Fix a number of byval / memcpy / memset related codegen issues.

1. x86-64 byval alignment should be max of 8 and alignment of type. Previously the code was not doing what the commit message was saying.
2. Do not use byte repeat move and store operations. These are slow.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55139 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/X86/X86ISelLowering.cpp

@@ -796,8 +796,14 @@ static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
 /// are at 4-byte boundaries.
 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
-  if (Subtarget->is64Bit())
+  if (Subtarget->is64Bit()) {
-    return getTargetData()->getABITypeAlignment(Ty);
+    // Max of 8 and alignment of type.
+    unsigned TyAlign = getTargetData()->getABITypeAlignment(Ty);
+    if (TyAlign > 8)
+      return TyAlign;
+    return 8;
+  }
+
   unsigned Align = 4;
   if (Subtarget->hasSSE1())
     getMaxByValAlign(Ty, Align);
@@ -5014,16 +5020,16 @@ X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
 
 SDValue
 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
-                                           SDValue Chain,
+                                        SDValue Chain,
-                                           SDValue Dst, SDValue Src,
+                                        SDValue Dst, SDValue Src,
-                                           SDValue Size, unsigned Align,
+                                        SDValue Size, unsigned Align,
                                         const Value *DstSV, uint64_t DstSVOff) {
   ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
 
   /// If not DWORD aligned or size is more than the threshold, call the library.
   /// The libc version is likely to be faster for these cases. It can use the
   /// address value and run time information about the CPU.
-  if ((Align & 3) == 0 ||
+  if ((Align & 3) != 0 ||
       !ConstantSize ||
       ConstantSize->getValue() > getSubtarget()->getMaxInlineSizeThreshold()) {
     SDValue InFlag(0, 0);
@@ -5065,27 +5071,27 @@ X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
 
     // If the value is a constant, then we can potentially use larger sets.
     switch (Align & 3) {
-      case 2:   // WORD aligned
+    case 2:   // WORD aligned
-        AVT = MVT::i16;
+      AVT = MVT::i16;
-        ValReg = X86::AX;
+      ValReg = X86::AX;
-        Val = (Val << 8) | Val;
+      Val = (Val << 8) | Val;
-        break;
+      break;
-      case 0:  // DWORD aligned
+    case 0:  // DWORD aligned
-        AVT = MVT::i32;
+      AVT = MVT::i32;
-        ValReg = X86::EAX;
+      ValReg = X86::EAX;
-        Val = (Val << 8)  | Val;
+      Val = (Val << 8)  | Val;
-        Val = (Val << 16) | Val;
+      Val = (Val << 16) | Val;
-        if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) {  // QWORD aligned
+      if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) {  // QWORD aligned
-          AVT = MVT::i64;
+        AVT = MVT::i64;
-          ValReg = X86::RAX;
+        ValReg = X86::RAX;
-          Val = (Val << 32) | Val;
+        Val = (Val << 32) | Val;
-        }
+      }
-        break;
+      break;
-      default:  // Byte aligned
+    default:  // Byte aligned
-        AVT = MVT::i8;
+      AVT = MVT::i8;
-        ValReg = X86::AL;
+      ValReg = X86::AL;
-        Count = DAG.getIntPtrConstant(SizeVal);
+      Count = DAG.getIntPtrConstant(SizeVal);
-        break;
+      break;
     }
 
     if (AVT.bitsGT(MVT::i8)) {
@@ -5153,13 +5159,11 @@ X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
 
 SDValue
 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
-                                           SDValue Chain,
+                                      SDValue Chain, SDValue Dst, SDValue Src,
-                                           SDValue Dst, SDValue Src,
+                                      SDValue Size, unsigned Align,
-                                           SDValue Size, unsigned Align,
+                                      bool AlwaysInline,
-                                           bool AlwaysInline,
+                                      const Value *DstSV, uint64_t DstSVOff,
-                                           const Value *DstSV, uint64_t DstSVOff,
+                                      const Value *SrcSV, uint64_t SrcSVOff) {  
-                                           const Value *SrcSV, uint64_t SrcSVOff){
-  
   // This requires the copy size to be a constant, preferrably
   // within a subtarget-specific limit.
   ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
@@ -5169,21 +5173,19 @@ X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
   if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
     return SDValue();
 
