compiler: Optimize most pure bifs to be guard bifs in try/catch

When put in a try, calling a guard bif with no side-effects and 6 other bifs ([PR-9042](erlang#9042)) are optimized to remove the try/catch. However, most pure bifs can be safely optimized in the same way in order to gain performance. This PR introduces a new internal instruction `call_pseudo_guard_bif` that enables the compiler to remove the try/catch around the bif. The instruction is designed to call the bif in the same way, but having more optimized error-handling in the runtime system. Example code: try float_to_list(A, utf8) catch _:_ -> error end. Before, assembly for the bif call after optimizations: {label,2}. {allocate,1,1}. {'try',{y,0},{f,3}}. {line,[{location,"zlc.erl",23}]}. {call_ext,1,{extfunc,erlang,float_to_list,1}}. {try_end,{y,0}}. {deallocate,1}. return. Now, assembly for the bif call after optimizations: {label,2}. {allocate,0,1}. {line,[{location,"zlc.erl",23}]}. {call_pseudo_guard_bif,1,{extfunc,erlang,float_to_list,1},{f,3}}. {deallocate,0}. return. Bifs that are optimized for try/catch in this change are listed in `pbif.tab`.
lucioleKi · Nov 22, 2024 · e67f514 · e67f514
1 parent 270cec6
commit e67f514
Show file tree

Hide file tree

Showing 25 changed files with 694 additions and 102 deletions.
diff --git a/erts/emulator/beam/emu/bif_instrs.tab b/erts/emulator/beam/emu/bif_instrs.tab
@@ -127,6 +127,102 @@ i_bif_body.call(Bif, Dst) {
     goto post_error_handling;
 }
 
+i_call_pseudo_guard_bif(Live, Bif, Exp, Fail) {
+    Export *export;
+    ErtsBifFunc bf;
+
+    Eterm result;
+    ErlHeapFragment *live_hf_end;
+
+    bf = (ErtsBifFunc) $Bif;
+    export = (Export*) $Exp;
+
+    if (!((FCALLS - 1) > 0 || (FCALLS-1) > neg_o_reds)) {
+        /*
+         * If we have run out of reductions, do a context
+         * switch before calling the BIF.
+         */
+        c_p->arity = GET_EXPORT_ARITY(export);
+        c_p->current = &export->info.mfa;
+        goto context_switch3;
+    }
+
+    if (ERTS_UNLIKELY(export->is_bif_traced)) {
+        $SAVE_CONTINUATION_POINTER($NEXT_INSTRUCTION);
+        $DISPATCH_EXPORT(export);
+    }
+
+    ERTS_MSACC_SET_BIF_STATE_CACHED_X(GET_EXPORT_MODULE(export), bf);
+
+    PRE_BIF_SWAPOUT(c_p);
+    ERTS_DBG_CHK_REDS(c_p, FCALLS);
+    c_p->fcalls = FCALLS - 1;
+    if (FCALLS <= 0) {
+        save_calls(c_p, export);
+    }
+    ASSERT(!ERTS_PROC_IS_EXITING(c_p));
+    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
+    live_hf_end = c_p->mbuf;
+    ERTS_CHK_MBUF_SZ(c_p);
+
+    ERTS_ASSERT_TRACER_REFS(&c_p->common);
+
+    result = (*bf)(c_p, reg, I);
+
+    ERTS_ASSERT_TRACER_REFS(&c_p->common);
+
+    /* Only heavy BIFs may GC. */
+    ASSERT(E == c_p->stop);
+
+    ERTS_CHK_MBUF_SZ(c_p);
+    ASSERT(!ERTS_PROC_IS_EXITING(c_p) || is_non_value(result));
+    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
+    ERTS_HOLE_CHECK(c_p);
+    ERTS_REQ_PROC_MAIN_LOCK(c_p);
+    if (ERTS_IS_GC_AFTER_BIF_DESIRED(c_p)) {
+        Uint arity = GET_EXPORT_ARITY(export);
+        result = erts_gc_after_bif_call_lhf(c_p, live_hf_end, result,
+                                            reg, arity);
+        E = c_p->stop;
+    }
+    HTOP = HEAP_TOP(c_p);
+    FCALLS = c_p->fcalls;
+    PROCESS_MAIN_CHK_LOCKS(c_p);
+    ERTS_DBG_CHK_REDS(c_p, FCALLS);
+
+    /*
+     * We have to update the cache if we are enabled in order
+     * to make sure no bookkeeping is done after we disabled
+     * msacc. We don't always do this as it is quite expensive.
+     */
+    if (ERTS_MSACC_IS_ENABLED_CACHED_X()) {
+        ERTS_MSACC_UPDATE_CACHE_X();
+    }
+    ERTS_MSACC_SET_STATE_CACHED_M_X(ERTS_MSACC_STATE_EMULATOR);
+    if (ERTS_LIKELY(is_value(result))) {
+        x(0) = result;
+        CHECK_TERM(x(0));
+        $NEXT0();
+    } else if (c_p->freason == TRAP) {
+        /*
+         * Set the continuation pointer to return to next
+         * instruction after the trap (either by a return from
+         * erlang code or by nif_bif.epilogue() when the BIF
+         * is done).
+         */
+        $SAVE_CONTINUATION_POINTER($NEXT_INSTRUCTION);
+        SET_I(c_p->i);
+        $DISPATCH();
+    }
+
+    /*
+     * Error handling.  SWAPOUT is not needed because it was done above.
+     */
+    ASSERT(c_p->stop == E);
+    $FAIL($Fail);
+    //| -no_next
+}
+
 //
 // length/1 is the only guard BIF that does not execute in constant
 // time. Here follows special instructions to allow the calculation of

