In Thumb2 mode, lower frame indix references to:

add <rd>, sp, #<imm8>
ldr <rd>, [sp, #<imm8>]
When the offset from sp is multiple of 4 and in range of 0-1020.
This saves code size by utilizing 16-bit instructions.

rdar://9321541

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@129971 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMFrameLowering.cpp b/lib/Target/ARM/ARMFrameLowering.cpp
index c99259df7addf8737ab2124e2377aed2d7566bae..e2e95d47b37b90b2d3cdd595f8759651d0a185db 100644 (file)
--- a/lib/Target/ARM/ARMFrameLowering.cpp
+++ b/lib/Target/ARM/ARMFrameLowering.cpp
@@ -445,8 +445,7 @@ ARMFrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
 
 int
 ARMFrameLowering::ResolveFrameIndexReference(const MachineFunction &MF,
-                                             int FI,
-                                             unsigned &FrameReg,
+                                             int FI, unsigned &FrameReg,
                                              int SPAdj) const {
   const MachineFrameInfo *MFI = MF.getFrameInfo();
   const ARMBaseRegisterInfo *RegInfo =
@@ -490,19 +489,23 @@ ARMFrameLowering::ResolveFrameIndexReference(const MachineFunction &MF,
       return FPOffset;
     } else if (MFI->hasVarSizedObjects()) {
       assert(RegInfo->hasBasePointer(MF) && "missing base pointer!");
-      // Try to use the frame pointer if we can, else use the base pointer
-      // since it's available. This is handy for the emergency spill slot, in
-      // particular.
       if (AFI->isThumb2Function()) {
+        // Try to use the frame pointer if we can, else use the base pointer
+        // since it's available. This is handy for the emergency spill slot, in
+        // particular.
         if (FPOffset >= -255 && FPOffset < 0) {
           FrameReg = RegInfo->getFrameRegister(MF);
           return FPOffset;
         }
-      } else
-        FrameReg = RegInfo->getBaseRegister();
+      }
     } else if (AFI->isThumb2Function()) {
+      // Use  add <rd>, sp, #<imm8> 
+      //      ldr <rd>, [sp, #<imm8>]
+      // if at all possible to save space.
+      if (Offset >= 0 && (Offset & 3) == 0 && Offset <= 1020)
+        return Offset;
       // In Thumb2 mode, the negative offset is very limited. Try to avoid
-      // out of range references.
+      // out of range references. ldr <rt>,[<rn>, #-<imm8>]
       if (FPOffset >= -255 && FPOffset < 0) {
         FrameReg = RegInfo->getFrameRegister(MF);
         return FPOffset;

diff --git a/lib/Target/ARM/ARMFrameLowering.h b/lib/Target/ARM/ARMFrameLowering.h
index a7b7f15449de08ef8630916f0b037a02b41f5a2f..61bb8afa40f25d5bf1eca213911f28b24ab4acaf 100644 (file)
--- a/lib/Target/ARM/ARMFrameLowering.h
+++ b/lib/Target/ARM/ARMFrameLowering.h
@@ -51,7 +51,8 @@ public:
   bool canSimplifyCallFramePseudos(const MachineFunction &MF) const;
   int getFrameIndexReference(const MachineFunction &MF, int FI,
                              unsigned &FrameReg) const;
-  int ResolveFrameIndexReference(const MachineFunction &MF, int FI,
+  int ResolveFrameIndexReference(const MachineFunction &MF,
+                                 int FI,
                                  unsigned &FrameReg, int SPAdj) const;
   int getFrameIndexOffset(const MachineFunction &MF, int FI) const;
 

diff --git a/test/CodeGen/Thumb2/2011-04-21-FILoweringBug.ll b/test/CodeGen/Thumb2/2011-04-21-FILoweringBug.ll
new file mode 100644 (file)
index 0000000..604a352
--- /dev/null
+++ b/test/CodeGen/Thumb2/2011-04-21-FILoweringBug.ll
@@ -0,0 +1,23 @@
+; RUN: llc < %s -mtriple=thumbv7-apple-darwin | FileCheck %s
+
+; Use sp, #imm to lower frame indices when the offset is multiple of 4
+; and in the range of 0-1020. This saves code size by utilizing
+; 16-bit instructions.
+; rdar://9321541
+
+define i32 @t() nounwind {
+entry:
+; CHECK: t:
+; CHECK: sub sp, #12
+; CHECK-NOT: sub
+; CHECK: add r0, sp, #4
+; CHECK: add r1, sp, #8
+; CHECK: mov r2, sp
+  %size = alloca i32, align 4
+  %count = alloca i32, align 4
+  %index = alloca i32, align 4
+  %0 = call i32 @foo(i32* %count, i32* %size, i32* %index) nounwind
+  ret i32 %0
+}
+
+declare i32 @foo(i32*, i32*, i32*)