diff options
author | Jonghyun Park <parjong@gmail.com> | 2016-12-14 08:36:06 +0900 |
---|---|---|
committer | Jan Vorlicek <janvorli@microsoft.com> | 2016-12-14 00:36:06 +0100 |
commit | e8e2adfb61834873477e0f154eab90252d6a6047 (patch) | |
tree | 7ac49b79e39419d6472bafd6193e45698e2f2979 | |
parent | 4f1a90f2f4a2bd0656f22cb76c56e7a7fb631d73 (diff) | |
download | coreclr-e8e2adfb61834873477e0f154eab90252d6a6047.tar.gz coreclr-e8e2adfb61834873477e0f154eab90252d6a6047.tar.bz2 coreclr-e8e2adfb61834873477e0f154eab90252d6a6047.zip |
[x86/Linux] Port jithelp.asm (#8491)
* [x86/Linux] Port jithelp.asm
This commit ports jithelp.asm for x86/Linux
The following Tailcall helpers are marked as NYI:
- JIT_TailCall
- JIT_TailCallReturnFromVSD
- JIT_TailCallVSDLeave
- JIT_TailCallLeave
* Revise macro and indentation
-rw-r--r-- | src/pal/inc/unixasmmacrosx86.inc | 11 | ||||
-rw-r--r-- | src/vm/CMakeLists.txt | 1 | ||||
-rw-r--r-- | src/vm/i386/jithelp.S | 749 | ||||
-rw-r--r-- | src/vm/i386/unixstubs.cpp | 20 |
4 files changed, 781 insertions, 0 deletions
diff --git a/src/pal/inc/unixasmmacrosx86.inc b/src/pal/inc/unixasmmacrosx86.inc index d7d530434a..6e7254e1b1 100644 --- a/src/pal/inc/unixasmmacrosx86.inc +++ b/src/pal/inc/unixasmmacrosx86.inc @@ -65,3 +65,14 @@ C_FUNC(\Name\()_End): .macro EPILOG_END pop ebp .endm + +.macro PREPARE_EXTERNAL_VAR Name, Reg +.att_syntax + call 0f +0: + popl %\Reg +1: + addl $_GLOBAL_OFFSET_TABLE_ + (1b - 0b), %\Reg + movl C_FUNC(\Name)@GOT(%\Reg), %\Reg +.intel_syntax noprefix +.endm diff --git a/src/vm/CMakeLists.txt b/src/vm/CMakeLists.txt index 0c7966bf51..9fdddd6c26 100644 --- a/src/vm/CMakeLists.txt +++ b/src/vm/CMakeLists.txt @@ -378,6 +378,7 @@ else(WIN32) elseif(CLR_CMAKE_TARGET_ARCH_I386) set(VM_SOURCES_WKS_ARCH_ASM ${ARCH_SOURCES_DIR}/asmhelpers.S + ${ARCH_SOURCES_DIR}/jithelp.S ${ARCH_SOURCES_DIR}/gmsasm.S ${ARCH_SOURCES_DIR}/umthunkstub.S ) diff --git a/src/vm/i386/jithelp.S b/src/vm/i386/jithelp.S new file mode 100644 index 0000000000..66ae9fb451 --- /dev/null +++ b/src/vm/i386/jithelp.S @@ -0,0 +1,749 @@ +// Licensed to the .NET Foundation under one or more agreements. +// The .NET Foundation licenses this file to you under the MIT license. +// See the LICENSE file in the project root for more information. + +.intel_syntax noprefix +#include "unixasmmacros.inc" +#include "asmconstants.h" + + +// *** +// JIT_WriteBarrier* - GC write barrier helper +// +// Purpose: +// Helper calls in order to assign an object to a field +// Enables book-keeping of the GC. +// +// Entry: +// EDX - address of ref-field (assigned to) +// the resp. other reg - RHS of assignment +// +// Exit: +// +// Uses: +// EDX is destroyed. +// +// Exceptions: +// +// ******************************************************************************* + +// The code here is tightly coupled with AdjustContextForWriteBarrier, if you change +// anything here, you might need to change AdjustContextForWriteBarrier as well +.macro WriteBarrierHelper rg +.align 4 + +// The entry point is the fully 'safe' one in which we check if EDX (the REF +// begin updated) is actually in the GC heap +NESTED_ENTRY JIT_CheckedWriteBarrier\rg, _TEXT, NoHandler + // check in the REF being updated is in the GC heap + push eax + PREPARE_EXTERNAL_VAR g_lowest_address, eax + cmp edx, [eax] + pop eax + jb LOCAL_LABEL(WriteBarrier_NotInHeap_\rg) + push eax + PREPARE_EXTERNAL_VAR g_highest_address, eax + cmp edx, [eax] + pop eax + jae LOCAL_LABEL(WriteBarrier_NotInHeap_\rg) + + // fall through