Changeset 96382 in vbox

Timestamp:

Aug 20, 2022 7:52:07 PM (2 years ago)

Author:

vboxsync

Message:

VMM/IEM: Implement haddps/haddpd/hsubps/hsubpd instructions, bugref:9898

Location:

trunk/src/VBox/VMM

Files:

• VMMAll/IEMAllAImpl.asm (modified) (1 diff)
• VMMAll/IEMAllAImplC.cpp (modified) (1 diff)
• VMMAll/IEMAllInstructionsTwoByte0f.cpp.h (modified) (2 diffs)
• include/IEMInternal.h (modified) (3 diffs)
• testcase/tstIEMCheckMc.cpp (modified) (1 diff)

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

@@ -4597 +4597 @@
 IEMIMPL_FP_F2 maxps
 IEMIMPL_FP_F2 maxpd
+IEMIMPL_FP_F2 haddps
+IEMIMPL_FP_F2 haddpd
+IEMIMPL_FP_F2 hsubps
+IEMIMPL_FP_F2 hsubpd
 
 

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

@@ -14759 +14759 @@
 }
 #endif
+
+
+/**
+ * HADDPS
+ */
+#ifdef IEM_WITHOUT_ASSEMBLY
+IEM_DECL_IMPL_DEF(void, iemAImpl_haddps_u128,(PX86FXSTATE pFpuState, PIEMSSERESULT pResult, PCX86XMMREG puSrc1, PCX86XMMREG puSrc2))
+{
+    pResult->MXCSR  = iemAImpl_addps_u128_worker(&pResult->uResult.ar32[0], pFpuState->MXCSR, &puSrc1->ar32[0], &puSrc1->ar32[1]);
+    pResult->MXCSR |= iemAImpl_addps_u128_worker(&pResult->uResult.ar32[1], pFpuState->MXCSR, &puSrc1->ar32[2], &puSrc1->ar32[3]);
+    pResult->MXCSR |= iemAImpl_addps_u128_worker(&pResult->uResult.ar32[2], pFpuState->MXCSR, &puSrc2->ar32[0], &puSrc2->ar32[1]);
+    pResult->MXCSR |= iemAImpl_addps_u128_worker(&pResult->uResult.ar32[3], pFpuState->MXCSR, &puSrc2->ar32[2], &puSrc2->ar32[3]);
+}
+#endif
+
+
+/**
+ * HADDPD
+ */
+#ifdef IEM_WITHOUT_ASSEMBLY
+IEM_DECL_IMPL_DEF(void, iemAImpl_haddpd_u128,(PX86FXSTATE pFpuState, PIEMSSERESULT pResult, PCX86XMMREG puSrc1, PCX86XMMREG puSrc2))
+{
+    pResult->MXCSR  = iemAImpl_addpd_u128_worker(&pResult->uResult.ar64[0], pFpuState->MXCSR, &puSrc1->ar64[0], &puSrc1->ar64[1]);
+    pResult->MXCSR |= iemAImpl_addpd_u128_worker(&pResult->uResult.ar64[1], pFpuState->MXCSR, &puSrc2->ar64[0], &puSrc2->ar64[1]);
+}
+#endif
+
+
+/**
+ * HSUBPS
+ */
+#ifdef IEM_WITHOUT_ASSEMBLY
+IEM_DECL_IMPL_DEF(void, iemAImpl_hsubps_u128,(PX86FXSTATE pFpuState, PIEMSSERESULT pResult, PCX86XMMREG puSrc1, PCX86XMMREG puSrc2))
+{
+    pResult->MXCSR  = iemAImpl_subps_u128_worker(&pResult->uResult.ar32[0], pFpuState->MXCSR, &puSrc1->ar32[0], &puSrc1->ar32[1]);
+    pResult->MXCSR |= iemAImpl_subps_u128_worker(&pResult->uResult.ar32[1], pFpuState->MXCSR, &puSrc1->ar32[2], &puSrc1->ar32[3]);
+    pResult->MXCSR |= iemAImpl_subps_u128_worker(&pResult->uResult.ar32[2], pFpuState->MXCSR, &puSrc2->ar32[0], &puSrc2->ar32[1]);
+    pResult->MXCSR |= iemAImpl_subps_u128_worker(&pResult->uResult.ar32[3], pFpuState->MXCSR, &puSrc2->ar32[2], &puSrc2->ar32[3]);
+}
+#endif
+
+
+/**
+ * HSUBPD
+ */
+#ifdef IEM_WITHOUT_ASSEMBLY
+IEM_DECL_IMPL_DEF(void, iemAImpl_hsubpd_u128,(PX86FXSTATE pFpuState, PIEMSSERESULT pResult, PCX86XMMREG puSrc1, PCX86XMMREG puSrc2))
+{
+    pResult->MXCSR  = iemAImpl_subpd_u128_worker(&pResult->uResult.ar64[0], pFpuState->MXCSR, &puSrc1->ar64[0], &puSrc1->ar64[1]);
+    pResult->MXCSR |= iemAImpl_subpd_u128_worker(&pResult->uResult.ar64[1], pFpuState->MXCSR, &puSrc2->ar64[0], &puSrc2->ar64[1]);
+}
+#endif

