ARM64-SVE: Add SVE registers to pal context (dotnet#103801)

* ARM64-SVE: Add SVE registers to pal context

* fix debug sizes

* Add SVE defines if missing from Linux host

* More missing defines

* More missing defines

* Add cast

* Move SVE registers after debug registers

* Fix dbgtargetcontext

* Remove SVE from debug context

* Move ffr

* Add SVE registers to asmconstants

* Remove Z registers from context

* backup/restore SVE in Context2.S

Change-Id: I2f6bc39068d9fed3f45b548089b144884607d97b

* Remove unused SVE128 struct

* Add XStateFeaturesMask

* restore instrsarm64sve.h changes

* Restore SIZEOF__CONTEXT for windows

* Fix AsmOffsets.cs for windows

* Fix AsmOffsets.cs for windows

* Restore missing ldr

* Check size of SVE data returned from the kernel

* 16 P registers

* Copy context based on XSTATE_MASK_SVE

* Move context handling inside XSTATE checks

* Set CONTEXT_XSTATE

* Remove __pad and fix sizes

* Fix context sizes

* Fix context sizes

* Only read/write OS context SVE registers on 128bit

src/coreclr/System.Private.CoreLib/src/System/Runtime/ExceptionServices/AsmOffsets.cs

@@ -25,9 +25,15 @@ class AsmOffsets
     public const int OFFSETOF__REGDISPLAY__ControlPC = 0xbe0;
 #endif // TARGET_UNIX
 #elif TARGET_ARM64
+#if TARGET_UNIX
+    public const int SIZEOF__REGDISPLAY = 0x9e0;
+    public const int OFFSETOF__REGDISPLAY__SP = 0x938;
+    public const int OFFSETOF__REGDISPLAY__ControlPC = 0x940;
+#else // TARGET_UNIX
     public const int SIZEOF__REGDISPLAY = 0x940;
     public const int OFFSETOF__REGDISPLAY__SP = 0x898;
     public const int OFFSETOF__REGDISPLAY__ControlPC = 0x8a0;
+#endif // TARGET_UNIX
 #elif TARGET_ARM
     public const int SIZEOF__REGDISPLAY = 0x410;
     public const int OFFSETOF__REGDISPLAY__SP = 0x3ec;
@@ -71,9 +77,15 @@ class AsmOffsets
     public const int OFFSETOF__REGDISPLAY__ControlPC = 0xbd8;
 #endif // TARGET_UNIX
 #elif TARGET_ARM64
+#if TARGET_UNIX
+    public const int SIZEOF__REGDISPLAY = 0x9d0;
+    public const int OFFSETOF__REGDISPLAY__SP = 0x930;
+    public const int OFFSETOF__REGDISPLAY__ControlPC = 0x938;
+#else // TARGET_UNIX
     public const int SIZEOF__REGDISPLAY = 0x930;
     public const int OFFSETOF__REGDISPLAY__SP = 0x890;
     public const int OFFSETOF__REGDISPLAY__ControlPC = 0x898;
+#endif // TARGET_UNIX
 #elif TARGET_ARM
     public const int SIZEOF__REGDISPLAY = 0x408;
     public const int OFFSETOF__REGDISPLAY__SP = 0x3e8;
@@ -111,9 +123,13 @@ class AsmOffsets
     public const int SIZEOF__PAL_LIMITED_CONTEXT = 0xc20;
 #else // TARGET_UNIX
     public const int SIZEOF__PAL_LIMITED_CONTEXT = 0x4d0;
-#endif // TARGET_UNIx
+#endif // TARGET_UNIX
 #elif TARGET_ARM64
+#if TARGET_UNIX
+    public const int SIZEOF__PAL_LIMITED_CONTEXT = 0x3e0;
+#else // TARGET_UNIX
     public const int SIZEOF__PAL_LIMITED_CONTEXT = 0x390;
+#endif // TARGET_UNIX
 #elif TARGET_ARM
     public const int SIZEOF__PAL_LIMITED_CONTEXT = 0x1a0;
 #elif TARGET_X86

src/coreclr/debug/inc/dbgtargetcontext.h

@@ -470,7 +470,12 @@ typedef DECLSPEC_ALIGN(16) struct {
 
 } DT_CONTEXT;
 
