Changeset 105306 in vbox

Timestamp:

Jul 12, 2024 1:20:30 PM (5 months ago)

Author:

vboxsync

Message:

VMM/IEM: Implement vpermq, vpermpd instruction emulations, bugref:9898

Location:

trunk/src/VBox/VMM

Files:

• VMMAll/IEMAllAImpl.asm (modified) (1 diff)
• VMMAll/IEMAllAImplC.cpp (modified) (1 diff)
• VMMAll/IEMAllInstVexMap3.cpp.h (modified) (1 diff)
• include/IEMInternal.h (modified) (1 diff)

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

@@ -4950 +4950 @@
 
 ;;
+; VPERMQ instruction.
+;
+; @param    A0      Pointer to the first media register size operand (output).
+; @param    A1      Pointer to the second media register size operand (input).
+; @param    A2      8-bit immediate for the round constant.
+;
+BEGINPROC_FASTCALL iemAImpl_vpermq_u256, 16
+        PROLOGUE_3_ARGS
+        IEMIMPL_AVX_PROLOGUE
+
+        movzx   A2, A2_8                ; must clear top bits
+        vmovdqu ymm1, [A1]
+        IEMIMPL_CALL_JUMP_TABLE_TARGET T1, A2, 8
+        vmovdqu [A0], ymm0
+
+        IEMIMPL_AVX_EPILOGUE
+        EPILOGUE_3_ARGS
+ %assign bImm 0
+ %rep 256
+.imm %+ bImm:
+        IBT_ENDBRxx_WITHOUT_NOTRACK
+        vpermq ymm0, ymm1, bImm
+        ret
+        int3
+  %assign bImm bImm + 1
+ %endrep
+.immEnd:
+ENDPROC iemAImpl_vpermq_u256
+
+
+;;
+; VPERMPD instruction.
+;
+; @param    A0      Pointer to the first media register size operand (output).
+; @param    A1      Pointer to the second media register size operand (input).
+; @param    A2      8-bit immediate for the round constant.
+;
+BEGINPROC_FASTCALL iemAImpl_vpermpd_u256, 16
+        PROLOGUE_3_ARGS
+        IEMIMPL_AVX_PROLOGUE
+
+        movzx   A2, A2_8                ; must clear top bits
+        vmovdqu ymm1, [A1]
+        IEMIMPL_CALL_JUMP_TABLE_TARGET T1, A2, 8
+        vmovdqu [A0], ymm0
+
+        IEMIMPL_AVX_EPILOGUE
+        EPILOGUE_3_ARGS
+ %assign bImm 0
+ %rep 256
+.imm %+ bImm:
+        IBT_ENDBRxx_WITHOUT_NOTRACK
+        vpermpd ymm0, ymm1, bImm
+        ret
+        int3
+  %assign bImm bImm + 1
+ %endrep
+.immEnd:
+ENDPROC iemAImpl_vpermpd_u256
+
+
+;;
 ; Media instruction working on one full sized source register, one full sized destination
 ; register, and one no-larger-than-XMM register (in the vps{ll,ra,rl}[dwq] instructions,

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

@@ -20416 +20416 @@
     return uMxCsrIn;
 }
+
+
+/**
+ * VPERMQ
+ */
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpermq_u256_fallback,(PRTUINT256U puDst, PCRTUINT256U puSrc, uint8_t bImm))
+{
+    puDst->au64[0] = puSrc->au64[bImm        & 0x3];
+    puDst->au64[1] = puSrc->au64[(bImm >> 2) & 0x3];
+    puDst->au64[2] = puSrc->au64[(bImm >> 4) & 0x3];
+    puDst->au64[3] = puSrc->au64[(bImm >> 6) & 0x3];
+}
+
+
+/**
+ * VPERMPD
+ */
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpermpd_u256_fallback,(PRTUINT256U puDst, PCRTUINT256U puSrc, uint8_t bImm))
+{
+    puDst->au64[0] = puSrc->au64[bImm        & 0x3];
+    puDst->au64[1] = puSrc->au64[(bImm >> 2) & 0x3];
+    puDst->au64[2] = puSrc->au64[(bImm >> 4) & 0x3];
+    puDst->au64[3] = puSrc->au64[(bImm >> 6) & 0x3];
+}