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

@@ -1049 +1049 @@
         IEM_MC_REF_XREG_U128(puDst, IEM_GET_MODRM_REG(pVCpu, bRm));
         IEM_MC_CALL_VOID_AIMPL_2(pfnU128, puDst, puSrc);
+
+        IEM_MC_ADVANCE_RIP();
+        IEM_MC_END();
+    }
+    return VINF_SUCCESS;
+}
+
+
+/**
+ * Common worker for SSE3 instructions on the forms:
+ *      hxxx      xmm1, xmm2/mem128
+ *
+ * Proper alignment of the 128-bit operand is enforced.
+ * Exceptions type 2. SSE3 cpuid checks.
+ *
+ * @sa iemOpCommonSse41_FullFull_To_Full, iemOpCommonSse2_FullFull_To_Full
+ */
+FNIEMOP_DEF_1(iemOpCommonSse3Fp_FullFull_To_Full, PFNIEMAIMPLFPSSEF2U128, pfnU128)
+{
+    uint8_t bRm; IEM_OPCODE_GET_NEXT_U8(&bRm);
+    if (IEM_IS_MODRM_REG_MODE(bRm))
+    {
+        /*
+         * Register, register.
+         */
+        IEMOP_HLP_DONE_DECODING_NO_LOCK_PREFIX();
+        IEM_MC_BEGIN(3, 1);
+        IEM_MC_LOCAL(IEMSSERESULT,          SseRes);
+        IEM_MC_ARG_LOCAL_REF(PIEMSSERESULT, pSseRes,        SseRes,     0);
+        IEM_MC_ARG(PCX86XMMREG,             pSrc1,                      1);
+        IEM_MC_ARG(PCX86XMMREG,             pSrc2,                      2);
+        IEM_MC_MAYBE_RAISE_SSE3_RELATED_XCPT();
+        IEM_MC_PREPARE_SSE_USAGE();
+        IEM_MC_REF_XREG_XMM_CONST(pSrc1, IEM_GET_MODRM_REG(pVCpu, bRm));
+        IEM_MC_REF_XREG_XMM_CONST(pSrc2, IEM_GET_MODRM_RM(pVCpu, bRm));
+        IEM_MC_CALL_SSE_AIMPL_3(pfnU128, pSseRes, pSrc1, pSrc2);
+        IEM_MC_STORE_SSE_RESULT(SseRes, IEM_GET_MODRM_REG(pVCpu, bRm));
+        IEM_MC_MAYBE_RAISE_SSE_AVX_SIMD_FP_OR_UD_XCPT();
+
+        IEM_MC_ADVANCE_RIP();
+        IEM_MC_END();
+    }
+    else
+    {
+        /*
+         * Register, memory.
+         */
+        IEM_MC_BEGIN(3, 2);
+        IEM_MC_LOCAL(IEMSSERESULT,          SseRes);
+        IEM_MC_LOCAL(X86XMMREG,             uSrc2);
+        IEM_MC_ARG_LOCAL_REF(PIEMSSERESULT, pSseRes,        SseRes,     0);
+        IEM_MC_ARG(PCX86XMMREG,             pSrc1,                      1);
+        IEM_MC_ARG_LOCAL_REF(PCX86XMMREG,   pSrc2, uSrc2,               2);
+        IEM_MC_LOCAL(RTGCPTR, GCPtrEffSrc);
+
+        IEM_MC_CALC_RM_EFF_ADDR(GCPtrEffSrc, bRm, 0);
+        IEMOP_HLP_DONE_DECODING_NO_LOCK_PREFIX();
+        IEM_MC_MAYBE_RAISE_SSE3_RELATED_XCPT();
+        IEM_MC_FETCH_MEM_XMM_ALIGN_SSE(uSrc2, pVCpu->iem.s.iEffSeg, GCPtrEffSrc);
+
+        IEM_MC_PREPARE_SSE_USAGE();
+        IEM_MC_REF_XREG_XMM_CONST(pSrc1, IEM_GET_MODRM_REG(pVCpu, bRm));
+        IEM_MC_CALL_SSE_AIMPL_3(pfnU128, pSseRes, pSrc1, pSrc2);
+        IEM_MC_STORE_SSE_RESULT(SseRes, IEM_GET_MODRM_REG(pVCpu, bRm));
+        IEM_MC_MAYBE_RAISE_SSE_AVX_SIMD_FP_OR_UD_XCPT();
 