to unchecked routine + // note that its entry point also happens to be aligned + +#ifdef WRITE_BARRIER_CHECK + // This entry point is used when you know the REF pointer being updated + // is in the GC heap +PATCH_LABEL JIT_DebugWriteBarrier\rg +#endif // WRITE_BARRIER_CHECK + +#ifdef _DEBUG + push edx + push ecx + push eax + + push \rg + push edx + call C_FUNC(WriteBarrierAssert) + + pop eax + pop ecx + pop edx +#endif // _DEBUG + + // in the !WRITE_BARRIER_CHECK case this will be the move for all + // addresses in the GCHeap, addresses outside the GCHeap will get + // taken care of below at WriteBarrier_NotInHeap_&rg + +#ifndef WRITE_BARRIER_CHECK + mov DWORD PTR [edx], \rg +#endif // !WRITE_BARRIER_CHECK + +#ifdef WRITE_BARRIER_CHECK + // Test dest here so if it is bad AV would happen before we change register/stack + // status. This makes job of AdjustContextForWriteBarrier easier. + cmp BYTE PTR [edx], 0 + // ALSO update the shadow GC heap if that is enabled + // Make ebp into the temporary src register. We need to do this so that we can use ecx + // in the calculation of the shadow GC address, but still have access to the src register + push ecx + push ebp + mov ebp, \rg + + // if g_GCShadow is 0, don't perform the check + push eax + PREPARE_EXTERNAL_VAR g_GCShadow, eax + cmp DWORD PTR [eax], 0 + pop eax + je LOCAL_LABEL(WriteBarrier_NoShadow_\rg) + + mov ecx, edx + push eax + PREPARE_EXTERNAL_VAR g_lowest_address, eax + sub ecx, [eax] + pop eax + jb LOCAL_LABEL(WriteBarrier_NoShadow_\rg) + push edx + PREPARE_EXTERNAL_VAR g_GCShadow, edx + mov [edx], edx + add ecx, [edx] + PREPARE_EXTERNAL_VAR g_GCShadowEnd, edx + mov [edx], edx + cmp ecx, [edx] + pop edx + ja LOCAL_LABEL(WriteBarrier_NoShadow_\rg) + + // TODO: In Orcas timeframe if we move to P4+ only on X86 we should enable + // mfence barriers on either side of these two writes to make sure that + // they stay as close together as possible + + // edx contains address in GC + // ecx contains address in ShadowGC + // ebp temporarially becomes the src register + + // When we're writing to the shadow GC heap we want to be careful to minimize + // the risk of a race that can occur here where the GC and ShadowGC don't match + mov DWORD PTR [edx], ebp + mov DWORD PTR [ecx], ebp + + // We need a scratch register to verify the shadow heap. We also need to + // construct a memory barrier so that the write to the shadow heap happens + // before the read from the GC heap. We can do both by using SUB/XCHG + // rather than PUSH. + // + // TODO: Should be changed to a push if the mfence described above is added. + // + sub esp, 4 + xchg [esp], eax + + // As part of our race avoidance (see above) we will now check whether the values + // in the GC and ShadowGC match. There is a possibility that we're wrong here but + // being overaggressive means we might mask a case where someone updates GC refs + // without going to a write barrier, but by its nature it will be indeterminant + // and we will find real bugs whereas the current implementation is indeterminant + // but only leads to investigations that find that this code is fundamentally flawed + mov eax, [edx] + cmp [ecx], eax + je LOCAL_LABEL(WriteBarrier_CleanupShadowCheck_\rg) + mov DWORD PTR [ecx], INVALIDGCVALUE + +LOCAL_LABEL(WriteBarrier_CleanupShadowCheck_\rg): + pop eax + + jmp LOCAL_LABEL(WriteBarrier_ShadowCheckEnd_\rg) + +LOCAL_LABEL(WriteBarrier_NoShadow_\rg): + // If we come here then we haven't written the value to the GC and need to. + // ebp