Changeset 104348 in vbox for trunk/src/VBox/VMM/VMMAll

Timestamp:

Apr 17, 2024 2:41:37 PM (10 months ago)

Author:

vboxsync

Message:

VMM/IEM: Implement native emitters for pcmpeqb, pcmpeqw and pcmpeqd (pcmpeqb is used quite a bit with an OL 9.2 amd64 guest), bugref:10652

Location:

trunk/src/VBox/VMM/VMMAll

Files:

• IEMAllInstTwoByte0f.cpp.h (modified) (3 diffs)
• target-x86/IEMAllN8veEmit-x86.h (modified) (1 diff)

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

@@ -6808 +6808 @@
 {
     IEMOP_MNEMONIC2(RM, PCMPEQB, pcmpeqb, Vx, Wx, DISOPTYPE_HARMLESS, IEMOPHINT_IGNORES_OP_SIZES);
-    return FNIEMOP_CALL_1(iemOpCommonSse2Opt_FullFull_To_Full, iemAImpl_pcmpeqb_u128);
+    SSE2_OPT_BODY_FullFull_To_Full(pcmpeqb, iemAImpl_pcmpeqb_u128, RT_ARCH_VAL_AMD64 | RT_ARCH_VAL_ARM64, RT_ARCH_VAL_AMD64 | RT_ARCH_VAL_ARM64);
 }
 
@@ -6828 +6828 @@
 {
     IEMOP_MNEMONIC2(RM, PCMPEQW, pcmpeqw, Vx, Wx, DISOPTYPE_HARMLESS, IEMOPHINT_IGNORES_OP_SIZES);
-    return FNIEMOP_CALL_1(iemOpCommonSse2Opt_FullFull_To_Full, iemAImpl_pcmpeqw_u128);
+    SSE2_OPT_BODY_FullFull_To_Full(pcmpeqw, iemAImpl_pcmpeqw_u128, RT_ARCH_VAL_AMD64 | RT_ARCH_VAL_ARM64, RT_ARCH_VAL_AMD64 | RT_ARCH_VAL_ARM64);
 }
 
@@ -6848 +6848 @@
 {
     IEMOP_MNEMONIC2(RM, PCMPEQD, pcmpeqd, Vx, Wx, DISOPTYPE_HARMLESS, IEMOPHINT_IGNORES_OP_SIZES);
-    return FNIEMOP_CALL_1(iemOpCommonSse2Opt_FullFull_To_Full, iemAImpl_pcmpeqd_u128);
+    SSE2_OPT_BODY_FullFull_To_Full(pcmpeqd, iemAImpl_pcmpeqd_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

@@ -2251 +2251 @@
 IEMNATIVE_NATIVE_EMIT_ADD_SUB_U128(psubq, true  /*a_fSub*/, kArmv8VecInstrArithSz_64, 0xfb);
 