+
+#if !defined(CROSS_COMPILE) && !defined(TARGET_WINDOWS)
+static_assert(sizeof(DT_CONTEXT) == offsetof(T_CONTEXT, XStateFeaturesMask), "DT_CONTEXT must not include the SVE registers on AMD64");
+#else
 static_assert(sizeof(DT_CONTEXT) == sizeof(T_CONTEXT), "DT_CONTEXT size must equal the T_CONTEXT size on ARM64");
+#endif
 
 #elif defined(DTCONTEXT_IS_LOONGARCH64)
 

src/coreclr/pal/inc/pal.h

@@ -1844,6 +1844,12 @@ typedef struct _IMAGE_ARM_RUNTIME_FUNCTION_ENTRY {
 #define CONTEXT_EXCEPTION_REQUEST 0x40000000L
 #define CONTEXT_EXCEPTION_REPORTING 0x80000000L
 
+#define CONTEXT_XSTATE (CONTEXT_ARM64 | 0x40L)
+
+#define XSTATE_SVE (0)
+
+#define XSTATE_MASK_SVE (UI64(1) << (XSTATE_SVE))
+
 //
 // This flag is set by the unwinder if it has unwound to a call
 // site, and cleared whenever it unwinds through a trap frame.
@@ -1944,7 +1950,18 @@ typedef struct DECLSPEC_ALIGN(16) _CONTEXT {
     /* +0x338 */ DWORD64 Bvr[ARM64_MAX_BREAKPOINTS];
     /* +0x378 */ DWORD Wcr[ARM64_MAX_WATCHPOINTS];
     /* +0x380 */ DWORD64 Wvr[ARM64_MAX_WATCHPOINTS];
-    /* +0x390 */
+
+    /* +0x390 */ DWORD64 XStateFeaturesMask;
+
+    //
+    // Sve Registers
+    //
+    // TODO-SVE: Support Vector register sizes >128bit
+    // For 128bit, Z and V registers fully overlap, so there is no need to load/store both.
+    /* +0x398 */ DWORD Vl;
+    /* +0x39c */ DWORD Ffr;
+    /* +0x3a0 */ DWORD P[16];
+    /* +0x3e0 */
 
 } CONTEXT, *PCONTEXT, *LPCONTEXT;
 

src/coreclr/pal/src/CMakeLists.txt

@@ -90,6 +90,11 @@ if(CLR_CMAKE_HOST_ARCH_AMD64 AND CLR_CMAKE_TARGET_LINUX AND NOT CLR_CMAKE_TARGET
   add_definitions(-DXSTATE_SUPPORTED)
 endif(CLR_CMAKE_HOST_ARCH_AMD64 AND CLR_CMAKE_TARGET_LINUX AND NOT CLR_CMAKE_TARGET_LINUX_MUSL)
 
+if(CLR_CMAKE_HOST_ARCH_ARM64 AND CLR_CMAKE_TARGET_LINUX AND NOT CLR_CMAKE_TARGET_LINUX_MUSL)
+  # Currently the _xstate is not available on Alpine Linux
+  add_definitions(-DXSTATE_SUPPORTED)
+endif(CLR_CMAKE_HOST_ARCH_ARM64 AND CLR_CMAKE_TARGET_LINUX AND NOT CLR_CMAKE_TARGET_LINUX_MUSL)
+
 if(CLR_CMAKE_TARGET_LINUX_MUSL)
   # Setting RLIMIT_NOFILE breaks debugging of coreclr on Alpine Linux for some reason
   add_definitions(-DDONT_SET_RLIMIT_NOFILE)

src/coreclr/pal/src/arch/arm64/asmconstants.h

@@ -18,6 +18,13 @@
 
 #define CONTEXT_FULL (CONTEXT_CONTROL | CONTEXT_INTEGER | CONTEXT_FLOATING_POINT)
 
+#define CONTEXT_XSTATE_BIT (6)
+#define CONTEXT_XSTATE (1 << CONTEXT_XSTATE_BIT)
+
+#define XSTATE_SVE_BIT (0)
+
+#define XSTATE_MASK_SVE (UI64(1) << (XSTATE_SVE))
+
 
