Changeset 96104 in vbox

Timestamp:

Aug 8, 2022 9:10:25 AM (2 years ago)

Author:

vboxsync

Message:

VMM/IEM: Implement [v]pmuludq instructions, bugref:9898

Location:

trunk/src/VBox/VMM

Files:

• VMMAll/IEMAllAImpl.asm (modified) (2 diffs)
• VMMAll/IEMAllAImplC.cpp (modified) (1 diff)
• VMMAll/IEMAllInstructionsTwoByte0f.cpp.h (modified) (1 diff)
• VMMAll/IEMAllInstructionsVexMap1.cpp.h (modified) (1 diff)
• include/IEMInternal.h (modified) (4 diffs)
• testcase/tstIEMCheckMc.cpp (modified) (2 diffs)

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

@@ -3660 +3660 @@
 IEMIMPL_MEDIA_F2 pmaddubsw, 1
 IEMIMPL_MEDIA_F2 pmulhrsw,  1
+IEMIMPL_MEDIA_F2 pmuludq,   1
 
 
@@ -4211 +4212 @@
 IEMIMPL_MEDIA_OPT_F3 vpsadbw
 IEMIMPL_MEDIA_OPT_F3 vpmuldq
+IEMIMPL_MEDIA_OPT_F3 vpmuludq
 
 

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

@@ -13197 +13197 @@
 
 /*
+ * PMULUDQ / VPMULUDQ
+ */
+#ifdef IEM_WITHOUT_ASSEMBLY
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_pmuludq_u64,(PCX86FXSTATE pFpuState, uint64_t *puDst, uint64_t const *puSrc))
+{
+    RTUINT64U uSrc1 = { *puDst };
+    RTUINT64U uSrc2 = { *puSrc };
+    ASMCompilerBarrier();
+    *puDst = (uint64_t)uSrc1.au32[0] * uSrc2.au32[0];
+    RT_NOREF(pFpuState);
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_pmuludq_u128,(PCX86FXSTATE pFpuState, PRTUINT128U puDst, PCRTUINT128U puSrc))
+{
+    RTUINT128U uSrc1 = *puDst;
+    RTUINT128U uSrc2 = *puSrc;
+    ASMCompilerBarrier();
+    puDst->au64[0] = (uint64_t)uSrc1.au32[0] * uSrc2.au32[0];
+    puDst->au64[1] = (uint64_t)uSrc1.au32[2] * uSrc2.au32[2];
+    RT_NOREF(pFpuState);
+}
+
+#endif
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpmuludq_u128_fallback,(PRTUINT128U puDst, PCRTUINT128U puSrc1, PCRTUINT128U puSrc2))
+{
+    RTUINT128U uSrc1 = *puSrc1; /* Could overlap with puDst */
+    RTUINT128U uSrc2 = *puSrc2; /* Could overlap with puDst */
+    ASMCompilerBarrier();
+    puDst->au64[0] = (uint64_t)uSrc1.au32[0] * uSrc2.au32[0];
+    puDst->au64[1] = (uint64_t)uSrc1.au32[2] * uSrc2.au32[2];
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpmuludq_u256_fallback,(PRTUINT256U puDst, PCRTUINT256U puSrc1, PCRTUINT256U puSrc2))
+{
+    RTUINT256U uSrc1 = *puSrc1; /* Could overlap with puDst */
+    RTUINT256U uSrc2 = *puSrc2; /* Could overlap with puDst */
+    ASMCompilerBarrier();
+    puDst->au64[0] = (uint64_t)uSrc1.au32[0] * uSrc2.au32[0];
+    puDst->au64[1] = (uint64_t)uSrc1.au32[2] * uSrc2.au32[2];
+    puDst->au64[2] = (uint64_t)uSrc1.au32[4] * uSrc2.au32[4];
+    puDst->au64[3] = (uint64_t)uSrc1.au32[6] * uSrc2.au32[6];
+}
+
+
+/*
  * CRC32 (SEE 4.2).
  */