+
+/**
+ * Common emitter for the pcmpeqb/pcmpeqw/pcmpeqd instructions.
+ */
+#ifdef RT_ARCH_AMD64
+# define IEMNATIVE_NATIVE_EMIT_PCMP_U128(a_Instr, a_enmOp, a_ArmElemSz, a_bOpcX86) \
+    DECL_INLINE_THROW(uint32_t) \
+    RT_CONCAT3(iemNativeEmit_,a_Instr,_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); \
+        PIEMNATIVEINSTR const pCodeBuf = iemNativeInstrBufEnsure(pReNative, off, 5); \
+        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++] = (a_bOpcX86); \
+        pCodeBuf[off++] = X86_MODRM_MAKE(X86_MOD_REG, idxSimdRegDst & 7, idxSimdRegSrc & 7); \
+        iemNativeSimdRegFreeTmp(pReNative, idxSimdRegDst); \
+        iemNativeSimdRegFreeTmp(pReNative, idxSimdRegSrc); \
+        IEMNATIVE_ASSERT_INSTR_BUF_ENSURE(pReNative, off); \
+        return off; \
+    } \
+    DECL_INLINE_THROW(uint32_t) \
+    RT_CONCAT3(iemNativeEmit_,a_Instr,_rv_u128)(PIEMRECOMPILERSTATE pReNative, uint32_t off, \
+                                                uint8_t const idxSimdGstRegDst, uint8_t const idxVarSrc) \
+    { \
+        uint8_t const idxSimdRegDst = iemNativeSimdRegAllocTmpForGuestSimdReg(pReNative, &off, IEMNATIVEGSTSIMDREG_SIMD(idxSimdGstRegDst), \
+                                                                              kIemNativeGstSimdRegLdStSz_Low128, kIemNativeGstRegUse_ForUpdate); \
+        uint8_t const idxSimdRegSrc = iemNativeVarSimdRegisterAcquire(pReNative, idxVarSrc, &off, true /*fInitialized*/); \
+        PIEMNATIVEINSTR const pCodeBuf = iemNativeInstrBufEnsure(pReNative, off, 5); \
+        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++] = (a_bOpcX86); \
+        pCodeBuf[off++] = X86_MODRM_MAKE(X86_MOD_REG, idxSimdRegDst & 7, idxSimdRegSrc & 7); \
+        iemNativeSimdRegFreeTmp(pReNative, idxSimdRegDst); \
+        iemNativeVarRegisterRelease(pReNative, idxVarSrc); \
+        IEMNATIVE_ASSERT_INSTR_BUF_ENSURE(pReNative, off); \
+        return off; \
+    } \
+    typedef int ignore_semicolon
+#elif defined(RT_ARCH_ARM64)
+# define IEMNATIVE_NATIVE_EMIT_PCMP_U128(a_Instr, a_enmOp, a_ArmElemSz, a_bOpcX86) \
+    DECL_INLINE_THROW(uint32_t) \
+    RT_CONCAT3(iemNativeEmit_,a_Instr,_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); \
+        PIEMNATIVEINSTR const pCodeBuf = iemNativeInstrBufEnsure(pReNative, off, 1); \
+        pCodeBuf[off++] = Armv8A64MkVecInstrCmp((a_enmOp), idxSimdRegDst, idxSimdRegDst, idxSimdRegSrc, (a_ArmElemSz)); \
+        iemNativeSimdRegFreeTmp(pReNative, idxSimdRegDst); \
+        iemNativeSimdRegFreeTmp(pReNative, idxSimdRegSrc); \
+        IEMNATIVE_ASSERT_INSTR_BUF_ENSURE(pReNative, off); \
+        return off; \
+    } \
+    DECL_INLINE_THROW(uint32_t) \
+    RT_CONCAT3(iemNativeEmit_,a_Instr,_rv_u128)(PIEMRECOMPILERSTATE pReNative, uint32_t off, \
+                                                uint8_t const idxSimdGstRegDst, uint8_t const idxVarSrc) \
+    { \
+        uint8_t const idxSimdRegDst = iemNativeSimdRegAllocTmpForGuestSimdReg(pReNative, &off, IEMNATIVEGSTSIMDREG_SIMD(idxSimdGstRegDst), \
+                                                                              kIemNativeGstSimdRegLdStSz_Low128, kIemNativeGstRegUse_ForUpdate); \
+        uint8_t const idxSimdRegSrc = iemNativeVarSimdRegisterAcquire(pReNative, idxVarSrc, &off, true /*fInitialized*/); \
+        PIEMNATIVEINSTR const pCodeBuf = iemNativeInstrBufEnsure(pReNative, off, 1); \
+        pCodeBuf[off++] = Armv8A64MkVecInstrCmp((a_enmOp), idxSimdRegDst, idxSimdRegDst, idxSimdRegSrc, (a_ArmElemSz)); \
+        iemNativeSimdRegFreeTmp(pReNative, idxSimdRegDst); \
+        iemNativeVarRegisterRelease(pReNative, idxVarSrc); \
+        IEMNATIVE_ASSERT_INSTR_BUF_ENSURE(pReNative, off); \
+        return off; \
+    } \
+    typedef int ignore_semicolon
+#else
+# error "Port me"
+#endif
+
+IEMNATIVE_NATIVE_EMIT_PCMP_U128(pcmpeqb, kArmv8VecInstrCmpOp_Eq, kArmv8VecInstrArithSz_8,  0x74);
+IEMNATIVE_NATIVE_EMIT_PCMP_U128(pcmpeqw, kArmv8VecInstrCmpOp_Eq, kArmv8VecInstrArithSz_16, 0x75);
+IEMNATIVE_NATIVE_EMIT_PCMP_U128(pcmpeqd, kArmv8VecInstrCmpOp_Eq, kArmv8VecInstrArithSz_32, 0x76);
+
 #endif /* IEMNATIVE_WITH_SIMD_REG_ALLOCATOR */