Forward function objects to jump pad targets

... and fix really ugly memory bugs occuring because some
code referenced function objects not the code objects
it was calling

4 files changed, 54 insertions, 12 deletions

diff --git a/TODO b/TODO
index 20ae1a5..b8dee38 100644
--- a/TODO
+++ b/TODO
@@ -1,6 +1,6 @@
 * apply the trace extractor to non-tail situations
-* implement the GC-based function opcode forwarding
 * forward scope-cannot-escape information
+* inline trivial functions when optimizer generates ... /rax :movqImmReg /rax :callqReg sequence
 * utf8
 * regex substitution
 * asm-based regex engine
diff --git a/compiler/elymasAsmLib.ey b/compiler/elymasAsmLib.ey
index 42d6ecb..4d51ce5 100644
--- a/compiler/elymasAsmLib.ey
+++ b/compiler/elymasAsmLib.ey
@@ -664,6 +664,7 @@
       /r10 :pushqReg
       /r11 :pushqReg
       /r12 :pushqReg
+      /r13 :pushqReg
 
       /r8 :movqImmOOBReg HEAPBASE # constant through mark
       /r9 :movqImmOOBReg BLOCKBASE # constant through mark
@@ -693,12 +694,13 @@
       /resetMarkLoop :loopLbl8
       @noResetNecessary
 
-      # start from current scope and mark all reachable blocks
-      /r14 /rdi :movqRegReg
+      # start from current coroutine and mark all reachable blocks
+      /r13 /rdi :movqRegReg # r13 must be checked first because we use it in markObject
+      /r13 /r13 :xorqRegReg
       /markObject :callqLbl32
 
-      # start from current coroutine and mark all reachable blocks
-      /r13 /rdi :movqRegReg
+      # start from current scope and mark all reachable blocks
+      /r14 /rdi :movqRegReg
       /markObject :callqLbl32
 
       # start from stack and mark all reachable blocks
@@ -760,6 +762,7 @@
       /freeLoop :loopLbl8
       @noFreeNecessary
 
+      /r13 :popqReg
       /r12 :popqReg
       /r11 :popqReg
       /r10 :popqReg
@@ -984,6 +987,7 @@
       16 /rsi /rdi :movqMemDisp8Reg
       /markObject :callqLbl32
       24 /rsi /rdi :movqMemDisp8Reg
+      /rsi /r13 :movqRegReg # optimistically inform the mark on the code object of our location
       /rsi :popqReg
       /markObject :jmpLbl32
 
@@ -1028,7 +1032,18 @@
       /markFunctionCodeFull :jnzLbl8
       1 [ /rax :jmpqReg ] * 27 /rdi :cmpbImmMemDisp8
       /markFunctionCodeFull :jnzLbl8
-      # TODO: This would be good moment to forward the incoming reference (which is on the stack somewhere), actually
+
+      /r13 /r13 :testqRegReg # check whether r13 contains something (a function object)
+      /noRewritePossible :jzLbl8
+      24 /r13 /rdi :cmpqMemDisp8Reg # ... which points to the forwarding function code object
+      /noRewritePossible :jnzLbl8
+
+      18 /rdi /rdi :movqMemDisp8Reg
+      16 /rdi :subqImm8Reg
+      /rdi 24 /r13 :movqRegMemDisp8 # if so, rewrite the originating function object while we are at it
+      /markObject :jmpLbl32
+
+      @noRewritePossible
       18 /rdi /rdi :movqMemDisp8Reg
       16 /rdi :subqImm8Reg
       /markObject :jmpLbl32
diff --git a/compiler/elymasGlobal.ey b/compiler/elymasGlobal.ey
index 5d9f8f9..b35aa7a 100644
--- a/compiler/elymasGlobal.ey
+++ b/compiler/elymasGlobal.ey
@@ -1482,7 +1482,9 @@
         :retn
       ] ::loadToRdi
 
