Changeset 102802 in vbox for trunk

Timestamp:

Jan 10, 2024 12:30:48 AM (13 months ago)

Author:

vboxsync

Message:

VMM/IEM: Implemented the unlocked cmpxchg* primitives in IEMAllAImplC.cpp in C code w/o so we can get on with running single CPU guests. Minimally tested. bugref:10547

File:

• trunk/src/VBox/VMM/VMMAll/IEMAllAImplC.cpp (modified) (4 diffs)

trunk/src/VBox/VMM/VMMAll/IEMAllAImplC.cpp

@@ -2132 +2132 @@
 IEM_DECL_IMPL_DEF(void, iemAImpl_cmpxchg_u16,       (uint16_t *pu16Dst, uint16_t *puAx,  uint16_t uSrcReg, uint32_t *pEFlags))
 {
-    iemAImpl_cmpxchg_u16_locked(pu16Dst, puAx, uSrcReg, pEFlags);
+# if 0
+    /* If correctly aligned, used the locked variation. */
+    if (!((uintptr_t)pu16Dst & 1))
+        iemAImpl_cmpxchg_u16_locked(pu16Dst, puAx, uSrcReg, pEFlags);
+    else
+# endif
+    {
+        /* Otherwise emulate it as best as we can. */
+        uint16_t const uOld = *puAx;
+        uint16_t const uDst = *pu16Dst;
+        if (uOld == uDst)
+        {
+            *pu16Dst = uSrcReg;
+            iemAImpl_cmp_u16(&uOld, uOld, pEFlags);
+        }
+        else
+        {
+            *puAx = uDst;
+            iemAImpl_cmp_u16(&uOld, uDst, pEFlags);
+        }
+    }
 }
 
@@ -2138 +2158 @@
 IEM_DECL_IMPL_DEF(void, iemAImpl_cmpxchg_u32,       (uint32_t *pu32Dst, uint32_t *puEax, uint32_t uSrcReg, uint32_t *pEFlags))
 {
-    iemAImpl_cmpxchg_u32_locked(pu32Dst, puEax, uSrcReg, pEFlags);
+# if 0
+    /* If correctly aligned, used the locked variation. */
+    if (!((uintptr_t)pu32Dst & 3))
+        iemAImpl_cmpxchg_u32_locked(pu32Dst, puEax, uSrcReg, pEFlags);
+    else
+# endif
+    {
+        /* Otherwise emulate it as best as we can. */
+        uint32_t const uOld = *puEax;
+        uint32_t const uDst = *pu32Dst;
+        if (uOld == uDst)
+        {
+            *pu32Dst = uSrcReg;
+            iemAImpl_cmp_u32(&uOld, uOld, pEFlags);
+        }
+        else
+        {
+            *puEax = uDst;
+            iemAImpl_cmp_u32(&uOld, uDst, pEFlags);
+        }
+    }
 }
 
