Changeset 100599 in vbox for trunk/src/VBox/VMM

Timestamp:

Jul 17, 2023 11:51:24 AM (17 months ago)

Author:

vboxsync

Message:

VMM/IEM: Implement vpaddusb/vpaddusw instruction emulations, bugref:9898

Location:

trunk/src/VBox/VMM

Files:

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

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

@@ -4301 +4301 @@
 IEMIMPL_MEDIA_OPT_F3 vpsubusb
 IEMIMPL_MEDIA_OPT_F3 vpsubusw
+IEMIMPL_MEDIA_OPT_F3 vpaddusb
+IEMIMPL_MEDIA_OPT_F3 vpaddusw
 
 

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

@@ -8582 +8582 @@
 #endif
 
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpaddusb_u128_fallback,(PRTUINT128U puDst,
+                                                         PCRTUINT128U puSrc1, PCRTUINT128U puSrc2))
+{
+    puDst->au8[0]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[0]  + puSrc2->au8[0]);
+    puDst->au8[1]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[1]  + puSrc2->au8[1]);
+    puDst->au8[2]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[2]  + puSrc2->au8[2]);
+    puDst->au8[3]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[3]  + puSrc2->au8[3]);
+    puDst->au8[4]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[4]  + puSrc2->au8[4]);
+    puDst->au8[5]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[5]  + puSrc2->au8[5]);
+    puDst->au8[6]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[6]  + puSrc2->au8[6]);
+    puDst->au8[7]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[7]  + puSrc2->au8[7]);
+    puDst->au8[8]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[8]  + puSrc2->au8[8]);
+    puDst->au8[9]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[9]  + puSrc2->au8[9]);
+    puDst->au8[10] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[10] + puSrc2->au8[10]);
+    puDst->au8[11] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[11] + puSrc2->au8[11]);
+    puDst->au8[12] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[12] + puSrc2->au8[12]);
+    puDst->au8[13] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[13] + puSrc2->au8[13]);
+    puDst->au8[14] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[14] + puSrc2->au8[14]);
+    puDst->au8[15] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[15] + puSrc2->au8[15]);
+}
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpaddusb_u256_fallback,(PRTUINT256U puDst,
+                                                         PCRTUINT256U puSrc1, PCRTUINT256U puSrc2))
+{
+    puDst->au8[0]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[0]  + puSrc2->au8[0]);
+    puDst->au8[1]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[1]  + puSrc2->au8[1]);
+    puDst->au8[2]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[2]  + puSrc2->au8[2]);
+    puDst->au8[3]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[3]  + puSrc2->au8[3]);
+    puDst->au8[4]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[4]  + puSrc2->au8[4]);
+    puDst->au8[5]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[5]  + puSrc2->au8[5]);
+    puDst->au8[6]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[6]  + puSrc2->au8[6]);
+    puDst->au8[7]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[7]  + puSrc2->au8[7]);
+    puDst->au8[8]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[8]  + puSrc2->au8[8]);
+    puDst->au8[9]  = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[9]  + puSrc2->au8[9]);
+    puDst->au8[10] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[10] + puSrc2->au8[10]);
+    puDst->au8[11] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[11] + puSrc2->au8[11]);
+    puDst->au8[12] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[12] + puSrc2->au8[12]);
+    puDst->au8[13] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[13] + puSrc2->au8[13]);
+    puDst->au8[14] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[14] + puSrc2->au8[14]);
+    puDst->au8[15] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[15] + puSrc2->au8[15]);
+    puDst->au8[16] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[16] + puSrc2->au8[16]);
+    puDst->au8[17] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[17] + puSrc2->au8[17]);
+    puDst->au8[18] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[18] + puSrc2->au8[18]);
+    puDst->au8[19] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[19] + puSrc2->au8[19]);
+    puDst->au8[20] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[20] + puSrc2->au8[20]);
+    puDst->au8[21] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[21] + puSrc2->au8[21]);
+    puDst->au8[22] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[22] + puSrc2->au8[22]);
+    puDst->au8[23] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[23] + puSrc2->au8[23]);
+    puDst->au8[24] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[24] + puSrc2->au8[24]);
+    puDst->au8[25] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[25] + puSrc2->au8[25]);
+    puDst->au8[26] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[26] + puSrc2->au8[26]);
+    puDst->au8[27] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[27] + puSrc2->au8[27]);
+    puDst->au8[28] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[28] + puSrc2->au8[28]);
+    puDst->au8[29] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[29] + puSrc2->au8[29]);
+    puDst->au8[30] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[30] + puSrc2->au8[30]);
+    puDst->au8[31] = SATURATED_UNSIGNED_WORD_TO_UNSIGNED_BYTE(puSrc1->au8[31] + puSrc2->au8[31]);
+}
+
 
