Fix riscv register mapping with llvm-libunwind (#111735)

* Fix riscv register mapping with llvm-libunwind * Use tail and cleanup custom regs * Delete emulated TLS from riscv and loongarch * Fix UniversalTransition * Fix emitOutputCall and use ret Co-authored-by: Filip Navara <filip.navara@gmail.com> * Use traditional TLS dialect for the time being --------- Co-authored-by: Filip Navara <filip.navara@gmail.com>
dotnet · Feb 7, 2025 · 42a58a2 · 42a58a2
1 parent 15f422d
commit 42a58a2
Show file tree

Hide file tree

Showing 19 changed files with 477 additions and 561 deletions.
diff --git a/src/coreclr/jit/emitriscv64.cpp b/src/coreclr/jit/emitriscv64.cpp
@@ -1624,8 +1624,8 @@ unsigned emitter::emitOutputCall(const insGroup* ig, BYTE* dst, instrDesc* id, c
 
         size_t addr = (size_t)(id->idAddr()->iiaAddr); // get addr.
 
-        int reg2 = ((int)addr & 1) + 10;
-        addr     = addr ^ 1;
+        int reg2 = (int)(addr & 1);
+        addr -= reg2;
 
         if (!emitComp->opts.compReloc)
         {

diff --git a/src/coreclr/nativeaot/Runtime/StackFrameIterator.cpp b/src/coreclr/nativeaot/Runtime/StackFrameIterator.cpp
@@ -508,6 +508,39 @@ void StackFrameIterator::InternalInit(Thread * pThreadToWalk, PTR_PAL_LIMITED_CO
     m_RegDisplay.pR4 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, R4);
     m_RegDisplay.pR5 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, R5);
 
+#elif defined(TARGET_RISCV64)
+    //
+    // preserved regs
+    //
+    m_RegDisplay.pS1 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S1);
+    m_RegDisplay.pS2 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S2);
+    m_RegDisplay.pS3 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S3);
+    m_RegDisplay.pS4 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S4);
+    m_RegDisplay.pS5 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S5);
+    m_RegDisplay.pS6 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S6);
+    m_RegDisplay.pS7 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S7);
+    m_RegDisplay.pS8 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S8);
+    m_RegDisplay.pS9 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S9);
+    m_RegDisplay.pS10 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S10);
+    m_RegDisplay.pS11 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, S11);
+    m_RegDisplay.pFP = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, FP);
+    m_RegDisplay.pRA = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, RA);
+
+    //
+    // preserved floating-point registers
+    //
+    int32_t preservedFpIndices[] = {8, 9, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27};
+    for (int i = 0; i < sizeof(preservedFpIndices) / sizeof(preservedFpIndices[0]); i++)
+    {
+        m_RegDisplay.F[preservedFpIndices[i]] = pCtx->F[preservedFpIndices[i]];
+    }
+
+    //
+    // scratch regs
+    //
+    m_RegDisplay.pA0 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, A0);
+    m_RegDisplay.pA1 = (PTR_uintptr_t)PTR_TO_MEMBER_TADDR(PAL_LIMITED_CONTEXT, pCtx, A1);
+
 #elif defined(UNIX_AMD64_ABI)
     //
     // preserved regs
@@ -776,7 +809,7 @@ void StackFrameIterator::InternalInit(Thread * pThreadToWalk, NATIVE_CONTEXT* pC
     m_RegDisplay.pS9 = (PTR_uintptr_t)PTR_TO_REG(pCtx, S9);
     m_RegDisplay.pS10 = (PTR_uintptr_t)PTR_TO_REG(pCtx, S10);
     m_RegDisplay.pS11 = (PTR_uintptr_t)PTR_TO_REG(pCtx, S11);
- 
+
     //
     // scratch regs
     //
@@ -1007,7 +1040,7 @@ void StackFrameIterator::UnwindFuncletInvokeThunk()
 #if defined(USE_PORTABLE_HELPERS) // @TODO: Currently no funclet invoke defined in a portable way
     return;
 #else // defined(USE_PORTABLE_HELPERS)
-    ASSERT((CategorizeUnadjustedReturnAddress(m_ControlPC) == InFuncletInvokeThunk) || 
+    ASSERT((CategorizeUnadjustedReturnAddress(m_ControlPC) == InFuncletInvokeThunk) ||
            (CategorizeUnadjustedReturnAddress(m_ControlPC) == InFilterFuncletInvokeThunk));
 
     PTR_uintptr_t SP;

diff --git a/src/coreclr/nativeaot/Runtime/loongarch64/AllocFast.S b/src/coreclr/nativeaot/Runtime/loongarch64/AllocFast.S
@@ -21,11 +21,7 @@
     LEAF_ENTRY RhpNewFast, _TEXT
 