-      /rdx 0 /rbp :movqRegMemDisp8 # save object reference
+      /rcx 0 /rbp :movqRegMemDisp8 # save scope reference
+      16 /rbx :subqImm8Reg
+      /rbx 8 /rbp :movqRegMemDisp8 # ... and code reference (in case the function object is ever updated)
 
       8 /r15 /rdi :movqMemDisp8Reg
       16 /rdi :addqImm8Reg
@@ -1515,7 +1517,8 @@
       [
         /rax :jmpqReg
       ] ::loadToRdi
-      /rdx 0 /rbp :movqRegMemDisp8 # save the address for the GC
+      16 /rcx :subqImm8Reg
+      /rcx 0 /rbp :movqRegMemDisp8 # save code address for the GC
 
       8 /r15 /rdi :movqMemDisp8Reg
       16 /rdi :addqImm8Reg
@@ -1527,20 +1530,24 @@
 
       @patchConstantPassive
       8 /r15 /rdi :movqMemDisp8Reg
-      # TODO: could reset string reference here
+      8 /rdi /rbp :movqMemDisp8Reg
+      16 1 /rbp /rdi /rbp :leaqMemIndexScaleDisp8Reg
       16 /rdi :addqImm8Reg
 
       [
         /rbx :popqReg
         /rax :movqImmOOBReg
       ] ::loadToRdi
-      /rdi :popqMem
+      /rdx :popqReg
+      /rdx /rdi :movqRegMem
       8 /rdi :addqImm8Reg
       [
         /rax :pushqReg
         /rbx :jmpqReg
       ] ::loadToRdi
 
+      /rdx 0 /rbp :movqRegMemDisp8 # save constant address for the GC
+
       8 /r15 /rdi :movqMemDisp8Reg
       16 /rdi :addqImm8Reg
 
@@ -1825,7 +1832,8 @@
       /rsi /rdi :movqRegMem
       /rsi 0 /rbp :movqRegMemDisp8
       8 /rdi :addqImm8Reg
-      8 /rbp :addqImm8Reg
+      16 /rbp :addqImm8Reg # we only save 1 reference here, but some cases need 2 after patching
+      0 8 neg /rbp :andqImm8MemDisp8
 
       [
         /rax :pushqReg
diff --git a/elymas/lib/sys/opt.ey b/elymas/lib/sys/opt.ey
index c15d0d5..009b6fa 100644
--- a/elymas/lib/sys/opt.ey
+++ b/elymas/lib/sys/opt.ey
@@ -360,6 +360,9 @@
   ] { | ::rawCodeAddress } [ 0 ] [ 0 ] '' * ==:STATICDEFININGFUNCTIONS
 
   { ==o ==executingScope
+    [ ] ==fragileReferences # things the GC should leave alone while we are working on it
+    { [ -01 ] fragileReferences cat =fragileReferences } /protectReference deffst
+
     0 ==containsScopeModifications
     { _ ==logic
       logic { ==entry 0 entry * ==action
@@ -1670,9 +1673,11 @@
                   executedObject sys .capturedScope { ==scope
                     [ NOP ] i 1 sub logic =[]
                     [ CALLSCOPED executedObject ::rawCodeAddress scope ::rawAddress ] i logic =[]
+                    executedObject ::rawCodeAddress 16 sub ::rawObject protectReference
                   } {
                     [ NOP ] i 1 sub logic =[]
                     [ CALL executedObject ::rawCodeAddress ] i logic =[]
+                    executedObject ::rawCodeAddress 16 sub ::rawObject protectReference
                   } ? *
                 } ? *
               }
@@ -1721,6 +1726,7 @@
                   } {
                     [ NOP ] i 1 sub logic =[]
                     [ RAWCONSTREP executedObject ::rawCodeAddress ] i logic =[]
+                    executedObject ::rawCodeAddress 16 sub ::rawObject protectReference
                   } ? *
                 } ? *
               }
@@ -1787,7 +1793,6 @@
             [ NOP ] i 2 sub logic =[]
             [ NOP ] i 1 sub logic =[]
             [ RAWCONSTLOOPTHISSCOPE headerBody 16 add loopBody 16 add ] i logic =[]
-            /RAWCONSTLOOPTHISSCOPE dump
           }
         ] conds
       } each
@@ -1884,6 +1889,8 @@
               [ NOP ] i 2 sub logic =[]
               [ NOP ] i 1 sub logic =[]
               [ INLINEQUESTIONSTARUNSCOPEDCONST 1 thirdLast * ::rawObject ::rawCodeAddress 1 secondLast * ::rawObject ::rawCodeAddress ] i logic =[]
+              1 thirdLast * ::rawObject ::rawCodeAddress 16 sub ::rawObject protectReference
+              1 secondLast * ::rawObject ::rawCodeAddress 16 sub ::rawObject protectReference
             } ? *
           }
         ] conds
@@ -2072,11 +2079,14 @@
           [ s 2 add _ 8 add range peek each ] 256 math .unbase ==pushedConstant
 