 /*
@@ -8736 +8794 @@
 
 #endif
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpaddusw_u128_fallback,(PRTUINT128U puDst,
+                                                         PCRTUINT128U puSrc1, PCRTUINT128U puSrc2))
+{
+    puDst->au16[0] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[0] + puSrc2->au16[0]);
+    puDst->au16[1] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[1] + puSrc2->au16[1]);
+    puDst->au16[2] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[2] + puSrc2->au16[2]);
+    puDst->au16[3] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[3] + puSrc2->au16[3]);
+    puDst->au16[4] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[4] + puSrc2->au16[4]);
+    puDst->au16[5] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[5] + puSrc2->au16[5]);
+    puDst->au16[6] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[6] + puSrc2->au16[6]);
+    puDst->au16[7] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[7] + puSrc2->au16[7]);
+}
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpaddusw_u256_fallback,(PRTUINT256U puDst,
+                                                         PCRTUINT256U puSrc1, PCRTUINT256U puSrc2))
+{
+    puDst->au16[0]  = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[0]  + puSrc2->au16[0]);
+    puDst->au16[1]  = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[1]  + puSrc2->au16[1]);
+    puDst->au16[2]  = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[2]  + puSrc2->au16[2]);
+    puDst->au16[3]  = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[3]  + puSrc2->au16[3]);
+    puDst->au16[4]  = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[4]  + puSrc2->au16[4]);
+    puDst->au16[5]  = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[5]  + puSrc2->au16[5]);
+    puDst->au16[6]  = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[6]  + puSrc2->au16[6]);
+    puDst->au16[7]  = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[7]  + puSrc2->au16[7]);
+    puDst->au16[8]  = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[8]  + puSrc2->au16[8]);
+    puDst->au16[9]  = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[9]  + puSrc2->au16[9]);
+    puDst->au16[10] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[10] + puSrc2->au16[10]);
+    puDst->au16[11] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[11] + puSrc2->au16[11]);
+    puDst->au16[12] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[12] + puSrc2->au16[12]);
+    puDst->au16[13] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[13] + puSrc2->au16[13]);
+    puDst->au16[14] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[14] + puSrc2->au16[14]);
+    puDst->au16[15] = SATURATED_UNSIGNED_DWORD_TO_UNSIGNED_WORD(puSrc1->au16[15] + puSrc2->au16[15]);
+}
 
 

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

@@ -4853 +4853 @@
 
 /*  Opcode VEX.0F 0xdc - invalid */
+
+
 /** Opcode VEX.66.0F 0xdc - vpaddusb Vx, Hx, Wx */
-FNIEMOP_STUB(iemOp_vpaddusb_Vx_Hx_Wx);
+FNIEMOP_DEF(iemOp_vpaddusb_Vx_Hx_Wx)
+{
+    IEMOP_MNEMONIC3(VEX_RVM, VPADDUSB, vpaddusb, Vx, Hx, Wx, DISOPTYPE_HARMLESS, 0);
+    IEMOPMEDIAOPTF3_INIT_VARS(vpaddusb);
+    return FNIEMOP_CALL_1(iemOpCommonAvxAvx2_Vx_Hx_Wx_Opt, IEM_SELECT_HOST_OR_FALLBACK(fAvx2, &s_Host, &s_Fallback));
+}
+
+
 /*  Opcode VEX.F3.0F 0xdc - invalid */
 /*  Opcode VEX.F2.0F 0xdc - invalid */
 
 /*  Opcode VEX.0F 0xdd - invalid */
+
+
 /** Opcode VEX.66.0F 0xdd - vpaddusw Vx, Hx, Wx */
-FNIEMOP_STUB(iemOp_vpaddusw_Vx_Hx_Wx);
+FNIEMOP_DEF(iemOp_vpaddusw_Vx_Hx_Wx)
+{
+    IEMOP_MNEMONIC3(VEX_RVM, VPADDUSW, vpaddusw, Vx, Hx, Wx, DISOPTYPE_HARMLESS, 0);
+    IEMOPMEDIAOPTF3_INIT_VARS(vpaddusw);
+    return FNIEMOP_CALL_1(iemOpCommonAvxAvx2_Vx_Hx_Wx_Opt, IEM_SELECT_HOST_OR_FALLBACK(fAvx2, &s_Host, &s_Fallback));
+}
+
+
 /*  Opcode VEX.F3.0F 0xdd - invalid */
 /*  Opcode VEX.F2.0F 0xdd - invalid */

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

@@ -2341 +2341 @@
 FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpsubusb_u128,   iemAImpl_vpsubusb_u128_fallback;
 FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpsubusw_u128,   iemAImpl_vpsubusw_u128_fallback;
+FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpaddusb_u128,   iemAImpl_vpaddusb_u128_fallback;
+FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpaddusw_u128,   iemAImpl_vpaddusw_u128_fallback;
 
 FNIEMAIMPLMEDIAOPTF2U128 iemAImpl_vpabsb_u128,     iemAImpl_vpabsb_u128_fallback;
@@ -2408 +2410 @@
 FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpsubusb_u256,   iemAImpl_vpsubusb_u256_fallback;
 FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpsubusw_u256,   iemAImpl_vpsubusw_u256_fallback;
+FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpaddusb_u256,   iemAImpl_vpaddusb_u256_fallback;
+FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpaddusw_u256,   iemAImpl_vpaddusw_u256_fallback;
 
 FNIEMAIMPLMEDIAOPTF2U256 iemAImpl_vpabsb_u256,     iemAImpl_vpabsb_u256_fallback;