Changeset 103696 in vbox

Timestamp:

Mar 6, 2024 7:13:30 AM (13 months ago)

Author:

vboxsync

svn:sync-xref-src-repo-rev:

162063

Message:

VMM/IEM: Implement vpmaddwd instruction dispatch & emulation, bugref:9898

Location:

trunk/src/VBox/VMM

Files:

• VMMAll/IEMAllAImpl.asm (modified) (1 diff)
• VMMAll/IEMAllAImplC.cpp (modified) (1 diff)
• VMMAll/IEMAllInstVexMap1.cpp.h (modified) (1 diff)
• include/IEMInternal.h (modified) (5 diffs)

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

@@ -4428 +4428 @@
 IEMIMPL_MEDIA_OPT_F3 vpermilps
 IEMIMPL_MEDIA_OPT_F3 vpermilpd
+IEMIMPL_MEDIA_OPT_F3 vpmaddwd
 
 ;;

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

@@ -11253 +11253 @@
 
 
+IEM_DECL_IMPL_DEF(void, iemAImpl_pmaddwd_u64_fallback,(PCX86FXSTATE pFpuState, uint64_t *puDst, uint64_t const *puSrc))
+{
+    RTUINT64U uSrc1 = { *puDst };
+    RTUINT64U uSrc2 = { *puSrc };
+    RTUINT64U uDst;
+
+    uDst.ai32[0] = (int32_t)uSrc1.ai16[0] * uSrc2.ai16[0] + (int32_t)uSrc1.ai16[1] * uSrc2.ai16[1];
+    uDst.ai32[1] = (int32_t)uSrc1.ai16[2] * uSrc2.ai16[2] + (int32_t)uSrc1.ai16[3] * uSrc2.ai16[3];
+    *puDst = uDst.u;
+    RT_NOREF(pFpuState);
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_pmaddwd_u128_fallback,(PCX86FXSTATE pFpuState, PRTUINT128U puDst, PCRTUINT128U puSrc))
+{
+    RTUINT128U uSrc1 = *puDst;
+
+    puDst->ai32[0] = (int32_t)uSrc1.ai16[0] * puSrc->ai16[0] + (int32_t)uSrc1.ai16[1] * puSrc->ai16[1];
+    puDst->ai32[1] = (int32_t)uSrc1.ai16[2] * puSrc->ai16[2] + (int32_t)uSrc1.ai16[3] * puSrc->ai16[3];
+    puDst->ai32[2] = (int32_t)uSrc1.ai16[4] * puSrc->ai16[4] + (int32_t)uSrc1.ai16[5] * puSrc->ai16[5];
+    puDst->ai32[3] = (int32_t)uSrc1.ai16[6] * puSrc->ai16[6] + (int32_t)uSrc1.ai16[7] * puSrc->ai16[7];
+    RT_NOREF(pFpuState);
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpmaddwd_u128_fallback,(PRTUINT128U puDst, PCRTUINT128U puSrc1, PCRTUINT128U puSrc2))
+{
+    puDst->ai32[0] = (int32_t)puSrc1->ai16[0] * puSrc2->ai16[0] + (int32_t)puSrc1->ai16[1] * puSrc2->ai16[1];
+    puDst->ai32[1] = (int32_t)puSrc1->ai16[2] * puSrc2->ai16[2] + (int32_t)puSrc1->ai16[3] * puSrc2->ai16[3];
+    puDst->ai32[2] = (int32_t)puSrc1->ai16[4] * puSrc2->ai16[4] + (int32_t)puSrc1->ai16[5] * puSrc2->ai16[5];
+    puDst->ai32[3] = (int32_t)puSrc1->ai16[6] * puSrc2->ai16[6] + (int32_t)puSrc1->ai16[7] * puSrc2->ai16[7];
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpmaddwd_u256_fallback,(PRTUINT256U puDst, PCRTUINT256U puSrc1, PCRTUINT256U puSrc2))
+{
+    puDst->ai32[0] = (int32_t)puSrc1->ai16[0] * puSrc2->ai16[0] + (int32_t)puSrc1->ai16[1] * puSrc2->ai16[1];
+    puDst->ai32[1] = (int32_t)puSrc1->ai16[2] * puSrc2->ai16[2] + (int32_t)puSrc1->ai16[3] * puSrc2->ai16[3];
+    puDst->ai32[2] = (int32_t)puSrc1->ai16[4] * puSrc2->ai16[4] + (int32_t)puSrc1->ai16[5] * puSrc2->ai16[5];
+    puDst->ai32[3] = (int32_t)puSrc1->ai16[6] * puSrc2->ai16[6] + (int32_t)puSrc1->ai16[7] * puSrc2->ai16[7];
+    puDst->ai32[4] = (int32_t)puSrc1->ai16[8] * puSrc2->ai16[8] + (int32_t)puSrc1->ai16[9] * puSrc2->ai16[9];
+    puDst->ai32[5] = (int32_t)puSrc1->ai16[10] * puSrc2->ai16[10] + (int32_t)puSrc1->ai16[11] * puSrc2->ai16[11];
+    puDst->ai32[6] = (int32_t)puSrc1->ai16[12] * puSrc2->ai16[12] + (int32_t)puSrc1->ai16[13] * puSrc2->ai16[13];
+    puDst->ai32[7] = (int32_t)puSrc1->ai16[14] * puSrc2->ai16[14] + (int32_t)puSrc1->ai16[15] * puSrc2->ai16[15];
+}
+
+
 /*
  * PMAXUB / VPMAXUB / PMAXUW / VPMAXUW / PMAXUD / VPMAXUD

