kernel: turn global GAP and module states back into static globals

... at least in plain GAP; keep them as before in HPC-GAP.

The whole concept was invented to allow HPC-GAP to have thread
local GAP and module state. In plain GAP, though, it introduced
some minor overhead.

Changing this here allows us to avoid the overhead, and it also
prevents GAPState from becoming larger than 32kb. This can happen
on platforms with a large jmp_buf (e.g., powerpc64).

Author: fingolfin

src/code.c

@@ -44,12 +44,15 @@
 
 GAP_STATIC_ASSERT(sizeof(StatHeader) == 8, "StatHeader has wrong size");
 
+#ifdef HPCGAP
 struct CodeModuleState {
-    Bag StackStat;
-    Int CountStat;
+#endif
+DECL_MODULE_STATE Bag StackStat;
+DECL_MODULE_STATE Int CountStat;
 
-    Bag StackExpr;
-    Int CountExpr;
+DECL_MODULE_STATE Bag StackExpr;
+DECL_MODULE_STATE Int CountExpr;
+#ifdef HPCGAP
 };
 
 static ModuleStateOffset CodeStateOffset = -1;
@@ -60,6 +63,9 @@ extern inline struct CodeModuleState * CShelper(void)
 }
 
 #define CS(x) (CShelper()->x)
+#else
+#define CS(x) (x)
+#endif
 
 
 /****************************************************************************
@@ -3250,8 +3256,10 @@ static StructInitInfo module = {
     .preSave = PreSave,
     .postRestore = PostRestore,
 
+#ifdef HPCGAP
     .moduleStateSize = sizeof(struct CodeModuleState),
     .moduleStateOffsetPtr = &CodeStateOffset,
+#endif
     .initModuleState = InitModuleState,
 };
 

src/collectors.cc

@@ -46,15 +46,19 @@ extern "C" {
 *F * * * * * * * * * * * * * module specific state  * * * * * * * * * * * * *
 */
 
+#ifdef HPCGAP
 struct CollectorsState_ {
-    Obj  SC_NW_STACK;
-    Obj  SC_LW_STACK;
-    Obj  SC_PW_STACK;
-    Obj  SC_EW_STACK;
-    Obj  SC_GE_STACK;
-    Obj  SC_CW_VECTOR;
-    Obj  SC_CW2_VECTOR;
-    UInt SC_MAX_STACK_SIZE;
+#endif
+
+DECL_MODULE_STATE Obj  SC_NW_STACK;
+DECL_MODULE_STATE Obj  SC_LW_STACK;
+DECL_MODULE_STATE Obj  SC_PW_STACK;
+DECL_MODULE_STATE Obj  SC_EW_STACK;
+DECL_MODULE_STATE Obj  SC_GE_STACK;
+DECL_MODULE_STATE Obj  SC_CW_VECTOR;
+DECL_MODULE_STATE Obj  SC_CW2_VECTOR;
+DECL_MODULE_STATE UInt SC_MAX_STACK_SIZE;
+#ifdef HPCGAP
 };
 
 static ModuleStateOffset CollectorsStateOffset = -1;
@@ -64,6 +68,17 @@ extern inline struct CollectorsState_ * CollectorsState(void)
     return (struct CollectorsState_ *)StateSlotsAtOffset(CollectorsStateOffset);
 }
 
+#define SC_NW_STACK (CollectorsState()->SC_NW_STACK)
+#define SC_LW_STACK (CollectorsState()->SC_LW_STACK)
+#define SC_PW_STACK (CollectorsState()->SC_PW_STACK)
+#define SC_EW_STACK (CollectorsState()->SC_EW_STACK)
+#define SC_GE_STACK (CollectorsState()->SC_GE_STACK)
+#define SC_CW_VECTOR (CollectorsState()->SC_CW_VECTOR)
+#define SC_CW2_VECTOR (CollectorsState()->SC_CW2_VECTOR)
+#define SC_MAX_STACK_SIZE (CollectorsState()->SC_MAX_STACK_SIZE)
+
+#endif
+
 
 /****************************************************************************
 **
@@ -84,7 +99,7 @@ extern inline struct CollectorsState_ * CollectorsState(void)
 */
 #define SC_PUSH_WORD( word, exp ) \
     if ( ++sp == max ) { \
-        CollectorsState()->SC_MAX_STACK_SIZE *= 2; \
+        SC_MAX_STACK_SIZE *= 2; \
         return -1; \
     } \
     *++nw = DATA_WORD(word); \
@@ -95,7 +110,7 @@ extern inline struct CollectorsState_ * CollectorsState(void)
 