diff --git a/erts/emulator/beam/emu/ops.tab b/erts/emulator/beam/emu/ops.tab
@@ -1541,6 +1541,15 @@ gc_bif3 Fail=f _Live Bif S1 S2 S3 Dst => i_bif3 S3 S2 S1 Fail Bif Dst
 unsupported_guard_bif/3
 unsupported_guard_bif _A _B _C | never() => _
 
+#
+# Pseudo-guard BIFs
+#
+
+call_pseudo_guard_bif Live Func Fail =>
+    i_call_pseudo_guard_bif Live Func Func Fail
+
+i_call_pseudo_guard_bif t b e f
+
 #
 # R13B03
 #

diff --git a/erts/emulator/beam/jit/arm/beam_asm.hpp b/erts/emulator/beam/jit/arm/beam_asm.hpp
@@ -1243,6 +1243,12 @@ class BeamModuleAssembler : public BeamAssembler,
                     const ArgWord &Bif,
                     const ArgRegister &Dst);
 
+    void emit_i_bif_pure(const ArgWord &Bif,
+                         const ArgExport &Exp,
+                         const ArgWord &Live,
+                         const ArgLabel &Fail,
+                         const ArgRegister &Dst);
+
     void emit_error(int code);
     void emit_error(int reason, const ArgSource &Src);
 
@@ -1299,7 +1305,7 @@ class BeamModuleAssembler : public BeamAssembler,
 
     void emit_validate_unicode(Label next, Label fail, a64::Gp value);
 
-    void ubif_comment(const ArgWord &Bif);
+    void ubif_comment(const ArgWord &Bif, const ArgExport &Exp);
 
     void emit_cmp_immed_to_bool(arm::CondCode cc,
                                 const ArgSource &LHS,

diff --git a/erts/emulator/beam/jit/arm/beam_asm_global.hpp.pl b/erts/emulator/beam/jit/arm/beam_asm_global.hpp.pl
@@ -43,6 +43,7 @@
     call_bif_shared
     call_light_bif_shared
     call_nif_yield_helper
+    call_pseudo_guard_bif_shared
     catch_end_shared
     call_nif_early
     call_nif_shared
@@ -191,6 +192,8 @@ sub gen_list {
 
     void emit_i_length_common(Label fail, int state_size);
 
+    void emit_call_bif_common(bool return_error);
+
     void emit_raise_badarg(const ErtsCodeMFA *mfa);
 
     void emit_bif_bit_size_helper(Label fail);

diff --git a/erts/emulator/beam/jit/arm/instr_bif.cpp b/erts/emulator/beam/jit/arm/instr_bif.cpp
@@ -30,11 +30,31 @@ extern "C"
 #include "erl_msacc.h"
 }
 
