JIT: Simplify internal GC tracking structures (#113071)

Track all integer registers for calls in `regPtrDsc`. This does not cost
any extra memory and it saves us from going back and forth between an
intermediate format. It also unblocks proper GC reporting for helper
calls that are GC reported with non-standard calling convention.

Author: jakobbotsch

src/coreclr/jit/emit.cpp

@@ -9983,7 +9983,6 @@ void emitter::emitStackPopLargeStk(BYTE* addr, bool isCall, unsigned char callIn
 
     unsigned argStkCnt;
     S_UINT16 argRecCnt(0); // arg count for ESP, ptr-arg count for EBP
-    unsigned gcrefRegs, byrefRegs;
 
 #ifdef JIT32_GCENCODER
     // For the general encoder, we always need to record calls, so we make this call
@@ -10025,26 +10024,19 @@ void emitter::emitStackPopLargeStk(BYTE* addr, bool isCall, unsigned char callIn
         return;
 #endif
 
-    // Do we have any interesting (i.e., callee-saved) registers live here?
+    // Do we have any interesting registers live here?
 
-    gcrefRegs = byrefRegs = 0;
+    unsigned gcrefRegs = emitThisGCrefRegs.GetIntRegSet() >> REG_INT_FIRST;
+    unsigned byrefRegs = emitThisByrefRegs.GetIntRegSet() >> REG_INT_FIRST;
 
-    // We make a bitmask whose bits correspond to callee-saved register indices (in the sequence
-    // of callee-saved registers only).
-    for (unsigned calleeSavedRegIdx = 0; calleeSavedRegIdx < CNT_CALL_GC_REGS; calleeSavedRegIdx++)
-    {
-        regMaskTP calleeSavedRbm = raRbmCalleeSaveOrder[calleeSavedRegIdx];
-        if (emitThisGCrefRegs & calleeSavedRbm)
-        {
-            gcrefRegs |= (1 << calleeSavedRegIdx);
-        }
-        if (emitThisByrefRegs & calleeSavedRbm)
-        {
-            byrefRegs |= (1 << calleeSavedRegIdx);
-        }
-    }
+    assert(regMaskTP::FromIntRegSet(SingleTypeRegSet(gcrefRegs << REG_INT_FIRST)) == emitThisGCrefRegs);
+    assert(regMaskTP::FromIntRegSet(SingleTypeRegSet(byrefRegs << REG_INT_FIRST)) == emitThisByrefRegs);
 
 #ifdef JIT32_GCENCODER
+    // x86 does not report GC refs/byrefs in return registers at call sites
+    gcrefRegs &= ~(1u << (REG_INTRET - REG_INT_FIRST));
+    byrefRegs &= ~(1u << (REG_INTRET - REG_INT_FIRST));
+
     // For the general encoder, we always have to record calls, so we don't take this early return.    /* Are there any
     // args to pop at this call site?
 

src/coreclr/jit/gcencode.cpp

@@ -3199,9 +3199,24 @@ size_t GCInfo::gcMakeRegPtrTable(BYTE* dest, int mask, const InfoHdr& header, un
 
                     callArgCnt = genRegPtrTemp->rpdPtrArg;
 
-                    unsigned gcrefRegMask = genRegPtrTemp->rpdCallGCrefRegs;
+                    unsigned gcrefRegMask = 0;
 
-                    byrefRegMask = genRegPtrTemp->rpdCallByrefRegs;
+                    byrefRegMask = 0;
+
+                    // The order here is fixed: it must agree with the order assumed in eetwain.
+                    // NB: x86 GC decoder does not report return registers at call sites.
+                    static const regNumber calleeSaveOrder[] = {REG_EDI, REG_ESI, REG_EBX, REG_EBP};
+                    for (unsigned i = 0; i < ArrLen(calleeSaveOrder); i++)
+                    {
+                        if ((genRegPtrTemp->rpdCallGCrefRegs & (1 << (calleeSaveOrder[i] - REG_INT_FIRST))) != 0)
+                        {
+                            gcrefRegMask |= 1u << i;
+                        }
+                        if ((genRegPtrTemp->rpdCallByrefRegs & (1 << (calleeSaveOrder[i] - REG_INT_FIRST))) != 0)
+                        {
+                            byrefRegMask |= 1u << i;
+                        }
+                    }
 
                     assert((gcrefRegMask & byrefRegMask) == 0);
 
