Changeset 104362 in vbox

Timestamp:

Apr 18, 2024 3:29:48 PM (10 months ago)

Author:

vboxsync

Message:

VMM/IEM: Don't unroll loops in TBs. bugref:10653

Location:

trunk/src/VBox/VMM

Files:

• VMMAll/IEMAllThrdRecompiler.cpp (modified) (2 diffs)
• VMMR3/IEMR3.cpp (modified) (1 diff)
• include/IEMInternal.h (modified) (2 diffs)

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

@@ -2152 +2152 @@
      * Case 1: We've branched (RIP changed).
      *
+     * Loop check:  If the new PC (GCPhysPC) is within a opcode range of this
+     *              TB, end the TB here as it is most likely a loop and if it
+     *              made sense to unroll it, the guest code compiler should've
+     *              done it already.
+     *
      * Sub-case 1a: Same page, no TLB load (fTbCrossedPage is false).
      *         Req: 1 extra range, no extra phys.
@@ -2210 +2215 @@
                 AssertCompile(RT_ELEMENTS(pTb->aGCPhysPages) == 2);
                 RTGCPHYS const GCPhysNew = pVCpu->iem.s.GCPhysInstrBuf & ~(RTGCPHYS)GUEST_PAGE_OFFSET_MASK;
+                uint8_t        idxPhysPage;
                 if ((pTb->GCPhysPc & ~(RTGCPHYS)GUEST_PAGE_OFFSET_MASK) == GCPhysNew)
-                    pTb->aRanges[idxRange].idxPhysPage = 0;
-                else if (   pTb->aGCPhysPages[0] == NIL_RTGCPHYS
-                         || pTb->aGCPhysPages[0] == GCPhysNew)
+                    pTb->aRanges[idxRange].idxPhysPage = idxPhysPage = 0;
+                else if (pTb->aGCPhysPages[0] == NIL_RTGCPHYS)
                 {
                     pTb->aGCPhysPages[0] = GCPhysNew;
                     pTb->aRanges[idxRange].idxPhysPage = 1;
+                    idxPhysPage = UINT8_MAX;
                 }
-                else if (   pTb->aGCPhysPages[1] == NIL_RTGCPHYS
-                         || pTb->aGCPhysPages[1] == GCPhysNew)
+                else if (pTb->aGCPhysPages[0] == GCPhysNew)
+                    pTb->aRanges[idxRange].idxPhysPage = idxPhysPage = 1;
+                else if (pTb->aGCPhysPages[1] == NIL_RTGCPHYS)
                 {
                     pTb->aGCPhysPages[1] = GCPhysNew;
                     pTb->aRanges[idxRange].idxPhysPage = 2;
+                    idxPhysPage = UINT8_MAX;
                 }
+                else if (pTb->aGCPhysPages[1] == GCPhysNew)
+                    pTb->aRanges[idxRange].idxPhysPage = idxPhysPage = 2;
                 else
                 {
                     Log8(("%04x:%08RX64: out of aGCPhysPages entires after branch\n", pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
                     return false;
+                }
+
+                /* Loop check: We weave the loop check in here to optimize the lookup. */
+                if (idxPhysPage != UINT8_MAX)
+                {
+                    uint32_t const offPhysPc = pVCpu->iem.s.offCurInstrStart;
+                    for (uint8_t idxLoopRange = 0; idxLoopRange < idxRange; idxLoopRange++)
+                        if (   pTb->aRanges[idxLoopRange].idxPhysPage == idxPhysPage
+                            &&   offPhysPc - (uint32_t)pTb->aRanges[idxLoopRange].offPhysPage
+                               < (uint32_t)pTb->aRanges[idxLoopRange].cbOpcodes)
+                        {
+                            Log8(("%04x:%08RX64: loop detected after branch\n", pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
+                            STAM_COUNTER_INC(&pVCpu->iem.s.StatTbLoopInTbDetected);
+                            return false;
+                        }
                 }
 

trunk/src/VBox/VMM/VMMR3/IEMR3.cpp

@@ -450 +450 @@
         STAMR3RegisterF(pVM, (void *)&pVCpu->iem.s.StatCheckNeedCsLimChecking, STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
                         "Needing CS.LIM checking TB after branch or on page crossing", "/IEM/CPU%u/re/CheckTbNeedCsLimChecking", idCpu);
+# ifdef VBOX_WITH_STATISTICS
+        STAMR3RegisterF(pVM, (void *)&pVCpu->iem.s.StatTbLoopInTbDetected, STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
+                        "Detected loop within TB", "/IEM/CPU%u/re/LoopInTbDetected", idCpu);
+#endif
 
         STAMR3RegisterF(pVM, (void *)&pVCpu->iem.s.StatNativeExecMemInstrBufAllocFailed, STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,

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

@@ -1909 +1909 @@
     /** Statistics: Times a jump or page crossing required a TB with CS.LIM checking. */
     STAMCOUNTER             StatCheckNeedCsLimChecking;
+    /** Statistics: Times a loop was detected within a TB.. */
+    STAMCOUNTER             StatTbLoopInTbDetected;
     /** Exec memory allocator statistics: Number of times allocaintg executable memory failed. */
     STAMCOUNTER             StatNativeExecMemInstrBufAllocFailed;
@@ -2069 +2071 @@
     STAMCOUNTER             StatNativeTbExitObsoleteTb;
 
-    uint64_t                au64Padding[5];
+    uint64_t                au64Padding[4];
     /** @} */