Changeset 95372 in vbox for trunk/src/VBox

Timestamp:

Jun 26, 2022 12:27:49 AM (3 years ago)

Author:

vboxsync

Message:

ValKit/bs3kit: Some more work on extended CPU/FPU state management. bugref:9898

Location:

trunk/src/VBox/ValidationKit/bootsectors/bs3kit

Files:

• Makefile.kmk (modified) (2 diffs)
• bs3-cmn-ExtCtxGetMxCsr.c (added)
• bs3-cmn-ExtCtxGetYmm.c (added)
• bs3-cmn-ExtCtxSave.asm (modified) (3 diffs)
• bs3-cmn-ExtCtxSetMxCsr.c (added)
• bs3-cmn-ExtCtxSetYmm.c (added)
• bs3-cmn-RegGetXcr0.asm (modified) (3 diffs)
• bs3-cmn-RegGetXcr0Asm.asm (deleted)
• bs3-cmn-RegSetXcr0.asm (modified) (3 diffs)
• bs3kit-autostubs.kmk (modified) (4 diffs)
• bs3kit-mangling-code-define.h (modified) (1 diff)
• bs3kit-mangling-code-undef.h (modified) (1 diff)
• bs3kit.h (modified) (1 diff)

trunk/src/VBox/ValidationKit/bootsectors/bs3kit/Makefile.kmk

@@ -139 +139 @@
        bs3-cmn-RegSetLdtr.asm \
        bs3-cmn-RegGetXcr0.asm \
-       bs3-cmn-RegGetXcr0Asm.asm \
        bs3-cmn-RegSetXcr0.asm \
        bs3-cmn-ExtCtxInit.c \
@@ -148 +147 @@
        bs3-cmn-ExtCtxFree.c \
        bs3-cmn-ExtCtxCopy.c \
+       bs3-cmn-ExtCtxGetMxCsr.c \
+       bs3-cmn-ExtCtxSetMxCsr.c \
+       bs3-cmn-ExtCtxGetYmm.c \
+       bs3-cmn-ExtCtxSetYmm.c \
        bs3-cmn-SelFar32ToFlat32.c \
        bs3-cmn-SelFar32ToFlat32NoClobber.asm \

trunk/src/VBox/ValidationKit/bootsectors/bs3kit/bs3-cmn-ExtCtxSave.asm

@@ -28 +28 @@
 
 extern BS3_CMN_NM(Bs3RegSetXcr0)
-extern BS3_CMN_NM(Bs3RegGetXcr0Asm)
 
 ;;
@@ -64 +63 @@
 
 .do_16_xsave:
-        call    BS3_CMN_NM(Bs3RegGetXcr0Asm)
+        ; xgetbv can be used in any ring!
+        xor     ecx, ecx
+        xgetbv
         mov     [es:bx + BS3EXTCTX.fXcr0Saved], eax
         mov     [es:bx + BS3EXTCTX.fXcr0Saved + 4], edx
@@ -106 +107 @@
 
 .do_xsave:
-        call    BS3_CMN_NM(Bs3RegGetXcr0Asm)
+        xor     ecx, ecx
+        xgetbv
         mov     [xBX + BS3EXTCTX.fXcr0Saved], eax
         mov     [xBX + BS3EXTCTX.fXcr0Saved + 4], edx

trunk/src/VBox/ValidationKit/bootsectors/bs3kit/bs3-cmn-RegGetXcr0.asm

@@ -28 +28 @@
 
 
-BS3_EXTERN_CMN Bs3Syscall
-%if TMPL_BITS == 16
-BS3_EXTERN_DATA16 g_bBs3CurrentMode
-%endif
-TMPL_BEGIN_TEXT
-
-
 ;;
 ; @cproto   BS3_CMN_PROTO_STUB(uint64_t, Bs3RegGetXcr0,(void));
@@ -41 +34 @@
 ; @remarks  Does not require 20h of parameter scratch space in 64-bit mode.
 ;
-; @uses     No GPRs, though 16-bit mode the upper 48-bits of rax and rdx are cleared.
+; @uses     No GPRs, though 16-bit mode the upper 48-bits of RAX, RDX and RCX are cleared.
 ;
 BS3_PROC_BEGIN_CMN Bs3RegGetXcr0, BS3_PBC_HYBRID_SAFE
-        BS3_CALL_CONV_PROLOG 0
         push    xBP
         mov     xBP, xSP
 TONLY64 push    rdx
 