@@ -4465,8 +4480,8 @@ void GCInfo::gcMakeRegPtrTable(
             assert(call->u1.cdArgMask == 0 && call->cdArgCnt == 0);
 
             // Other than that, we just have to deal with the regmasks.
-            regMaskSmall gcrefRegMask = call->cdGCrefRegs & RBM_CALL_GC_REGS.GetIntRegSet();
-            regMaskSmall byrefRegMask = call->cdByrefRegs & RBM_CALL_GC_REGS.GetIntRegSet();
+            regMaskSmall gcrefRegMask = call->cdGCrefRegs;
+            regMaskSmall byrefRegMask = call->cdByrefRegs;
 
             assert((gcrefRegMask & byrefRegMask) == 0);
 
@@ -4552,11 +4567,8 @@ void GCInfo::gcMakeRegPtrTable(
                 {
                     // This is a true call site.
 
-                    regMaskSmall gcrefRegMask =
-                        genRegMaskFromCalleeSavedMask(genRegPtrTemp->rpdCallGCrefRegs).GetIntRegSet();
-
-                    regMaskSmall byrefRegMask =
-                        genRegMaskFromCalleeSavedMask(genRegPtrTemp->rpdCallByrefRegs).GetIntRegSet();
+                    regMaskSmall gcrefRegMask = regMaskSmall(genRegPtrTemp->rpdCallGCrefRegs << REG_INT_FIRST);
+                    regMaskSmall byrefRegMask = regMaskSmall(genRegPtrTemp->rpdCallByrefRegs << REG_INT_FIRST);
 
                     assert((gcrefRegMask & byrefRegMask) == 0);
 

src/coreclr/jit/jitgcinfo.h

@@ -162,7 +162,13 @@ class GCInfo
                 regMaskSmall rpdDel; // regptr bitset being removed
             } rpdCompiler;
 
-            unsigned short rpdPtrArg; // arg offset or popped arg count
+            struct
+            {
+                // Registers after call containing GC/byref (index 0 = REG_INT_FIRST)
+                unsigned int   rpdCallGCrefRegs;
+                unsigned int   rpdCallByrefRegs;
+                unsigned short rpdPtrArg; // arg offset or popped arg count
+            };
         };
 
 #ifndef JIT32_GCENCODER
@@ -182,11 +188,8 @@ class GCInfo
             return (GCtype)rpdGCtype;
         }
 
-        unsigned short rpdIsThis : 1;                       // is it the 'this' pointer
-        unsigned short rpdCall   : 1;                       // is this a true call site?
-        unsigned short           : 1;                       // Padding bit, so next two start on a byte boundary
-        unsigned short rpdCallGCrefRegs : CNT_CALL_GC_REGS; // Callee-saved and return registers containing GC pointers.
-        unsigned short rpdCallByrefRegs : CNT_CALL_GC_REGS; // Callee-saved and return registers containing byrefs.
+        unsigned short rpdIsThis : 1; // is it the 'this' pointer
+        unsigned short rpdCall   : 1; // is this a true call site?
 
 #ifndef JIT32_GCENCODER
         bool rpdIsCallInstr()

src/coreclr/jit/regset.cpp

@@ -944,27 +944,6 @@ regNumber genRegArgNext(regNumber argReg)
     }
 }
 
-/*****************************************************************************
- *
- *  The following table determines the order in which callee registers
- *  are encoded in GC information at call sites.
- */
-
-const regMaskTP raRbmCalleeSaveOrder[] = {RBM_CALL_GC_REGS_ORDER};
-
-regMaskTP genRegMaskFromCalleeSavedMask(unsigned short calleeSaveMask)
-{
-    regMaskTP res = 0;
-    for (int i = 0; i < CNT_CALL_GC_REGS; i++)
-    {
-        if ((calleeSaveMask & (1 << i)) != 0)
-        {
-            res |= raRbmCalleeSaveOrder[i];
-        }
-    }
-    return res;
-}
-
 /*****************************************************************************
  *
  *  Initializes the spill code. Should be called once per function compiled.