trunk/src/VBox/VMM/VMMAll/IEMAllInstVexMap3.cpp.h

@@ -483 +483 @@
 
 /** Opcode VEX.66.0F3A 0x00. */
-FNIEMOP_STUB(iemOp_vpermq_Vqq_Wqq_Ib);
+FNIEMOP_DEF(iemOp_vpermq_Vqq_Wqq_Ib)
+{
+    uint8_t bRm; IEM_OPCODE_GET_NEXT_U8(&bRm);
+    if (IEM_IS_MODRM_REG_MODE(bRm))
+    {
+        /*
+         * Register, register.
+         */
+        uint8_t bImm; IEM_OPCODE_GET_NEXT_U8(&bImm);
+        IEM_MC_BEGIN(IEM_MC_F_NOT_286_OR_OLDER, 0);
+        IEMOP_HLP_DONE_VEX_DECODING_L0_AND_NO_VVVV_EX(fAvx2);
+        IEM_MC_LOCAL(RTUINT256U,            uDst);
+        IEM_MC_LOCAL(RTUINT256U,            uSrc);
+        IEM_MC_ARG_LOCAL_REF(PRTUINT256U,   puDst, uDst,         0);
+        IEM_MC_ARG_LOCAL_REF(PCRTUINT256U,  puSrc, uSrc,         1);
+        IEM_MC_ARG_CONST(uint8_t,           bImmArg, /*=*/ bImm, 2);
+        IEM_MC_MAYBE_RAISE_AVX_RELATED_XCPT();
+        IEM_MC_PREPARE_AVX_USAGE();
+        IEM_MC_FETCH_YREG_U256(uSrc, IEM_GET_MODRM_RM(pVCpu, bRm));
+        IEM_MC_CALL_VOID_AIMPL_3(IEM_SELECT_HOST_OR_FALLBACK(fAvx2, iemAImpl_vpermq_u256, iemAImpl_vpermq_u256_fallback),
+                                 puDst, puSrc, bImmArg);
+        IEM_MC_STORE_YREG_U256_ZX_VLMAX(IEM_GET_MODRM_REG(pVCpu, bRm), uDst);
+        IEM_MC_ADVANCE_RIP_AND_FINISH();
+        IEM_MC_END();
+    }
+    else
+    {
+        /*
+         * Register, memory.
+         */
+        IEM_MC_BEGIN(IEM_MC_F_NOT_286_OR_OLDER, 0);
+        IEM_MC_LOCAL(RTGCPTR,               GCPtrEffSrc);
+        IEM_MC_CALC_RM_EFF_ADDR(GCPtrEffSrc, bRm, 0);
+        uint8_t bImm; IEM_OPCODE_GET_NEXT_U8(&bImm);
+        IEMOP_HLP_DONE_VEX_DECODING_L0_AND_NO_VVVV_EX(fAvx2);
+        IEM_MC_MAYBE_RAISE_AVX_RELATED_XCPT();
+        IEM_MC_PREPARE_AVX_USAGE();
+
+        IEM_MC_LOCAL(RTUINT256U,            uSrc);
+        IEM_MC_ARG_LOCAL_REF(PCRTUINT256U,  puSrc, uSrc,         1);
+        IEM_MC_FETCH_MEM_U256_NO_AC(uSrc, pVCpu->iem.s.iEffSeg, GCPtrEffSrc);
+        IEM_MC_LOCAL(RTUINT256U,            uDst);
+        IEM_MC_ARG_LOCAL_REF(PRTUINT256U,   puDst, uDst,         0);
+        IEM_MC_ARG_CONST(uint8_t,           bImmArg, /*=*/ bImm, 2);
+        IEM_MC_CALL_VOID_AIMPL_3(IEM_SELECT_HOST_OR_FALLBACK(fAvx2, iemAImpl_vpermq_u256, iemAImpl_vpermq_u256_fallback),
+                                 puDst, puSrc, bImmArg);
+        IEM_MC_STORE_YREG_U256_ZX_VLMAX(IEM_GET_MODRM_REG(pVCpu, bRm), uDst);
+        IEM_MC_ADVANCE_RIP_AND_FINISH();
+        IEM_MC_END();
+    }
+}
+
+
 /** Opcode VEX.66.0F3A 0x01. */
-FNIEMOP_STUB(iemOp_vpermpd_Vqq_Wqq_Ib);
+FNIEMOP_DEF(iemOp_vpermpd_Vqq_Wqq_Ib)
+{
+    uint8_t bRm; IEM_OPCODE_GET_NEXT_U8(&bRm);
+    if (IEM_IS_MODRM_REG_MODE(bRm))
+    {
+        /*
+         * Register, register.