-  MVT AVT;
+  /// If not DWORD aligned, call the library.
-  unsigned BytesLeft = 0;
+  if ((Align & 3) != 0)
-  if (Align >= 8 && Subtarget->is64Bit())
+    return SDValue();
+
+  // DWORD aligned
+  MVT AVT = MVT::i32;
+  if (Subtarget->is64Bit() && ((Align & 0x7) == 0))  // QWORD aligned
     AVT = MVT::i64;
-  else if (Align >= 4)
-    AVT = MVT::i32;
-  else if (Align >= 2)
-    AVT = MVT::i16;
-  else
-    AVT = MVT::i8;
 
   unsigned UBytes = AVT.getSizeInBits() / 8;
   unsigned CountVal = SizeVal / UBytes;
   SDValue Count = DAG.getIntPtrConstant(CountVal);
-  BytesLeft = SizeVal % UBytes;
+  unsigned BytesLeft = SizeVal % UBytes;
 
   SDValue InFlag(0, 0);
   Chain  = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,

test/CodeGen/X86/2004-02-12-Memcpy.ll

@@ -1,11 +1,11 @@
-; RUN: llvm-as < %s | llc -march=x86 -mtriple=i686-pc-linux-gnu | grep movs | count 3
+; RUN: llvm-as < %s | llc -march=x86 -mtriple=i686-pc-linux-gnu | grep movs | count 1
 
 @A = global [32 x i32] zeroinitializer
 @B = global [32 x i32] zeroinitializer
 
 declare void @llvm.memcpy.i32(i8*, i8*, i32, i32)
 
-define void @main() {
+define void @main() nounwind {
   ; dword copy
   call void @llvm.memcpy.i32(i8* bitcast ([32 x i32]* @A to i8*),
                            i8* bitcast ([32 x i32]* @B to i8*),

test/CodeGen/X86/byval3.ll

@@ -1,4 +1,4 @@
-; RUN: llvm-as < %s | llc -march=x86-64 | grep rep.movsl | count 2
+; RUN: llvm-as < %s | llc -march=x86-64 | grep rep.movsq | count 2
 ; RUN: llvm-as < %s | llc -march=x86 | grep rep.movsl | count 2
 
 %struct.s = type { i32, i32, i32, i32, i32, i32, i32, i32,
@@ -7,7 +7,7 @@
                    i32, i32, i32, i32, i32, i32, i32, i32,
                    i32 }
 
-define void @g(i32 %a1, i32 %a2, i32 %a3, i32 %a4, i32 %a5, i32 %a6) {
+define void @g(i32 %a1, i32 %a2, i32 %a3, i32 %a4, i32 %a5, i32 %a6) nounwind {
 entry:
         %d = alloca %struct.s, align 16
         %tmp = getelementptr %struct.s* %d, i32 0, i32 0

test/CodeGen/X86/byval4.ll

@@ -1,4 +1,4 @@
-; RUN: llvm-as < %s | llc -march=x86-64 | grep rep.movsw | count 2
+; RUN: llvm-as < %s | llc -march=x86-64 | grep rep.movsq | count 2
 ; RUN: llvm-as < %s | llc -march=x86 | grep rep.movsl	 | count 2
 
 %struct.s = type { i16, i16, i16, i16, i16, i16, i16, i16,
@@ -13,7 +13,7 @@
 
 
 define void @g(i16 signext  %a1, i16 signext  %a2, i16 signext  %a3,
-	 i16 signext  %a4, i16 signext  %a5, i16 signext  %a6) {
+	 i16 signext  %a4, i16 signext  %a5, i16 signext  %a6) nounwind {
 entry:
         %a = alloca %struct.s, align 16
         %tmp = getelementptr %struct.s* %a, i32 0, i32 0

test/CodeGen/X86/byval5.ll

@@ -1,4 +1,4 @@
-; RUN: llvm-as < %s | llc -march=x86-64 | grep rep.movsb | count 2
+; RUN: llvm-as < %s | llc -march=x86-64 | grep rep.movsq | count 2
 ; RUN: llvm-as < %s | llc -march=x86 | grep rep.movsl	 | count 2
 
