Changeset 104322 in vbox

Timestamp:

Apr 12, 2024 3:18:43 PM (8 months ago)

Author:

vboxsync

Message:

VMM/IEM: Have a single TB prologue which serves as an entry point to the TB, bugref:10653

Location:

trunk/src/VBox/VMM

Files:

• VMMAll/IEMAllN8veHlpA-arm64.S (modified) (3 diffs)
• VMMAll/IEMAllN8veHlpA.asm (modified) (2 diffs)
• VMMAll/IEMAllN8veRecompiler.cpp (modified) (3 diffs)
• VMMAll/IEMAllThrdRecompiler.cpp (modified) (1 diff)
• include/IEMN8veRecompiler.h (modified) (4 diffs)

trunk/src/VBox/VMM/VMMAll/IEMAllN8veHlpA-arm64.S

@@ -32 +32 @@
 #include <iprt/asmdefs-arm.h>
 
+#ifdef AssertCompile
+# undef AssertCompile
+#endif
+#define AssertCompile(a_Expr)
+#define INCLUDED_FROM_ARM64_ASSEMBLY
+#include "IEMN8veRecompiler.h"
+
+#define IEM_HLP_FUNCTION_ALIGNMENT 0x20
 
 BEGINCODE
 
 .extern NAME(iemThreadedFunc_BltIn_LogCpuStateWorker)
+
+#ifdef IEMNATIVE_WITH_RECOMPILER_PROLOGUE_SINGLETON
+/**
+ * This is the common prologue of a TB, saving all volatile registers
+ * and creating the stack frame for saving temporary values.
+ *
+ * @param    pVCpu    (x0)      The cross-context vCPU structure pointer.
+ * @param    pCpumCtx (x1)      The cross-context CPUM context structure pointer.
+ * @param    pTbStart (x2)      The TB instruction start pointer.
+ */
+ALIGNCODE(IEM_HLP_FUNCTION_ALIGNMENT)
+BEGINPROC_HIDDEN iemNativeTbEntry
+# ifdef RT_OS_DARWIN
+        pacibsp
+# endif
+        stp     x19, x20, [sp, #-IEMNATIVE_FRAME_SAVE_REG_SIZE]! /* Allocate space for saving registers and place x19+x20 at the bottom. */
+        stp     x21, x22, [sp, #0x10]                            /* Save x21 thru x28 (SP remains unchanged). */
+        stp     x23, x24, [sp, #0x20]
+        stp     x25, x26, [sp, #0x30]
+        stp     x27, x28, [sp, #0x40]
+        stp     x29, x30, [sp, #0x50]                            /* Save the BP and LR (ret address) registers at the top of the frame. */
+        add     x29, sp, #(IEMNATIVE_FRAME_SAVE_REG_SIZE - 16)   /* Set BP to point to the old BP stack address */
+        sub     sp, sp, #IEMNATIVE_FRAME_VAR_SIZE                /* Allocate the variable area from SP. */
+        mov     IEMNATIVE_REG_FIXED_PVMCPU_ASM,   x0
+        mov     IEMNATIVE_REG_FIXED_PCPUMCTX_ASM, x1
+# ifdef VBOX_WITH_IEM_NATIVE_RECOMPILER_LONGJMP
+        str     x29, [IEMNATIVE_REG_FIXED_PVMCPU_ASM, #0x7c8]    /* Save the frame pointer to pvTbFramePointerR3 */ /** @todo Get rid of this hardcoded assumption */
+# endif
+        /*
+         * Everything is done, jump to the start of the TB.
+         */
+        br      x2
+#endif
+
 
 /**
@@ -53 +95 @@
  *       it ourselves, I think.
  */
+ALIGNCODE(IEM_HLP_FUNCTION_ALIGNMENT)
 BEGINPROC_HIDDEN iemNativeTbLongJmp
         /*
@@ -81 +124 @@
  * so the impact of inserting LogCpuState is minimal to the other TB code.
  */
+ALIGNCODE(IEM_HLP_FUNCTION_ALIGNMENT)
 BEGINPROC_HIDDEN iemNativeHlpAsmSafeWrapLogCpuState
 #ifdef RT_OS_DARWIN

trunk/src/VBox/VMM/VMMAll/IEMAllN8veHlpA.asm

@@ -32 +32 @@
 %include "VBox/asmdefs.mac"
 
