Changeset 95986 in vbox

Timestamp:

Aug 2, 2022 1:19:54 PM (3 years ago)

Author:

vboxsync

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

152746

Message:

VMM/IEM: Add implementation for vpmaxub instruction, bugref:9898

Location:

trunk/src/VBox/VMM

Files:

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

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

@@ -4086 +4086 @@
 IEMIMPL_MEDIA_F3 vpshufb
 IEMIMPL_MEDIA_F3 vpand
+IEMIMPL_MEDIA_F3 vpmaxub
 IEMIMPL_MEDIA_F3 vpandn
 IEMIMPL_MEDIA_F3 vpor

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

@@ -9555 +9555 @@
 #endif
 
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpmaxub_u128_fallback,(PX86XSAVEAREA pExtState, PRTUINT128U puDst,
+                                                        PCRTUINT128U puSrc1, PCRTUINT128U puSrc2))
+{
+    puDst->au8[ 0] = RT_MAX(puSrc1->au8[ 0], puSrc2->au8[ 0]);
+    puDst->au8[ 1] = RT_MAX(puSrc1->au8[ 1], puSrc2->au8[ 1]);
+    puDst->au8[ 2] = RT_MAX(puSrc1->au8[ 2], puSrc2->au8[ 2]);
+    puDst->au8[ 3] = RT_MAX(puSrc1->au8[ 3], puSrc2->au8[ 3]);
+    puDst->au8[ 4] = RT_MAX(puSrc1->au8[ 4], puSrc2->au8[ 4]);
+    puDst->au8[ 5] = RT_MAX(puSrc1->au8[ 5], puSrc2->au8[ 5]);
+    puDst->au8[ 6] = RT_MAX(puSrc1->au8[ 6], puSrc2->au8[ 6]);
+    puDst->au8[ 7] = RT_MAX(puSrc1->au8[ 7], puSrc2->au8[ 7]);
+    puDst->au8[ 8] = RT_MAX(puSrc1->au8[ 8], puSrc2->au8[ 8]);
+    puDst->au8[ 9] = RT_MAX(puSrc1->au8[ 9], puSrc2->au8[ 9]);
+    puDst->au8[10] = RT_MAX(puSrc1->au8[10], puSrc2->au8[10]);
+    puDst->au8[11] = RT_MAX(puSrc1->au8[11], puSrc2->au8[11]);
+    puDst->au8[12] = RT_MAX(puSrc1->au8[12], puSrc2->au8[12]);
+    puDst->au8[13] = RT_MAX(puSrc1->au8[13], puSrc2->au8[13]);
+    puDst->au8[14] = RT_MAX(puSrc1->au8[14], puSrc2->au8[14]);
+    puDst->au8[15] = RT_MAX(puSrc1->au8[15], puSrc2->au8[15]);
+    RT_NOREF(pExtState);
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vpmaxub_u256_fallback,(PX86XSAVEAREA pExtState, PRTUINT256U puDst,
+                                                        PCRTUINT256U puSrc1, PCRTUINT256U puSrc2))
+{
+    puDst->au8[ 0] = RT_MAX(puSrc1->au8[ 0], puSrc2->au8[ 0]);
+    puDst->au8[ 1] = RT_MAX(puSrc1->au8[ 1], puSrc2->au8[ 1]);
+    puDst->au8[ 2] = RT_MAX(puSrc1->au8[ 2], puSrc2->au8[ 2]);
+    puDst->au8[ 3] = RT_MAX(puSrc1->au8[ 3], puSrc2->au8[ 3]);
+    puDst->au8[ 4] = RT_MAX(puSrc1->au8[ 4], puSrc2->au8[ 4]);
+    puDst->au8[ 5] = RT_MAX(puSrc1->au8[ 5], puSrc2->au8[ 5]);
+    puDst->au8[ 6] = RT_MAX(puSrc1->au8[ 6], puSrc2->au8[ 6]);
+    puDst->au8[ 7] = RT_MAX(puSrc1->au8[ 7], puSrc2->au8[ 7]);
+    puDst->au8[ 8] = RT_MAX(puSrc1->au8[ 8], puSrc2->au8[ 8]);
+    puDst->au8[ 9] = RT_MAX(puSrc1->au8[ 9], puSrc2->au8[ 9]);
+    puDst->au8[10] = RT_MAX(puSrc1->au8[10], puSrc2->au8[10]);
+    puDst->au8[11] = RT_MAX(puSrc1->au8[11], puSrc2->au8[11]);
+    puDst->au8[12] = RT_MAX(puSrc1->au8[12], puSrc2->au8[12]);
+    puDst->au8[13] = RT_MAX(puSrc1->au8[13], puSrc2->au8[13]);
+    puDst->au8[14] = RT_MAX(puSrc1->au8[14], puSrc2->au8[14]);
+    puDst->au8[15] = RT_MAX(puSrc1->au8[15], puSrc2->au8[15]);
+    puDst->au8[16] = RT_MAX(puSrc1->au8[16], puSrc2->au8[16]);
+    puDst->au8[17] = RT_MAX(puSrc1->au8[17], puSrc2->au8[17]);
+    puDst->au8[18] = RT_MAX(puSrc1->au8[18], puSrc2->au8[18]);
+    puDst->au8[19] = RT_MAX(puSrc1->au8[19], puSrc2->au8[19]);
+    puDst->au8[20] = RT_MAX(puSrc1->au8[20], puSrc2->au8[20]);
+    puDst->au8[21] = RT_MAX(puSrc1->au8[21], puSrc2->au8[21]);
+    puDst->au8[22] = RT_MAX(puSrc1->au8[22], puSrc2->au8[22]);
+    puDst->au8[23] = RT_MAX(puSrc1->au8[23], puSrc2->au8[23]);
+    puDst->au8[24] = RT_MAX(puSrc1->au8[24], puSrc2->au8[24]);
+    puDst->au8[25] = RT_MAX(puSrc1->au8[25], puSrc2->au8[25]);
+    puDst->au8[26] = RT_MAX(puSrc1->au8[26], puSrc2->au8[26]);
+    puDst->au8[27] = RT_MAX(puSrc1->au8[27], puSrc2->au8[27]);
+    puDst->au8[28] = RT_MAX(puSrc1->au8[28], puSrc2->au8[28]);
+    puDst->au8[29] = RT_MAX(puSrc1->au8[29], puSrc2->au8[29]);
+    puDst->au8[30] = RT_MAX(puSrc1->au8[30], puSrc2->au8[30]);
+    puDst->au8[31] = RT_MAX(puSrc1->au8[31], puSrc2->au8[31]);
+    RT_NOREF(pExtState);
+}
+
 