           [ PUSH pushedConstant ] emitLogic
+          pushedConstant ::rawObject protectReference
         }
 
         { callConstantPattern match }' {
           [ s 2 add _ 8 add range peek each ] 256 math .unbase ==calledAddress
           calledAddress ==j
+          calledAddress 16 sub ::rawObject protectReference
+
           { { j }' { =j }' peek generalMatch }' /callTargetMatch deffst
           [
             { calledAddress 105553116266496 lt }' { # HEAPBASE # FIXME: should use global constant
@@ -2088,6 +2098,7 @@
 
               [ PUSH calledConstant ] emitLogic
               [ CALL "*" | ::rawCodeAddress ] emitLogic
+              calledConstant 16 sub ::rawObject protectReference
             }
 
             { constantNormalFunctionScopedUntypedPattern callTargetMatch }' {
@@ -2095,18 +2106,22 @@
               [ calledAddress 23 add _ 8 add range peek each ] 256 math .unbase ==finalAddress
 
               [ CALLSCOPED finalAddress functionScope ] emitLogic
+              finalAddress 16 sub ::rawObject protectReference
+              functionScope ::rawObject protectReference
             }
 
             { constantNormalFunctionUnscopedUntypedPattern callTargetMatch }' {
               [ calledAddress 2 add _ 8 add range peek each ] 256 math .unbase ==finalAddress
 
               [ CALL finalAddress ] emitLogic
+              finalAddress 16 sub ::rawObject protectReference
             }
 
             { constantPassivePattern callTargetMatch }' {
               [ calledAddress 3 add _ 8 add range peek each ] 256 math .unbase ==pushedConstant
 
               [ PUSH pushedConstant ] emitLogic
+              pushedConstant ::rawObject protectReference
             }
 
             { staticLoadPattern callTargetMatch }' {
@@ -2132,6 +2147,7 @@
                   [ calledAddress exampleObjectOffset add _ 8 add range peek each ] 256 math .unbase ::rawObject ==exampleObject
 
                   [ STATICTYPED offsetInScope parentCount 0 exampleObject ] emitLogic
+                  exampleObject protectReference
                 }
 
                 { staticLoadFromExtensionPattern callTargetMatch }' {
@@ -2140,6 +2156,7 @@
                   [ calledAddress exampleObjectOffset add _ 8 add range peek each ] 256 math .unbase ::rawObject ==exampleObject
                   
                   [ STATICTYPED offsetInScope parentCount 1 exampleObject ] emitLogic
+                  exampleObject protectReference
                 }
 
                 { 1 }' {
@@ -2172,12 +2189,14 @@
               [ calledAddress 9 add _ 8 add range peek each ] 256 math .unbase ==pushedConstant
 
               [ PUSH pushedConstant ] emitLogic
+              pushedConstant ::rawObject protectReference
             }
 
             { customFunctionObjectCreationHeaderPattern callTargetMatch }' {
               [ calledAddress 9 add _ 8 add range peek each ] 256 math .unbase ==functionBody # function code object address
 
               [ FUNCTIONCREATE calledAddress functionBody ] emitLogic
+              functionBody ::rawObject protectReference
             }
 
             { customFunctionHeaderPattern callTargetMatch }' {