contains rg + // We restore ebp/ecx immediately after this, and if either of them is the src + // register it will regain its value as the src register. + mov DWORD PTR [edx], ebp +LOCAL_LABEL(WriteBarrier_ShadowCheckEnd_\rg): + pop ebp + pop ecx +#endif // WRITE_BARRIER_CHECK + + push eax + push ebx + mov eax, \rg + PREPARE_EXTERNAL_VAR g_ephemeral_low, ebx + cmp eax, [ebx] + pop ebx + pop eax + jb LOCAL_LABEL(WriteBarrier_NotInEphemeral_\rg) + push eax + push ebx + mov eax, \rg + PREPARE_EXTERNAL_VAR g_ephemeral_high, ebx + cmp eax, [ebx] + pop ebx + pop eax + jae LOCAL_LABEL(WriteBarrier_NotInEphemeral_\rg) + + shr edx, 10 + push eax + PREPARE_EXTERNAL_VAR g_card_table, eax + add edx, [eax] + pop eax + cmp BYTE PTR [edx], 0FFh + jne LOCAL_LABEL(WriteBarrier_UpdateCardTable_\rg) + ret + +LOCAL_LABEL(WriteBarrier_UpdateCardTable_\rg): + mov BYTE PTR [edx], 0FFh + ret + +LOCAL_LABEL(WriteBarrier_NotInHeap_\rg): + // If it wasn't in the heap then we haven't updated the dst in memory yet + mov DWORD PTR [edx], \rg + +LOCAL_LABEL(WriteBarrier_NotInEphemeral_\rg): + // If it is in the GC Heap but isn't in the ephemeral range we've already + // updated the Heap with the Object*. + ret +NESTED_END JIT_CheckedWriteBarrier\rg, _TEXT + +.endm + + +// *** +// JIT_ByRefWriteBarrier* - GC write barrier helper +// +// Purpose: +// Helper calls in order to assign an object to a byref field +// Enables book-keeping of the GC. +// +// Entry: +// EDI - address of ref-field (assigned to) +// ESI - address of the data (source) +// ECX can be trashed +// +// Exit: +// +// Uses: +// EDI and ESI are incremented by a DWORD +// +// Exceptions: +// +// ******************************************************************************* +// +// The code here is tightly coupled with AdjustContextForWriteBarrier, if you change +// anything here, you might need to change AdjustContextForWriteBarrier as well +// +.macro ByRefWriteBarrierHelper +.align 4 + +LEAF_ENTRY JIT_ByRefWriteBarrier, _TEXT + // test for dest in range + mov ecx, [esi] + push eax + PREPARE_EXTERNAL_VAR g_lowest_address, eax + cmp edi, [eax] + pop eax + jb LOCAL_LABEL(ByRefWriteBarrier_NotInHeap) + push eax + PREPARE_EXTERNAL_VAR g_highest_address, eax + cmp edi, [eax] + pop eax + jae LOCAL_LABEL(ByRefWriteBarrier_NotInHeap) + +#ifndef WRITE_BARRIER_CHECK + // write barrier + mov [edi], ecx +#endif // !WRITE_BARRIER_CHECK + +#ifdef WRITE_BARRIER_CHECK + // Test dest here so if it is bad AV would happen before we change register/stack + // status. This makes job of AdjustContextForWriteBarrier easier. + cmp BYTE PTR [edi], 0 + + // ALSO update the shadow GC heap if that is enabled + + // use edx for address in GC Shadow, + push edx + + // if g_GCShadow is 0, don't do the update + push ebx + PREPARE_EXTERNAL_VAR g_GCShadow, ebx + cmp DWORD PTR [ebx], 0 + pop ebx + je LOCAL_LABEL(ByRefWriteBarrier_NoShadow) + + mov edx, edi + push ebx + PREPARE_EXTERNAL_VAR g_lowest_address, ebx + sub edx, [ebx] // U/V + pop ebx + jb LOCAL_LABEL(ByRefWriteBarrier_NoShadow) + push eax + PREPARE_EXTERNAL_VAR g_GCShadow, eax + mov eax, [eax] + add edx, [eax] + PREPARE_EXTERNAL_VAR g_GCShadowEnd, eax + mov eax, [eax] + cmp edx, [eax] + pop eax + ja LOCAL_LABEL(ByRefWriteBarrier_NoShadow) + + // TODO: In Orcas timeframe if we move to P4+ only on X86 we should enable + // mfence barriers on either side of these two writes to make sure that + // they stay as close together as possible + + // edi contains address in GC + // edx contains