Handle more than 64 registers - The finale (#103387)

* Add predicate registers

* Increase REGNUM_BITS to 7

* Assign float registers if node is mask

* Remove some TODO-SVE present for predicate register implementation

* Make sure to use vector registers if there is no mask

* handle some more printing of predicate registers

* jit format

* try to fix gcc failure

* Revert "try to fix gcc failure"

This reverts commit 5452f6a.

* proper gcc-14 build error fix

src/coreclr/jit/codegenarm64test.cpp

@@ -6254,7 +6254,7 @@ void CodeGen::genArm64EmitterUnitTestsSve()
 
     // IF_SVE_CW_4A
     theEmitter->emitIns_R_R_R(INS_sve_mov, EA_SCALABLE, REG_V0, REG_P0, REG_V30, INS_OPTS_SCALABLE_H,
-                              INS_SCALABLE_OPTS_PREDICATE_MERGE_MOV); // MOV     <Zd>.<T>, <Pv>/M, <Zn>.<T>
+                              INS_SCALABLE_OPTS_PREDICATE_MERGE); // MOV     <Zd>.<T>, <Pv>/M, <Zn>.<T>
     theEmitter->emitIns_R_R_R_R(INS_sve_sel, EA_SCALABLE, REG_V29, REG_P15, REG_V28, REG_V4, INS_OPTS_SCALABLE_D,
                                 INS_SCALABLE_OPTS_UNPREDICATED); // SEL     <Zd>.<T>, <Pv>, <Zn>.<T>, <Zm>.<T>
     theEmitter->emitIns_R_R_R_R(INS_sve_sel, EA_SCALABLE, REG_V5, REG_P13, REG_V27, REG_V5, INS_OPTS_SCALABLE_S,

src/coreclr/jit/emit.h

@@ -758,7 +758,7 @@ class emitter
         // x86:         38 bits
         // amd64:       38 bits
         // arm:         32 bits
-        // arm64:       44 bits
+        // arm64:       46 bits
         // loongarch64: 28 bits
         // risc-v:      28 bits
 
@@ -828,7 +828,7 @@ class emitter
         // x86:         48 bits
         // amd64:       48 bits
         // arm:         48 bits
-        // arm64:       53 bits
+        // arm64:       55 bits
         // loongarch64: 46 bits
         // risc-v:      46 bits
 
@@ -840,7 +840,7 @@ class emitter
 #if defined(TARGET_ARM)
 #define ID_EXTRA_BITFIELD_BITS (16)
 #elif defined(TARGET_ARM64)
-#define ID_EXTRA_BITFIELD_BITS (21)
+#define ID_EXTRA_BITFIELD_BITS (23)
 #elif defined(TARGET_LOONGARCH64) || defined(TARGET_RISCV64)
 #define ID_EXTRA_BITFIELD_BITS (14)
 #elif defined(TARGET_XARCH)
@@ -881,7 +881,7 @@ class emitter
         // x86:         54/50 bits
         // amd64:       55/50 bits
         // arm:         54/50 bits
-        // arm64:       60/55 bits
+        // arm64:       62/57 bits
         // loongarch64: 53/48 bits
         // risc-v:      53/48 bits
 
@@ -897,7 +897,7 @@ class emitter
         // x86:         10/14 bits
         // amd64:       9/14 bits
         // arm:         10/14 bits
-        // arm64:        4/9 bits
+        // arm64:        2/7 bits
         // loongarch64: 11/16 bits
         // risc-v:      11/16 bits
 

src/coreclr/jit/emitarm64.cpp

@@ -4250,11 +4250,9 @@ void emitter::emitIns_Mov(
 
         case INS_sve_mov:
         {
-            // TODO-SVE: Remove check for insOptsNone() when predicate registers
-            //  are present.
-            if (insOptsNone(opt) && isPredicateRegister(dstReg) && isPredicateRegister(srcReg))
+            if (isPredicateRegister(dstReg) && isPredicateRegister(srcReg))
             {
-                // assert(insOptsNone(opt));
+                assert(insOptsNone(opt));
 