+;; @todo r=aeichner The following is copied from IEMInternal.h
+;%define VBOX_WITH_IEM_NATIVE_RECOMPILER_LONGJMP - not enabled right now on amd64
+
+;; @todo r=aeichner The following defines are copied from IEMN8veRecompiler.h
+
+; /** @def IEMNATIVE_WITH_RECOMPILER_PROLOGUE_SINGLETON
+;  * Enables having only a single prologue for native TBs. */
+%define IEMNATIVE_WITH_RECOMPILER_PROLOGUE_SINGLETON
+
+; /** An stack alignment adjustment (between non-volatile register pushes and
+;  *  the stack variable area, so the latter better aligned). */
+%define IEMNATIVE_FRAME_ALIGN_SIZE              8
+
+; /** The size of the area for stack variables and spills and stuff.
+; * @note This limit is duplicated in the python script(s).  We add 0x40 for
+; *       alignment padding. */
+%define IEMNATIVE_FRAME_VAR_SIZE                (0xc0 + 0x40)
+
+; This needs to correspond to IEMNATIVE_REG_FIXED_PVMCPU in IEMN8veRecompiler.h
+%define IEMNATIVE_REG_FIXED_PVMCPU_ASM          xBX
+
+; /** Number of stack arguments slots for calls made from the frame. */
+%ifdef RT_OS_WINDOWS
+%define IEMNATIVE_FRAME_STACK_ARG_COUNT         4
+%else
+%define IEMNATIVE_FRAME_STACK_ARG_COUNT         2
+%endif
+; /** Number of any shadow arguments (spill area) for calls we make. */
+%ifdef RT_OS_WINDOWS
+%define IEMNATIVE_FRAME_SHADOW_ARG_COUNT        4
+%else
+%define IEMNATIVE_FRAME_SHADOW_ARG_COUNT        0
+%endif
+
 
 BEGINCODE
@@ -37 +71 @@
 extern NAME(iemThreadedFunc_BltIn_LogCpuStateWorker)
 extern NAME(iemNativeHlpCheckTlbLookup)
+
+%ifdef IEMNATIVE_WITH_RECOMPILER_PROLOGUE_SINGLETON
+;;
+; This is the common prologue of a TB, saving all volatile registers
+; and creating the stack frame for saving temporary values.
+;
+; @param    pVCpu    (gcc:rdi, msc:rcx)      The cross-context vCPU structure pointer.
+; @param    pTbStart (gcc:rsi, msc:rdx)      The TB instruction start pointer.
+;
+BEGINPROC   iemNativeTbEntry
+        push rbp
+        mov rbp, rsp
+        push rbx
+%ifdef ASM_CALL64_MSC
+        mov IEMNATIVE_REG_FIXED_PVMCPU_ASM, rcx
+        push rsi
+        push rdi
+%else
+        mov IEMNATIVE_REG_FIXED_PVMCPU_ASM, rdi
+%endif
+        push r12
+        push r13
+        push r14
+        push r15
+%ifdef VBOX_WITH_IEM_NATIVE_RECOMPILER_LONGJMP
+%error "Port me"
+%endif
+        sub rsp,  IEMNATIVE_FRAME_ALIGN_SIZE \
+                + IEMNATIVE_FRAME_VAR_SIZE \
+                + IEMNATIVE_FRAME_STACK_ARG_COUNT * 8 \
+                + IEMNATIVE_FRAME_SHADOW_ARG_COUNT * 8 \
+
+%ifdef ASM_CALL64_MSC
+        jmp rdx
+%else
+        jmp rsi
+%endif
+ENDPROC     iemNativeTbEntry
+%endif
 
 

trunk/src/VBox/VMM/VMMAll/IEMAllN8veRecompiler.cpp

@@ -6402 +6402 @@
 
 
+#ifndef IEMNATIVE_WITH_RECOMPILER_PROLOGUE_SINGLETON
 /**
  * Emits a standard prolog.
@@ -6515 +6516 @@
     return off;
 }
+#endif
 
 
@@ -9110 +9112 @@
     IEMNATIVE_TRY_SETJMP(pReNative, rc)
     {
+#ifndef IEMNATIVE_WITH_RECOMPILER_PROLOGUE_SINGLETON
         /*
          * Emit prolog code (fixed).
          */
         off = iemNativeEmitProlog(pReNative, off);
+#endif
 