address in ShadowGC + // ecx is the value to assign + + // When we're writing to the shadow GC heap we want to be careful to minimize + // the risk of a race that can occur here where the GC and ShadowGC don't match + mov DWORD PTR [edi], ecx + mov DWORD PTR [edx], ecx + + // We need a scratch register to verify the shadow heap. We also need to + // construct a memory barrier so that the write to the shadow heap happens + // before the read from the GC heap. We can do both by using SUB/XCHG + // rather than PUSH. + // + // TODO: Should be changed to a push if the mfence described above is added. + // + sub esp, 4 + xchg [esp], eax + + // As part of our race avoidance (see above) we will now check whether the values + // in the GC and ShadowGC match. There is a possibility that we're wrong here but + // being overaggressive means we might mask a case where someone updates GC refs + // without going to a write barrier, but by its nature it will be indeterminant + // and we will find real bugs whereas the current implementation is indeterminant + // but only leads to investigations that find that this code is fundamentally flawed + + mov eax, [edi] + cmp [edx], eax + je LOCAL_LABEL(ByRefWriteBarrier_CleanupShadowCheck) + mov DWORD PTR [edx], INVALIDGCVALUE +LOCAL_LABEL(ByRefWriteBarrier_CleanupShadowCheck): + pop eax + jmp LOCAL_LABEL(ByRefWriteBarrier_ShadowCheckEnd) + +LOCAL_LABEL(ByRefWriteBarrier_NoShadow): + // If we come here then we haven't written the value to the GC and need to. + mov DWORD PTR [edi], ecx + +LOCAL_LABEL(ByRefWriteBarrier_ShadowCheckEnd): + pop edx +#endif // WRITE_BARRIER_CHECK + + // test for *src in ephemeral segement + push eax + PREPARE_EXTERNAL_VAR g_ephemeral_low, eax + cmp ecx, [eax] + pop eax + jb LOCAL_LABEL(ByRefWriteBarrier_NotInEphemeral) + push eax + PREPARE_EXTERNAL_VAR g_ephemeral_high, eax + cmp ecx, [eax] + pop eax + jae LOCAL_LABEL(ByRefWriteBarrier_NotInEphemeral) + + mov ecx, edi + add esi, 4 + add edi, 4 + + shr ecx, 10 + push eax + PREPARE_EXTERNAL_VAR g_card_table, eax + add ecx, [eax] + pop eax + cmp BYTE PTR [ecx], 0FFh + jne LOCAL_LABEL(ByRefWriteBarrier_UpdateCardTable) + ret +LOCAL_LABEL(ByRefWriteBarrier_UpdateCardTable): + mov BYTE PTR [ecx], 0FFh + ret + +LOCAL_LABEL(ByRefWriteBarrier_NotInHeap): + // If it wasn't in the heap then we haven't updated the dst in memory yet + mov [edi], ecx +LOCAL_LABEL(ByRefWriteBarrier_NotInEphemeral): + // If it is in the GC Heap but isn't in the ephemeral range we've already + // updated the Heap with the Object*. + add esi, 4 + add edi, 4 + ret +NESTED_END JIT_ByRefWriteBarrier, _TEXT + +.endm + +// WriteBarrierStart and WriteBarrierEnd are used to determine bounds of +// WriteBarrier functions so can determine if got AV in them. +// +LEAF_ENTRY JIT_WriteBarrierStart, _TEXT + ret +LEAF_END JIT_WriteBarrierStart, _TEXT + +#ifdef FEATURE_USE_ASM_GC_WRITE_BARRIERS +// ******************************************************************************* +// Write barrier wrappers with fcall calling convention +// +.macro UniversalWriteBarrierHelper name +.align 4 + +LEAF_ENTRY JIT_\name, _TEXT + mov eax, edx + mov edx, ecx + jmp C_FUNC(JIT_\name\()EAX) +LEAF_END JIT_\name, _TEXT + +.endm + +// Only define these if we're using the ASM GC write barriers; if this flag is not defined, +// we'll use C++ versions of these write barriers. +UniversalWriteBarrierHelper CheckedWriteBarrier +UniversalWriteBarrierHelper WriteBarrier +#endif // FEATURE_USE_ASM_GC_WRITE_BARRIERS + +WriteBarrierHelper