Rework global alignment computation again. Now we do round up

alignment of globals to the preferred alignment, but only when
there is no section specified on the global (by far the common
case).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102515 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/CodeGen/AsmPrinter.h b/include/llvm/CodeGen/AsmPrinter.h
index b5c287a..243ddbb 100644 (file)
--- a/include/llvm/CodeGen/AsmPrinter.h
+++ b/include/llvm/CodeGen/AsmPrinter.h
@@ -204,12 +204,8 @@ namespace llvm {
     /// EmitAlignment - Emit an alignment directive to the specified power of
     /// two boundary.  For example, if you pass in 3 here, you will get an 8
     /// byte alignment.  If a global value is specified, and if that global has
-    /// an explicit alignment requested, it will unconditionally override the
-    /// alignment request.
-    ///
-    /// The algorithm is:
-    ///     Align = NumBits;
-    ///     if (GV && GV->hasalignment) Align = GV->getAlignment();
+    /// an explicit alignment requested, it will override the alignment request
+    /// if required for correctness.
     ///
     void EmitAlignment(unsigned NumBits, const GlobalValue *GV = 0) const;
     
@@ -218,7 +214,6 @@ namespace llvm {
     /// it if appropriate.
     void EmitBasicBlockStart(const MachineBasicBlock *MBB) const;
     
-    
     /// EmitGlobalConstant - Print a general LLVM constant to the .s file.
     void EmitGlobalConstant(const Constant *CV, unsigned AddrSpace = 0);
     

diff --git a/lib/CodeGen/AsmPrinter/AsmPrinter.cpp b/lib/CodeGen/AsmPrinter/AsmPrinter.cpp
index 68e70aa..a694033 100644 (file)
--- a/lib/CodeGen/AsmPrinter/AsmPrinter.cpp
+++ b/lib/CodeGen/AsmPrinter/AsmPrinter.cpp
@@ -61,6 +61,35 @@ static gcp_map_type &getGCMap(void *&P) {
 }
 
 
+/// getGVAlignmentLog2 - Return the alignment to use for the specified global
+/// value in log2 form.  This rounds up to the preferred alignment if possible
+/// and legal.
+static unsigned getGVAlignmentLog2(const GlobalValue *GV, const TargetData &TD,
+                                   unsigned InBits = 0) {
+  unsigned NumBits = 0;
+  if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
+    NumBits = TD.getPreferredAlignmentLog(GVar);
+  
+  // If InBits is specified, round it to it.
+  if (InBits > NumBits)
+    NumBits = InBits;
+  
+  // If the GV has a specified alignment, take it into account.
+  if (GV->getAlignment() == 0)
+    return NumBits;
+  
+  unsigned GVAlign = Log2_32(GV->getAlignment());
+  
+  // If the GVAlign is larger than NumBits, or if we are required to obey
+  // NumBits because the GV has an assigned section, obey it.
+  if (GVAlign > NumBits || GV->hasSection())
+    NumBits = GVAlign;
+  return NumBits;
+}
+
+
+
+
 AsmPrinter::AsmPrinter(TargetMachine &tm, MCStreamer &Streamer)
   : MachineFunctionPass(&ID),
     TM(tm), MAI(tm.getMCAsmInfo()),
@@ -227,16 +256,12 @@ void AsmPrinter::EmitGlobalVariable(const GlobalVariable *GV) {
   SectionKind GVKind = TargetLoweringObjectFile::getKindForGlobal(GV, TM);
 
   const TargetData *TD = TM.getTargetData();
-  unsigned Size = TD->getTypeAllocSize(GV->getType()->getElementType());
+  uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
   
   // If the alignment is specified, we *must* obey it.  Overaligning a global
   // with a specified alignment is a prompt way to break globals emitted to
   // sections and expected to be contiguous (e.g. ObjC metadata).
-  unsigned AlignLog;
-  if (unsigned GVAlign = GV->getAlignment())
-    AlignLog = Log2_32(GVAlign);
-  else
-    AlignLog = TD->getPreferredAlignmentLog(GV);
+  unsigned AlignLog = getGVAlignmentLog2(GV, *TD);
   
   // Handle common and BSS local symbols (.lcomm).
   if (GVKind.isCommon() || GVKind.isBSSLocal()) {
@@ -1010,7 +1035,7 @@ bool AsmPrinter::EmitSpecialLLVMGlobal(const GlobalVariable *GV) {
   unsigned Align = Log2_32(TD->getPointerPrefAlignment());
   if (GV->getName() == "llvm.global_ctors") {
     OutStreamer.SwitchSection(getObjFileLowering().getStaticCtorSection());
-    EmitAlignment(Align, 0);
+    EmitAlignment(Align);
     EmitXXStructorList(GV->getInitializer());
     
     if (TM.getRelocationModel() == Reloc::Static &&
@@ -1024,7 +1049,7 @@ bool AsmPrinter::EmitSpecialLLVMGlobal(const GlobalVariable *GV) {
   
   if (GV->getName() == "llvm.global_dtors") {
     OutStreamer.SwitchSection(getObjFileLowering().getStaticDtorSection());
-    EmitAlignment(Align, 0);
+    EmitAlignment(Align);
     EmitXXStructorList(GV->getInitializer());
 