 #define SC_PUSH_GEN( gen, exp ) \
     if ( ++sp == max ) { \
-        CollectorsState()->SC_MAX_STACK_SIZE *= 2; \
+        SC_MAX_STACK_SIZE *= 2; \
         return -1; \
     } \
     *++nw = DATA_WORD(gen); \
@@ -385,22 +400,22 @@ static Int SingleCollectWord(Obj sc, Obj vv, Obj w)
     exps = (UInt)1 << (ebits-1);
 
     // <nw> contains the stack of words to insert
-    vnw = CollectorsState()->SC_NW_STACK;
+    vnw = SC_NW_STACK;
 
     // <lw> contains the word end of the word in <nw>
-    vlw = CollectorsState()->SC_LW_STACK;
+    vlw = SC_LW_STACK;
 
     // <pw> contains the position of the word in <nw> to look at
-    vpw = CollectorsState()->SC_PW_STACK;
+    vpw = SC_PW_STACK;
 
     // <ew> contains the unprocessed exponents at position <pw>
-    vew = CollectorsState()->SC_EW_STACK;
+    vew = SC_EW_STACK;
 
     // <ge> contains the global exponent of the word
-    vge = CollectorsState()->SC_GE_STACK;
+    vge = SC_GE_STACK;
 
     // get the maximal stack size
-    max = CollectorsState()->SC_MAX_STACK_SIZE;
+    max = SC_MAX_STACK_SIZE;
 
     // ensure that the stacks are large enough
     const UInt desiredStackSize = sizeof(Obj) * (max + 2);
@@ -893,22 +908,22 @@ static Int CombiCollectWord(Obj sc, Obj vv, Obj w)
     exps = (UInt)1 << (ebits-1);
 
     // <nw> contains the stack of words to insert
-    vnw = CollectorsState()->SC_NW_STACK;
+    vnw = SC_NW_STACK;
 
     // <lw> contains the word end of the word in <nw>
-    vlw = CollectorsState()->SC_LW_STACK;
+    vlw = SC_LW_STACK;
 
     // <pw> contains the position of the word in <nw> to look at
-    vpw = CollectorsState()->SC_PW_STACK;
+    vpw = SC_PW_STACK;
 
     // <ew> contains the unprocessed exponents at position <pw>
-    vew = CollectorsState()->SC_EW_STACK;
+    vew = SC_EW_STACK;
 
     // <ge> contains the global exponent of the word
-    vge = CollectorsState()->SC_GE_STACK;
+    vge = SC_GE_STACK;
 
     // get the maximal stack size
-    max = CollectorsState()->SC_MAX_STACK_SIZE;
+    max = SC_MAX_STACK_SIZE;
 
     // ensure that the stacks are large enough
     const UInt desiredStackSize = sizeof(Obj) * (max + 2);
@@ -1267,7 +1282,7 @@ static Obj ReducedComm(FinPowConjCol * fc, Obj sc, Obj w, Obj u)
     Obj                 vc2;        // collect vector
 
     // use 'cwVector' to collect word <u>*<w> to
-    vcw = CollectorsState()->SC_CW_VECTOR;
+    vcw = SC_CW_VECTOR;
     num = SC_NUMBER_RWS_GENERATORS(sc);
 
     // check that it has the correct length, unpack <u> into it
@@ -1283,7 +1298,7 @@ static Obj ReducedComm(FinPowConjCol * fc, Obj sc, Obj w, Obj u)
     }
 
     // use 'cw2Vector' to collect word <w>*<u> to
-    vc2 = CollectorsState()->SC_CW2_VECTOR;
+    vc2 = SC_CW2_VECTOR;
 
     // check that it has the correct length, unpack <w> into it
     if ( fc->vectorWord( vc2, w, num ) == -1 ) {
@@ -1325,7 +1340,7 @@ static Obj ReducedForm(FinPowConjCol * fc, Obj sc, Obj w)
     Obj                 type;   // type of the return objue
 
     // use 'cwVector' to collect word <w> to
-    vcw = CollectorsState()->SC_CW_VECTOR;
+    vcw = SC_CW_VECTOR;
     num = SC_NUMBER_RWS_GENERATORS(sc);
 
     // check that it has the correct length
@@ -1358,7 +1373,7 @@ static Obj ReducedLeftQuotient(FinPowConjCol * fc, Obj sc, Obj w, Obj u)
     Obj                 vc2;        // collect vector
 
