Changeset 96049 in vbox

Timestamp:

Aug 5, 2022 10:17:21 AM (2 years ago)

Author:

vboxsync

Message:

VMM/IEM: Implement [v]phsub{w,d} instructions, bugref:9898

Location:

trunk/src/VBox/VMM

Files:

• VMMAll/IEMAllAImpl.asm (modified) (2 diffs)
• VMMAll/IEMAllAImplC.cpp (modified) (1 diff)
• VMMAll/IEMAllInstructionsThree0f38.cpp.h (modified) (2 diffs)
• VMMAll/IEMAllInstructionsVexMap2.cpp.h (modified) (2 diffs)
• include/IEMInternal.h (modified) (4 diffs)

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

@@ -3654 +3654 @@
 IEMIMPL_MEDIA_F2 phaddw,  1
 IEMIMPL_MEDIA_F2 phaddd,  1
+IEMIMPL_MEDIA_F2 phsubw,  1
+IEMIMPL_MEDIA_F2 phsubd,  1
 
 
@@ -4195 +4197 @@
 IEMIMPL_MEDIA_OPT_F3 vphaddw
 IEMIMPL_MEDIA_OPT_F3 vphaddd
+IEMIMPL_MEDIA_OPT_F3 vphsubw
+IEMIMPL_MEDIA_OPT_F3 vphsubd
 
 