+         */
+        uint8_t bImm; IEM_OPCODE_GET_NEXT_U8(&bImm);
+        IEM_MC_BEGIN(IEM_MC_F_NOT_286_OR_OLDER, 0);
+        IEMOP_HLP_DONE_VEX_DECODING_L0_AND_NO_VVVV_EX(fAvx2);
+        IEM_MC_LOCAL(RTUINT256U,            uDst);
+        IEM_MC_LOCAL(RTUINT256U,            uSrc);
+        IEM_MC_ARG_LOCAL_REF(PRTUINT256U,   puDst, uDst,         0);
+        IEM_MC_ARG_LOCAL_REF(PCRTUINT256U,  puSrc, uSrc,         1);
+        IEM_MC_ARG_CONST(uint8_t,           bImmArg, /*=*/ bImm, 2);
+        IEM_MC_MAYBE_RAISE_AVX_RELATED_XCPT();
+        IEM_MC_PREPARE_AVX_USAGE();
+        IEM_MC_FETCH_YREG_U256(uSrc, IEM_GET_MODRM_RM(pVCpu, bRm));
+        IEM_MC_CALL_VOID_AIMPL_3(IEM_SELECT_HOST_OR_FALLBACK(fAvx2, iemAImpl_vpermpd_u256, iemAImpl_vpermpd_u256_fallback),
+                                 puDst, puSrc, bImmArg);
+        IEM_MC_STORE_YREG_U256_ZX_VLMAX(IEM_GET_MODRM_REG(pVCpu, bRm), uDst);
+        IEM_MC_ADVANCE_RIP_AND_FINISH();
+        IEM_MC_END();
+    }
+    else
+    {
+        /*
+         * Register, memory.
+         */
+        IEM_MC_BEGIN(IEM_MC_F_NOT_286_OR_OLDER, 0);
+        IEM_MC_LOCAL(RTGCPTR,               GCPtrEffSrc);
+        IEM_MC_CALC_RM_EFF_ADDR(GCPtrEffSrc, bRm, 0);
+        uint8_t bImm; IEM_OPCODE_GET_NEXT_U8(&bImm);
+        IEMOP_HLP_DONE_VEX_DECODING_L0_AND_NO_VVVV_EX(fAvx2);
+        IEM_MC_MAYBE_RAISE_AVX_RELATED_XCPT();
+        IEM_MC_PREPARE_AVX_USAGE();
+
+        IEM_MC_LOCAL(RTUINT256U,            uSrc);
+        IEM_MC_ARG_LOCAL_REF(PCRTUINT256U,  puSrc, uSrc,         1);
+        IEM_MC_FETCH_MEM_U256_NO_AC(uSrc, pVCpu->iem.s.iEffSeg, GCPtrEffSrc);
+        IEM_MC_LOCAL(RTUINT256U,            uDst);
+        IEM_MC_ARG_LOCAL_REF(PRTUINT256U,   puDst, uDst,         0);
+        IEM_MC_ARG_CONST(uint8_t,           bImmArg, /*=*/ bImm, 2);
+        IEM_MC_CALL_VOID_AIMPL_3(IEM_SELECT_HOST_OR_FALLBACK(fAvx2, iemAImpl_vpermpd_u256, iemAImpl_vpermpd_u256_fallback),
+                                 puDst, puSrc, bImmArg);
+        IEM_MC_STORE_YREG_U256_ZX_VLMAX(IEM_GET_MODRM_REG(pVCpu, bRm), uDst);
+        IEM_MC_ADVANCE_RIP_AND_FINISH();
+        IEM_MC_END();
+    }
+}
 
 

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

@@ -4000 +4000 @@
 IEM_DECL_IMPL_DEF(void, iemAImpl_sha256rnds2_u128,(PRTUINT128U puDst, PCRTUINT128U puSrc, PCRTUINT128U puXmm0Constants));
 IEM_DECL_IMPL_DEF(void, iemAImpl_sha256rnds2_u128_fallback,(PRTUINT128U puDst, PCRTUINT128U puSrc, PCRTUINT128U puXmm0Constants));
+
+FNIEMAIMPLMEDIAOPTF2U256IMM8 iemAImpl_vpermq_u256,      iemAImpl_vpermq_u256_fallback;
+FNIEMAIMPLMEDIAOPTF2U256IMM8 iemAImpl_vpermpd_u256,     iemAImpl_vpermpd_u256_fallback;
 
 typedef struct IEMPCMPISTRXSRC