-void BeamModuleAssembler::ubif_comment(const ArgWord &Bif) {
+void BeamModuleAssembler::ubif_comment(const ArgWord &Bif,
+                                       const ArgExport &Exp) {
     if (logger.file()) {
-        ErtsCodeMFA *mfa = ubif2mfa((void *)Bif.get());
-        if (mfa) {
-            comment("UBIF: %T/%d", mfa->function, mfa->arity);
+        BeamFile_ImportEntry *import;
+        const Export *exp;
+        bool is_pure;
+
+        import = &beam->imports.entries[Exp.val];
+        exp = erts_active_export_entry(import->module,
+                                       import->function,
+                                       import->arity);
+
+        is_pure = exp->bif_number != -1 &&
+                  bif_table[exp->bif_number].kind == BIF_KIND_PURE;
+
+        if (is_pure) {
+            comment("fake UBIF: %T:%T/%d",
+                    import->module,
+                    import->function,
+                    import->arity);
+        } else {
+            ErtsCodeMFA *mfa = ubif2mfa((void *)Bif.get());
+            if (mfa) {
+                comment("UBIF: %T/%d", mfa->function, mfa->arity);
+            }
         }
     }
 }
@@ -108,25 +128,27 @@ void BeamGlobalAssembler::emit_i_bif_body_shared() {
 void BeamModuleAssembler::emit_i_bif1(const ArgSource &Src1,
                                       const ArgLabel &Fail,
                                       const ArgWord &Bif,
+                                      const ArgExport &Exp,
                                       const ArgRegister &Dst) {
     auto src1 = load_source(Src1);
 
     a.str(src1.reg, getXRef(0));
 
-    ubif_comment(Bif);
+    ubif_comment(Bif, Exp);
     emit_i_bif(Fail, Bif, Dst);
 }
 
 void BeamModuleAssembler::emit_i_bif2(const ArgSource &Src1,
                                       const ArgSource &Src2,
                                       const ArgLabel &Fail,
                                       const ArgWord &Bif,
+                                      const ArgExport &Exp,
                                       const ArgRegister &Dst) {
     auto [src1, src2] = load_sources(Src1, TMP1, Src2, TMP2);
 
     a.stp(src1.reg, src2.reg, getXRef(0));
 
-    ubif_comment(Bif);
+    ubif_comment(Bif, Exp);
     emit_i_bif(Fail, Bif, Dst);
 }
 
@@ -135,14 +157,15 @@ void BeamModuleAssembler::emit_i_bif3(const ArgSource &Src1,
                                       const ArgSource &Src3,
                                       const ArgLabel &Fail,
                                       const ArgWord &Bif,
+                                      const ArgExport &Exp,
                                       const ArgRegister &Dst) {
     auto [src1, src2] = load_sources(Src1, TMP1, Src2, TMP2);
     auto src3 = load_source(Src3, TMP3);
 
     a.stp(src1.reg, src2.reg, getXRef(0));
     a.str(src3.reg, getXRef(2));
 
-    ubif_comment(Bif);
+    ubif_comment(Bif, Exp);
     emit_i_bif(Fail, Bif, Dst);
 }
 
@@ -168,12 +191,13 @@ void BeamModuleAssembler::emit_i_bif(const ArgLabel &Fail,
 
 void BeamModuleAssembler::emit_nofail_bif1(const ArgSource &Src1,
                                            const ArgWord &Bif,
+                                           const ArgExport &Exp,
                                            const ArgRegister &Dst) {
     auto src1 = load_source(Src1);
 
     a.str(src1.reg, getXRef(0));
 
-    ubif_comment(Bif);
+    ubif_comment(Bif, Exp);
     mov_arg(ARG4, Bif);
     fragment_call(ga->get_i_bif_guard_shared());
     mov_arg(Dst, ARG1);
@@ -182,12 +206,13 @@ void BeamModuleAssembler::emit_nofail_bif1(const ArgSource &Src1,
 void BeamModuleAssembler::emit_nofail_bif2(const ArgSource &Src1,
                                            const ArgSource &Src2,
                                            const ArgWord &Bif,
+                                           const ArgExport &Exp,
                                            const ArgRegister &Dst) {
     auto [src1, src2] = load_sources(Src1, TMP1, Src2, TMP2);
 
     a.stp(src1.reg, src2.reg, getXRef(0));
 
-    ubif_comment(Bif);
+    ubif_comment(Bif, Exp);
     mov_arg(ARG4, Bif);
     fragment_call(ga->get_i_bif_guard_shared());
     mov_arg(Dst, ARG1);
@@ -364,13 +389,8 @@ static Eterm debug_call_light_bif(Process *c_p,
 