EAX +WriteBarrierHelper EBX +WriteBarrierHelper ECX +WriteBarrierHelper ESI +WriteBarrierHelper EDI +WriteBarrierHelper EBP + +ByRefWriteBarrierHelper + +LEAF_ENTRY JIT_WriteBarrierLast, _TEXT + ret +LEAF_END JIT_WriteBarrierLast, _TEXT + +// This is the first function outside the "keep together range". Used by BBT scripts. +LEAF_ENTRY JIT_WriteBarrierEnd, _TEXT + ret +LEAF_END JIT_WriteBarrierEnd, _TEXT + +// *********************************************************************/ +// In cases where we support it we have an optimized GC Poll callback. +// Normall (when we're not trying to suspend for GC, the CORINFO_HELP_POLL_GC +// helper points to this nop routine. When we're ready to suspend for GC, +// we whack the Jit Helper table entry to point to the real helper. When we're +// done with GC we whack it back. +LEAF_ENTRY JIT_PollGC_Nop, _TEXT + ret +LEAF_END JIT_PollGC_Nop, _TEXT + +// *********************************************************************/ +// llshl - long shift left +// +// Purpose: +// Does a Long Shift Left (signed and unsigned are identical) +// Shifts a long left any number of bits. +// +// NOTE: This routine has been adapted from the Microsoft CRTs. +// +// Entry: +// EDX:EAX - long value to be shifted +// ECX - number of bits to shift by +// +// Exit: +// EDX:EAX - shifted value +// +.align 16 +LEAF_ENTRY JIT_LLsh, _TEXT + cmp ecx, 32 + jae LOCAL_LABEL(LLshMORE32) + + // Handle shifts of between bits 0 and 31 + shld edx, eax, cl + shl eax, cl + ret + +LOCAL_LABEL(LLshMORE32): + // Handle shifts of between bits 32 and 63 + // The x86 shift instructions only use the lower 5 bits. + mov edx, eax + xor eax, eax + shl edx, cl + ret +LEAF_END JIT_LLsh, _TEXT + +// *********************************************************************/ +// LRsh - long shift right +// +// Purpose: +// Does a signed Long Shift Right +// Shifts a long right any number of bits. +// +// NOTE: This routine has been adapted from the Microsoft CRTs. +// +// Entry: +// EDX:EAX - long value to be shifted +// ECX - number of bits to shift by +// +// Exit: +// EDX:EAX - shifted value +// +.align 16 +LEAF_ENTRY JIT_LRsh, _TEXT + cmp ecx, 32 + jae LOCAL_LABEL(LRshMORE32) + + // Handle shifts of between bits 0 and 31 + shrd eax, edx, cl + sar edx, cl + ret + +LOCAL_LABEL(LRshMORE32): + // Handle shifts of between bits 32 and 63 + // The x86 shift instructions only use the lower 5 bits. + mov eax, edx + sar edx, 31 + sar eax, cl + ret +LEAF_END JIT_LRsh, _TEXT + +// *********************************************************************/ +// LRsz: +// Purpose: +// Does a unsigned Long Shift Right +// Shifts a long right any number of bits. +// +// NOTE: This routine has been adapted from the Microsoft CRTs. +// +// Entry: +// EDX:EAX - long value to be shifted +// ECX - number of bits to shift by +// +// Exit: +// EDX:EAX - shifted value +// +.align 16 +LEAF_ENTRY JIT_LRsz, _TEXT + cmp ecx, 32 + jae LOCAL_LABEL(LRszMORE32) + + // Handle shifts of between bits 0 and 31 + shrd eax, edx, cl + shr edx, cl + ret + +LOCAL_LABEL(LRszMORE32): + // Handle shifts of between bits 32 and 63 + // The x86 shift instructions only use the lower 5 bits. + mov eax, edx + xor edx, edx + shr eax, cl + ret +LEAF_END JIT_LRsz, _TEXT + +// *********************************************************************/ +// JIT_Dbl2LngP4x87 +// +// Purpose: +// converts a double to a long truncating toward zero (C semantics) +// +// uses stdcall calling conventions +// +// This code is faster