src/coreclr/jit/target.h

@@ -379,6 +379,11 @@ struct regMaskTP
 #endif
     }
 
+    static regMaskTP FromIntRegSet(SingleTypeRegSet intRegs)
+    {
+        return regMaskTP(intRegs);
+    }
+
     void operator|=(const regMaskTP& second)
     {
         low |= second.getLow();
@@ -1066,16 +1071,6 @@ inline SingleTypeRegSet getSingleTypeRegMask(regNumber reg, var_types regType)
     return regMask;
 }
 
-/*****************************************************************************
- *
- *  These arrays list the callee-saved register numbers (and bitmaps, respectively) for
- *  the current architecture.
- */
-extern const regMaskTP raRbmCalleeSaveOrder[CNT_CALL_GC_REGS];
-
-// This method takes a "compact" bitset of the callee-saved registers, and "expands" it to a full register mask.
-regMaskTP genRegMaskFromCalleeSavedMask(unsigned short);
-
 /*****************************************************************************
  *
  *  Assumes that "reg" is of the given "type". Return the next unused reg number after "reg"

src/coreclr/jit/targetamd64.h

@@ -300,24 +300,19 @@
 #ifdef UNIX_AMD64_ABI
   #define CNT_CALLEE_SAVED         (5 + REG_ETW_FRAMED_EBP_COUNT)
   #define CNT_CALLEE_ENREG         (CNT_CALLEE_SAVED)
-  #define CNT_CALL_GC_REGS         (CNT_CALLEE_SAVED + 2)
 
   #define CNT_CALLEE_TRASH_INT_INIT   (9)
   #define CNT_CALLEE_TRASH_HIGHINT    (8)
 
   #define CNT_CALLEE_SAVED_FLOAT   (0)
   #define CNT_CALLEE_TRASH_FLOAT_INIT (16)
   #define CNT_CALLEE_TRASH_HIGHFLOAT    (16)
-  /* NOTE: Sync with variable name defined in compiler.h */
-  #define RBM_CALL_GC_REGS_ORDER   RBM_EBX,RBM_ETW_FRAMED_EBP_LIST RBM_R12,RBM_R13,RBM_R14,RBM_R15,RBM_INTRET,RBM_INTRET_1
-  #define RBM_CALL_GC_REGS         (RBM_EBX|RBM_ETW_FRAMED_EBP|RBM_R12|RBM_R13|RBM_R14|RBM_R15|RBM_INTRET|RBM_INTRET_1)
 
   // For SysV we have more volatile registers so we do not save any callee saves for EnC.
   #define RBM_ENC_CALLEE_SAVED     0
 #else // !UNIX_AMD64_ABI
   #define CNT_CALLEE_SAVED         (7 + REG_ETW_FRAMED_EBP_COUNT)
   #define CNT_CALLEE_ENREG         (CNT_CALLEE_SAVED)
-  #define CNT_CALL_GC_REGS         (CNT_CALLEE_SAVED + 1)
 
   #define CNT_CALLEE_TRASH_INT_INIT   (7)
   #define CNT_CALLEE_TRASH_HIGHINT    (8)
@@ -326,9 +321,6 @@
   #define CNT_CALLEE_SAVED_FLOAT        (10)
   #define CNT_CALLEE_TRASH_FLOAT_INIT   (6)
   #define CNT_CALLEE_TRASH_HIGHFLOAT    (16)
-  /* NOTE: Sync with variable name defined in compiler.h */
-  #define RBM_CALL_GC_REGS_ORDER   RBM_EBX,RBM_ESI,RBM_EDI,RBM_ETW_FRAMED_EBP_LIST RBM_R12,RBM_R13,RBM_R14,RBM_R15,RBM_INTRET
-  #define RBM_CALL_GC_REGS         (RBM_EBX|RBM_ESI|RBM_EDI|RBM_ETW_FRAMED_EBP|RBM_R12|RBM_R13|RBM_R14|RBM_R15|RBM_INTRET)
 
