Changeset 95360 in vbox

Timestamp:

Jun 23, 2022 9:45:55 PM (2 years ago)

Author:

vboxsync

Message:

VMM/IEM: Implemented the POPCNT instruction. bugref:9898

Location:

trunk/src/VBox/VMM

Files:

• VMMAll/IEMAllAImpl.asm (modified) (1 diff)
• VMMAll/IEMAllAImplC.cpp (modified) (2 diffs)
• VMMAll/IEMAllInstructionsTwoByte0f.cpp.h (modified) (1 diff)
• include/IEMInternal.h (modified) (1 diff)
• testcase/tstIEMCheckMc.cpp (modified) (1 diff)

trunk/src/VBox/VMM/VMMAll/IEMAllAImpl.asm

@@ -1045 +1045 @@
 IEMIMPL_BIT_OP2 tzcnt, (X86_EFL_ZF | X86_EFL_CF), (X86_EFL_OF | X86_EFL_SF | X86_EFL_AF | X86_EFL_PF), 0
 IEMIMPL_BIT_OP2 lzcnt, (X86_EFL_ZF | X86_EFL_CF), (X86_EFL_OF | X86_EFL_SF | X86_EFL_AF | X86_EFL_PF), 0
+
+
+;;
+; Macro for implementing POPCNT.
+;
+; This will generate code for the 16, 32 and 64 bit accesses, except on 32-bit
+; system where the 64-bit accesses requires hand coding.
+;
+; All the functions takes a pointer to the destination memory operand in A0,
+; the source register operand in A1 and a pointer to eflags in A2.
+;
+; ASSUMES Intel and AMD set EFLAGS the same way.
+;
+; ASSUMES the instruction does not support memory destination.
+;
+; @param        1       The instruction mnemonic.
+; @param        2       The modified flags.
+; @param        3       The undefined flags.
+;
+%macro IEMIMPL_BIT_OP3 3
+BEGINCODE
+BEGINPROC_FASTCALL iemAImpl_ %+ %1 %+ _u16, 12
+        PROLOGUE_3_ARGS
+        IEM_MAYBE_LOAD_FLAGS           A2, %2, %3
+        %1      T0_16, A1_16
+        mov     [A0], T0_16
+        IEM_SAVE_FLAGS                 A2, %2, %3
+        EPILOGUE_3_ARGS
+ENDPROC iemAImpl_ %+ %1 %+ _u16
+
+BEGINPROC_FASTCALL iemAImpl_ %+ %1 %+ _u32, 12
+        PROLOGUE_3_ARGS
+        IEM_MAYBE_LOAD_FLAGS           A2, %2, %3
+        %1      T0_32, A1_32
+        mov     [A0], T0_32
+        IEM_SAVE_FLAGS                 A2, %2, %3
+        EPILOGUE_3_ARGS
+ENDPROC iemAImpl_ %+ %1 %+ _u32
+
+ %ifdef RT_ARCH_AMD64
+BEGINPROC_FASTCALL iemAImpl_ %+ %1 %+ _u64, 16
+        PROLOGUE_3_ARGS
+        IEM_MAYBE_LOAD_FLAGS           A2, %2, %3
+        %1      T0, A1
+        mov     [A0], T0
+        IEM_SAVE_FLAGS                 A2, %2, %3
+        EPILOGUE_3_ARGS_EX 8
+ENDPROC iemAImpl_ %+ %1 %+ _u64
+ %endif ; RT_ARCH_AMD64
+%endmacro
+IEMIMPL_BIT_OP3 popcnt, (X86_EFL_ZF | X86_EFL_CF | X86_EFL_OF | X86_EFL_SF | X86_EFL_AF | X86_EFL_PF), 0
 
 