on a P4 than the Dbl2Lng code above, but is +// slower on a PIII. Hence we choose this code when on a P4 or above. +// +LEAF_ENTRY JIT_Dbl2LngP4x87, _TEXT + // get some local space + sub esp, 8 + + #define arg1 [esp + 0Ch] + fld QWORD PTR arg1 // fetch arg + fnstcw WORD PTR arg1 // store FPCW + movzx eax, WORD PTR arg1 // zero extend - wide + or ah, 0Ch // turn on OE and DE flags + mov DWORD PTR [esp], eax // store new FPCW bits + fldcw WORD PTR [esp] // reload FPCW with new bits + fistp QWORD PTR [esp] // convert + + // reload FP result + mov eax, DWORD PTR [esp] + mov edx, DWORD PTR [esp + 4] + + // reload original FPCW value + fldcw WORD PTR arg1 + #undef arg1 + + // restore stack + add esp, 8 + + ret 8 +LEAF_END JIT_Dbl2LngP4x87, _TEXT + +// *********************************************************************/ +// JIT_Dbl2LngSSE3 +// +// Purpose: +// converts a double to a long truncating toward zero (C semantics) +// +// uses stdcall calling conventions +// +// This code is faster than the above P4 x87 code for Intel processors +// equal or later than Core2 and Atom that have SSE3 support +// +LEAF_ENTRY JIT_Dbl2LngSSE3, _TEXT + // get some local space + sub esp, 8 + + fld QWORD PTR [esp + 0Ch] // fetch arg + fisttp QWORD PTR [esp] // convert + mov eax, DWORD PTR [esp] // reload FP result + mov edx, DWORD PTR [esp + 4] + + // restore stack + add esp, 8 + + ret 8 +LEAF_END JIT_Dbl2LngSSE3, _TEXT + +// *********************************************************************/ +// JIT_Dbl2IntSSE2 +// +// Purpose: +// converts a double to a long truncating toward zero (C semantics) +// +// uses stdcall calling conventions +// +// This code is even faster than the P4 x87 code for Dbl2LongP4x87, +// but only returns a 32 bit value (only good for int). +// +LEAF_ENTRY JIT_Dbl2IntSSE2, _TEXT + movsd xmm0, [esp + 4] + cvttsd2si eax, xmm0 + ret 8 +LEAF_END JIT_Dbl2IntSSE2, _TEXT + +// *********************************************************************/ +// This is the small write barrier thunk we use when we know the +// ephemeral generation is higher in memory than older generations. +// The 0x0F0F0F0F values are bashed by the two functions above. +// This the generic version - wherever the code says ECX, +// the specific register is patched later into a copy +// Note: do not replace ECX by EAX - there is a smaller encoding for +// the compares just for EAX, which won't work for other registers. +// +// READ THIS!!!!!! +// it is imperative that the addresses of of the values that we overwrite +// (card table, ephemeral region ranges, etc) are naturally aligned since +// there are codepaths that will overwrite these values while the EE is running. +// +LEAF_ENTRY JIT_WriteBarrierReg_PreGrow, _TEXT + mov DWORD PTR [edx], ecx + cmp ecx, 0F0F0F0F0h + jb LOCAL_LABEL(NoWriteBarrierPre) + + shr edx, 10 + nop // padding for alignment of constant + cmp BYTE PTR [edx + 0F0F0F0F0h], 0FFh + jne LOCAL_LABEL(WriteBarrierPre) + +LOCAL_LABEL(NoWriteBarrierPre): + ret + nop // padding for alignment of constant + nop // padding for alignment of constant + +LOCAL_LABEL(WriteBarrierPre): + mov BYTE PTR [edx+0F0F0F0F0h], 0FFh + ret +LEAF_END JIT_WriteBarrierReg_PreGrow, _TEXT + +// *********************************************************************/ +// This is the larger write barrier thunk we use when we know that older +// generations may be higher in memory than the ephemeral generation +// The 0x0F0F0F0F values are bashed by the two functions above. +// This the generic