   // Callee-preserved registers we always save and allow use of for EnC code, since there are quite few volatile registers.
   #define RBM_ENC_CALLEE_SAVED     (RBM_RSI | RBM_RDI)

src/coreclr/jit/targetarm.h

@@ -89,13 +89,9 @@
   #define RBM_LOW_REGS            (RBM_R0|RBM_R1|RBM_R2|RBM_R3|RBM_R4|RBM_R5|RBM_R6|RBM_R7)
   #define RBM_HIGH_REGS           (RBM_R8|RBM_R9|RBM_R10|RBM_R11|RBM_R12|RBM_SP|RBM_LR|RBM_PC)
 
-  #define RBM_CALL_GC_REGS_ORDER   RBM_R4,RBM_R5,RBM_R6,RBM_R7,RBM_R8,RBM_R9,RBM_R10,RBM_R11,RBM_INTRET
-  #define RBM_CALL_GC_REGS         (RBM_R4|RBM_R5|RBM_R6|RBM_R7|RBM_R8|RBM_R9|RBM_R10|RBM_R11|RBM_INTRET)
-
   #define CNT_CALLEE_SAVED        (8)
   #define CNT_CALLEE_TRASH        (6)
   #define CNT_CALLEE_ENREG        (CNT_CALLEE_SAVED-1)
-  #define CNT_CALL_GC_REGS        (CNT_CALLEE_SAVED+1)
 
   #define CNT_CALLEE_SAVED_FLOAT  (16)
   #define CNT_CALLEE_TRASH_FLOAT  (16)

src/coreclr/jit/targetarm64.h

@@ -108,13 +108,9 @@
                                    REG_V12, REG_V13, REG_V14, REG_V15, \
                                    REG_V3,  REG_V2, REG_V1,  REG_V0
 
-  #define RBM_CALL_GC_REGS_ORDER   RBM_R19,RBM_R20,RBM_R21,RBM_R22,RBM_R23,RBM_R24,RBM_R25,RBM_R26,RBM_R27,RBM_R28,RBM_INTRET,RBM_INTRET_1
-  #define RBM_CALL_GC_REGS         (RBM_R19|RBM_R20|RBM_R21|RBM_R22|RBM_R23|RBM_R24|RBM_R25|RBM_R26|RBM_R27|RBM_R28|RBM_INTRET|RBM_INTRET_1)
-
   #define CNT_CALLEE_SAVED        (11)
   #define CNT_CALLEE_TRASH        (17)
   #define CNT_CALLEE_ENREG        (CNT_CALLEE_SAVED-1)
-  #define CNT_CALL_GC_REGS        (CNT_CALLEE_SAVED+2)
 
   #define CNT_CALLEE_SAVED_FLOAT  (8)
   #define CNT_CALLEE_TRASH_FLOAT  (24)

src/coreclr/jit/targetx86.h

@@ -143,16 +143,9 @@
   #define REG_VAR_ORDER            REG_EAX,REG_EDX,REG_ECX,REG_ESI,REG_EDI,REG_EBX
   #define MAX_VAR_ORDER_SIZE       6
 
-  // The order here is fixed: it must agree with an order assumed in eetwain...
-  // NB: x86 GC decoder does not report return registers at call sites.
-  #define RBM_CALL_GC_REGS_ORDER   RBM_EDI,RBM_ESI,RBM_EBX,RBM_EBP
-  #define RBM_CALL_GC_REGS         (RBM_EDI|RBM_ESI|RBM_EBX|RBM_EBP)
-
   #define CNT_CALLEE_SAVED        (4)
   #define CNT_CALLEE_TRASH        (3)
   #define CNT_CALLEE_ENREG        (CNT_CALLEE_SAVED-1)
-  // NB: x86 GC decoder does not report return registers at call sites.
-  #define CNT_CALL_GC_REGS        (CNT_CALLEE_SAVED)
 
   #define CNT_CALLEE_SAVED_FLOAT  (0)
   #define CNT_CALLEE_TRASH_FLOAT  (6)