@@ -2145 +2185 @@
 IEM_DECL_IMPL_DEF(void, iemAImpl_cmpxchg_u64,       (uint64_t *pu64Dst, uint64_t *puRax, uint64_t *puSrcReg, uint32_t *pEFlags))
 {
-    iemAImpl_cmpxchg_u64_locked(pu64Dst, puRax, puSrcReg, pEFlags);
-}
-# else
+#  if 0
+    /* If correctly aligned, used the locked variation. */
+    if (!((uintptr_t)pu32Dst & 7))
+        iemAImpl_cmpxchg_u64_locked(pu64Dst, puRax, puSrcReg, pEFlags);
+    else
+#  endif
+    {
+        /* Otherwise emulate it as best as we can. */
+        uint64_t const uOld = *puRax;
+        uint64_t const uSrc = *puSrcReg;
+        uint64_t const uDst = *pu64Dst;
+        if (uOld == uDst)
+        {
+            *pu64Dst = uSrc;
+            iemAImpl_cmp_u64(&uOld, uOld, pEFlags);
+        }
+        else
+        {
+            *puRax = uDst;
+            iemAImpl_cmp_u64(&uOld, uDst, pEFlags);
+        }
+    }
+}
+# else  /* ARCH_BITS != 32 */
 IEM_DECL_IMPL_DEF(void, iemAImpl_cmpxchg_u64,       (uint64_t *pu64Dst, uint64_t *puRax, uint64_t uSrcReg, uint32_t *pEFlags))
 {
-    iemAImpl_cmpxchg_u64_locked(pu64Dst, puRax, uSrcReg, pEFlags);
-}
+#  if 0
+    /* If correctly aligned, used the locked variation. */
+    if (!((uintptr_t)pu64Dst & 7))
+        iemAImpl_cmpxchg_u64_locked(pu64Dst, puRax, uSrcReg, pEFlags);
+    else
+#  endif
+    {
+        /* Otherwise emulate it as best as we can. */
+        uint64_t const uOld = *puRax;
+        uint64_t const uDst = *pu64Dst;
+        if (uOld == uDst)
+        {
+            *pu64Dst = uSrcReg;
+            iemAImpl_cmp_u64(&uOld, uOld, pEFlags);
+        }
+        else
+        {
+            *puRax = uDst;
+            iemAImpl_cmp_u64(&uOld, uDst, pEFlags);
+        }
+    }
+}
+# endif /* ARCH_BITS != 32 */
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_cmpxchg8b,(uint64_t *pu64Dst, PRTUINT64U pu64EaxEdx, PRTUINT64U pu64EbxEcx, uint32_t *pEFlags))
+{
+# if 0
+    /* If correctly aligned, used the locked variation. */
+    if (!((uintptr_t)pu64Dst & 7))
+        iemAImpl_cmpxchg8b_locked(pu64Dst, pu64EaxEdx, pu64EbxEcx, pEFlags);
+    else
 # endif
-
-
-IEM_DECL_IMPL_DEF(void, iemAImpl_cmpxchg8b,(uint64_t *pu64Dst, PRTUINT64U pu64EaxEdx, PRTUINT64U pu64EbxEcx, uint32_t *pEFlags))
-{
-    iemAImpl_cmpxchg8b_locked(pu64Dst, pu64EaxEdx, pu64EbxEcx, pEFlags);
+    {
+        /* Otherwise emulate it as best as we can. */
+        uint64_t const uNew = pu64EbxEcx->u;
+        uint64_t const uOld = pu64EaxEdx->u;
+        uint64_t const uDst = *pu64Dst;
+        if (uDst == uOld)
+        {
+            *pu64Dst = uNew;
+            *pEFlags |= X86_EFL_ZF;
+        }
+        else
+        {
+            pu64EaxEdx->u = uDst;
+            *pEFlags &= ~X86_EFL_ZF;
+        }
+    }
 }
 
@@ -2164 +2266 @@
                                              uint32_t *pEFlags))
 {
-    iemAImpl_cmpxchg16b_locked(pu128Dst, pu128RaxRdx, pu128RbxRcx, pEFlags);
+# if 0
+    /* If correctly aligned, used the locked variation. */
+    if (!((uintptr_t)pu64Dst & 15))
+        iemAImpl_cmpxchg16b_locked(pu128Dst, pu128RaxRdx, pu128RbxRcx, pEFlags);
+    else
+# endif
+    {
+        /* Otherwise emulate it as best as we can. */
+# ifdef RT_COMPILER_WITH_128BIT_INT_TYPES
+        uint128_t const uNew = pu128RbxRcx->u;
+        uint128_t const uOld = pu128RaxRdx->u;
+        uint128_t const uDst = pu128Dst->u;
+        if (uDst == uOld)
+        {
+            pu128Dst->u = uNew;
+            *pEFlags |= X86_EFL_ZF;
+        }
+        else
+        {
+            pu128RaxRdx->u = uDst;
+            *pEFlags &= ~X86_EFL_ZF;
+        }
+# else
+        RTUINT128U const uNew = *pu128RbxRcx;
+        RTUINT128U const uOld = *pu128RaxRdx;
+        RTUINT128U const uDst = *pu128Dst;
+        if (   uDst.s.Lo == uOld.s.Lo
+            && uDst.s.Hi == uOld.s.Hi)
+        {
+            *pu128Dst = uNew;
+            *pEFlags |= X86_EFL_ZF;
+        }
+        else
+        {
+            *pu128RaxRdx = uDst;
+            *pEFlags &= ~X86_EFL_ZF;
+        }
+# endif
+    }
 }