         /*

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

@@ -2689 +2689 @@
         iemThreadedLogCurInstr(pVCpu, "EXn", 0);
 # endif
-# ifdef RT_ARCH_AMD64
+
+# ifndef IEMNATIVE_WITH_RECOMPILER_PROLOGUE_SINGLETON
+#  ifdef RT_ARCH_AMD64
         VBOXSTRICTRC const rcStrict = ((PFNIEMTBNATIVE)pTb->Native.paInstructions)(pVCpu);
+#  else
+        VBOXSTRICTRC const rcStrict = ((PFNIEMTBNATIVE)pTb->Native.paInstructions)(pVCpu, &pVCpu->cpum.GstCtx);
+#  endif
 # else
-        VBOXSTRICTRC const rcStrict = ((PFNIEMTBNATIVE)pTb->Native.paInstructions)(pVCpu, &pVCpu->cpum.GstCtx);
+#  ifdef VBOX_WITH_IEM_NATIVE_RECOMPILER_LONGJMP
+        AssertCompileMemberOffset(VMCPUCC, iem.s.pvTbFramePointerR3, 0x7c8); /* This is assumed in iemNativeTbEntry */
+#  endif
+#  ifdef RT_ARCH_AMD64
+        VBOXSTRICTRC const rcStrict = iemNativeTbEntry(pVCpu, (uintptr_t)pTb->Native.paInstructions);
+#  else
+        VBOXSTRICTRC const rcStrict = iemNativeTbEntry(pVCpu, &pVCpu->cpum.GstCtx, (uintptr_t)pTb->Native.paInstructions);
+#  endif
 # endif
+
 # ifdef VBOX_WITH_IEM_NATIVE_RECOMPILER_LONGJMP
         pVCpu->iem.s.pvTbFramePointerR3 = NULL;

trunk/src/VBox/VMM/include/IEMN8veRecompiler.h

@@ -68 +68 @@
 /** Always count instructions for now. */
 # define IEMNATIVE_WITH_INSTRUCTION_COUNTING
+#endif
+
+/** @def IEMNATIVE_WITH_RECOMPILER_PROLOGUE_SINGLETON
+ * Enables having only a single prologue for native TBs. */
+#if 1 || defined(DOXYGEN_RUNNING)
+# define IEMNATIVE_WITH_RECOMPILER_PROLOGUE_SINGLETON
 #endif
 
@@ -199 +205 @@
 #elif defined(RT_ARCH_ARM64) || defined(DOXYGEN_RUNNING)
 # define IEMNATIVE_REG_FIXED_PVMCPU         ARMV8_A64_REG_X28
+# define IEMNATIVE_REG_FIXED_PVMCPU_ASM     x28
 # define IEMNATIVE_REG_FIXED_PCPUMCTX       ARMV8_A64_REG_X27
+# define IEMNATIVE_REG_FIXED_PCPUMCTX_ASM   x27
 # define IEMNATIVE_REG_FIXED_TMP0           ARMV8_A64_REG_X15
 # if defined(IEMNATIVE_WITH_DELAYED_PC_UPDATING) && 0 /* debug the updating with a shadow RIP. */
@@ -417 +425 @@
 
 
+/** @todo r=aeichner Can this be made prettier? */
+#ifndef INCLUDED_FROM_ARM64_ASSEMBLY
+
 /** Native code generator label types. */
 typedef enum
@@ -2480 +2491 @@
 
 
+#ifdef IEMNATIVE_WITH_RECOMPILER_PROLOGUE_SINGLETON
+# ifdef RT_ARCH_AMD64
+extern "C" IEM_DECL_NATIVE_HLP_DEF(VBOXSTRICTRC, iemNativeTbEntry, (PVMCPUCC pVCpu, uintptr_t pTbInsn));
+# elif defined(RT_ARCH_ARM64)
+extern "C" IEM_DECL_NATIVE_HLP_DEF(VBOXSTRICTRC, iemNativeTbEntry, (PVMCPUCC pVCpu, PCPUMCTX pCpumCtx, uintptr_t pTbInsn));
+# endif
+#endif
+
+#endif /* !INCLUDED_FROM_ARM64_ASSEMBLY */
+
 /** @} */