Implement UnknownSizeFrame for locals with unknown size

src/coreclr/jit/codegen.h

@@ -430,6 +430,10 @@ class CodeGen final : public CodeGenInterface
 
     void genPushCalleeSavedRegisters(regNumber initReg, bool* pInitRegZeroed);
 
+#if defined(TARGET_ARM64)
+    void genUnknownSizeFrame();
+#endif
+
 #elif defined(TARGET_LOONGARCH64) || defined(TARGET_RISCV64)
     bool genInstrWithConstant(instruction ins,
                               emitAttr    attr,

src/coreclr/jit/codegenarmarch.cpp

@@ -4865,9 +4865,54 @@ void CodeGen::genPushCalleeSavedRegisters()
     m_compiler->compFrameInfo.calleeSaveSpOffset = calleeSaveSpOffset;
     m_compiler->compFrameInfo.calleeSaveSpDelta  = calleeSaveSpDelta;
     m_compiler->compFrameInfo.offsetSpToSavedFp  = offsetSpToSavedFp;
+
+    if (m_compiler->compUsesUnknownSizeFrame)
+    {
+        genUnknownSizeFrame();
+    }
 #endif // TARGET_ARM64
 }
 
+#if defined(TARGET_ARM64)
+/*****************************************************************************
+ *
+ * Generates code for creating the UnknownSizeFrame stack space.
+ *
+ * See Compiler::UnknownSizeFrame for implementation details. The space contains
+ * stack allocations for Vector<T>.
+ */
+
+void CodeGen::genUnknownSizeFrame()
+{
+    assert(m_compiler->compLocallocUsed && m_compiler->compUsesUnknownSizeFrame);
+    assert(m_compiler->unkSizeFrame.isFinalized);
+    unsigned totalVectorCount = m_compiler->unkSizeFrame.FrameSizeInVectors();
+
+    // We reserve REG_UNKBASE for addressing SVE locals. This will always point at the top of
+    // of the UnknownSizeFrame and we index into it.
+    // TODO-SVE: We may want this to point into the middle of the frame to reduce address
+    // computations (we have a signed 9-bit indexing immediate).
+    inst_Mov(TYP_I_IMPL, REG_UNKBASE, REG_SP, false);
+
+    if (0 < totalVectorCount && totalVectorCount <= 32)
+    {
+        GetEmitter()->emitIns_R_R_I(INS_sve_addvl, EA_8BYTE, REG_SP, REG_SP, -(ssize_t)totalVectorCount);
+    }
+    else
+    {
+        // Generate `sp = sp - totalVectorCount * VL`
+        assert(totalVectorCount != 0);
+        regNumber rsvd = rsGetRsvdReg();
+        // mov   rsvd, #totalVectorCount
+        // rdvl  scratch, #1
+        // msub  sp, rsvd, scratch, sp
+        instGen_Set_Reg_To_Imm(EA_8BYTE, rsvd, totalVectorCount);
+        GetEmitter()->emitIns_R_I(INS_sve_rdvl, EA_8BYTE, REG_SCRATCH, 1);
+        GetEmitter()->emitIns_R_R_R_R(INS_msub, EA_8BYTE, REG_SP, rsvd, REG_SCRATCH, REG_SP);
+    }
+}
+#endif
+
 /*****************************************************************************
  *
  *  Generates code for a function epilog.

src/coreclr/jit/codegencommon.cpp

@@ -3708,6 +3708,11 @@ void CodeGen::genCheckUseBlockInit()
             continue;
         }
 
+        if (m_compiler->lvaIsAllocatedOnUnknownSizeFrame(varNum))
+        {
+            continue;
+        }
+
         if (m_compiler->fgVarIsNeverZeroInitializedInProlog(varNum))
         {
             varDsc->lvMustInit = 0;
@@ -4065,6 +4070,12 @@ void CodeGen::genZeroInitFrame(int untrLclHi, int untrLclLo, regNumber initReg,
 
             noway_assert(varDsc->lvOnFrame);
 
+            if (m_compiler->lvaIsAllocatedOnUnknownSizeFrame(varNum))
+            {
+                // This local will belong on the UnknownSizeFrame, which will handle zeroing instead.
+                continue;
+            }
+
             // lvMustInit can only be set for GC types or TYP_STRUCT types
             // or when compInitMem is true
             // or when in debug code
@@ -4807,6 +4818,11 @@ void CodeGen::genFinalizeFrame()
             regSet.rsSetRegsModified(maskPairRegs);
         }
     }
+
+    if (m_compiler->compUsesUnknownSizeFrame)
+    {
+        regSet.rsSetRegsModified(RBM_UNKBASE);
+    }
 #endif
 