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

@@ -12555 +12555 @@
 
 /*
+ * PHSUBW / VPHSUBW / PHSUBD / VPHSUBD
+ */
+IEM_DECL_IMPL_DEF(void, iemAImpl_phsubw_u64_fallback,(PCX86FXSTATE pFpuState, uint64_t *puDst, uint64_t const *puSrc))
+{
+    RTUINT64U uSrc1 = { *puDst };
+    RTUINT64U uSrc2 = { *puSrc };
+    RTUINT64U uDst = { 0 }; /* Shut up MSVC. */
+
+    uDst.ai16[0] = uSrc1.ai16[0] - uSrc1.ai16[1];
+    uDst.ai16[1] = uSrc1.ai16[2] - uSrc1.ai16[3];
+    uDst.ai16[2] = uSrc2.ai16[0] - uSrc2.ai16[1];
+    uDst.ai16[3] = uSrc2.ai16[2] - uSrc2.ai16[3];
+    *puDst = uDst.u;
+    RT_NOREF(pFpuState);
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_phsubw_u128_fallback,(PCX86FXSTATE pFpuState, PRTUINT128U puDst, PCRTUINT128U puSrc))
+{
+    RTUINT128U uSrc1 = *puDst;
+
+    puDst->ai16[0] = uSrc1.ai16[0] - uSrc1.ai16[1];
+    puDst->ai16[1] = uSrc1.ai16[2] - uSrc1.ai16[3];
+    puDst->ai16[2] = uSrc1.ai16[4] - uSrc1.ai16[5];
+    puDst->ai16[3] = uSrc1.ai16[6] - uSrc1.ai16[7];
+
+    puDst->ai16[4] = puSrc->ai16[0] - puSrc->ai16[1];
+    puDst->ai16[5] = puSrc->ai16[2] - puSrc->ai16[3];
+    puDst->ai16[6] = puSrc->ai16[4] - puSrc->ai16[5];
+    puDst->ai16[7] = puSrc->ai16[6] - puSrc->ai16[7];
+    RT_NOREF(pFpuState);
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_phsubd_u64_fallback,(PCX86FXSTATE pFpuState, uint64_t *puDst, uint64_t const *puSrc))
+{
+    RTUINT64U uSrc1 = { *puDst };
+    RTUINT64U uSrc2 = { *puSrc };
+    RTUINT64U uDst = { 0 }; /* Shut up MSVC. */
+
+    uDst.ai32[0] = uSrc1.ai32[0] - uSrc1.ai32[1];
+    uDst.ai32[1] = uSrc2.ai32[0] - uSrc2.ai32[1];
+    *puDst = uDst.u;
+    RT_NOREF(pFpuState);
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_phsubd_u128_fallback,(PCX86FXSTATE pFpuState, PRTUINT128U puDst, PCRTUINT128U puSrc))
+{
+    RTUINT128U uSrc1 = *puDst;
+
+    puDst->ai32[0] = uSrc1.ai32[0] - uSrc1.ai32[1];
+    puDst->ai32[1] = uSrc1.ai32[2] - uSrc1.ai32[3];
+
+    puDst->ai32[2] = puSrc->ai32[0] - puSrc->ai32[1];
+    puDst->ai32[3] = puSrc->ai32[2] - puSrc->ai32[3];
+    RT_NOREF(pFpuState);
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vphsubw_u128_fallback,(PRTUINT128U puDst, PCRTUINT128U puSrc1, PCRTUINT128U puSrc2))
+{
+    RTUINT128U uDst; /* puDst can be the same as one of the source operands. */
+
+    uDst.ai16[0] = puSrc1->ai16[0] - puSrc1->ai16[1];
+    uDst.ai16[1] = puSrc1->ai16[2] - puSrc1->ai16[3];
+    uDst.ai16[2] = puSrc1->ai16[4] - puSrc1->ai16[5];
+    uDst.ai16[3] = puSrc1->ai16[6] - puSrc1->ai16[7];
+
+    uDst.ai16[4] = puSrc2->ai16[0] - puSrc2->ai16[1];
+    uDst.ai16[5] = puSrc2->ai16[2] - puSrc2->ai16[3];
+    uDst.ai16[6] = puSrc2->ai16[4] - puSrc2->ai16[5];
+    uDst.ai16[7] = puSrc2->ai16[6] - puSrc2->ai16[7];