     // use 'cwVector' to collect word <w> to
-    vcw = CollectorsState()->SC_CW_VECTOR;
+    vcw = SC_CW_VECTOR;
     num = SC_NUMBER_RWS_GENERATORS(sc);
 
     // check that it has the correct length, unpack <w> into it
@@ -1368,7 +1383,7 @@ static Obj ReducedLeftQuotient(FinPowConjCol * fc, Obj sc, Obj w, Obj u)
     }
 
     // use 'cw2Vector' to collect word <u> to
-    vc2 = CollectorsState()->SC_CW2_VECTOR;
+    vc2 = SC_CW2_VECTOR;
 
     // check that it has the correct length, unpack <u> into it
     if ( fc->vectorWord( vc2, u, num ) == -1 ) {
@@ -1402,7 +1417,7 @@ static Obj ReducedProduct(FinPowConjCol * fc, Obj sc, Obj w, Obj u)
     Obj                 vcw;        // collect vector
 
     // use 'cwVector' to collect word <w> to
-    vcw = CollectorsState()->SC_CW_VECTOR;
+    vcw = SC_CW_VECTOR;
     num = SC_NUMBER_RWS_GENERATORS(sc);
 
     // check that it has the correct length, unpack <w> into it
@@ -1441,8 +1456,8 @@ static Obj ReducedPowerSmallInt(FinPowConjCol * fc, Obj sc, Obj w, Obj vpow)
     pow = INT_INTOBJ(vpow);
 
     // use 'cwVector' and 'cw2Vector to collect words to
-    vcw  = CollectorsState()->SC_CW_VECTOR;
-    vc2  = CollectorsState()->SC_CW2_VECTOR;
+    vcw  = SC_CW_VECTOR;
+    vc2  = SC_CW2_VECTOR;
     num  = SC_NUMBER_RWS_GENERATORS(sc);
     type = SC_DEFAULT_TYPE(sc);
 
@@ -1528,8 +1543,8 @@ static Obj ReducedQuotient(FinPowConjCol * fc, Obj sc, Obj w, Obj u)
     Obj                 vc2;        // collect vector
 
     // use 'cwVector' to collect word <w> to
-    vcw  = CollectorsState()->SC_CW_VECTOR;
-    vc2  = CollectorsState()->SC_CW2_VECTOR;
+    vcw  = SC_CW_VECTOR;
+    vc2  = SC_CW2_VECTOR;
     num  = SC_NUMBER_RWS_GENERATORS(sc);
     type = SC_DEFAULT_TYPE(sc);
 
@@ -1647,8 +1662,7 @@ static Obj FuncFinPowConjCol_ReducedQuotient ( Obj self, Obj sc, Obj w, Obj u )
 */
 static Obj FuncSET_SCOBJ_MAX_STACK_SIZE(Obj self, Obj size)
 {
-    CollectorsState()->SC_MAX_STACK_SIZE =
-        GetPositiveSmallInt(SELF_NAME, size);
+    SC_MAX_STACK_SIZE = GetPositiveSmallInt(SELF_NAME, size);
     return 0;
 }
 
@@ -1739,25 +1753,25 @@ static Int InitLibrary (
 static Int InitModuleState(void)
 {
     // register global bags with the garbage collector
-    InitGlobalBag( &CollectorsState()->SC_NW_STACK, "SC_NW_STACK" );
-    InitGlobalBag( &CollectorsState()->SC_LW_STACK, "SC_LW_STACK" );
-    InitGlobalBag( &CollectorsState()->SC_PW_STACK, "SC_PW_STACK" );
-    InitGlobalBag( &CollectorsState()->SC_EW_STACK, "SC_EW_STACK" );
-    InitGlobalBag( &CollectorsState()->SC_GE_STACK, "SC_GE_STACK" );
-    InitGlobalBag( &CollectorsState()->SC_CW_VECTOR, "SC_CW_VECTOR" );
-    InitGlobalBag( &CollectorsState()->SC_CW2_VECTOR, "SC_CW2_VECTOR" );
+    InitGlobalBag(&SC_NW_STACK, "SC_NW_STACK");
+    InitGlobalBag(&SC_LW_STACK, "SC_LW_STACK");
+    InitGlobalBag(&SC_PW_STACK, "SC_PW_STACK");
+    InitGlobalBag(&SC_EW_STACK, "SC_EW_STACK");
+    InitGlobalBag(&SC_GE_STACK, "SC_GE_STACK");
+    InitGlobalBag(&SC_CW_VECTOR, "SC_CW_VECTOR");
+    InitGlobalBag(&SC_CW2_VECTOR, "SC_CW2_VECTOR");
 