 #define CONTEXT_ContextFlags 0
 #define CONTEXT_Cpsr         CONTEXT_ContextFlags+4
@@ -54,6 +61,7 @@
 #define CONTEXT_Lr           CONTEXT_Fp+8
 #define CONTEXT_Sp           CONTEXT_Lr+8
 #define CONTEXT_Pc           CONTEXT_Sp+8
+
 #define CONTEXT_NEON_OFFSET  CONTEXT_Pc+8
 #define CONTEXT_V0           0
 #define CONTEXT_V1           CONTEXT_V0+16
@@ -89,7 +97,42 @@
 #define CONTEXT_V31          CONTEXT_V30+16
 #define CONTEXT_FLOAT_CONTROL_OFFSET  CONTEXT_V31+16
 #define CONTEXT_Fpcr         0
-#define CONTEXT_Fpsr         CONTEXT_Fpcr+8
-#define CONTEXT_Size         ((CONTEXT_NEON_OFFSET + CONTEXT_Fpsr + 8 + 0xf) & ~0xf)
+#define CONTEXT_Fpsr         CONTEXT_Fpcr+4
+#define CONTEXT_NEON_SIZE    CONTEXT_FLOAT_CONTROL_OFFSET+CONTEXT_Fpsr+4
+
+#define CONTEXT_DEBUG_OFFSET CONTEXT_NEON_OFFSET+CONTEXT_NEON_SIZE
+#define CONTEXT_DEBUG_SIZE   120 // (8*4)+(8*8)+(2*4)+(2*8)
+
+#define CONTEXT_XSTATEFEATURESMASK_OFFSET CONTEXT_DEBUG_OFFSET+CONTEXT_DEBUG_SIZE
+
+// TODO-SVE: Support Vector register sizes >128bit
+
+#define CONTEXT_SVE_OFFSET   CONTEXT_XSTATEFEATURESMASK_OFFSET+8
+#define CONTEXT_VL_OFFSET    0
+
+// SVE register offsets are multiples of the vector length
+#define CONTEXT_SVE_REGS_OFFSET   CONTEXT_VL_OFFSET+4
+#define CONTEXT_FFR_VL       0
+#define CONTEXT_P0_VL        CONTEXT_FFR_VL+1
+#define CONTEXT_P1_VL        CONTEXT_P0_VL+1
+#define CONTEXT_P2_VL        CONTEXT_P1_VL+1
+#define CONTEXT_P3_VL        CONTEXT_P2_VL+1
+#define CONTEXT_P4_VL        CONTEXT_P3_VL+1
+#define CONTEXT_P5_VL        CONTEXT_P4_VL+1
+#define CONTEXT_P6_VL        CONTEXT_P5_VL+1
+#define CONTEXT_P7_VL        CONTEXT_P6_VL+1
+#define CONTEXT_P8_VL        CONTEXT_P7_VL+1
+#define CONTEXT_P9_VL        CONTEXT_P8_VL+1
+#define CONTEXT_P10_VL       CONTEXT_P9_VL+1
+#define CONTEXT_P11_VL       CONTEXT_P10_VL+1
+#define CONTEXT_P12_VL       CONTEXT_P11_VL+1
+#define CONTEXT_P13_VL       CONTEXT_P12_VL+1
+#define CONTEXT_P14_VL       CONTEXT_P13_VL+1
+#define CONTEXT_P15_VL       CONTEXT_P14_VL+1
+
+#define CONTEXT_SVE_REGS_SIZE     ((CONTEXT_P15_VL+1) * 4)
+#define CONTEXT_SVE_SIZE     CONTEXT_SVE_REGS_SIZE + 8
+
+#define CONTEXT_Size         CONTEXT_SVE_OFFSET + CONTEXT_SVE_SIZE
 
