temporary register. That register thus samples the value of the variable at the correct point in the left-to-right evaluation, so the function is called with the correct values.
* compiler: allow v reg source in t-reg optimizationKaz Kylheku2023-04-161-13/+15
| | | | | | | | | | | | This change is now possible due to the previous bugfix. * stdlib/optimize.tl (basic-blocks rename): If the source register is a v-reg, do not allow the propagation past an end instruction. This is a precaution because the end instruction could be the end of the frame in which the v-register is valid; we don't want to propagate it outside of that frame.
* compiler: bugfix: wrong propagation into close insn.Kaz Kylheku2023-04-161-1/+3
| | | | | | | | | * stdlib/optimize.tl (basic-blocks rename): When we encounter a close instruction, we must leave it alone. The registers named in the argument area of the instruction do not belong to the current instruction stream or basic block; they belong to the function body.
* compiler: tighten cases in liveness calculationKaz Kylheku2023-04-151-8/+8
| | | | | | | | * stdlib/optimize.tl (basic-blocks local-liveness): Handle all instructions explicitly with no catch-all behavior. Make a copy of the live-info even for instructions that have no source or destination operands, so that they don't mistakenly marked as having defs or refs.
* compiler: keep track of multiple defs in live-info.Kaz Kylheku2023-04-101-17/+22
| | | | | | | | | | | * stdlib/optimize.tl (live-info): Slot def replaced by def0 and def1. (basic-blocks local-liveness): The local function def becomes defs: it can take two defs. These become def0 and def1. In the catch instruction case, we use both arguments, capture the resulting live-info and use it to call refs. (basic-blocks rename): Check whether either def0 or def1 is the source or destination.
* compiler: streamline live-info object creation.Kaz Kylheku2023-04-101-9/+8
| | | | | | | | | | | | | | * stdlib/optimize.tl (basic-blocks local-liveness): When processing a pure def, we don't copy the live-info unconditionally, which is waseteful since if the destination register is a t-reg, we will invoke (new live-info) to make yet another live info. Instead, let's destructively mutate the incoming live info from the instruction below, and return a copy that is made before that is done. In the def-ref case, the local copy is entirely superfluous, because in all cases we return a new object. We also eliminate redundant (set [bb.li-hash insn] li) evaluations.
* compiler: bug in liveness calculation over catch insnKaz Kylheku2023-04-101-1/+4
| | | | | | | | | * stdlib/optimize.tl (basic-blocks local-liveness): The exception symbol and argument registers in the catch instruction are clobbers, not references. We must treat them as defs. Unfortunately, the instruction has two clobbers but live-info has only one def slot, which should be fixed.
* compiler: improve t-reg copy elimination.Kaz Kylheku2023-04-101-7/+18
| | | | | | | | | | | | | * optimize.tl (rename): Instead of a mapping operation, we perform the substitution only until we hit an instruction that defines either the src or dst register. (basic-blocks do-peephole-block): Drop the conditions for doing the rename: that neither register can be defined somewhere in the rest of the block. This restriction is too limiting. We have to be careful now; we cannot delete the first instruction, and must only set the recalc flag and add to the rescan list if the substitution did something, to avoid looping.
generated by cgit v1.2.3 (git 2.25.1) at 2025-03-22 16:34:21 +0000