[PATCH] D93931: [X86] Don't fold negative offset into 32-bit absolute address (e.g. movl $foo-1, %eax)

Fangrui Song via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Tue Dec 29 21:42:59 PST 2020


MaskRay created this revision.
MaskRay added reviewers: craig.topper, pengfei, RKSimon.
Herald added a subscriber: hiraditya.
MaskRay requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

When building abseil-cpp `bin/absl_hash_test` with Clang in -fno-pic
mode, an instruction like `movl $foo-2147483648, $eax` may be produced
(subtracting a number from a static variable address). This is invalid
if foo's address is smaller than 2147483648 because ELF R_X86_64_32 does
not allow a negative value (errored by GNU ld/gold/LLD).

Actually any negative offset is not allowed because the symbol address
can be zero (e.g. set by `-Wl,--defsym=foo=0`). So disallow such folding
in isOffsetSuitableForCodeModel.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D93931

Files:
  llvm/lib/Target/X86/X86ISelLowering.cpp
  llvm/lib/Target/X86/X86ISelLowering.h
  llvm/test/CodeGen/X86/fold-add.ll


Index: llvm/test/CodeGen/X86/fold-add.ll
===================================================================
--- llvm/test/CodeGen/X86/fold-add.ll
+++ llvm/test/CodeGen/X86/fold-add.ll
@@ -54,10 +54,12 @@
   ret i64 add (i64 ptrtoint (i32* @foo to i64), i64 1701208431)
 }
 
+;; Test we don't emit movl foo-1, %eax. ELF R_X86_64_32 does not allow
+;; a negative value.
 define dso_local i64 @neg_1() #0 {
 ; CHECK-LABEL: neg_1:
 ; CHECK:       # %bb.0:
-; STATIC-NEXT:   movl $foo-1, %eax
+; STATIC-NEXT:   leaq foo-1(%rip), %rax
 ; PIC-NEXT:      leaq foo-1(%rip), %rax
 ; MSTATIC-NEXT:  movabsq $foo, %rax
 ; MSTATIC-NEXT:  decq %rax
@@ -71,7 +73,7 @@
 define dso_local i64 @neg_0x80000000() #0 {
 ; CHECK-LABEL: neg_0x80000000:
 ; CHECK:       # %bb.0:
-; STATIC-NEXT:   movl $foo-2147483648, %eax
+; STATIC-NEXT:   leaq foo-2147483648(%rip), %rax
 ; PIC-NEXT:      leaq foo-2147483648(%rip), %rax
 ; MSTATIC-NEXT:  movabsq $foo, %rax
 ; MSTATIC-NEXT:  addq $-2147483648, %rax
Index: llvm/lib/Target/X86/X86ISelLowering.h
===================================================================
--- llvm/lib/Target/X86/X86ISelLowering.h
+++ llvm/lib/Target/X86/X86ISelLowering.h
@@ -855,7 +855,8 @@
     /// Returns true of the given offset can be
     /// fit into displacement field of the instruction.
     bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
-                                      bool hasSymbolicDisplacement = true);
+                                      bool hasSymbolicDisplacement,
+                                      bool formsAbsoluteAddress = false);
 
     /// Determines whether the callee is required to pop its
     /// own arguments. Callee pop is necessary to support tail calls.
Index: llvm/lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- llvm/lib/Target/X86/X86ISelLowering.cpp
+++ llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -4872,7 +4872,8 @@
 }
 
 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
-                                       bool hasSymbolicDisplacement) {
+                                       bool hasSymbolicDisplacement,
+                                       bool formsAbsoluteAddress) {
   // Offset should fit into 32 bit immediate field.
   if (!isInt<32>(Offset))
     return false;
@@ -4888,8 +4889,11 @@
 
   // For small code model we assume that latest object is 16MB before end of 31
   // bits boundary. We may also accept pretty large negative constants knowing
-  // that all objects are in the positive half of address space.
-  if (M == CodeModel::Small && Offset < 16*1024*1024)
+  // that all objects are in the positive half of address space, if Offset is
+  // not used to form an absolute address. movl foo-1, %eax is not allowed
+  // because if the address of foo is 0, this fixup is not representable.
+  if (M == CodeModel::Small && Offset < 16 * 1024 * 1024 &&
+      (Offset >= 0 || !formsAbsoluteAddress))
     return true;
 
   // For kernel code model we know that all object resist in the negative half
@@ -19109,7 +19113,7 @@
     // offset into the global address reference. Otherwise, ADD it on later.
     int64_t GlobalOffset = 0;
     if (OpFlags == X86II::MO_NO_FLAG &&
-        X86::isOffsetSuitableForCodeModel(Offset, M)) {
+        X86::isOffsetSuitableForCodeModel(Offset, M, true, true)) {
       std::swap(GlobalOffset, Offset);
     }
     Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, GlobalOffset, OpFlags);


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D93931.314065.patch
Type: text/x-patch
Size: 3522 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201230/dbcfa425/attachment.bin>


More information about the llvm-commits mailing list