         // a1 = GetThread()
-#ifdef FEATURE_EMULATED_TLS
-        GETTHREAD_ETLS_1
-#else
         INLINE_GETTHREAD  $a1
-#endif
 
         //
         // a0 contains MethodTable pointer
@@ -125,11 +121,7 @@ LOCAL_LABEL(NewOutOfMemory):
         // a1 == element count
         // a2 == string size
 
-#ifdef FEATURE_EMULATED_TLS
-        GETTHREAD_ETLS_3
-#else
         INLINE_GETTHREAD $a3
-#endif
 
         // Load potential new object address into t3.
         ld.d  $t3, $a3, OFFSETOF__Thread__m_alloc_context__alloc_ptr

diff --git a/src/coreclr/nativeaot/Runtime/loongarch64/GcProbe.S b/src/coreclr/nativeaot/Runtime/loongarch64/GcProbe.S
@@ -102,11 +102,7 @@
 .macro FixupHijackedCallstack
 
     // a2 <- GetThread()
-#ifdef FEATURE_EMULATED_TLS
-    GETTHREAD_ETLS_2
-#else
     INLINE_GETTHREAD  $a2
-#endif
 
     //
     // Fix the stack by restoring the original return address

diff --git a/src/coreclr/nativeaot/Runtime/loongarch64/InteropThunksHelpers.S b/src/coreclr/nativeaot/Runtime/loongarch64/InteropThunksHelpers.S
@@ -21,14 +21,7 @@
         // Custom calling convention:
         //      tp pointer to the current thunk's data block (data contains 2 pointer values: context + target pointers)
 
-#ifdef FEATURE_EMULATED_TLS
-        // This doesn't save and restore the floating point argument registers. If we encounter a
-        // target system that uses TLS emulation and modify these registers during this call we
-        // need to save and restore them, too
-        GETTHUNKDATA_ETLS_9
-#else
         INLINE_GET_TLS_VAR  $t0, C_FUNC(tls_thunkData)
-#endif
 
         // t0  = base address of TLS data
         // tp = address of context cell in thunk's data

diff --git a/src/coreclr/nativeaot/Runtime/loongarch64/PInvoke.S b/src/coreclr/nativeaot/Runtime/loongarch64/PInvoke.S
@@ -26,11 +26,7 @@ NESTED_ENTRY RhpPInvoke, _TEXT, NoHandler
 
         // get TLS global variable address
 
-#ifdef FEATURE_EMULATED_TLS
-        GETTHREAD_ETLS_1
-#else
         INLINE_GETTHREAD  $a1
-#endif
 
         st.d  $a1, $a0, OFFSETOF__PInvokeTransitionFrame__m_pThread
         st.d  $a0, $a1, OFFSETOF__Thread__m_pTransitionFrame

diff --git a/src/coreclr/nativeaot/Runtime/riscv64/AllocFast.S b/src/coreclr/nativeaot/Runtime/riscv64/AllocFast.S
@@ -19,19 +19,17 @@
     LEAF_ENTRY RhpNewFast, _TEXT
 
         // a1 = GetThread()
-#ifdef FEATURE_EMULATED_TLS
-        GETTHREAD_ETLS_1
-#else
+        // Save MethodTable pointer. INLINE_GETTHREAD will trash a0.
+        mv                t2, a0
         INLINE_GETTHREAD  a1
-#endif
 
         //
-        // a0 contains MethodTable pointer
+        // t2 contains MethodTable pointer
         //
-        lw  a2, OFFSETOF__MethodTable__m_uBaseSize(a0)
+        lw  a2, OFFSETOF__MethodTable__m_uBaseSize(t2)
 
         //
-        // a0: MethodTable pointer
+        // t2: MethodTable pointer
         // a1: Thread pointer
         // a2: base size
         //
@@ -49,13 +47,14 @@
         sd    a2, OFFSETOF__Thread__m_alloc_context__alloc_ptr(a1)
 
         // Set the new objects MethodTable pointer
-        sd    a0, OFFSETOF__Object__m_pEEType(t3)
+        sd    t2, OFFSETOF__Object__m_pEEType(t3)
 
         mv    a0, t3
         ret
 
 LOCAL_LABEL(RhpNewFast_RarePath):
         mv    a1, zero
+        mv    a0, t2
         tail  RhpNewObject
     LEAF_END RhpNewFast, _TEXT
 
@@ -123,11 +122,9 @@ LOCAL_LABEL(NewOutOfMemory):
         // a1 == element count
         // a2 == string size
 