trunk/src/VBox/VMM/VMMAll/IEMAllInstVexMap1.cpp.h

@@ -5522 +5522 @@
 
 /*  Opcode VEX.0F 0xf5 - invalid */
+
+
 /** Opcode VEX.66.0F 0xf5 - vpmaddwd Vx, Hx, Wx */
-FNIEMOP_STUB(iemOp_vpmaddwd_Vx_Hx_Wx);
+FNIEMOP_DEF(iemOp_vpmaddwd_Vx_Hx_Wx)
+{
+    IEMOP_MNEMONIC3(VEX_RVM, VPMADDWD, vpmaddwd, Vx, Hx, Wx, DISOPTYPE_HARMLESS, 0);
+    IEMOPMEDIAOPTF3_INIT_VARS(vpmaddwd);
+    return FNIEMOP_CALL_1(iemOpCommonAvxAvx2_Vx_Hx_Wx_Opt, IEM_SELECT_HOST_OR_FALLBACK(fAvx2, &s_Host, &s_Fallback));
+}
+
+
 /*  Opcode VEX.F2.0F 0xf5 - invalid */
 

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

@@ -3092 +3092 @@
 FNIEMAIMPLMEDIAF2U64     iemAImpl_psubd_u64;
 FNIEMAIMPLMEDIAF2U64     iemAImpl_psubq_u64;
-FNIEMAIMPLMEDIAF2U64     iemAImpl_pmaddwd_u64;
+FNIEMAIMPLMEDIAF2U64     iemAImpl_pmaddwd_u64, iemAImpl_pmaddwd_u64_fallback;
 FNIEMAIMPLMEDIAF2U64     iemAImpl_pmullw_u64, iemAImpl_pmulhw_u64;
 FNIEMAIMPLMEDIAF2U64     iemAImpl_pminub_u64, iemAImpl_pmaxub_u64;
@@ -3137 +3137 @@
 FNIEMAIMPLMEDIAF2U128    iemAImpl_pmulhw_u128;
 FNIEMAIMPLMEDIAF2U128    iemAImpl_pmulld_u128, iemAImpl_pmulld_u128_fallback;
-FNIEMAIMPLMEDIAF2U128    iemAImpl_pmaddwd_u128;
+FNIEMAIMPLMEDIAF2U128    iemAImpl_pmaddwd_u128, iemAImpl_pmaddwd_u128_fallback;
 FNIEMAIMPLMEDIAF2U128    iemAImpl_pminub_u128;
 FNIEMAIMPLMEDIAF2U128    iemAImpl_pminud_u128, iemAImpl_pminud_u128_fallback;
@@ -3165 +3165 @@
 FNIEMAIMPLMEDIAF2U128    iemAImpl_pmulhrsw_u128, iemAImpl_pmulhrsw_u128_fallback;
 FNIEMAIMPLMEDIAF2U128    iemAImpl_pmuludq_u128;
+FNIEMAIMPLMEDIAF2U128    iemAImpl_pmaddwd_u128, iemAImpl_pmaddwd_u128_fallback;
 FNIEMAIMPLMEDIAOPTF2U128 iemAImpl_packsswb_u128, iemAImpl_packuswb_u128;
 FNIEMAIMPLMEDIAOPTF2U128 iemAImpl_packssdw_u128, iemAImpl_packusdw_u128;
@@ -3251 +3252 @@
 FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpsrld_u128,     iemAImpl_vpsrld_u128_fallback;
 FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpsrlq_u128,     iemAImpl_vpsrlq_u128_fallback;
+FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpmaddwd_u128, iemAImpl_vpmaddwd_u128_fallback;
 
 FNIEMAIMPLMEDIAOPTF2U128 iemAImpl_vpabsb_u128,     iemAImpl_vpabsb_u128_fallback;
@@ -3330 +3332 @@
 FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpsrld_u256,     iemAImpl_vpsrld_u256_fallback;
 FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpsrlq_u256,     iemAImpl_vpsrlq_u256_fallback;
+FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpmaddwd_u256, iemAImpl_vpmaddwd_u256_fallback;
 
 FNIEMAIMPLMEDIAOPTF2U256 iemAImpl_vpabsb_u256,     iemAImpl_vpabsb_u256_fallback;