-%if TMPL_BITS == 16
-        ; If V8086 mode we have to go thru a syscall.
-        test    byte [BS3_DATA16_WRT(g_bBs3CurrentMode)], BS3_MODE_CODE_V86
-        jnz     .via_system_call
-        cmp     byte [BS3_DATA16_WRT(g_bBs3CurrentMode)], BS3_MODE_RM
-        je      .direct_access
-%endif
-        ; If not in ring-0, we have to make a system call.
-        mov     ax, ss
-        and     ax, X86_SEL_RPL
-        jnz     .via_system_call
-
-.direct_access:
+        ; Read the value.
 TNOT16  push    sCX
         xor     ecx, ecx
         xgetbv
 TNOT16  pop     sCX
-        jmp     .return
 
-.via_system_call:
-        mov     xAX, BS3_SYSCALL_GET_XCR0
-        call    Bs3Syscall
-
-.return:
+        ; Move the edx:eax value into the appropriate return register(s).
 %if TMPL_BITS == 16
         ; value [dx cx bx ax]
@@ -84 +59 @@
         shr     rdx, 32
         or      rax, rdx
-        pop     rdx
 %endif
+
+TONLY64 pop     rdx
         pop     xBP
-        BS3_CALL_CONV_EPILOG 0
         BS3_HYBRID_RET
 BS3_PROC_END_CMN   Bs3RegGetXcr0

trunk/src/VBox/ValidationKit/bootsectors/bs3kit/bs3-cmn-RegSetXcr0.asm

@@ -40 +40 @@
 ; @param    uValue      The value to set.
 
-; @remarks  Does not require 20h of parameter scratch space in 64-bit mode,
-;           only 8 bytes for dumping rcx.
+; @remarks  Does not require 20h of parameter scratch space in 64-bit mode.
 ;
 ; @uses     No GPRs.
 ;
 BS3_PROC_BEGIN_CMN Bs3RegSetXcr0, BS3_PBC_HYBRID_SAFE
-        BS3_CALL_CONV_PROLOG 1
         push    xBP
         mov     xBP, xSP
@@ -54 +52 @@
 
         ; Load the value
+%if TMPL_BITS == 64
+        mov     eax, ecx
+        mov     rdx, rcx
+        shr     rdx, 32
+%else
         mov     sAX, [xBP + xCB + cbCurRetAddr]
         mov     sDX, [xBP + xCB + cbCurRetAddr + 4]
+%endif
 
 %if TMPL_BITS == 16
@@ -86 +90 @@
         pop     sSI
         pop     xBP
-        BS3_CALL_CONV_EPILOG 1
         BS3_HYBRID_RET
 BS3_PROC_END_CMN   Bs3RegSetXcr0

trunk/src/VBox/ValidationKit/bootsectors/bs3kit/bs3kit-autostubs.kmk

@@ -37 +37 @@
 $(call BS3KIT_FN_GEN_CMN_FARSTUB,bs3kit-common-16,Bs3SelFar32ToFlat32NoClobber,6)
 $(call BS3KIT_FN_GEN_CMN_FARSTUB,bs3kit-common-16,Bs3RegCtxSaveEx,8)
+$(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3ExtCtxSetYmm)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3TestCheckRegCtxEx)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3TestFailed)
@@ -60 +61 @@
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3ExtCtxInit)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3TrapSetHandler)
+$(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3ExtCtxGetYmm)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3Printf)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3PrintfV)
@@ -71 +73 @@
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3SlabFree)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3TestSubErrorCount)
+$(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3ExtCtxGetMxCsr)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3SelFar32ToFlat32)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3SelProtFar32ToFlat32)
@@ -78 +81 @@
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3TrapSetDpl)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3ExtCtxFree)
+$(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3ExtCtxSetMxCsr)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3MemAlloc)
 $(call BS3KIT_FN_GEN_CMN_NEARSTUB,bs3kit-common-16,Bs3MemAllocZ)