                 opt  = INS_OPTS_SCALABLE_B;
                 attr = EA_SCALABLE;

src/coreclr/jit/emitarm64sve.cpp

@@ -3786,9 +3786,7 @@ void emitter::emitInsSve_R_R_R(instruction     ins,
                 // MOV is an alias for CPY, and is always the preferred disassembly.
                 ins = INS_sve_mov;
             }
-            // TODO-SVE: Change the below check to INS_SCALABLE_OPTS_PREDICATE_MERGE
-            // once predicate registers are present.
-            else if (sopt == INS_SCALABLE_OPTS_PREDICATE_MERGE_MOV)
+            else if (sopt == INS_SCALABLE_OPTS_PREDICATE_MERGE)
             {
                 assert(isVectorRegister(reg1));
                 assert(isPredicateRegister(reg2));
@@ -5909,7 +5907,7 @@ void emitter::emitInsSve_R_R_R_R(instruction     ins,
                 {
                     // mov is a preferred alias for sel
                     return emitInsSve_R_R_R(INS_sve_mov, attr, reg1, reg2, reg3, opt,
-                                            INS_SCALABLE_OPTS_PREDICATE_MERGE_MOV);
+                                            INS_SCALABLE_OPTS_PREDICATE_MERGE);
                 }
 
                 assert(insOptsScalableStandard(opt));

src/coreclr/jit/instr.h

@@ -386,7 +386,6 @@ enum insScalableOpts : unsigned
     INS_SCALABLE_OPTS_TO_PREDICATE,      // Variants moving to a predicate from a vector (e.g. pmov)
     INS_SCALABLE_OPTS_TO_VECTOR,         // Variants moving to a vector from a predicate (e.g. pmov)
     INS_SCALABLE_OPTS_BROADCAST,         // Used to distinguish mov from cpy, where mov is an alias for both
-    INS_SCALABLE_OPTS_PREDICATE_MERGE_MOV, // Use to distinguish mov (predicated) from other variants
 };
 
 // Maps directly to the pattern used in SVE instructions such as cntb.

src/coreclr/jit/lsra.cpp

@@ -11518,7 +11518,13 @@ void LinearScan::dumpRegRecordTitleIfNeeded()
     if ((lastDumpedRegisters != registersToDump) || (rowCountSinceLastTitle > MAX_ROWS_BETWEEN_TITLES))
     {
         lastUsedRegNumIndex = 0;
-        int lastRegNumIndex = compiler->compFloatingPointUsed ? REG_FP_LAST : REG_INT_LAST;
+        int lastRegNumIndex = compiler->compFloatingPointUsed ?
+#ifdef HAS_MORE_THAN_64_REGISTERS
+                                                              REG_MASK_LAST
+#else
+                                                              REG_FP_LAST
+#endif
+                                                              : REG_INT_LAST;
         for (int regNumIndex = 0; regNumIndex <= lastRegNumIndex; regNumIndex++)
         {
             if (registersToDump.IsRegNumInMask((regNumber)regNumIndex))
@@ -12129,7 +12135,7 @@ void LinearScan::verifyFinalAllocation()
 
             case RefTypeKill:
                 dumpLsraAllocationEvent(LSRA_EVENT_KILL_REGS, nullptr, REG_NA, currentBlock, NONE,
-                                        currentRefPosition.registerAssignment);
+                                        currentRefPosition.getKillRegisterAssignment());
                 break;
 
             case RefTypeFixedReg:

src/coreclr/jit/lsraarm64.cpp

@@ -1588,30 +1588,37 @@ int LinearScan::BuildHWIntrinsic(GenTreeHWIntrinsic* intrinsicTree, int* pDstCou
         }
         else if (HWIntrinsicInfo::IsMaskedOperation(intrin.id))
         {
-            SingleTypeRegSet predMask = RBM_ALLMASK.GetPredicateRegSet();
-            if (intrin.id == NI_Sve_ConditionalSelect)
+            if (!varTypeIsMask(intrin.op1->TypeGet()) && !HWIntrinsicInfo::IsExplicitMaskedOperation(intrin.id))
             {
-                // If this is conditional select, make sure to check the embedded
-                // operation to determine the predicate mask.
-                assert(intrinsicTree->GetOperandCount() == 3);
-                assert(!HWIntrinsicInfo::IsLowMaskedOperation(intrin.id));
-
-                if (intrin.op2->OperIs(GT_HWINTRINSIC))
+                srcCount += BuildOperandUses(intrin.op1);
+            }
+            else
+            {
+                SingleTypeRegSet predMask = RBM_ALLMASK.GetPredicateRegSet();
+                if (intrin.id == NI_Sve_ConditionalSelect)
                 {
-                    GenTreeHWIntrinsic* embOp2Node = intrin.op2->AsHWIntrinsic();
-                    const HWIntrinsic   intrinEmb(embOp2Node);
-                    if (HWIntrinsicInfo::IsLowMaskedOperation(intrinEmb.id))
+                    // If this is conditional select, make sure to check the embedded
+                    // operation to determine the predicate mask.
+                    assert(intrinsicTree->GetOperandCount() == 3);
+                    assert(!HWIntrinsicInfo::IsLowMaskedOperation(intrin.id));
+
+                    if (intrin.op2->OperIs(GT_HWINTRINSIC))
                     {
-                        predMask = RBM_LOWMASK.GetPredicateRegSet();
+                        GenTreeHWIntrinsic* embOp2Node = intrin.op2->AsHWIntrinsic();
+                        const HWIntrinsic   intrinEmb(embOp2Node);
+                        if (HWIntrinsicInfo::IsLowMaskedOperation(intrinEmb.id))
+                        {
+                            predMask = RBM_LOWMASK.GetPredicateRegSet();
+                        }
                     }
                 }
-            }
-            else if (HWIntrinsicInfo::IsLowMaskedOperation(intrin.id))
-            {
-                predMask = RBM_LOWMASK.GetPredicateRegSet();
-            }
+                else if (HWIntrinsicInfo::IsLowMaskedOperation(intrin.id))
+                {
+                    predMask = RBM_LOWMASK.GetPredicateRegSet();
+                }
 
-            srcCount += BuildOperandUses(intrin.op1, predMask);
+                srcCount += BuildOperandUses(intrin.op1, predMask);
+            }
         }
         else if (intrinsicTree->OperIsMemoryLoadOrStore())
         {

src/coreclr/jit/registerarm64.h

@@ -94,28 +94,31 @@ REGDEF(V29,  29+VBASE, VMASK(29), "d29", "s29")
 REGDEF(V30,  30+VBASE, VMASK(30), "d30", "s30")
 REGDEF(V31,  31+VBASE, VMASK(31), "d31", "s31")
 
-// TODO-SVE: Fix once we add predicate registers
-REGALIAS(P0, V0)
-REGALIAS(P1, V1)
-REGALIAS(P2, V2)
-REGALIAS(P3, V3)
-REGALIAS(P4, V4)
-REGALIAS(P5, V5)
-REGALIAS(P6, V6)
-REGALIAS(P7, V7)
-REGALIAS(P8, V8)
-REGALIAS(P9, V9)
-REGALIAS(P10, V10)
-REGALIAS(P11, V11)
-REGALIAS(P12, V12)
-REGALIAS(P13, V13)
-REGALIAS(P14, V14)
-REGALIAS(P15, V15)
+#define PBASE 64
+#define PMASK(x) (1ULL << x)
 