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

@@ -1773 +1773 @@
                                                                       a_Type uSrc2, uint32_t *pfEFlags)) \
 { \
-    /* uSrc1 is considered virtually zero extended to 512 bits width. */ \
     uint32_t      fEfl      = *pfEFlags & ~(X86_EFL_OF | X86_EFL_SF | X86_EFL_ZF | X86_EFL_AF | X86_EFL_PF | X86_EFL_CF); \
     a_Type        uResult; \
@@ -1798 +1797 @@
 EMIT_BZHI(32, uint32_t, RT_NOTHING)
 #endif
+
+/*
+ * POPCNT
+ */
+RT_ALIGNAS_VAR(64) static uint8_t const g_abBitCounts6[64] =
+{
+    0, 1, 1, 2,  1, 2, 2, 3,  1, 2, 2, 3,  2, 3, 3, 4,
+    1, 2, 2, 3,  2, 3, 3, 4,  2, 3, 3, 4,  3, 4, 4, 5,
+    1, 2, 2, 3,  2, 3, 3, 4,  2, 3, 3, 4,  3, 4, 4, 5,
+    2, 3, 3, 4,  3, 4, 4, 5,  3, 4, 4, 5,  4, 5, 5, 6,
+};
+
+/** @todo Use native popcount where possible and employ some more efficient
+ *        algorithm here (or in asm.h fallback)! */
+
+DECLINLINE(uint8_t) iemPopCountU16(uint16_t u16)
+{
+    return g_abBitCounts6[ u16        & 0x3f]
+        +  g_abBitCounts6[(u16 >> 6)  & 0x3f]
+        +  g_abBitCounts6[(u16 >> 12) & 0x3f];
+}
+
+DECLINLINE(uint8_t) iemPopCountU32(uint32_t u32)
+{
+    return g_abBitCounts6[ u32        & 0x3f]
+        +  g_abBitCounts6[(u32 >> 6)  & 0x3f]
+        +  g_abBitCounts6[(u32 >> 12) & 0x3f]
+        +  g_abBitCounts6[(u32 >> 18) & 0x3f]
+        +  g_abBitCounts6[(u32 >> 24) & 0x3f]
+        +  g_abBitCounts6[(u32 >> 30) & 0x3f];
+}
+
+DECLINLINE(uint8_t) iemPopCountU64(uint64_t u64)
+{
+    return g_abBitCounts6[ u64        & 0x3f]
+        +  g_abBitCounts6[(u64 >> 6)  & 0x3f]
+        +  g_abBitCounts6[(u64 >> 12) & 0x3f]
+        +  g_abBitCounts6[(u64 >> 18) & 0x3f]
+        +  g_abBitCounts6[(u64 >> 24) & 0x3f]
+        +  g_abBitCounts6[(u64 >> 30) & 0x3f]
+        +  g_abBitCounts6[(u64 >> 36) & 0x3f]
+        +  g_abBitCounts6[(u64 >> 42) & 0x3f]
+        +  g_abBitCounts6[(u64 >> 48) & 0x3f]
+        +  g_abBitCounts6[(u64 >> 54) & 0x3f]
+        +  g_abBitCounts6[(u64 >> 60) & 0x3f];
+}
+
+#define EMIT_POPCNT(a_cBits, a_Type, a_Suffix) \
+IEM_DECL_IMPL_DEF(void, RT_CONCAT3(iemAImpl_popcnt_u,a_cBits,a_Suffix),(a_Type *puDst, a_Type uSrc, uint32_t *pfEFlags)) \
+{ \
+    uint32_t    fEfl = *pfEFlags & ~(X86_EFL_OF | X86_EFL_SF | X86_EFL_ZF | X86_EFL_AF | X86_EFL_PF | X86_EFL_CF); \
+    a_Type      uResult; \
+    if (uSrc) \
+        uResult = iemPopCountU ## a_cBits(uSrc); \
+    else \
+    { \
+        fEfl |= X86_EFL_ZF; \
+        uResult = 0; \
+    } \
+    *puDst    = uResult; \
+    *pfEFlags = fEfl; \
+}
+
+EMIT_POPCNT(64, uint64_t, _fallback)
+EMIT_POPCNT(32, uint32_t, _fallback)
+EMIT_POPCNT(16, uint16_t, _fallback)
+#if defined(RT_ARCH_X86) || defined(IEM_WITHOUT_ASSEMBLY)
+EMIT_POPCNT(64, uint64_t, RT_NOTHING)
+#endif
+#if (!defined(RT_ARCH_X86) && !defined(RT_ARCH_AMD64)) || defined(IEM_WITHOUT_ASSEMBLY)
+EMIT_POPCNT(32, uint32_t, RT_NOTHING)
+EMIT_POPCNT(16, uint16_t, RT_NOTHING)
+#endif
+
 