     if (TM.getRelocationModel() == Reloc::Static &&
@@ -1153,20 +1178,13 @@ void AsmPrinter::EmitLabelOffsetDifference(const MCSymbol *Hi, uint64_t Offset,
 // EmitAlignment - Emit an alignment directive to the specified power of
 // two boundary.  For example, if you pass in 3 here, you will get an 8
 // byte alignment.  If a global value is specified, and if that global has
-// an explicit alignment requested, it will unconditionally override the
-// alignment request.  However, if ForcedAlignBits is specified, this value
-// has final say: the ultimate alignment will be the max of ForcedAlignBits
-// and the alignment computed with NumBits and the global.
-//
-// The algorithm is:
-//     Align = NumBits;
-//     if (GV && GV->hasalignment) Align = GV->getAlignment();
+// an explicit alignment requested, it will override the alignment request
+// if required for correctness.
 //
 void AsmPrinter::EmitAlignment(unsigned NumBits, const GlobalValue *GV) const {
-  if (GV && GV->getAlignment())
-    NumBits = Log2_32(GV->getAlignment());
+  if (GV) NumBits = getGVAlignmentLog2(GV, *TM.getTargetData(), NumBits);
   
-  if (NumBits == 0) return;   // No need to emit alignment.
+  if (NumBits == 0) return;   // 1-byte aligned: no need to emit alignment.
   
   if (getCurrentSection()->getKind().isText())
     OutStreamer.EmitCodeAlignment(1 << NumBits);

diff --git a/test/CodeGen/X86/alignment.ll b/test/CodeGen/X86/alignment.ll
index 06a4f3f..9678e6d 100644 (file)
--- a/test/CodeGen/X86/alignment.ll
+++ b/test/CodeGen/X86/alignment.ll
@@ -6,7 +6,7 @@
 
 ; CHECK:       .bss
 ; CHECK:       .globl  GlobalA
-; CHECK:       .align  8
+; CHECK:       .align  16
 ; CHECK: GlobalA:
 ; CHECK:       .zero   384
 
@@ -15,4 +15,29 @@
 ; PR6921
 @GlobalB = common global { [384 x i8] } zeroinitializer, align 8
 
-; CHECK:       .comm   GlobalB,384,8
\ No newline at end of file
+; CHECK:       .comm   GlobalB,384,16
+
+
+@GlobalC = common global { [384 x i8] } zeroinitializer, align 2
+
+; CHECK:       .comm   GlobalC,384,16
+
+
+
+; This cannot get rounded up to the preferred alignment (16) if they have an
+; explicit alignment specified *and* a section specified.
+@GlobalAS = global { [384 x i8] } zeroinitializer, align 8, section "foo"
+
+; CHECK:       .globl  GlobalAS
+; CHECK:       .align  8
+; CHECK: GlobalAS:
+; CHECK:       .zero   384
+
+; Common variables should not get rounded up to the preferred alignment (16) if
+; they have an explicit alignment specified and a section specified.
+; PR6921
+@GlobalBS = common global { [384 x i8] } zeroinitializer, align 8, section "foo"
+; CHECK:       .comm   GlobalBS,384,8
+
+@GlobalCS = common global { [384 x i8] } zeroinitializer, align 2, section "foo"
+; CHECK:       .comm   GlobalCS,384,2
\ No newline at end of file

diff --git a/test/CodeGen/X86/unaligned-load.ll b/test/CodeGen/X86/unaligned-load.ll
index b26097f..a99af06 100644 (file)
--- a/test/CodeGen/X86/unaligned-load.ll
+++ b/test/CodeGen/X86/unaligned-load.ll
@@ -29,8 +29,8 @@ return:
 declare void @llvm.memcpy.i64(i8* nocapture, i8* nocapture, i64, i32) nounwind
 
 ; CORE2: .section
-; CORE2: .align  3
+; CORE2: .align  4
 ; CORE2-NEXT: _.str1:
 ; CORE2-NEXT: .asciz "DHRYSTONE PROGRAM, SOME STRING"
-; CORE2: .align 3
+; CORE2: .align 4
 ; CORE2-NEXT: _.str3:

diff --git a/test/FrontendC/cstring-align.c b/test/FrontendC/cstring-align.c
index 715d0f3..b9ec281 100644 (file)
--- a/test/FrontendC/cstring-align.c
+++ b/test/FrontendC/cstring-align.c
@@ -1,6 +1,5 @@
 // RUN: %llvmgcc %s -c -Os -m32 -emit-llvm -o - | llc -march=x86 -mtriple=i386-apple-darwin10 | FileCheck %s -check-prefix=DARWIN32
 // RUN: %llvmgcc %s -c -Os -m64 -emit-llvm -o - | llc -march=x86-64 -mtriple=x86_64-apple-darwin10 | FileCheck %s -check-prefix=DARWIN64
-// XFAIL: *
 // XTARGET: darwin
 
 extern void func(const char *, const char *);
@@ -9,10 +8,10 @@ void long_function_name() {
   func("%s: the function name", __func__);
 }
 
-// DARWIN64: .align 3
+// DARWIN64: .align 4
 // DARWIN64: ___func__.
 // DARWIN64: .asciz "long_function_name"
 
-// DARWIN32: .align 2
+// DARWIN32: .align 4
 // DARWIN32: ___func__.
 // DARWIN32: .asciz "long_function_name"