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

@@ -10565 +10565 @@
 
 /** Opcode      0x0f 0xf4 - pmuludq Pq, Qq */
-FNIEMOP_STUB(iemOp_pmuludq_Pq_Qq);
+FNIEMOP_DEF(iemOp_pmuludq_Pq_Qq)
+{
+    IEMOP_MNEMONIC2(RM, PMULUDQ, pmuludq, Pq, Qq, DISOPTYPE_HARMLESS | DISOPTYPE_SSE, 0);
+    return FNIEMOP_CALL_1(iemOpCommonMmx_FullFull_To_Full, iemAImpl_pmuludq_u64);
+}
+
+
 /** Opcode 0x66 0x0f 0xf4 - pmuludq Vx, W */
-FNIEMOP_STUB(iemOp_pmuludq_Vx_Wx);
+FNIEMOP_DEF(iemOp_pmuludq_Vx_Wx)
+{
+    IEMOP_MNEMONIC2(RM, PMULUDQ, pmuludq, Vx, Wx, DISOPTYPE_HARMLESS | DISOPTYPE_SSE, 0);
+    return FNIEMOP_CALL_1(iemOpCommonSse2_FullFull_To_Full, iemAImpl_pmuludq_u128);
+}
+
+
 /*  Opcode 0xf2 0x0f 0xf4 - invalid */
 

trunk/src/VBox/VMM/VMMAll/IEMAllInstructionsVexMap1.cpp.h

@@ -4475 +4475 @@
 
 /*  Opcode VEX.0F 0xf4 - invalid */
+
+
 /** Opcode VEX.66.0F 0xf4 - vpmuludq Vx, Hx, W */
-FNIEMOP_STUB(iemOp_vpmuludq_Vx_Hx_W);
+FNIEMOP_DEF(iemOp_vpmuludq_Vx_Hx_W)
+{
+    IEMOP_MNEMONIC3(VEX_RVM, VPMULUDQ, vpmuludq, Vx, Hx, Wx, DISOPTYPE_HARMLESS, 0);
+    IEMOPMEDIAOPTF3_INIT_VARS(vpmuludq);
+    return FNIEMOP_CALL_1(iemOpCommonAvxAvx2_Vx_Hx_Wx_Opt, IEM_SELECT_HOST_OR_FALLBACK(fAvx2, &s_Host, &s_Fallback));
+}
+
+
 /*  Opcode VEX.F2.0F 0xf4 - invalid */
 

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

@@ -1795 +1795 @@
 FNIEMAIMPLMEDIAF2U64     iemAImpl_pmaddubsw_u64, iemAImpl_pmaddubsw_u64_fallback;
 FNIEMAIMPLMEDIAF2U64     iemAImpl_pmulhrsw_u64, iemAImpl_pmulhrsw_u64_fallback;
+FNIEMAIMPLMEDIAF2U64     iemAImpl_pmuludq_u64;
 FNIEMAIMPLMEDIAOPTF2U64  iemAImpl_psllw_u64, iemAImpl_psrlw_u64, iemAImpl_psraw_u64;
 FNIEMAIMPLMEDIAOPTF2U64  iemAImpl_pslld_u64, iemAImpl_psrld_u64, iemAImpl_psrad_u64;
@@ -1848 +1849 @@
 FNIEMAIMPLMEDIAF2U128    iemAImpl_pmaddubsw_u128, iemAImpl_pmaddubsw_u128_fallback;
 FNIEMAIMPLMEDIAF2U128    iemAImpl_pmulhrsw_u128, iemAImpl_pmulhrsw_u128_fallback;
+FNIEMAIMPLMEDIAF2U128    iemAImpl_pmuludq_u128;
 FNIEMAIMPLMEDIAOPTF2U128 iemAImpl_packsswb_u128, iemAImpl_packuswb_u128;
 FNIEMAIMPLMEDIAOPTF2U128 iemAImpl_packssdw_u128, iemAImpl_packusdw_u128;
@@ -1914 +1916 @@
 FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpsadbw_u128,    iemAImpl_vpsadbw_u128_fallback;
 FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpmuldq_u128,    iemAImpl_vpmuldq_u128_fallback;
+FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpmuludq_u128,   iemAImpl_vpmuludq_u128_fallback;
 
 FNIEMAIMPLMEDIAOPTF2U128 iemAImpl_vpabsb_u128,     iemAImpl_vpabsb_u128_fallback;
@@ -1975 +1978 @@
 FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpsadbw_u256,    iemAImpl_vpsadbw_u256_fallback;
 FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpmuldq_u256,    iemAImpl_vpmuldq_u256_fallback;
+FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpmuludq_u256,   iemAImpl_vpmuludq_u256_fallback;
 
 FNIEMAIMPLMEDIAOPTF2U256 iemAImpl_vpabsb_u256,     iemAImpl_vpabsb_u256_fallback;

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

@@ -447 +447 @@
 #define iemAImpl_pavgw_u64              NULL
 #define iemAImpl_psadbw_u64             NULL
+#define iemAImpl_pmuludq_u64            NULL
 
 #define iemAImpl_pmaxub_u128            NULL
@@ -455 +456 @@
 #define iemAImpl_pavgw_u128             NULL
 #define iemAImpl_psadbw_u128            NULL
+#define iemAImpl_pmuludq_u128           NULL
 
 /** @}  */