 %struct.s = type { i8, i8, i8, i8, i8, i8, i8, i8,

test/CodeGen/X86/memset-2.ll

@@ -0,0 +1,45 @@
+; RUN: llvm-as < %s | llc -march=x86 | not grep rep
+; RUN: llvm-as < %s | llc -march=x86 | grep memset
+
+declare void @llvm.memset.i32(i8*, i8, i32, i32) nounwind
+
+define fastcc i32 @cli_scanzip(i32 %desc) nounwind {
+entry:
+	br label %bb8.i.i.i.i
+
+bb8.i.i.i.i:		; preds = %bb8.i.i.i.i, %entry
+	icmp eq i32 0, 0		; <i1>:0 [#uses=1]
+	br i1 %0, label %bb61.i.i.i, label %bb8.i.i.i.i
+
+bb32.i.i.i:		; preds = %bb61.i.i.i
+	ptrtoint i8* %tail.0.i.i.i to i32		; <i32>:1 [#uses=1]
+	sub i32 0, %1		; <i32>:2 [#uses=1]
+	icmp sgt i32 %2, 19		; <i1>:3 [#uses=1]
+	br i1 %3, label %bb34.i.i.i, label %bb61.i.i.i
+
+bb34.i.i.i:		; preds = %bb32.i.i.i
+	load i32* null, align 4		; <i32>:4 [#uses=1]
+	icmp eq i32 %4, 101010256		; <i1>:5 [#uses=1]
+	br i1 %5, label %bb8.i11.i.i.i, label %bb61.i.i.i
+
+bb8.i11.i.i.i:		; preds = %bb8.i11.i.i.i, %bb34.i.i.i
+	icmp eq i32 0, 0		; <i1>:6 [#uses=1]
+	br i1 %6, label %cli_dbgmsg.exit49.i, label %bb8.i11.i.i.i
+
+cli_dbgmsg.exit49.i:		; preds = %bb8.i11.i.i.i
+	icmp eq [32768 x i8]* null, null		; <i1>:7 [#uses=1]
+	br i1 %7, label %bb1.i28.i, label %bb8.i.i
+
+bb61.i.i.i:		; preds = %bb61.i.i.i, %bb34.i.i.i, %bb32.i.i.i, %bb8.i.i.i.i
+	%tail.0.i.i.i = getelementptr [1024 x i8]* null, i32 0, i32 0		; <i8*> [#uses=2]
+	load i8* %tail.0.i.i.i, align 1		; <i8>:8 [#uses=1]
+	icmp eq i8 %8, 80		; <i1>:9 [#uses=1]
+	br i1 %9, label %bb32.i.i.i, label %bb61.i.i.i
+
+bb1.i28.i:		; preds = %cli_dbgmsg.exit49.i
+	call void @llvm.memset.i32( i8* null, i8 0, i32 88, i32 1 ) nounwind
+	unreachable
+
+bb8.i.i:		; preds = %bb8.i.i, %cli_dbgmsg.exit49.i
+	br label %bb8.i.i
+}

test/CodeGen/X86/memset64-on-x86-32.ll

@@ -1,12 +1,10 @@
-; RUN: llvm-as < %s | llc -march=x86 | grep stosb
+; RUN: llvm-as < %s | llc -mtriple=i386-apple-darwin | grep stosl
+; RUN: llvm-as < %s | llc -mtriple=x86_64-apple-darwin | grep movq | count 10
 
-target triple = "i386-apple-darwin9"
+define void @bork() nounwind {
-        %struct.S = type { [80 x i8] }
-
-define %struct.S* @bork() {
 entry:
-        call void @llvm.memset.i64( i8* null, i8 0, i64 80, i32 1 )
+        call void @llvm.memset.i64( i8* null, i8 0, i64 80, i32 4 )
-        ret %struct.S* null
+        ret void
 }
 
 declare void @llvm.memset.i64(i8*, i8, i64, i32) nounwind