trunk/src/VBox/ValidationKit/bootsectors/bs3kit/bs3kit-mangling-code-define.h

@@ -34 +34 @@
 #define Bs3ExtCtxCopy BS3_CMN_MANGLER(Bs3ExtCtxCopy)
 #define Bs3ExtCtxFree BS3_CMN_MANGLER(Bs3ExtCtxFree)
+#define Bs3ExtCtxGetMxCsr BS3_CMN_MANGLER(Bs3ExtCtxGetMxCsr)
 #define Bs3ExtCtxGetSize BS3_CMN_MANGLER(Bs3ExtCtxGetSize)
+#define Bs3ExtCtxGetYmm BS3_CMN_MANGLER(Bs3ExtCtxGetYmm)
 #define Bs3ExtCtxInit BS3_CMN_MANGLER(Bs3ExtCtxInit)
 #define Bs3ExtCtxRestore BS3_CMN_MANGLER(Bs3ExtCtxRestore)
 #define Bs3ExtCtxSave BS3_CMN_MANGLER(Bs3ExtCtxSave)
+#define Bs3ExtCtxSetMxCsr BS3_CMN_MANGLER(Bs3ExtCtxSetMxCsr)
+#define Bs3ExtCtxSetYmm BS3_CMN_MANGLER(Bs3ExtCtxSetYmm)
 #define Bs3GetCpuVendor BS3_CMN_MANGLER(Bs3GetCpuVendor)
 #define Bs3GetModeName BS3_CMN_MANGLER(Bs3GetModeName)

trunk/src/VBox/ValidationKit/bootsectors/bs3kit/bs3kit-mangling-code-undef.h

@@ -34 +34 @@
 #undef Bs3ExtCtxCopy
 #undef Bs3ExtCtxFree
+#undef Bs3ExtCtxGetMxCsr
 #undef Bs3ExtCtxGetSize
+#undef Bs3ExtCtxGetYmm
 #undef Bs3ExtCtxInit
 #undef Bs3ExtCtxRestore
 #undef Bs3ExtCtxSave
+#undef Bs3ExtCtxSetMxCsr
+#undef Bs3ExtCtxSetYmm
 #undef Bs3GetCpuVendor
 #undef Bs3GetModeName

trunk/src/VBox/ValidationKit/bootsectors/bs3kit/bs3kit.h

@@ -2880 +2880 @@
 BS3_CMN_PROTO_STUB(PBS3EXTCTX, Bs3ExtCtxCopy,(PBS3EXTCTX pDst, PCBS3EXTCTX pSrc));
 
+/**
+ * Gets the MXCSR register value from @a pExtCtx.
+ *
+ * @returns MXCSR value, 0 if not part of context.
+ * @param   pExtCtx         The extended CPU context.
+ */
+BS3_CMN_PROTO_STUB(uint32_t, Bs3ExtCtxGetMxCsr,(PCBS3EXTCTX pExtCtx));
+
+/**
+ * Sets the MXCSR register value in @a pExtCtx.
+ *
+ * @param   pExtCtx         The extended CPU context.
+ * @param   uValue          The new MXCSR value.
+ */
+BS3_CMN_PROTO_STUB(void, Bs3ExtCtxSetMxCsr,(PBS3EXTCTX pExtCtx, uint32_t uValue));
+
+/**
+ * Gets the value of YMM register number @a iReg from @a pExtCtx.
+ *
+ * @returns MXCSR value, 0 if not part of context.
+ * @param   pExtCtx         The extended CPU context.
+ * @param   iReg            The register to get.
+ * @param   pValue          Where to return the value.  Parts not in the
+ *                          extended state are zeroed.  For absent or invalid
+ *                          @a iReg values this is set to zero.
+ */
+BS3_CMN_PROTO_STUB(PRTUINT256U, Bs3ExtCtxGetYmm,(PCBS3EXTCTX pExtCtx, uint8_t iReg, PRTUINT256U pValue));
+
+/**
+ * Sets the value of YMM register number @a iReg in @a pExtCtx to @a pValue.
+ *
+ * @returns true if set (even if only partially). False if not set (not
+ *          supported by state format, unsupported/invalid iReg).
+ * @param   pExtCtx         The extended CPU context.
+ * @param   iReg            The register to set.
+ * @param   pValue          The new register value.
+ * @param   f128Bit         A little peculiarity grown from mixing SSE and AVX
+ *                          tests:  When set to true only the lower 128 bits of
+ *                          the YMM register will be set. The upper bits are
+ *                          untouched.  When false, the whole 256-bits value
+ *                          will be set.
+ */
+BS3_CMN_PROTO_STUB(bool, Bs3ExtCtxSetYmm,(PBS3EXTCTX pExtCtx, uint8_t iReg, PCRTUINT256U pValue, bool f128Bit));
+
 
 /** @name Debug register accessors for V8086 mode (works everwhere).