+/*
+REGDEF(name,  rnum,         mask,  xname,  wname) */
+REGDEF(P0,    0+PBASE,  PMASK(0),  "p0" ,  "na")
+REGDEF(P1,    1+PBASE,  PMASK(1),  "p1" ,  "na")
+REGDEF(P2,    2+PBASE,  PMASK(2),  "p2" ,  "na")
+REGDEF(P3,    3+PBASE,  PMASK(3),  "p3" ,  "na")
+REGDEF(P4,    4+PBASE,  PMASK(4),  "p4" ,  "na")
+REGDEF(P5,    5+PBASE,  PMASK(5),  "p5" ,  "na")
+REGDEF(P6,    6+PBASE,  PMASK(6),  "p6" ,  "na")
+REGDEF(P7,    7+PBASE,  PMASK(7),  "p7" ,  "na")
+REGDEF(P8,    8+PBASE,  PMASK(8),  "p8" ,  "na")
+REGDEF(P9,    9+PBASE,  PMASK(9),  "p9" ,  "na")
+REGDEF(P10,  10+PBASE, PMASK(10),  "p10",  "na")
+REGDEF(P11,  11+PBASE, PMASK(11),  "p11",  "na")
+REGDEF(P12,  12+PBASE, PMASK(12),  "p12",  "na")
+REGDEF(P13,  13+PBASE, PMASK(13),  "p13",  "na")
+REGDEF(P14,  14+PBASE, PMASK(14),  "p14",  "na")
+REGDEF(P15,  15+PBASE, PMASK(15),  "p15",  "na")
 
 
-// The registers with values 64 (NBASE) and above are not real register numbers
-#define NBASE 64
+// The registers with values 80 (NBASE) and above are not real register numbers
+#define NBASE 80
 
 REGDEF(SP,    0+NBASE, 0x0000,    "sp",  "wsp?")
 // This must be last!

src/coreclr/jit/target.h

@@ -230,7 +230,7 @@ typedef uint64_t regMaskSmall;
 #endif
 
 #ifdef TARGET_ARM64
-// #define HAS_MORE_THAN_64_REGISTERS 1
+#define HAS_MORE_THAN_64_REGISTERS 1
 #endif // TARGET_ARM64
 
 // TODO: Rename regMaskSmall as RegSet64 (at least for 64-bit)
@@ -267,14 +267,7 @@ struct regMaskTP
     static constexpr regMaskTP CreateFromRegNum(regNumber reg, regMaskSmall mask)
     {
 #ifdef HAS_MORE_THAN_64_REGISTERS
-        if (reg < 64)
-        {
-            return regMaskTP(mask, RBM_NONE);
-        }
-        else
-        {
-            return regMaskTP(RBM_NONE, mask);
-        }
+        return (reg < 64) ? regMaskTP(mask, RBM_NONE) : regMaskTP(RBM_NONE, mask);
 #else
         return regMaskTP(mask, RBM_NONE);
 #endif

src/coreclr/jit/targetarm64.h

@@ -60,7 +60,7 @@
 
   static_assert_no_msg(REG_PREDICATE_HIGH_LAST == REG_PREDICATE_LAST);
 
-  #define REGNUM_BITS              6       // number of bits in a REG_*
+  #define REGNUM_BITS              7       // number of bits in a REG_*
   #define REGSIZE_BYTES            8       // number of bytes in one general purpose register
   #define FP_REGSIZE_BYTES         16      // number of bytes in one FP/SIMD register
   #define FPSAVE_REGSIZE_BYTES     8       // number of bytes in one FP/SIMD register that are saved/restored, for callee-saved registers

src/coreclr/jit/utils.cpp

@@ -308,7 +308,7 @@ const char* dspRegRange(regMaskTP regMask, size_t& minSiz, const char* sep, regN
     {
         regMaskTP regBit = genRegMask(regNum);
 
-        if ((regMask & regBit) != 0)
+        if ((regMask & regBit).IsNonEmpty())
         {
             // We have a register to display. It gets displayed now if:
             // 1. This is the first register to display of a new range of registers (possibly because