This is an archive of the discontinued LLVM Phabricator instance.

Differential D119086

[COFF] Change the thunk that get inserted when calling a function in a dealyloaded DLL to leave 32 unused bytes on top of the stack on x64
Needs ReviewPublic

Authored by a_Tom on Feb 6 2022, 9:15 AM.

Details

Reviewers
ruiu
Summary

When linking a DLL with /delayload, lld insert a thunk that will call __delayLoadHelper2 on first invocation of a function located in the delayloaded DLL.
Prior to the call to __delayLoadHelper2, registers are saved on the stack (and restored after) but it seems that, at least on x64, the __delayLoadHelper2 function writes into the top of the stack, overwritting the value of the saved registers.
This change makes the thunk allocate 32 more bytes on the stack but doesn't use them so that __delayLoadHelper2 won't overwrite anything.

Diff Detail

Event Timeline

a_Tom requested review of this revision.Feb 6 2022, 9:15 AM
a_Tom created this revision.
Herald added a project: Restricted Project. · View Herald TranscriptFeb 6 2022, 9:15 AM
Herald added a subscriber: llvm-commits. · View Herald Transcript
a_Tom edited the summary of this revision. (Show Details)Feb 6 2022, 9:16 AM
Harbormaster completed remote builds in B147829: Diff 406259.Feb 6 2022, 9:25 AM

Revision Contents

PathSize
lld/
COFF/
22 lines

Diff 406259

lld/COFF/DLL.cpp

Show First 20 Lines • Show All 195 Lines • ▼ Show 20 Linesstatic const uint8_t thunkX64[] = {
0xE9, 0, 0, 0, 0, // jmp __tailMerge_<lib> 0xE9, 0, 0, 0, 0, // jmp __tailMerge_<lib>
}; };
static const uint8_t tailMergeX64[] = { static const uint8_t tailMergeX64[] = {
0x51, // push rcx 0x51, // push rcx
0x52, // push rdx 0x52, // push rdx
0x41, 0x50, // push r8 0x41, 0x50, // push r8
0x41, 0x51, // push r9 0x41, 0x51, // push r9
0x48, 0x83, 0xEC, 0x48, // sub rsp, 48h 0x48, 0x83, 0xEC, 0x68, // sub rsp, 68h
0x66, 0x0F, 0x7F, 0x04, 0x24, // movdqa xmmword ptr [rsp], xmm0 0x66, 0x0F, 0x7F, 0x44, 0x24, 0x20, // movdqa xmmword ptr [rsp+20h], xmm0
0x66, 0x0F, 0x7F, 0x4C, 0x24, 0x10, // movdqa xmmword ptr [rsp+10h], xmm1 0x66, 0x0F, 0x7F, 0x4C, 0x24, 0x30, // movdqa xmmword ptr [rsp+30h], xmm1
0x66, 0x0F, 0x7F, 0x54, 0x24, 0x20, // movdqa xmmword ptr [rsp+20h], xmm2 0x66, 0x0F, 0x7F, 0x54, 0x24, 0x40, // movdqa xmmword ptr [rsp+40h], xmm2
0x66, 0x0F, 0x7F, 0x5C, 0x24, 0x30, // movdqa xmmword ptr [rsp+30h], xmm3 0x66, 0x0F, 0x7F, 0x5C, 0x24, 0x50, // movdqa xmmword ptr [rsp+50h], xmm3
0x48, 0x8B, 0xD0, // mov rdx, rax 0x48, 0x8B, 0xD0, // mov rdx, rax
0x48, 0x8D, 0x0D, 0, 0, 0, 0, // lea rcx, [___DELAY_IMPORT_...] 0x48, 0x8D, 0x0D, 0, 0, 0, 0, // lea rcx, [___DELAY_IMPORT_...]
Lint: Pre-merge checks
Inline Actions

clang-format: please reformat the code

-    0x48, 0x8D, 0x0D, 0, 0, 0, 0,       // lea     rcx, [___DELAY_IMPORT_...]
-    0xE8, 0, 0, 0, 0,                   // call    __delayLoadHelper2
+    0x48, 0x8D, 0x0D, 0,    0,    0,    0, // lea     rcx, [___DELAY_IMPORT_...]
+    0xE8, 0,    0,    0,    0,             // call    __delayLoadHelper2
Lint: Pre-merge checks: clang-format: please reformat the code ``` - 0x48, 0x8D, 0x0D, 0, 0, 0, 0, // lea…
0xE8, 0, 0, 0, 0, // call __delayLoadHelper2 0xE8, 0, 0, 0, 0, // call __delayLoadHelper2
0x66, 0x0F, 0x6F, 0x04, 0x24, // movdqa xmm0, xmmword ptr [rsp] 0x66, 0x0F, 0x6F, 0x44, 0x24, 0x20, // movdqa xmm0, xmmword ptr [rsp+20h]
0x66, 0x0F, 0x6F, 0x4C, 0x24, 0x10, // movdqa xmm1, xmmword ptr [rsp+10h] 0x66, 0x0F, 0x6F, 0x4C, 0x24, 0x30, // movdqa xmm1, xmmword ptr [rsp+30h]
0x66, 0x0F, 0x6F, 0x54, 0x24, 0x20, // movdqa xmm2, xmmword ptr [rsp+20h] 0x66, 0x0F, 0x6F, 0x54, 0x24, 0x40, // movdqa xmm2, xmmword ptr [rsp+40h]
0x66, 0x0F, 0x6F, 0x5C, 0x24, 0x30, // movdqa xmm3, xmmword ptr [rsp+30h] 0x66, 0x0F, 0x6F, 0x5C, 0x24, 0x50, // movdqa xmm3, xmmword ptr [rsp+50h]
0x48, 0x83, 0xC4, 0x48, // add rsp, 48h 0x48, 0x83, 0xC4, 0x48, // add rsp, 48h
0x41, 0x59, // pop r9 0x41, 0x59, // pop r9
0x41, 0x58, // pop r8 0x41, 0x58, // pop r8
0x5A, // pop rdx 0x5A, // pop rdx
0x59, // pop rcx 0x59, // pop rcx
0xFF, 0xE0, // jmp rax 0xFF, 0xE0, // jmp rax
}; };
▲ Show 20 Lines • Show All 87 Lines • ▼ Show 20 Lines
class TailMergeChunkX64 : public NonSectionChunk { class TailMergeChunkX64 : public NonSectionChunk {
public: public:
TailMergeChunkX64(Chunk *d, Defined *h) : desc(d), helper(h) {} TailMergeChunkX64(Chunk *d, Defined *h) : desc(d), helper(h) {}
size_t getSize() const override { return sizeof(tailMergeX64); } size_t getSize() const override { return sizeof(tailMergeX64); }
void writeTo(uint8_t *buf) const override { void writeTo(uint8_t *buf) const override {
memcpy(buf, tailMergeX64, sizeof(tailMergeX64)); memcpy(buf, tailMergeX64, sizeof(tailMergeX64));
write32le(buf + 39, desc->getRVA() - rva - 43); write32le(buf + 40, desc->getRVA() - rva - 44);
write32le(buf + 44, helper->getRVA() - rva - 48); write32le(buf + 45, helper->getRVA() - rva - 49);
} }
Chunk *desc = nullptr; Chunk *desc = nullptr;
Defined *helper = nullptr; Defined *helper = nullptr;
}; };
class ThunkChunkX86 : public NonSectionChunk { class ThunkChunkX86 : public NonSectionChunk {
public: public:
▲ Show 20 Lines • Show All 428 Lines • Show Last 20 Lines