 #ifdef DEBUG
@@ -5131,6 +5147,11 @@ void CodeGen::genFnProlog()
             continue;
         }
 
+        if (m_compiler->lvaIsAllocatedOnUnknownSizeFrame(varNum))
+        {
+            continue;
+        }
+
         signed int loOffs = varDsc->GetStackOffset();
         signed int hiOffs = varDsc->GetStackOffset() + m_compiler->lvaLclStackHomeSize(varNum);
 

src/coreclr/jit/compiler.h

@@ -4304,6 +4304,186 @@ class Compiler
 
     int lvaOSRLocalTier0FrameOffset(unsigned varNum);
 
+    //------------------------- UnknownSizeFrame ---------------------------------
+
+    void lvaInitUnknownSizeFrame();
+    void lvaAllocUnknownSizeLocal(unsigned varNum);
+    bool lvaIsAllocatedOnUnknownSizeFrame(unsigned varNum);
+
+    bool compUsesUnknownSizeFrame;
+
+#if defined(FEATURE_SIMD) && defined(TARGET_ARM64)
+    // For ARM64, the UnknownSizeFrame lives at the end of the statically
+    // allocated stack space. This means it belongs to the 'alloca' space on the
+    // frame, and it is essentially the first dynamically allocated stack
+    // variable.
+    //
+    // Currently, the only locals with unknown size are SIMD types supporting
+    // Vector<T>, TYP_SIMD and TYP_MASK. We do not know the size of these types
+    // at compile time, so we need to execute the rdvl/addvl instruction to
+    // learn this size and allocate the UnknownSizeFrame.
+    //
+    // We reserve the x19 register to point to the top of the UnknownSizeFrame
+    // and use this as the base address for local variables with unknown size.
+    // Reserving a register is simpler than using fp/sp, as fp may point
+    // to different locations depending on various properties of the frame, and
+    // the value of sp may change at runtime.
+    //
+    // Typically, a vector is loaded using a base address and some index which
+    // the instruction will scale by VL, for example: `ldr z0, [x19, #3 MUL VL]`.
+    // A mask is loaded with `ldr p0, [x19, #3 MUL VL]`, but in this case the
+    // `MUL VL` indicates we are scaling with the length of the predicate
+    // register rather than the vector. A predicate register is defined to have
+    // 1/8th the length of a vector register.
+    //
+    // We know that sizeof(TYP_SIMD) and sizeof(TYP_MASK) are invariant despite
+    // being unknown at compile time, so we allocate them in single homogeneous
+    // blocks per type. An individual local can be referenced from the start of
+    // its block by an index into the block.
+    //
+    // The difference in addressing-mode index scaling means we have to be
+    // careful where we place the mask locals block with respect to the vector
+    // locals block. If we place the mask locals after the vector locals, we'll
+    // need to offset the load index by (8 * nVector) to account for the vector
+    // locals.
+    //
+    // Instead, we choose to pad the mask locals block to VL and place it at the
+    // beginning of the frame (closest to fp). This way we'll need to offset
+    // vector load indices by `roundUp(nMask, 8) / 8`. This is less likely to
+    // put pressure on the immediate encoding range and result in requiring an
+    // address computation.
+    //
+    // The maximum wasted space from the padding is 7/8ths VL (224 bytes with
+    // the architectural maximum 256 byte vectors), which occurs when 1 mask
+    // local is spilled to the frame. Alternatively this is 28 bytes for 32 byte
+    // vectors, for an example closer to today's implementations.
+    //
+    // The padding also makes it simple to allocate the UnknownSizeFrame since
+    // the UnknownSizeFrame will be aligned to VL. The total number of vectors
+    // to allocate is `(roundUp(nMask, 8) / 8) + nVector`. The stack pointer
+    // can be adjusted with a single instruction `addvl sp, sp, #totalVectors`.
+    //
+    // See the diagram below for a visual representation of this scheme.
+    //
+    //                 ...
+    //  |          static space            |
+    //  |        (totalFrameSize)          |
+    //  +----------------------------------+ x19, begin UnknownSizeFrame
+    //  |         mask locals block        |                 ^
+    //  |          (nMask * VL/8)          |                 |
+    //  +----------------------------------+                 |
+    //  |      padding to VL alignment     |                 |
+    //  +----------------------------------+ (roundUp(nMask, 8)/8 + nVector)*VL
+    //  |                                  |                 |
+    //  |       vector locals block        |                 |