+
+    puDst->au64[0] = uDst.au64[0];
+    puDst->au64[1] = uDst.au64[1];
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vphsubw_u256_fallback,(PRTUINT256U puDst, PCRTUINT256U puSrc1, PCRTUINT256U puSrc2))
+{
+    RTUINT256U uDst; /* puDst can be the same as one of the source operands. */
+
+    uDst.ai16[ 0] = puSrc1->ai16[ 0] - puSrc1->ai16[ 1];
+    uDst.ai16[ 1] = puSrc1->ai16[ 2] - puSrc1->ai16[ 3];
+    uDst.ai16[ 2] = puSrc1->ai16[ 4] - puSrc1->ai16[ 5];
+    uDst.ai16[ 3] = puSrc1->ai16[ 6] - puSrc1->ai16[ 7];
+    uDst.ai16[ 4] = puSrc2->ai16[ 0] - puSrc2->ai16[ 1];
+    uDst.ai16[ 5] = puSrc2->ai16[ 2] - puSrc2->ai16[ 3];
+    uDst.ai16[ 6] = puSrc2->ai16[ 4] - puSrc2->ai16[ 5];
+    uDst.ai16[ 7] = puSrc2->ai16[ 6] - puSrc2->ai16[ 7];
+
+    uDst.ai16[ 8] = puSrc1->ai16[ 8] - puSrc1->ai16[ 9];
+    uDst.ai16[ 9] = puSrc1->ai16[10] - puSrc1->ai16[11];
+    uDst.ai16[10] = puSrc1->ai16[12] - puSrc1->ai16[13];
+    uDst.ai16[11] = puSrc1->ai16[14] - puSrc1->ai16[15];
+    uDst.ai16[12] = puSrc2->ai16[ 8] - puSrc2->ai16[ 9];
+    uDst.ai16[13] = puSrc2->ai16[10] - puSrc2->ai16[11];
+    uDst.ai16[14] = puSrc2->ai16[12] - puSrc2->ai16[13];
+    uDst.ai16[15] = puSrc2->ai16[14] - puSrc2->ai16[15];
+
+    puDst->au64[0] = uDst.au64[0];
+    puDst->au64[1] = uDst.au64[1];
+    puDst->au64[2] = uDst.au64[2];
+    puDst->au64[3] = uDst.au64[3];
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vphsubd_u128_fallback,(PRTUINT128U puDst, PCRTUINT128U puSrc1, PCRTUINT128U puSrc2))
+{
+    RTUINT128U uDst; /* puDst can be the same as one of the source operands. */
+
+    uDst.ai32[0] = puSrc1->ai32[0] - puSrc1->ai32[1];
+    uDst.ai32[1] = puSrc1->ai32[2] - puSrc1->ai32[3];
+
+    uDst.ai32[2] = puSrc2->ai32[0] - puSrc2->ai32[1];
+    uDst.ai32[3] = puSrc2->ai32[2] - puSrc2->ai32[3];
+
+    puDst->au64[0] = uDst.au64[0];
+    puDst->au64[1] = uDst.au64[1];
+}
+
+
+IEM_DECL_IMPL_DEF(void, iemAImpl_vphsubd_u256_fallback,(PRTUINT256U puDst, PCRTUINT256U puSrc1, PCRTUINT256U puSrc2))
+{
+    RTUINT256U uDst; /* puDst can be the same as one of the source operands. */
+
+    uDst.ai32[0] = puSrc1->ai32[ 0] - puSrc1->ai32[ 1];
+    uDst.ai32[1] = puSrc1->ai32[ 2] - puSrc1->ai32[ 3];
+    uDst.ai32[2] = puSrc2->ai32[ 0] - puSrc2->ai32[ 1];
+    uDst.ai32[3] = puSrc2->ai32[ 2] - puSrc2->ai32[ 3];
+
+    uDst.ai32[4] = puSrc1->ai32[ 4] - puSrc1->ai32[ 5];
+    uDst.ai32[5] = puSrc1->ai32[ 6] - puSrc1->ai32[ 7];
+    uDst.ai32[6] = puSrc2->ai32[ 4] - puSrc2->ai32[ 5];
+    uDst.ai32[7] = puSrc2->ai32[ 6] - puSrc2->ai32[ 7];
+
+    puDst->au64[0] = uDst.au64[0];
+    puDst->au64[1] = uDst.au64[1];
+    puDst->au64[2] = uDst.au64[2];
+    puDst->au64[3] = uDst.au64[3];
+}
+
+
+/*
  * CRC32 (SEE 4.2).
  */