         IEM_MC_ADVANCE_RIP();
@@ -5626 +5691 @@
 
 /*  Opcode      0x0f 0x7c - invalid */
+
+
 /** Opcode 0x66 0x0f 0x7c - haddpd Vpd, Wpd */
-FNIEMOP_STUB(iemOp_haddpd_Vpd_Wpd);
+FNIEMOP_DEF(iemOp_haddpd_Vpd_Wpd)
+{
+    IEMOP_MNEMONIC2(RM, HADDPD, haddpd, Vpd, Wpd, DISOPTYPE_HARMLESS, 0);
+    return FNIEMOP_CALL_1(iemOpCommonSse3Fp_FullFull_To_Full, iemAImpl_haddpd_u128);
+}
+
+
 /*  Opcode 0xf3 0x0f 0x7c - invalid */
+
+
 /** Opcode 0xf2 0x0f 0x7c - haddps Vps, Wps */
-FNIEMOP_STUB(iemOp_haddps_Vps_Wps);
+FNIEMOP_DEF(iemOp_haddps_Vps_Wps)
+{
+    IEMOP_MNEMONIC2(RM, HADDPS, haddps, Vps, Wps, DISOPTYPE_HARMLESS, 0);
+    return FNIEMOP_CALL_1(iemOpCommonSse3Fp_FullFull_To_Full, iemAImpl_haddps_u128);
+}
+
 
