Changeset 105173 in vbox for trunk/src/VBox/VMM

Timestamp:

Jul 7, 2024 1:04:51 PM (7 months ago)

Author:

vboxsync

Message:

VMM/IEM: Implement native emitter for packuswb, bugref:10652

Location:

trunk/src/VBox/VMM/VMMAll

Files:

• IEMAllInstTwoByte0f.cpp.h (modified) (1 diff)
• target-x86/IEMAllN8veEmit-x86.h (modified) (1 diff)

trunk/src/VBox/VMM/VMMAll/IEMAllInstTwoByte0f.cpp.h

@@ -5872 +5872 @@
 {
     IEMOP_MNEMONIC2(RM, PACKUSWB, packuswb, Vx, Wx, DISOPTYPE_HARMLESS | DISOPTYPE_X86_SSE, 0);
-    return FNIEMOP_CALL_1(iemOpCommonSse2Opt_FullFull_To_Full, iemAImpl_packuswb_u128);
+    SSE2_OPT_BODY_FullFull_To_Full(packuswb, iemAImpl_packuswb_u128, RT_ARCH_VAL_AMD64 | RT_ARCH_VAL_ARM64, RT_ARCH_VAL_AMD64 | RT_ARCH_VAL_ARM64);
 }
 

trunk/src/VBox/VMM/VMMAll/target-x86/IEMAllN8veEmit-x86.h

@@ -2354 +2354 @@
 }
 
+
+/**
+ * Common emitter for the PACKUSWB instructions - guest register / guest register variant.
+ */
+DECL_INLINE_THROW(uint32_t)
+iemNativeEmit_packuswb_rr_u128(PIEMRECOMPILERSTATE pReNative, uint32_t off,
+                               uint8_t const idxSimdGstRegDst, uint8_t const idxSimdGstRegSrc)
+{
+    uint8_t const idxSimdRegDst = iemNativeSimdRegAllocTmpForGuestSimdReg(pReNative, &off, IEMNATIVEGSTSIMDREG_SIMD(idxSimdGstRegDst),
+                                                                          kIemNativeGstSimdRegLdStSz_Low128, kIemNativeGstRegUse_ForUpdate);
+    uint8_t const idxSimdRegSrc = iemNativeSimdRegAllocTmpForGuestSimdReg(pReNative, &off, IEMNATIVEGSTSIMDREG_SIMD(idxSimdGstRegSrc),
+                                                                          kIemNativeGstSimdRegLdStSz_Low128, kIemNativeGstRegUse_ReadOnly);
+
+#ifdef RT_ARCH_AMD64
+    PIEMNATIVEINSTR const pCodeBuf = iemNativeInstrBufEnsure(pReNative, off, 5);
+
+    /* packuswb xmm, xmm */
+    pCodeBuf[off++] = X86_OP_PRF_SIZE_OP;
+    if (idxSimdRegDst >= 8 || idxSimdRegSrc >= 8)
+        pCodeBuf[off++] =   (idxSimdRegSrc >= 8 ? X86_OP_REX_B : 0)
+                          | (idxSimdRegDst >= 8 ? X86_OP_REX_R : 0);
+    pCodeBuf[off++] = 0x0f;
+    pCodeBuf[off++] = 0x67;
+    pCodeBuf[off++] = X86_MODRM_MAKE(X86_MOD_REG, idxSimdRegDst & 7, idxSimdRegSrc & 7);
+
+#elif defined(RT_ARCH_ARM64)
+    PIEMNATIVEINSTR const pCodeBuf = iemNativeInstrBufEnsure(pReNative, off, 2);
+
+    pCodeBuf[off++] = Armv8A64MkVecInstrQxtn(kArmv8VecInstrQxtnOp_Sqxtun, false /*fUpper*/, idxSimdRegDst, idxSimdRegDst, kArmv8VecInstrArithSz_8);
+    pCodeBuf[off++] = Armv8A64MkVecInstrQxtn(kArmv8VecInstrQxtnOp_Sqxtun, true  /*fUpper*/, idxSimdRegDst, idxSimdRegSrc, kArmv8VecInstrArithSz_8);
+
+#else
+# error "port me"
+#endif
+
+    iemNativeSimdRegFreeTmp(pReNative, idxSimdRegDst);
+    iemNativeSimdRegFreeTmp(pReNative, idxSimdRegSrc);
+
+    IEMNATIVE_ASSERT_INSTR_BUF_ENSURE(pReNative, off);
+    return off;
+}
+
+
+/**
+ * Common emitter for the PACKUSWB instructions - guest register / recompiler variable variant.
+ */
+DECL_INLINE_THROW(uint32_t)
+iemNativeEmit_packuswb_rv_u128(PIEMRECOMPILERSTATE pReNative, uint32_t off,
+                               uint8_t const idxSimdGstRegDst, uint8_t const idxVarSrc)
+{
+    IEMNATIVE_ASSERT_VAR_IDX(pReNative, idxVarSrc);
+    IEMNATIVE_ASSERT_VAR_SIZE(pReNative, idxVarSrc, sizeof(RTUINT128U));
+
+    uint8_t const idxSimdRegDst = iemNativeSimdRegAllocTmpForGuestSimdReg(pReNative, &off, IEMNATIVEGSTSIMDREG_SIMD(idxSimdGstRegDst),
+                                                                          kIemNativeGstSimdRegLdStSz_Low128, kIemNativeGstRegUse_ForUpdate);
+    uint8_t const idxSimdRegSrc = iemNativeVarSimdRegisterAcquire(pReNative, idxVarSrc, &off, true /*fInitialized*/);
+
+
+#ifdef RT_ARCH_AMD64
+    PIEMNATIVEINSTR const pCodeBuf = iemNativeInstrBufEnsure(pReNative, off, 5);
+
+    /* packuswb xmm, xmm */
+    pCodeBuf[off++] = X86_OP_PRF_SIZE_OP;
+    if (idxSimdRegDst >= 8 || idxSimdRegSrc >= 8)
+        pCodeBuf[off++] =   (idxSimdRegSrc >= 8 ? X86_OP_REX_B : 0)
+                          | (idxSimdRegDst >= 8 ? X86_OP_REX_R : 0);
+    pCodeBuf[off++] = 0x0f;
+    pCodeBuf[off++] = 0x67;
+    pCodeBuf[off++] = X86_MODRM_MAKE(X86_MOD_REG, idxSimdRegDst & 7, idxSimdRegSrc & 7);
+
+#elif defined(RT_ARCH_ARM64)
+    PIEMNATIVEINSTR const pCodeBuf = iemNativeInstrBufEnsure(pReNative, off, 2);
+
+    pCodeBuf[off++] = Armv8A64MkVecInstrQxtn(kArmv8VecInstrQxtnOp_Sqxtun, false /*fUpper*/, idxSimdRegDst, idxSimdRegDst, kArmv8VecInstrArithSz_8);
+    pCodeBuf[off++] = Armv8A64MkVecInstrQxtn(kArmv8VecInstrQxtnOp_Sqxtun, true  /*fUpper*/, idxSimdRegDst, idxSimdRegSrc, kArmv8VecInstrArithSz_8);
+
+#else
+# error "port me"
+#endif
+
+    iemNativeSimdRegFreeTmp(pReNative, idxSimdRegDst);
+    iemNativeVarRegisterRelease(pReNative, idxVarSrc);
+
+    IEMNATIVE_ASSERT_INSTR_BUF_ENSURE(pReNative, off);
+    return off;
+}
+
+
 #endif /* IEMNATIVE_WITH_SIMD_REG_ALLOCATOR */