 #if !defined(RT_ARCH_AMD64) || defined(IEM_WITHOUT_ASSEMBLY)

trunk/src/VBox/VMM/VMMAll/IEMAllInstructionsTwoByte0f.cpp.h

@@ -7801 +7801 @@
 /** Opcode      0x0f 0xb8 - JMPE (reserved for emulator on IPF) */
 FNIEMOP_UD_STUB(iemOp_jmpe);
+
+
 /** Opcode 0xf3 0x0f 0xb8 - POPCNT Gv, Ev */
-FNIEMOP_STUB(iemOp_popcnt_Gv_Ev);
+FNIEMOP_DEF(iemOp_popcnt_Gv_Ev)
+{
+    IEMOP_MNEMONIC2(RM, POPCNT, popcnt, Gv, Ev, DISOPTYPE_HARMLESS, 0);
+    if (!IEM_GET_GUEST_CPU_FEATURES(pVCpu)->fPopCnt)
+        return iemOp_InvalidNeedRM(pVCpu);
+#ifndef TST_IEM_CHECK_MC
+# if defined(RT_ARCH_X86) || defined(RT_ARCH_AMD64)
+    static const IEMOPBINSIZES s_Native =
+    {   NULL, NULL, iemAImpl_popcnt_u16, NULL, iemAImpl_popcnt_u32, NULL, iemAImpl_popcnt_u64, NULL };
+# endif
+    static const IEMOPBINSIZES s_Fallback =
+    {   NULL, NULL, iemAImpl_popcnt_u16_fallback, NULL, iemAImpl_popcnt_u32_fallback, NULL, iemAImpl_popcnt_u64_fallback, NULL };
+#endif
+    return FNIEMOP_CALL_1(iemOpHlpBinaryOperator_rv_rm, IEM_SELECT_HOST_OR_FALLBACK(fPopCnt, &s_Native, &s_Fallback));
+}
 
 

trunk/src/VBox/VMM/include/IEMInternal.h

@@ -1394 +1394 @@
 FNIEMAIMPLBINU32 iemAImpl_tzcnt_u32, iemAImpl_tzcnt_u32_amd, iemAImpl_tzcnt_u32_intel;
 FNIEMAIMPLBINU64 iemAImpl_tzcnt_u64, iemAImpl_tzcnt_u64_amd, iemAImpl_tzcnt_u64_intel;
+FNIEMAIMPLBINU16 iemAImpl_popcnt_u16, iemAImpl_popcnt_u16_fallback;
+FNIEMAIMPLBINU32 iemAImpl_popcnt_u32, iemAImpl_popcnt_u32_fallback;
+FNIEMAIMPLBINU64 iemAImpl_popcnt_u64, iemAImpl_popcnt_u64_fallback;
 /** @}  */
 

trunk/src/VBox/VMM/testcase/tstIEMCheckMc.cpp

@@ -238 +238 @@
 #undef  IEMTARGETCPU_EFL_BEHAVIOR_SELECT_EX
 #define IEMTARGETCPU_EFL_BEHAVIOR_SELECT_EX(a_aaArray, a_fNative)   NULL
+
+#undef  IEM_SELECT_HOST_OR_FALLBACK
+#define IEM_SELECT_HOST_OR_FALLBACK(a_fCpumFeatureMember, a_pfnNative, a_pfnFallback)   NULL
 
 #define iemAImpl_fpu_r32_to_r80         NULL