trunk/src/VBox/VMM/VMMAll/IEMAllInstructionsThree0f38.cpp.h

@@ -325 +325 @@
 /** Opcode 0x66 0x0f 0x38 0x04. */
 FNIEMOP_STUB(iemOp_pmaddubsw_Vx_Wx);
+
+
 /** Opcode      0x0f 0x38 0x05. */
-FNIEMOP_STUB(iemOp_phsubw_Pq_Qq);
+FNIEMOP_DEF(iemOp_phsubw_Pq_Qq)
+{
+    IEMOP_MNEMONIC2(RM, PHSUBW, phsubw, Pq, Qq, DISOPTYPE_HARMLESS, IEMOPHINT_IGNORES_OP_SIZES);
+    return FNIEMOP_CALL_2(iemOpCommonMmx_FullFull_To_Full_Ex,
+                          IEM_SELECT_HOST_OR_FALLBACK(fSsse3, iemAImpl_phsubw_u64,&iemAImpl_phsubw_u64_fallback),
+                          IEM_GET_GUEST_CPU_FEATURES(pVCpu)->fSsse3);
+}
+
+
 /** Opcode 0x66 0x0f 0x38 0x05. */
-FNIEMOP_STUB(iemOp_phsubw_Vx_Wx);
+FNIEMOP_DEF(iemOp_phsubw_Vx_Wx)
+{
+    IEMOP_MNEMONIC2(RM, PHSUBW, phsubw, Vx, Wx, DISOPTYPE_HARMLESS, IEMOPHINT_IGNORES_OP_SIZES);
+    return FNIEMOP_CALL_1(iemOpCommonSsse3_FullFull_To_Full,
+                          IEM_SELECT_HOST_OR_FALLBACK(fSsse3, iemAImpl_phsubw_u128, iemAImpl_phsubw_u128_fallback));
+
+}
+
+
 /** Opcode      0x0f 0x38 0x06. */
-FNIEMOP_STUB(iemOp_phsubd_Pq_Qq);
+FNIEMOP_DEF(iemOp_phsubd_Pq_Qq)
+{
+    IEMOP_MNEMONIC2(RM, PHSUBD, phsubd, Pq, Qq, DISOPTYPE_HARMLESS, IEMOPHINT_IGNORES_OP_SIZES);
+    return FNIEMOP_CALL_2(iemOpCommonMmx_FullFull_To_Full_Ex,
+                          IEM_SELECT_HOST_OR_FALLBACK(fSsse3, iemAImpl_phsubd_u64,&iemAImpl_phsubd_u64_fallback),
+                          IEM_GET_GUEST_CPU_FEATURES(pVCpu)->fSsse3);
+}
+
+
+
 /** Opcode 0x66 0x0f 0x38 0x06. */
-FNIEMOP_STUB(iemOp_phsubdq_Vx_Wx);
+FNIEMOP_DEF(iemOp_phsubd_Vx_Wx)
+{
+    IEMOP_MNEMONIC2(RM, PHSUBD, phsubd, Vx, Wx, DISOPTYPE_HARMLESS, IEMOPHINT_IGNORES_OP_SIZES);
+    return FNIEMOP_CALL_1(iemOpCommonSsse3_FullFull_To_Full,
+                          IEM_SELECT_HOST_OR_FALLBACK(fSsse3, iemAImpl_phsubd_u128, iemAImpl_phsubd_u128_fallback));
+
+}
+
+
 /** Opcode      0x0f 0x38 0x07. */
 FNIEMOP_STUB(iemOp_phsubsw_Pq_Qq);
@@ -1369 +1404 @@
     /* 0x04 */  iemOp_pmaddubsw_Pq_Qq,      iemOp_pmaddubsw_Vx_Wx,      iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
     /* 0x05 */  iemOp_phsubw_Pq_Qq,         iemOp_phsubw_Vx_Wx,         iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
-    /* 0x06 */  iemOp_phsubd_Pq_Qq,         iemOp_phsubdq_Vx_Wx,        iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
+    /* 0x06 */  iemOp_phsubd_Pq_Qq,         iemOp_phsubd_Vx_Wx,         iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
     /* 0x07 */  iemOp_phsubsw_Pq_Qq,        iemOp_phsubsw_Vx_Wx,        iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
     /* 0x08 */  iemOp_psignb_Pq_Qq,         iemOp_psignb_Vx_Wx,         iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,

trunk/src/VBox/VMM/VMMAll/IEMAllInstructionsVexMap2.cpp.h

@@ -67 +67 @@
 FNIEMOP_STUB(iemOp_vpmaddubsw_Vx_Hx_Wx);
 /* Opcode VEX.0F38 0x05 - invalid. */
+
+
 /** Opcode VEX.66.0F38 0x05. */
-FNIEMOP_STUB(iemOp_vphsubw_Vx_Hx_Wx);
+FNIEMOP_DEF(iemOp_vphsubw_Vx_Hx_Wx)
+{
+    IEMOP_MNEMONIC3(VEX_RVM, VPHSUBW, vphsubw, Vx, Hx, Wx, DISOPTYPE_HARMLESS, 0);
+    IEMOPMEDIAOPTF3_INIT_VARS(vphsubw);
+    return FNIEMOP_CALL_1(iemOpCommonAvxAvx2_Vx_Hx_Wx_Opt, IEM_SELECT_HOST_OR_FALLBACK(fAvx2, &s_Host, &s_Fallback));
+}
+
+
 /*  Opcode VEX.0F38 0x06 - invalid. */
+
+
 /** Opcode VEX.66.0F38 0x06. */
-FNIEMOP_STUB(iemOp_vphsubdq_Vx_Hx_Wx);
+FNIEMOP_DEF(iemOp_vphsubd_Vx_Hx_Wx)
+{
+    IEMOP_MNEMONIC3(VEX_RVM, VPHSUBD, vphsubd, Vx, Hx, Wx, DISOPTYPE_HARMLESS, 0);
+    IEMOPMEDIAOPTF3_INIT_VARS(vphsubd);
+    return FNIEMOP_CALL_1(iemOpCommonAvxAvx2_Vx_Hx_Wx_Opt, IEM_SELECT_HOST_OR_FALLBACK(fAvx2, &s_Host, &s_Fallback));
+}
+
+
 /*  Opcode VEX.0F38 0x07 - invalid. */
 /** Opcode VEX.66.0F38 0x07. */
@@ -1544 +1562 @@
     /* 0x04 */  iemOp_InvalidNeedRM,        iemOp_vpmaddubsw_Vx_Hx_Wx,  iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
     /* 0x05 */  iemOp_InvalidNeedRM,        iemOp_vphsubw_Vx_Hx_Wx,     iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
-    /* 0x06 */  iemOp_InvalidNeedRM,        iemOp_vphsubdq_Vx_Hx_Wx,    iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
+    /* 0x06 */  iemOp_InvalidNeedRM,        iemOp_vphsubd_Vx_Hx_Wx,     iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
     /* 0x07 */  iemOp_InvalidNeedRM,        iemOp_vphsubsw_Vx_Hx_Wx,    iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,
     /* 0x08 */  iemOp_InvalidNeedRM,        iemOp_vpsignb_Vx_Hx_Wx,     iemOp_InvalidNeedRM,        iemOp_InvalidNeedRM,