 #endif

src/coreclr/pal/src/arch/arm64/context2.S

@@ -1,7 +1,7 @@
 // Licensed to the .NET Foundation under one or more agreements.
 // The .NET Foundation licenses this file to you under the MIT license.
 //
-// Implementation of _CONTEXT_CaptureContext for the ARM platform.
+// Implementation of _CONTEXT_CaptureContext for the ARM64 platform.
 // This function is processor dependent.  It is used by exception handling,
 // and is always apply to the current thread.
 //
@@ -12,6 +12,7 @@
 // Incoming:
 //  x0: Context*
 //
+.arch_extension sve
 LEAF_ENTRY CONTEXT_CaptureContext, _TEXT
     PROLOG_STACK_ALLOC 32
     .cfi_adjust_cfa_offset 32
@@ -69,7 +70,6 @@ LOCAL_LABEL(Done_CONTEXT_CONTROL):
     stp x26, x27, [x0, CONTEXT_X26]
     str x28,      [x0, CONTEXT_X28]
 
-
 LOCAL_LABEL(Done_CONTEXT_INTEGER):
     ldr w1, [x0, CONTEXT_ContextFlags]
     // clangs assembler doesn't seem to support the mov Wx, imm32 yet
@@ -104,6 +104,41 @@ LOCAL_LABEL(Done_CONTEXT_INTEGER):
     sub x0, x0,   CONTEXT_FLOAT_CONTROL_OFFSET + CONTEXT_NEON_OFFSET
 
 LOCAL_LABEL(Done_CONTEXT_FLOATING_POINT):
+    ldr w1, [x0, CONTEXT_ContextFlags]
+    // clangs assembler doesn't seem to support the mov Wx, imm32 yet
+    movz w2, #0x40, lsl #16
+    movk w2, #0x40
+    mov w3, w2
+    and w2, w1, w2
+    cmp w2, w3
+    b.ne LOCAL_LABEL(Done_CONTEXT_SVE)
+
+    ldr x1, [x0, CONTEXT_XSTATEFEATURESMASK_OFFSET]
+    tbz x1, #XSTATE_SVE_BIT, LOCAL_LABEL(Done_CONTEXT_SVE)
+
+    add x0, x0,   CONTEXT_SVE_OFFSET
+    str p0, [x0, CONTEXT_P0_VL, MUL VL]
+    str p1, [x0, CONTEXT_P1_VL, MUL VL]
+    str p2, [x0, CONTEXT_P2_VL, MUL VL]
+    str p3, [x0, CONTEXT_P3_VL, MUL VL]
+    str p4, [x0, CONTEXT_P4_VL, MUL VL]
+    str p5, [x0, CONTEXT_P5_VL, MUL VL]
+    str p6, [x0, CONTEXT_P6_VL, MUL VL]
+    str p7, [x0, CONTEXT_P7_VL, MUL VL]
+    str p8, [x0, CONTEXT_P8_VL, MUL VL]
+    str p9, [x0, CONTEXT_P9_VL, MUL VL]
+    str p10, [x0, CONTEXT_P10_VL, MUL VL]
+    str p11, [x0, CONTEXT_P11_VL, MUL VL]
+    str p12, [x0, CONTEXT_P12_VL, MUL VL]
+    str p13, [x0, CONTEXT_P13_VL, MUL VL]
+    str p14, [x0, CONTEXT_P14_VL, MUL VL]
+    str p15, [x0, CONTEXT_P15_VL, MUL VL]
+    rdffr p0.b
+    str p0, [x0, CONTEXT_FFR_VL, MUL VL]
+    ldr p0, [x0, 0, MUL VL]
+    sub x0, x0,   CONTEXT_SVE_OFFSET
+
+LOCAL_LABEL(Done_CONTEXT_SVE):
 
     EPILOG_STACK_FREE 32
     ret
@@ -124,6 +159,7 @@ LEAF_ENTRY RtlCaptureContext, _TEXT
     orr w1, w1, #0x4
     orr w1, w1, #0x8
     str w1, [x0, CONTEXT_ContextFlags]
+    str xzr, [x0, CONTEXT_XSTATEFEATURESMASK_OFFSET]
     ldr x1, [sp]
     EPILOG_STACK_FREE 16
     b C_FUNC(CONTEXT_CaptureContext)
@@ -133,6 +169,7 @@ LEAF_END RtlCaptureContext, _TEXT
 //  x0: Context*
 //  x1: Exception*
 //
+.arch_extension sve
 LEAF_ENTRY RtlRestoreContext, _TEXT
 