     const UInt maxStackSize = 256;
     const UInt desiredStackSize = sizeof(Obj) * (maxStackSize + 2);
-    CollectorsState()->SC_NW_STACK = NewKernelBuffer(desiredStackSize);
-    CollectorsState()->SC_LW_STACK = NewKernelBuffer(desiredStackSize);
-    CollectorsState()->SC_PW_STACK = NewKernelBuffer(desiredStackSize);
-    CollectorsState()->SC_EW_STACK = NewKernelBuffer(desiredStackSize);
-    CollectorsState()->SC_GE_STACK = NewKernelBuffer(desiredStackSize);
+    SC_NW_STACK = NewKernelBuffer(desiredStackSize);
+    SC_LW_STACK = NewKernelBuffer(desiredStackSize);
+    SC_PW_STACK = NewKernelBuffer(desiredStackSize);
+    SC_EW_STACK = NewKernelBuffer(desiredStackSize);
+    SC_GE_STACK = NewKernelBuffer(desiredStackSize);
 
-    CollectorsState()->SC_CW_VECTOR = NEW_STRING(0);
-    CollectorsState()->SC_CW2_VECTOR = NEW_STRING(0);
-    CollectorsState()->SC_MAX_STACK_SIZE = maxStackSize;
+    SC_CW_VECTOR = NEW_STRING(0);
+    SC_CW2_VECTOR = NEW_STRING(0);
+    SC_MAX_STACK_SIZE = maxStackSize;
 
     return 0;
 }
@@ -1779,8 +1793,13 @@ static StructInitInfo module = {
  /* preSave     = */ 0,
  /* postSave    = */ 0,
  /* postRestore = */ 0,
+#ifdef HPCGAP
  /* moduleStateSize      = */ sizeof(CollectorsState_),
  /* moduleStateOffsetPtr = */ &CollectorsStateOffset,
+#else
+ /* moduleStateSize      = */ 0,
+ /* moduleStateOffsetPtr = */ 0,
+#endif
  /* initModuleState      = */ InitModuleState,
  /* destroyModuleState   = */ 0,
 };

src/cyclotom.c

@@ -142,10 +142,11 @@ static inline void SET_NOF_CYC(Obj cyc, Obj val)
 
 // #define XXX_CYC(cyc,len)        (EXPOS_CYC(cyc,len)[0])
 
-
+#ifdef HPCGAP
 static ModuleStateOffset CycStateOffset = -1;
 
 struct CycModuleState {
+#endif
 
 /****************************************************************************
 **
@@ -158,7 +159,7 @@ struct CycModuleState {
 **  It is created in 'InitCyc' with room for up to 1000 coefficients  and  is
 **  resized when need arises.
 */
-Obj ResultCyc;
+DECL_MODULE_STATE Obj ResultCyc;
 
 /****************************************************************************
 **
@@ -178,9 +179,10 @@ Obj ResultCyc;
 **  is called to compute $e_n^i$ and can then do this easier by just  putting
 **  1 at the <i>th place in 'ResultCyc' and then calling 'Cyclotomic'.
 */
-Obj  LastECyc;
-UInt LastNCyc;
+DECL_MODULE_STATE Obj  LastECyc;
+DECL_MODULE_STATE UInt LastNCyc;
 
+#ifdef HPCGAP
 }; // end of struct CycModuleState
 
 extern inline struct CycModuleState *CycState(void)
@@ -189,10 +191,10 @@ extern inline struct CycModuleState *CycState(void)
 }
 
 // For convenience and readability
-#define ResultCyc   CycState()->ResultCyc
-#define LastECyc    CycState()->LastECyc
-#define LastNCyc    CycState()->LastNCyc
-
+#define ResultCyc   (CycState()->ResultCyc)
+#define LastECyc    (CycState()->LastECyc)
+#define LastNCyc    (CycState()->LastNCyc)
+#endif
 
 static void GrowResultCyc(UInt size)
 {
@@ -2201,8 +2203,10 @@ static StructInitInfo module = {
     .initKernel = InitKernel,
     .initLibrary = InitLibrary,
 
+#ifdef HPCGAP
     .moduleStateSize = sizeof(struct CycModuleState),
     .moduleStateOffsetPtr = &CycStateOffset,
+#endif
     .initModuleState = InitModuleState,
 };