Arm64: keep SP in base operand for large-offset adds

src/coreclr/jit/codegenarm64.cpp

@@ -2151,17 +2151,18 @@ void CodeGen::instGen_Set_Reg_To_Base_Plus_Imm(emitAttr       size,
                                                insFlags flags DEBUGARG(size_t targetHandle)
                                                    DEBUGARG(GenTreeFlags gtFlags))
 {
-    // If the imm values < 12 bits, we can use a single "add rsvd, reg2, #imm".
-    // Otherwise, use "mov rsvd, #imm", followed up "add rsvd, reg2, rsvd".
+    // If the immediate can be encoded by add/sub, use a single instruction.
+    // Otherwise, use "mov dstReg, #imm", followed by "add dstReg, baseReg, dstReg".
+    // Keep baseReg as the second operand in the fallback since it can be SP.
 
-    if (imm < 4096)
+    if (emitter::emitIns_valid_imm_for_add(imm, size))
     {
-        GetEmitter()->emitIns_R_R_I(INS_add, EA_PTRSIZE, dstReg, baseReg, imm);
+        GetEmitter()->emitIns_R_R_I(INS_add, size, dstReg, baseReg, imm);
     }
     else
     {
         instGen_Set_Reg_To_Imm(size, dstReg, imm);
-        GetEmitter()->emitIns_R_R_R(INS_add, size, dstReg, dstReg, baseReg);
+        GetEmitter()->emitIns_R_R_R(INS_add, size, dstReg, baseReg, dstReg);
     }
 }
 

src/coreclr/jit/emitarm64.cpp

@@ -15017,8 +15017,8 @@ void emitter::emitInsLoadStoreOp(instruction ins, emitAttr attr, regNumber dataR
                     // First load/store tmpReg with the large offset constant
                     codeGen->instGen_Set_Reg_To_Imm(EA_PTRSIZE, tmpReg, offset);
                     // Then add the base register
-                    //      rd = rd + base
-                    emitIns_R_R_R(INS_add, addType, tmpReg, tmpReg, memBase->GetRegNum());
+                    //      rd = base + rd
+                    emitIns_R_R_R(INS_add, addType, tmpReg, memBase->GetRegNum(), tmpReg);
 
                     noway_assert(emitInsIsLoad(ins) || (tmpReg != dataReg));
                     noway_assert(tmpReg != index->GetRegNum());

src/tests/JIT/Regression/JitBlue/GitHub_130437/GitHub_130437.csproj

@@ -0,0 +1,14 @@
+<Project Sdk="Microsoft.NET.Sdk.IL">
+  <PropertyGroup>
+    <LanguageTargets>$(RepoRoot)src/coreclr/.nuget/Microsoft.NET.Sdk.IL/Microsoft.NET.Sdk.IL.targets.template</LanguageTargets>
+    <Optimize>True</Optimize>
+    <RequiresProcessIsolation>true</RequiresProcessIsolation>
+    <CLRTestTargetUnsupported Condition="'$(TargetArchitecture)' != 'arm64'">true</CLRTestTargetUnsupported>
+  </PropertyGroup>
+  <ItemGroup>
+    <Compile Include="$(MSBuildProjectName).il" />
+    <CLRTestEnvironmentVariable Include="DOTNET_TieredCompilation" Value="0" />
+    <CLRTestEnvironmentVariable Include="DOTNET_JitStressModeNames" Value="STRESS_LCL_FLDS" />
+    <CLRTestEnvironmentVariable Include="DOTNET_JitSaveFpLrWithCalleeSavedRegisters" Value="2" />
+  </ItemGroup>
+</Project>