+    //  |         (nVector * VL)           |                 |
+    //  |                                  |                 v
+    //  +----------------------------------+ end UnknownSizeFrame
+    //  |                                  |
+    //  |       rest of alloca space       |
+    //                 ...                   sp
+    struct UnknownSizeFrame
+    {
+        // Number of allocated vectors/masks. These also represent the end of
+        // the allocation space for each block. The allocator for each block is
+        // a simple bump allocator.
+        unsigned nVector = 0;
+        unsigned nMask = 0;
+
+#ifdef DEBUG
+        bool isFinalized = false;
+#endif
+
+        // Returns the size of the mask block in number of vector lengths.
+        unsigned MaskBlockSizeInVectors()
+        {
+            assert(roundUp(0U, 8U) == 0);
+            return roundUp(nMask, 8) / 8;
+        }
+
+        // Returns the size of the vector block in number of vector lengths.
+        unsigned VectorBlockSize()
+        {
+            return nVector;
+        }
+
+        // Returns the size of the total UnknownSizeFrame in number of vector
+        // lengths.
+        unsigned FrameSizeInVectors()
+        {
+            return MaskBlockSizeInVectors() + VectorBlockSize();
+        }
+
+        // Allocate a mask, returning an index of the mask in the mask block.
+        unsigned AllocMask()
+        {
+            assert(!isFinalized);
+            unsigned idx = nMask;
+            nMask++;
+            return idx;
+        }
+
+        // Allocate a vector, returning an index of the vector in the vector
+        // block.
+        unsigned AllocVector()
+        {
+            assert(!isFinalized);
+            unsigned idx = nVector;
+            nVector++;
+            return idx;
+        }
+
+        // Returns a negative offset relative to the base of the UnknownSizeFrame
+        // for addressing an allocated vector or mask local.
+        // If `isMask == true`, given an index that was assigned to mask local,
+        // the returned offset is an index measured in units of VL/8.
+        // Otherwise given an index that was assigned to a vector local, the
+        // returned offset is measured in units of VL.
+        // The index parameter should have been obtained through AllocMask() or
+        // AllocVector().
+        int GetOffset(unsigned index, bool isMask = false)
+        {
+            // We can't compute addresses if we haven't finished allocating.
+            assert(isFinalized);
+
+            unsigned offset = UINT32_MAX;
+            if (isMask)
+            {
+                assert(index < nMask);
+                offset = index;
+            }
+            else
+            {
+                assert(index < nVector);
+                offset = MaskBlockSizeInVectors() + index;
+            }
+            assert(offset != UINT32_MAX);
+            // The index is always offset by 1 as we are writing from below fp
+            // upwards.
+            return -(int)(offset + 1);
+        }
+
+        // Given a local on the UnknownSizeFrame, compute the offset used for addressing
+        // this local relative to the base address of the UnknownSizeFrame. This offset
+        // can be used with addvl/addpl for TYP_SIMD/TYP_MASK respectively. The offset
+        // needs to be scaled by VL/PL to produce an absolute address value.
+        int GetAddressingOffset(LclVarDsc* varDsc)
+        {
+            return GetOffset(varDsc->GetStackOffset(), varDsc->TypeIs(TYP_MASK));
+        }
+
+        // This system ensures we don't try and generate an address on the frame
+        // without finishing all allocations.
+        void Finalize()
+        {
+#ifdef DEBUG
+            isFinalized = true;
+#endif
+        }
+
+    } unkSizeFrame;
+#endif
+
     //------------------------ For splitting types ----------------------------
 
     void lvaInitTypeRef();

src/coreclr/jit/compiler.hpp

@@ -2721,6 +2721,7 @@ inline
     bool fConservative = false;
     if (varNum >= 0)
     {
+        assert(!lvaIsAllocatedOnUnknownSizeFrame(varNum));
         LclVarDsc* varDsc          = lvaGetDesc(varNum);
         bool       isPrespilledArg = false;
 #if defined(TARGET_ARM) && defined(PROFILING_SUPPORTED)
@@ -2779,6 +2780,7 @@ inline
                 tmpDsc = codeGen->regSet.tmpFindNum(varNum, RegSet::TEMP_USAGE_USED);
             }
             assert(tmpDsc != nullptr);
+            assert(!varTypeHasUnknownSize(tmpDsc->tdTempType()));
             varOffset = tmpDsc->tdTempOffs();
         }
         else