-#ifdef FEATURE_EMULATED_TLS
-        GETTHREAD_ETLS_3
-#else
+        // Save MethodTable pointer. INLINE_GETTHREAD will trash a0.
+        mv               t2, a0
         INLINE_GETTHREAD a3
-#endif
 
         // Load potential new object address into t3.
         ld   t3, OFFSETOF__Thread__m_alloc_context__alloc_ptr(a3)
@@ -145,7 +142,7 @@ LOCAL_LABEL(NewOutOfMemory):
         sd   a2, OFFSETOF__Thread__m_alloc_context__alloc_ptr(a3)
 
         // Set the new object's MethodTable pointer and element count.
-        sd   a0, OFFSETOF__Object__m_pEEType(t3)
+        sd   t2, OFFSETOF__Object__m_pEEType(t3)
         sd   a1, OFFSETOF__Array__m_Length(t3)
 
         // Return the object allocated in a0.
@@ -158,11 +155,12 @@ LOCAL_LABEL(StringSizeOverflow):
         // 32-bit value. We are going to tail-call to a managed helper that will throw
         // an OOM exception that the caller of this allocator understands.
 
-        // a0 holds MethodTable pointer already
+        mv    a0, t2
         li    a1, 1                  // Indicate that we should throw OverflowException
         tail  C_FUNC(RhExceptionHandling_FailedAllocation)
 
 LOCAL_LABEL(RhNewString_Rare):
+        mv    a0, t2
         tail  C_FUNC(RhpNewArrayRare)
     LEAF_END RhNewString, _TEXT
 
@@ -178,9 +176,9 @@ LOCAL_LABEL(RhNewString_Rare):
         li  a2, 0x7fffffff
         bltu  a2, a1, LOCAL_LABEL(ArraySizeOverflow)  // Branch if a2 < a1 (check for overflow)
 
-        ld   a2, OFFSETOF__MethodTable__m_usComponentSize(a0) // Load component size
+        lhu  a2, OFFSETOF__MethodTable__m_usComponentSize(a0) // Load component size
         mul  a2, a1, a2                // a2 = a1 * component size
-        ld   a3, OFFSETOF__MethodTable__m_uBaseSize(a0)       // Load base size
+        lw   a3, OFFSETOF__MethodTable__m_uBaseSize(a0)       // Load base size
         add  a2, a2, a3                // a2 = a2 + base size
         addi a2, a2, 7                 // a2 = a2 + 7
         andi a2, a2, ~0x7              // Clear the bits[2:0] of a2 (align to 8 bytes)
@@ -189,6 +187,8 @@ LOCAL_LABEL(RhNewString_Rare):
         // a1 == element count
         // a2 == array size
 
+        // Save MethodTable pointer. INLINE_GETTHREAD will trash a0.
+        mv               t2, a0
         INLINE_GETTHREAD a3
 
         // Load potential new object address into t3.
@@ -207,7 +207,7 @@ LOCAL_LABEL(RhNewString_Rare):
         sd   a2, OFFSETOF__Thread__m_alloc_context__alloc_ptr(a3)
 
         // Set the new object's MethodTable pointer and element count.
-        sd   a0, OFFSETOF__Object__m_pEEType(t3)
+        sd   t2, OFFSETOF__Object__m_pEEType(t3)
         sd   a1, OFFSETOF__Array__m_Length(t3)
 
         // Return the object allocated in a0.
@@ -220,11 +220,12 @@ LOCAL_LABEL(ArraySizeOverflow):
         // 32-bit value. We are going to tail-call to a managed helper that will throw
         // an overflow exception that the caller of this allocator understands.
 
-        // a0 holds MethodTable pointer already
+        mv   a0, t2
         li   a1, 1 // Indicate that we should throw OverflowException
         tail  C_FUNC(RhExceptionHandling_FailedAllocation)
 
 LOCAL_LABEL(RhpNewArray_Rare):
+        mv    a0, t2
         tail  C_FUNC(RhpNewArrayRare)
     LEAF_END RhpNewArray, _TEXT
 

diff --git a/src/coreclr/nativeaot/Runtime/riscv64/ExceptionHandling.S b/src/coreclr/nativeaot/Runtime/riscv64/ExceptionHandling.S
@@ -591,7 +591,7 @@ LOCAL_LABEL(NoAbort):
 