 /*

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

@@ -4062 +4062 @@
 
 /*  Opcode VEX.0F 0xde - invalid */
-/** Opcode VEX.66.0F 0xde - vpmaxub Vx, Hx, W */
-FNIEMOP_STUB(iemOp_vpmaxub_Vx_Hx_W);
+
+
+/** Opcode VEX.66.0F 0xde - vpmaxub Vx, Hx, Wx */
+FNIEMOP_DEF(iemOp_vpmaxub_Vx_Hx_Wx)
+{
+    IEMOP_MNEMONIC3(VEX_RVM, VPMAXUB, vpmaxub, Vx, Hx, Wx, DISOPTYPE_HARMLESS, 0);
+    IEMOPMEDIAF3_INIT_VARS(vpmaxub);
+    return FNIEMOP_CALL_1(iemOpCommonAvxAvx2_Vx_Hx_Wx, IEM_SELECT_HOST_OR_FALLBACK(fAvx2, &s_Host, &s_Fallback));
+}
+
+
 /*  Opcode VEX.F3.0F 0xde - invalid */
 /*  Opcode VEX.F2.0F 0xde - invalid */
@@ -4675 +4684 @@
     /* 0xdc */  iemOp_InvalidNeedRM,        iemOp_vpaddusb_Vx_Hx_Wx,    iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
     /* 0xdd */  iemOp_InvalidNeedRM,        iemOp_vpaddusw_Vx_Hx_Wx,    iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
-    /* 0xde */  iemOp_InvalidNeedRM,        iemOp_vpmaxub_Vx_Hx_W,      iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
+    /* 0xde */  iemOp_InvalidNeedRM,        iemOp_vpmaxub_Vx_Hx_Wx,     iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
     /* 0xdf */  iemOp_InvalidNeedRM,        iemOp_vpandn_Vx_Hx_Wx,      iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
 

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

@@ -1828 +1828 @@
 FNIEMAIMPLMEDIAF3U128    iemAImpl_vpsubd_u128,     iemAImpl_vpsubd_u128_fallback;
 FNIEMAIMPLMEDIAF3U128    iemAImpl_vpsubq_u128,     iemAImpl_vpsubq_u128_fallback;
+FNIEMAIMPLMEDIAF3U128    iemAImpl_vpmaxub_u128,    iemAImpl_vpmaxub_u128_fallback;
 FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpacksswb_u128,  iemAImpl_vpacksswb_u128_fallback;
 FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vpackssdw_u128,  iemAImpl_vpackssdw_u128_fallback;
@@ -1854 +1855 @@
 FNIEMAIMPLMEDIAF3U256    iemAImpl_vpsubd_u256,     iemAImpl_vpsubd_u256_fallback;
 FNIEMAIMPLMEDIAF3U256    iemAImpl_vpsubq_u256,     iemAImpl_vpsubq_u256_fallback;
+FNIEMAIMPLMEDIAF3U256    iemAImpl_vpmaxub_u256,    iemAImpl_vpmaxub_u256_fallback;
 FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpacksswb_u256,  iemAImpl_vpacksswb_u256_fallback;
 FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vpackssdw_u256,  iemAImpl_vpackssdw_u256_fallback;