 /*  Opcode      0x0f 0x7d - invalid */
+
+
 /** Opcode 0x66 0x0f 0x7d - hsubpd Vpd, Wpd */
-FNIEMOP_STUB(iemOp_hsubpd_Vpd_Wpd);
+FNIEMOP_DEF(iemOp_hsubpd_Vpd_Wpd)
+{
+    IEMOP_MNEMONIC2(RM, HSUBPD, hsubpd, Vpd, Wpd, DISOPTYPE_HARMLESS, 0);
+    return FNIEMOP_CALL_1(iemOpCommonSse3Fp_FullFull_To_Full, iemAImpl_hsubpd_u128);
+}
+
+
 /*  Opcode 0xf3 0x0f 0x7d - invalid */
+
+
 /** Opcode 0xf2 0x0f 0x7d - hsubps Vps, Wps */
-FNIEMOP_STUB(iemOp_hsubps_Vps_Wps);
+FNIEMOP_DEF(iemOp_hsubps_Vps_Wps)
+{
+    IEMOP_MNEMONIC2(RM, HSUBPS, hsubps, Vps, Wps, DISOPTYPE_HARMLESS, 0);
+    return FNIEMOP_CALL_1(iemOpCommonSse3Fp_FullFull_To_Full, iemAImpl_hsubps_u128);
+}
 
 

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

@@ -2438 +2438 @@
 FNIEMAIMPLFPSSEF2U128 iemAImpl_maxps_u128;
 FNIEMAIMPLFPSSEF2U128 iemAImpl_maxpd_u128;
+FNIEMAIMPLFPSSEF2U128 iemAImpl_haddps_u128;
+FNIEMAIMPLFPSSEF2U128 iemAImpl_haddpd_u128;
+FNIEMAIMPLFPSSEF2U128 iemAImpl_hsubps_u128;
+FNIEMAIMPLFPSSEF2U128 iemAImpl_hsubpd_u128;
 
 FNIEMAIMPLFPSSEF2U128R32 iemAImpl_addss_u128_r32;
@@ -2466 +2470 @@
 FNIEMAIMPLFPAVXF3U128 iemAImpl_vmaxps_u128, iemAImpl_vmaxps_u128_fallback;
 FNIEMAIMPLFPAVXF3U128 iemAImpl_vmaxpd_u128, iemAImpl_vmaxpd_u128_fallback;
+FNIEMAIMPLFPAVXF3U128 iemAImpl_vhaddps_u128, iemAImpl_vhaddps_u128_fallback;
+FNIEMAIMPLFPAVXF3U128 iemAImpl_vhaddpd_u128, iemAImpl_vhaddpd_u128_fallback;
+FNIEMAIMPLFPAVXF3U128 iemAImpl_vhsubps_u128, iemAImpl_vhsubps_u128_fallback;
+FNIEMAIMPLFPAVXF3U128 iemAImpl_vhsubpd_u128, iemAImpl_vhsubpd_u128_fallback;
 
 FNIEMAIMPLFPAVXF3U128R32 iemAImpl_vaddss_u128_r32, iemAImpl_vaddss_u128_r32_fallback;
@@ -2492 +2500 @@
 FNIEMAIMPLFPAVXF3U256 iemAImpl_vmaxps_u256, iemAImpl_vmaxps_u256_fallback;
 FNIEMAIMPLFPAVXF3U256 iemAImpl_vmaxpd_u256, iemAImpl_vmaxpd_u256_fallback;
+FNIEMAIMPLFPAVXF3U256 iemAImpl_vhaddps_u256, iemAImpl_vhaddps_u256_fallback;
+FNIEMAIMPLFPAVXF3U256 iemAImpl_vhaddpd_u256, iemAImpl_vhaddpd_u256_fallback;
+FNIEMAIMPLFPAVXF3U256 iemAImpl_vhsubps_u256, iemAImpl_vhsubps_u256_fallback;
+FNIEMAIMPLFPAVXF3U256 iemAImpl_vhsubpd_u256, iemAImpl_vhsubpd_u256_fallback;
 /** @} */
 

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

@@ -474 +474 @@
 #define iemAImpl_maxps_u128             NULL
 #define iemAImpl_maxpd_u128             NULL
+#define iemAImpl_haddps_u128            NULL
+#define iemAImpl_haddpd_u128            NULL
+#define iemAImpl_hsubps_u128            NULL
+#define iemAImpl_hsubpd_u128            NULL
 
 #define iemAImpl_addss_u128_r32         NULL