version - wherever the code says ECX, +// the specific register is patched later into a copy +// Note: do not replace ECX by EAX - there is a smaller encoding for +// the compares just for EAX, which won't work for other registers. +// NOTE: we need this aligned for our validation to work properly +.align 4 +LEAF_ENTRY JIT_WriteBarrierReg_PostGrow, _TEXT + mov DWORD PTR [edx], ecx + cmp ecx, 0F0F0F0F0h + jb LOCAL_LABEL(NoWriteBarrierPost) + cmp ecx, 0F0F0F0F0h + jae LOCAL_LABEL(NoWriteBarrierPost) + + shr edx, 10 + nop // padding for alignment of constant + cmp BYTE PTR [edx + 0F0F0F0F0h], 0FFh + jne LOCAL_LABEL(WriteBarrierPost) + +LOCAL_LABEL(NoWriteBarrierPost): + ret + nop // padding for alignment of constant + nop // padding for alignment of constant + +LOCAL_LABEL(WriteBarrierPost): + mov BYTE PTR [edx + 0F0F0F0F0h], 0FFh + ret +LEAF_END JIT_WriteBarrierReg_PostGrow,_TEXT + +// PatchedCodeStart and PatchedCodeEnd are used to determine bounds of patched code. +// + +LEAF_ENTRY JIT_PatchedCodeStart, _TEXT + ret +LEAF_END JIT_PatchedCodeStart, _TEXT + +// ********************************************************************** +// Write barriers generated at runtime + +LEAF_ENTRY JIT_PatchedWriteBarrierStart, _TEXT + ret +LEAF_END JIT_PatchedWriteBarrierStart, _TEXT + +.macro PatchedWriteBarrierHelper rg +.align 8 +LEAF_ENTRY JIT_WriteBarrier\rg, _TEXT + // Just allocate space that will be filled in at runtime + .space 0CCH, 48 +LEAF_END JIT_WriteBarrier\rg, _TEXT + +.endm + +PatchedWriteBarrierHelper EAX +PatchedWriteBarrierHelper EBX +PatchedWriteBarrierHelper ECX +PatchedWriteBarrierHelper ESI +PatchedWriteBarrierHelper EDI +PatchedWriteBarrierHelper EBP + +LEAF_ENTRY JIT_PatchedWriteBarrierLast, _TEXT + ret +LEAF_END JIT_PatchedWriteBarrierLast, _TEXT + +LEAF_ENTRY JIT_PatchedCodeLast, _TEXT + ret +LEAF_END JIT_PatchedCodeLast, _TEXT + +// This is the first function outside the "keep together range". Used by BBT scripts. +LEAF_ENTRY JIT_PatchedCodeEnd, _TEXT + ret +LEAF_END JIT_PatchedCodeEnd, _TEXT + +// Note that the debugger skips this entirely when doing SetIP, +// since COMPlusCheckForAbort should always return 0. Excep.cpp:LeaveCatch +// asserts that to be true. If this ends up doing more work, then the +// debugger may need additional support. +// void __stdcall JIT_EndCatch(); +NESTED_ENTRY JIT_EndCatch, _TEXT, NoHandler + // make temp storage for return address, and push the address of that + // as the last arg to COMPlusEndCatch + mov ecx, [esp] + push ecx + push esp + + // push the rest of COMPlusEndCatch's args, right-to-left + push esi + push edi + push ebx + push ebp + + // returns old esp value in eax, stores jump address + call C_FUNC(COMPlusEndCatch) + // now eax = new esp, [esp] = new eip + + pop edx // edx = new eip + mov esp, eax // esp = new esp + jmp edx // eip = new eip +NESTED_END JIT_EndCatch, _TEXT diff --git a/src/vm/i386/unixstubs.cpp b/src/vm/i386/unixstubs.cpp index 5f098c1814..9fe7127946 100644 --- a/src/vm/i386/unixstubs.cpp +++ b/src/vm/i386/unixstubs.cpp @@ -84,3 +84,23 @@ EXTERN_C VOID BackPatchWorkerAsmStub() { PORTABILITY_ASSERT("BackPatchWorkerAsmStub"); } + +EXTERN_C VOID JIT_TailCall() +{ + PORTABILITY_ASSERT("JIT_TailCall"); +} + +EXTERN_C VOID JIT_TailCallReturnFromVSD() +{ + PORTABILITY_ASSERT("JIT_TailCallReturnFromVSD"); +} + +EXTERN_C VOID JIT_TailCallVSDLeave() +{ + PORTABILITY_ASSERT("JIT_TailCallVSDLeave"); +} + +EXTERN_C VOID JIT_TailCallLeave() +{ + PORTABILITY_ASSERT("JIT_TailCallLeave"); +} |