r105228 reduced the memcpy / memset inline limit to 4 with -Os to avoid blowing

up freebsd bootloader. However, this doesn't make much sense for Darwin, whose
-Os is meant to optimize for size only if it doesn't hurt performance.
rdar://8821501


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122936 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 57 insertions, 26 deletions

diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 110812c437..be2de0c00d 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -50,6 +50,7 @@
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringExtras.h"
+#include "llvm/ADT/Triple.h"
 #include <algorithm>
 #include <cmath>
 using namespace llvm;
@@ -3286,8 +3287,14 @@ static bool FindOptimalMemOpLowering(std::vector<EVT> &MemOps,
   // the size of a call to memcpy or memset (3 arguments + call).
   if (Limit != ~0U) {
     const Function *F = DAG.getMachineFunction().getFunction();
-    if (F->hasFnAttr(Attribute::OptimizeForSize))
-      Limit = 4;
+    if (F->hasFnAttr(Attribute::OptimizeForSize)) {
+      Triple T(((LLVMTargetMachine&)TLI.getTargetMachine()).getTargetTriple());
+      if (T.getOS() != Triple::Darwin)
+        // A pretty terrible hack to defat the wild guess. On Darwin, -Os means
+        // optimize for size without hurting performance so we don't want to
+        // bump down the limit.
+        Limit = 4;
+    }
   }
 
   unsigned NumMemOps = 0;
diff --git a/test/CodeGen/X86/2010-06-25-CoalescerSubRegDefDead.ll b/test/CodeGen/X86/2010-06-25-CoalescerSubRegDefDead.ll
index c6421a247e..6db3ce1f42 100644
--- a/test/CodeGen/X86/2010-06-25-CoalescerSubRegDefDead.ll
+++ b/test/CodeGen/X86/2010-06-25-CoalescerSubRegDefDead.ll
@@ -1,4 +1,4 @@
-; RUN: llc -O1 -mtriple=x86_64-apple-darwin10 -relocation-model=pic -disable-fp-elim < %s | FileCheck %s
+; RUN: llc -O1 -mtriple=x86_64-unknown-linux-gnu -relocation-model=pic -disable-fp-elim < %s | FileCheck %s
 ; <rdar://problem/8124405>
 
 %struct.type = type { %struct.subtype*, i32, i8, i32, i8, i32, i32, i32, i32, i32, i8, i32, i32, i32, i32, i32, [256 x i32], i32, [257 x i32], [257 x i32], i32*, i16*, i8*, i32, i32, i32, i32, i32, [256 x i8], [16 x i8], [256 x i8], [4096 x i8], [16 x i32], [18002 x i8], [18002 x i8], [6 x [258 x i8]], [6 x [258 x i32]], [6 x [258 x i32]], [6 x [258 x i32]], [6 x i32], i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32*, i32*, i32* }
@@ -21,9 +21,9 @@ bb:
 ; statement. It can be an ADD or LEA instruction, it's not important which one
 ; it is.
 ;
-;      CHECK: ## %bb
-; CHECK-NEXT: addq $64036, %rdi
-;      CHECK: rep;stosl
+; CHECK: # %bb
+; CHECK: addq $64036, %rdi
+; CHECK: rep;stosl
 
   %tmp5 = bitcast i32* %tmp4 to i8*
   call void @llvm.memset.p0i8.i64(i8* %tmp5, i8 0, i64 84, i32 4, i1 false)
diff --git a/test/CodeGen/X86/memcpy.ll b/test/CodeGen/X86/memcpy.ll
index 7bc31bec16..72342cbacb 100644
--- a/test/CodeGen/X86/memcpy.ll
+++ b/test/CodeGen/X86/memcpy.ll
@@ -1,4 +1,5 @@
-; RUN: llc < %s -march=x86-64 | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu | FileCheck %s -check-prefix=LINUX
+; RUN: llc < %s -mtriple=x86_64-apple-darwin | FileCheck %s -check-prefix=DARWIN
 
 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i32, i1) nounwind
 
@@ -9,8 +10,8 @@ entry:
 	tail call void @llvm.memcpy.p0i8.p0i8.i64( i8* %a, i8* %b, i64 %n, i32 1, i1 0 )
 	ret i8* %a
         
-; CHECK: test1:
-; CHECK: memcpy
+; LINUX: test1:
+; LINUX: memcpy
 }
 
 ; Variable memcpy's should lower to calls.
@@ -21,18 +22,41 @@ entry:
 	tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %tmp14, i8* %tmp25, i64 %n, i32 8, i1 0 )
 	ret i8* %tmp14
         
-; CHECK: test2:
-; CHECK: memcpy
+; LINUX: test2:
+; LINUX: memcpy
 }
 
 ; Large constant memcpy's should lower to a call when optimizing for size.
 ; PR6623
+
+; On the other hand, Darwin's definition of -Os is optimizing for size without
+; hurting performance so it should just ignore optsize when expanding memcpy.
+; rdar://8821501
 define void @test3(i8* nocapture %A, i8* nocapture %B) nounwind optsize noredzone {
 entry:
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %A, i8* %B, i64 64, i32 1, i1 false)
   ret void
-; CHECK: test3:
-; CHECK: memcpy
+; LINUX: test3:
+; LINUX: memcpy
+
+; DARWIN: test3:
+; DARWIN-NOT: memcpy
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
+; DARWIN: movq
 }
 
 ; Large constant memcpy's should be inlined when not optimizing for size.
@@ -40,18 +64,18 @@ define void @test4(i8* nocapture %A, i8* nocapture %B) nounwind noredzone {
 entry:
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %A, i8* %B, i64 64, i32 1, i1 false)
   ret void
-; CHECK: test4:
-; CHECK: movq
-; CHECK: movq
-; CHECK: movq
-; CHECK: movq
-; CHECK: movq
-; CHECK: movq
-; CHECK: movq
-; CHECK: movq
-; CHECK: movq
-; CHECK: movq
-; CHECK: movq
-; CHECK: movq
+; LINUX: test4:
+; LINUX movq
+; LINUX movq
+; LINUX movq
+; LINUX movq
+; LINUX movq
+; LINUX movq
+; LINUX movq
+; LINUX movq
+; LINUX movq
+; LINUX movq
+; LINUX movq
+; LINUX movq
 }