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

@@ -1789 +1789 @@
 FNIEMAIMPLMEDIAF2U64     iemAImpl_phaddw_u64, iemAImpl_phaddw_u64_fallback;
 FNIEMAIMPLMEDIAF2U64     iemAImpl_phaddd_u64, iemAImpl_phaddd_u64_fallback;
+FNIEMAIMPLMEDIAF2U64     iemAImpl_phsubw_u64, iemAImpl_phsubw_u64_fallback;
+FNIEMAIMPLMEDIAF2U64     iemAImpl_phsubd_u64, iemAImpl_phsubd_u64_fallback;
 FNIEMAIMPLMEDIAOPTF2U64  iemAImpl_psllw_u64, iemAImpl_psrlw_u64, iemAImpl_psraw_u64;
 FNIEMAIMPLMEDIAOPTF2U64  iemAImpl_pslld_u64, iemAImpl_psrld_u64, iemAImpl_psrad_u64;
@@ -1835 +1837 @@
 FNIEMAIMPLMEDIAF2U128    iemAImpl_phaddw_u128, iemAImpl_phaddw_u128_fallback;
 FNIEMAIMPLMEDIAF2U128    iemAImpl_phaddd_u128, iemAImpl_phaddd_u128_fallback;
+FNIEMAIMPLMEDIAF2U128    iemAImpl_phsubw_u128, iemAImpl_phsubw_u128_fallback;
+FNIEMAIMPLMEDIAF2U128    iemAImpl_phsubd_u128, iemAImpl_phsubd_u128_fallback;
 FNIEMAIMPLMEDIAOPTF2U128 iemAImpl_packsswb_u128, iemAImpl_packuswb_u128;
 FNIEMAIMPLMEDIAOPTF2U128 iemAImpl_packssdw_u128, iemAImpl_packusdw_u128;
@@ -1891 +1895 @@
 FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vphaddw_u128,    iemAImpl_vphaddw_u128_fallback;
 FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vphaddd_u128,    iemAImpl_vphaddd_u128_fallback;
+FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vphsubw_u128,    iemAImpl_vphsubw_u128_fallback;
+FNIEMAIMPLMEDIAOPTF3U128 iemAImpl_vphsubd_u128,    iemAImpl_vphsubd_u128_fallback;
 
 FNIEMAIMPLMEDIAOPTF2U128 iemAImpl_vpabsb_u128,     iemAImpl_vpabsb_u128_fallback;
@@ -1944 +1950 @@
 FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vphaddw_u256,    iemAImpl_vphaddw_u256_fallback;
 FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vphaddd_u256,    iemAImpl_vphaddd_u256_fallback;
+FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vphsubw_u256,    iemAImpl_vphsubw_u256_fallback;
+FNIEMAIMPLMEDIAOPTF3U256 iemAImpl_vphsubd_u256,    iemAImpl_vphsubd_u256_fallback;
 
 FNIEMAIMPLMEDIAOPTF2U256 iemAImpl_vpabsb_u256,     iemAImpl_vpabsb_u256_fallback;