src/tests/JIT/Regression/JitBlue/GitHub_130437/GitHub_130437.il

@@ -0,0 +1,207 @@
+// Licensed to the .NET Foundation under one or more agreements.
+// The .NET Foundation licenses this file to you under the MIT license.
+
+.assembly extern mscorlib {}
+.assembly extern System.Numerics.Vectors {}
+.assembly extern System.Runtime.Intrinsics {}
+
+.assembly GitHub_130437 {}
+.module GitHub_130437.dll
+
+.class private explicit ansi sealed beforefieldinit LargeFrame
+       extends [mscorlib]System.ValueType
+{
+    .pack 1
+    .size 8192
+}
+
+.class private auto ansi beforefieldinit GitHub_130437
+       extends [mscorlib]System.Object
+{
+    .method public hidebysig static int32
+        Main() cil managed
+    {
+        .entrypoint
+        .maxstack 2
+
+        call bool [System.Runtime.Intrinsics]System.Runtime.Intrinsics.Arm.Sve2::get_IsSupported()
+        brtrue.s SVE_SUPPORTED
+
+        ldc.i4.s 100
+        ret
+
+    SVE_SUPPORTED:
+        call int64 GitHub_130437::RunBasicScenario_UnsafeRead()
+        ldc.i4.s 44
+        conv.i8
+        ceq
+        brfalse.s FAIL
+
+        ldc.i4.s 100
+        ret
+
+    FAIL:
+        ldc.i4.m1
+        ret
+    }
+
+    .method private hidebysig static int64
+        RunBasicScenario_UnsafeRead() cil managed noinlining
+    {
+        .maxstack 4
+        .locals init (
+            // The mask local must be allocated above the large frame local so the
+            // SP-based predicate spill/reload uses a large positive offset.
+            [0] valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64> mask,
+            [1] valuetype LargeFrame largeFrame,
+            [2] uint32[] baseArray,
+            [3] int64[] maskArray,
+            [4] int64[] offsetsArray,
+            [5] uint32* pinned basePinned,
+            [6] int64* pinned maskPinned,
+            [7] int64* pinned offsetsPinned,
+            [8] valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64> result1,
+            [9] valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64> result2,
+            [10] int32 index)
+
+        ldc.i4 1024
+        newarr [mscorlib]System.UInt32
+        stloc.2
+
+        ldloc.2
+        ldc.i4.0
+        ldc.i4.s 20
+        stelem.i4
+
+        ldc.i4.s 64
+        newarr [mscorlib]System.Int64
+        stloc.3
+
+        ldc.i4.0
+        stloc.s index
+
+    FILL_MASK:
+        ldloc.s index
+        ldc.i4.s 64
+        bge.s FILL_MASK_DONE
+
+        ldloc.3
+        ldloc.s index
+        ldc.i4.1
+        conv.i8
+        stelem.i8
+
+        ldloc.s index
+        ldc.i4.1
+        add
+        stloc.s index
+        br.s FILL_MASK
+
+    FILL_MASK_DONE:
+        ldc.i4.s 64
+        newarr [mscorlib]System.Int64
+        stloc.s offsetsArray
+
+        ldloc.2
+        ldc.i4.0
+        ldelema [mscorlib]System.UInt32
+        stloc.s basePinned
+
+        ldloc.3
+        ldc.i4.0
+        ldelema [mscorlib]System.Int64
+        stloc.s maskPinned
+
+        ldloc.s offsetsArray
+        ldc.i4.0
+        ldelema [mscorlib]System.Int64
+        stloc.s offsetsPinned
+
+        ldloc.s maskPinned
+        ldobj valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64>
+        stloc.0
+
+        ldloca.s largeFrame
+        call void GitHub_130437::TouchLargeFrame(valuetype LargeFrame&)
+
+        ldloc.0
+        ldloc.s basePinned
+        ldloc.s offsetsPinned
+        ldobj valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64>
+        call valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64> [System.Runtime.Intrinsics]System.Runtime.Intrinsics.Arm.Sve2::GatherVectorUInt32WithByteOffsetsZeroExtendNonTemporal(valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64>, uint32*, valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64>)
+        stloc.s result1
+
+        ldloc.0
+        ldloc.s basePinned
+        ldloc.s offsetsPinned
+        ldobj valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64>
+        call valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64> [System.Runtime.Intrinsics]System.Runtime.Intrinsics.Arm.Sve2::GatherVectorUInt32WithByteOffsetsZeroExtendNonTemporal(valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64>, uint32*, valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64>)
+        stloc.s result2
+
+        ldloca.s result1
+        ldc.i4.0
+        call instance !0 valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64>::get_Item(int32)
+
+        ldloca.s result2
+        ldc.i4.0
+        call instance !0 valuetype [System.Numerics.Vectors]System.Numerics.Vector`1<int64>::get_Item(int32)
+        add
+
+        ldloca.s largeFrame
+        call int64 GitHub_130437::SumLargeFrame(valuetype LargeFrame&)
+        add
+        ret
+    }
+
+    .method private hidebysig static void
+        TouchLargeFrame(valuetype LargeFrame& frame) cil managed noinlining
+    {
+        .maxstack 3
+
+        ldarg.0
+        conv.u
+        ldc.i4.0
+        add
+        ldc.i4.1
+        stind.i1
+
+        ldarg.0
+        conv.u
+        ldc.i4 4096
+        add
+        ldc.i4.2
+        stind.i1
+
+        ldarg.0
+        conv.u
+        ldc.i4 8191
+        add
+        ldc.i4.3
+        stind.i1
+
+        ret
+    }
+
+    .method private hidebysig static int64
+        SumLargeFrame(valuetype LargeFrame& frame) cil managed noinlining
+    {
+        .maxstack 3
+
+        ldarg.0
+        conv.u
+        ldc.i4.0
+        add
+        ldind.u1
+        conv.i8
+
+        ldarg.0
+        conv.u
+        ldc.i4 8191
+        add
+        ldind.u1
+        conv.i8
+        add
+
+        ret
+    }
+}