 #ifdef HAS_ADDRESS_SANITIZER
@@ -154,6 +191,34 @@ LOCAL_LABEL(Restore_CONTEXT_FLOATING_POINT):
     // since we potentially clobber x0 below, we'll bank it in x16
     mov x16, x0
 
+    ldr w17, [x16, CONTEXT_ContextFlags]
+    tbz w17, #CONTEXT_XSTATE_BIT, LOCAL_LABEL(No_Restore_CONTEXT_SVE)
+
+    ldr w17, [x16, CONTEXT_XSTATEFEATURESMASK_OFFSET]
+    tbz w17, #XSTATE_SVE_BIT, LOCAL_LABEL(No_Restore_CONTEXT_SVE)
+
+    add x16, x16,   CONTEXT_SVE_OFFSET
+    ldr p0, [x16, CONTEXT_FFR_VL, MUL VL]
+    wrffr p0.b
+    ldr p0, [x16, CONTEXT_P0_VL, MUL VL]
+    ldr p1, [x16, CONTEXT_P1_VL, MUL VL]
+    ldr p2, [x16, CONTEXT_P2_VL, MUL VL]
+    ldr p3, [x16, CONTEXT_P3_VL, MUL VL]
+    ldr p4, [x16, CONTEXT_P4_VL, MUL VL]
+    ldr p5, [x16, CONTEXT_P5_VL, MUL VL]
+    ldr p6, [x16, CONTEXT_P6_VL, MUL VL]
+    ldr p7, [x16, CONTEXT_P7_VL, MUL VL]
+    ldr p8, [x16, CONTEXT_P8_VL, MUL VL]
+    ldr p9, [x16, CONTEXT_P9_VL, MUL VL]
+    ldr p10, [x16, CONTEXT_P10_VL, MUL VL]
+    ldr p11, [x16, CONTEXT_P11_VL, MUL VL]
+    ldr p12, [x16, CONTEXT_P12_VL, MUL VL]
+    ldr p13, [x16, CONTEXT_P13_VL, MUL VL]
+    ldr p14, [x16, CONTEXT_P14_VL, MUL VL]
+    ldr p15, [x16, CONTEXT_P15_VL, MUL VL]
+    sub x16, x16,   CONTEXT_SVE_OFFSET
+
+LOCAL_LABEL(No_Restore_CONTEXT_SVE):
     ldr w17, [x16, CONTEXT_ContextFlags]
     tbz w17, #CONTEXT_FLOATING_POINT_BIT, LOCAL_LABEL(No_Restore_CONTEXT_FLOATING_POINT)
 
@@ -230,3 +295,12 @@ LEAF_ENTRY RestoreCompleteContext, _TEXT
 LEAF_END RestoreCompleteContext, _TEXT
 
 #endif // __APPLE__
+
+// Incoming:
+//  None
+//
+.arch_extension sve
+    LEAF_ENTRY CONTEXT_GetSveLengthFromOS, _TEXT
+        rdvl    x0, 1
+        ret     lr
+    LEAF_END CONTEXT_GetSveLengthFromOS, _TEXT

src/coreclr/pal/src/exception/signal.cpp

@@ -863,7 +863,7 @@ static void inject_activation_handler(int code, siginfo_t *siginfo, void *contex
 
         ULONG contextFlags = CONTEXT_CONTROL | CONTEXT_INTEGER | CONTEXT_FLOATING_POINT;
 
-#if defined(HOST_AMD64)
+#if defined(HOST_AMD64) || defined(HOST_ARM64)
         contextFlags |= CONTEXT_XSTATE;
 #endif
 
@@ -1053,7 +1053,7 @@ static bool common_signal_handler(int code, siginfo_t *siginfo, void *sigcontext
 
     ULONG contextFlags = CONTEXT_CONTROL | CONTEXT_INTEGER | CONTEXT_FLOATING_POINT;
 
-#if defined(HOST_AMD64)
+#if defined(HOST_AMD64) || defined(HOST_ARM64)
     contextFlags |= CONTEXT_XSTATE;
 #endif