         // Allocate space for the call funclet frame
         ALLOC_CALL_FUNCLET_FRAME 0x80
-        
+
         // Save floating-point registers
         fsd  fs0,  0x00(sp)
         fsd  fs1,  0x08(sp)
@@ -605,7 +605,7 @@ LOCAL_LABEL(NoAbort):
         fsd  fs9,  0x48(sp)
         fsd  fs10, 0x50(sp)
         fsd  fs11, 0x58(sp)
-        
+
         // Save integer registers
         sd a0, 0x60(sp)  // Save a0 to 0x60
         sd a1, 0x68(sp)  // Save a1 to 0x68
@@ -664,7 +664,7 @@ LOCAL_LABEL(NoAbort):
 
         // Free call funclet frame
         FREE_CALL_FUNCLET_FRAME 0x80
-        
+
         // Return
         EPILOG_RETURN
 

diff --git a/src/coreclr/nativeaot/Runtime/riscv64/GcProbe.S b/src/coreclr/nativeaot/Runtime/riscv64/GcProbe.S
@@ -84,11 +84,7 @@
 .macro FixupHijackedCallstack
 
     // a2 <- GetThread()
-#ifdef FEATURE_EMULATED_TLS
-    GETTHREAD_ETLS_2
-#else
     INLINE_GETTHREAD a2
-#endif
 
     // Fix the stack by restoring the original return address
     ld  ra, OFFSETOF__Thread__m_pvHijackedReturnAddress(a2)

diff --git a/src/coreclr/nativeaot/Runtime/riscv64/InteropThunksHelpers.S b/src/coreclr/nativeaot/Runtime/riscv64/InteropThunksHelpers.S
@@ -21,13 +21,7 @@
         // Custom calling convention:
         //      tp pointer to the current thunk's data block (data contains 2 pointer values: context + target pointers)
 
-#ifdef FEATURE_EMULATED_TLS
-        // If using TLS emulation, fetch the TLS data block address
-        GETTHUNKDATA_ETLS_9
-#else
-        // Inline function to get TLS variable
         INLINE_GET_TLS_VAR t0, C_FUNC(tls_thunkData)
-#endif
 
         // t0 = base address of TLS data
         // tp = address of context cell in thunk's data

diff --git a/src/coreclr/nativeaot/Runtime/riscv64/PInvoke.S b/src/coreclr/nativeaot/Runtime/riscv64/PInvoke.S
@@ -19,31 +19,32 @@
 NESTED_ENTRY RhpPInvoke, _TEXT, NoHandler
         sd    fp, OFFSETOF__PInvokeTransitionFrame__m_FramePointer(a0)
         sd    ra, OFFSETOF__PInvokeTransitionFrame__m_RIP(a0)
-        sd    sp, OFFSETOF__PInvokeTransitionFrame__m_PreservedRegs(a0)
+        sd    t0, OFFSETOF__PInvokeTransitionFrame__m_PreservedRegs(a0)
         li    t0, PTFF_SAVE_SP
         sd    t0, OFFSETOF__PInvokeTransitionFrame__m_Flags(a0)
 
         // get TLS global variable address
-#ifdef FEATURE_EMULATED_TLS
-        GETTHREAD_ETLS_1
-#else
+        // Save MethodTable pointer. INLINE_GETTHREAD will trash a0.
+        mv               t2, a0
         INLINE_GETTHREAD a1
-#endif
 
-        sd    a1, OFFSETOF__PInvokeTransitionFrame__m_pThread(a0)
-        sd    a0, OFFSETOF__Thread__m_pTransitionFrame(a1)
-        jr  ra
+        sd    a1, OFFSETOF__PInvokeTransitionFrame__m_pThread(t2)
+        sd    t2, OFFSETOF__Thread__m_pTransitionFrame(a1)
+        ret
 NESTED_END RhpPInvoke, _TEXT
 
 LEAF_ENTRY RhpPInvokeReturn, _TEXT
         ld    t0, OFFSETOF__PInvokeTransitionFrame__m_pThread(a0)
-        sd    zero, OFFSETOF__Thread__m_pTransitionFrame(t0)
+        mv    t1, zero
+        sd    t1, OFFSETOF__Thread__m_pTransitionFrame(t0)
 
-        PREPARE_EXTERNAL_VAR_INDIRECT_W RhpTrapThreads, a5
+        PREPARE_EXTERNAL_VAR_INDIRECT_W RhpTrapThreads, t0
+
+        bnez    t1, 0f  // If TrapThreadsFlags_None is non-zero, branch
+        ret
+
+0:
+        // Passing transition frame pointer in a0
+        tail    RhpWaitForGC2
 
-        bne   t0, zero, 1f   // TrapThreadsFlags_None = 0
-        jr  ra
-1:
-        // passing transition frame pointer in a0
-        tail  C_FUNC(RhpWaitForGC2)
 LEAF_END RhpPInvokeReturn, _TEXT