 /* It is important that the below code is as optimized as possible.
  * When doing any changes, make sure to look at the estone bif_dispatch
- * benchmark to make sure you don't introduce any regressions.
- *
- * ARG3 = entry
- * ARG4 = export entry
- * ARG8 = BIF pointer
- */
-void BeamGlobalAssembler::emit_call_light_bif_shared() {
+ * benchmark to make sure you don't introduce any regressions. */
+void BeamGlobalAssembler::emit_call_bif_common(bool return_error) {
     arm::Mem entry_mem = TMP_MEM1q, export_mem = TMP_MEM2q,
              mbuf_mem = TMP_MEM3q;
 
@@ -528,11 +548,16 @@ void BeamGlobalAssembler::emit_call_light_bif_shared() {
 
             a.bind(error);
             {
-                /* raise_exception_shared expects current PC in ARG2 and MFA in
-                 * ARG4. */
-                a.ldp(ARG2, ARG4, entry_mem);
-                add(ARG4, ARG4, offsetof(Export, info.mfa));
-                a.b(labels[raise_exception_shared]);
+                if (return_error) {
+                    a.mov(XREG0, THE_NON_VALUE);
+                    a.ret(a64::x30);
+                } else {
+                    /* raise_exception_shared expects current PC in
+                     * ARG2 and MFA in ARG4. */
+                    a.ldp(ARG2, ARG4, entry_mem);
+                    add(ARG4, ARG4, offsetof(Export, info.mfa));
+                    a.b(labels[raise_exception_shared]);
+                }
             }
         }
 
@@ -577,6 +602,30 @@ void BeamGlobalAssembler::emit_call_light_bif_shared() {
     }
 }
 
+/*
+ * ARG3 = entry
+ * ARG4 = export entry
+ * ARG8 = BIF pointer
+ *
+ * If successful, the result is returned in XREG0.
+ * In case of error, an exception is raised.
+ */
+void BeamGlobalAssembler::emit_call_light_bif_shared() {
+    emit_call_bif_common(false);
+}
+
+/*
+ * ARG3 = entry
+ * ARG4 = export entry
+ * ARG8 = BIF pointer
+ *
+ * If successful, the result is returned in XREG0.
+ * In case of error, XREG0 is THE_NON_VALUE on return.
+ */
+void BeamGlobalAssembler::emit_call_pseudo_guard_bif_shared() {
+    emit_call_bif_common(true);
+}
+
 void BeamModuleAssembler::emit_call_light_bif(const ArgWord &Bif,
                                               const ArgExport &Exp) {
     Label entry = a.newLabel();
@@ -594,6 +643,26 @@ void BeamModuleAssembler::emit_call_light_bif(const ArgWord &Bif,
     fragment_call(ga->get_call_light_bif_shared());
 }
 
+void BeamModuleAssembler::emit_i_call_pseudo_guard_bif(const ArgWord &Live,
+                                                       const ArgWord &Bif,
+                                                       const ArgExport &Exp,
+                                                       const ArgLabel &Fail) {
+    Label entry = a.newLabel();
+    BeamFile_ImportEntry *e = &beam->imports.entries[Exp.get()];
+
+    a.bind(entry);
+
+    mov_arg(ARG4, Exp);
+    mov_arg(ARG8, Bif);
+    a.adr(ARG3, entry);
+
+    if (logger.file()) {
+        comment("BIF: %T:%T/%d", e->module, e->function, e->arity);
+    }
+    fragment_call(ga->get_call_pseudo_guard_bif_shared());
+    emit_branch_if_not_value(XREG0, resolve_beam_label(Fail, dispUnknown));
+}
+
 void BeamModuleAssembler::emit_send() {
     Label entry = a.newLabel();
 

diff --git a/erts/emulator/beam/jit/arm/instr_guard_bifs.cpp b/erts/emulator/beam/jit/arm/instr_guard_bifs.cpp
@@ -802,12 +802,13 @@ void BeamModuleAssembler::emit_bif_hd(const ArgSource &Src,
  */
 
 void BeamModuleAssembler::emit_bif_is_map_key(const ArgWord &Bif,
+                                              const ArgExport &Exp,
                                               const ArgLabel &Fail,
                                               const ArgSource &Key,
                                               const ArgSource &Src,
                                               const ArgRegister &Dst) {
     if (!exact_type<BeamTypeId::Map>(Src)) {
-        emit_i_bif2(Key, Src, Fail, Bif, Dst);
+        emit_i_bif2(Key, Src, Fail, Bif, Exp, Dst);
         return;
     }