[arm fast-isel] Make the fast-isel implementation of memcpy respect alignment.

rdar://12821569

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@169460 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 116 insertions, 14 deletions

diff --git a/lib/Target/ARM/ARMFastISel.cpp b/lib/Target/ARM/ARMFastISel.cpp
index 23d372ccea..21bb147644 100644
--- a/lib/Target/ARM/ARMFastISel.cpp
+++ b/lib/Target/ARM/ARMFastISel.cpp
@@ -186,7 +186,8 @@ class ARMFastISel : public FastISel {
     bool ARMComputeAddress(const Value *Obj, Address &Addr);
     void ARMSimplifyAddress(Address &Addr, EVT VT, bool useAM3);
     bool ARMIsMemCpySmall(uint64_t Len);
-    bool ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len);
+    bool ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
+                               unsigned Alignment);
     unsigned ARMEmitIntExt(EVT SrcVT, unsigned SrcReg, EVT DestVT, bool isZExt);
     unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
     unsigned ARMMaterializeInt(const Constant *C, EVT VT);
@@ -2422,21 +2423,30 @@ bool ARMFastISel::ARMIsMemCpySmall(uint64_t Len) {
 }
 
 bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src,
-                                        uint64_t Len) {
+                                        uint64_t Len, unsigned Alignment) {
   // Make sure we don't bloat code by inlining very large memcpy's.
   if (!ARMIsMemCpySmall(Len))
     return false;
 
-  // We don't care about alignment here since we just emit integer accesses.
   while (Len) {
     MVT VT;
-    if (Len >= 4)
-      VT = MVT::i32;
-    else if (Len >= 2)
-      VT = MVT::i16;
-    else {
-      assert(Len == 1);
-      VT = MVT::i8;
+    if (!Alignment || Alignment >= 4) {
+      if (Len >= 4)
+        VT = MVT::i32;
+      else if (Len >= 2)
+        VT = MVT::i16;
+      else {
+        assert (Len == 1 && "Expected a length of 1!");
+        VT = MVT::i8;
+      }
+    } else {
+      // Bound based on alignment.
+      if (Len >= 2 && Alignment == 2)
+        VT = MVT::i16;
+      else {
+        assert (Alignment == 1 && "Expected an alignment of 1!");
+        VT = MVT::i8;
+      }
     }
 
     bool RV;
@@ -2515,7 +2525,8 @@ bool ARMFastISel::SelectIntrinsicCall(const IntrinsicInst &I) {
         if (!ARMComputeAddress(MTI.getRawDest(), Dest) ||
             !ARMComputeAddress(MTI.getRawSource(), Src))
           return false;
-        if (ARMTryEmitSmallMemCpy(Dest, Src, Len))
+        unsigned Alignment = MTI.getAlignment();
+        if (ARMTryEmitSmallMemCpy(Dest, Src, Len, Alignment))
           return true;
       }
     }
diff --git a/test/CodeGen/ARM/fast-isel-intrinsic.ll b/test/CodeGen/ARM/fast-isel-intrinsic.ll
index b73fceff6c..7d38cc2a7f 100644
--- a/test/CodeGen/ARM/fast-isel-intrinsic.ll
+++ b/test/CodeGen/ARM/fast-isel-intrinsic.ll
@@ -35,7 +35,7 @@ define void @t1() nounwind ssp {
 ; THUMB-LONG: movt r3, :upper16:L_memset$non_lazy_ptr
 ; THUMB-LONG: ldr r3, [r3]
 ; THUMB-LONG: blx r3
-  call void @llvm.memset.p0i8.i32(i8* getelementptr inbounds ([60 x i8]* @message1, i32 0, i32 5), i8 64, i32 10, i32 1, i1 false)
+  call void @llvm.memset.p0i8.i32(i8* getelementptr inbounds ([60 x i8]* @message1, i32 0, i32 5), i8 64, i32 10, i32 4, i1 false)
   ret void
 }
 
@@ -73,7 +73,7 @@ define void @t2() nounwind ssp {
 ; THUMB-LONG: movt r3, :upper16:L_memcpy$non_lazy_ptr
 ; THUMB-LONG: ldr r3, [r3]
 ; THUMB-LONG: blx r3
-  call void @llvm.memcpy.p0i8.p0i8.i32(i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 4), i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 16), i32 17, i32 1, i1 false)
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 4), i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 16), i32 17, i32 4, i1 false)
   ret void
 }
 
@@ -125,6 +125,7 @@ define void @t4() nounwind ssp {
 ; ARM: ldrh r1, [r0, #24]
 ; ARM: strh r1, [r0, #12]
 ; ARM: bx lr
+; THUMB: t4
 ; THUMB: movw r0, :lower16:L_temp$non_lazy_ptr
 ; THUMB: movt r0, :upper16:L_temp$non_lazy_ptr
 ; THUMB: ldr r0, [r0]
@@ -135,8 +136,98 @@ define void @t4() nounwind ssp {
 ; THUMB: ldrh r1, [r0, #24]
 ; THUMB: strh r1, [r0, #12]
 ; THUMB: bx lr
-  call void @llvm.memcpy.p0i8.p0i8.i32(i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 4), i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 16), i32 10, i32 1, i1 false)
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 4), i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 16), i32 10, i32 4, i1 false)
   ret void
 }
 
 declare void @llvm.memmove.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind
+
+define void @t5() nounwind ssp {
+; ARM: t5
+; ARM: movw r0, :lower16:L_temp$non_lazy_ptr
+; ARM: movt r0, :upper16:L_temp$non_lazy_ptr
+; ARM: ldr r0, [r0]
+; ARM: ldrh r1, [r0, #16]
+; ARM: strh r1, [r0, #4]
+; ARM: ldrh r1, [r0, #18]
+; ARM: strh r1, [r0, #6]
+; ARM: ldrh r1, [r0, #20]
+; ARM: strh r1, [r0, #8]
+; ARM: ldrh r1, [r0, #22]
+; ARM: strh r1, [r0, #10]
+; ARM: ldrh r1, [r0, #24]
+; ARM: strh r1, [r0, #12]
+; ARM: bx lr
+; THUMB: t5
+; THUMB: movw r0, :lower16:L_temp$non_lazy_ptr
+; THUMB: movt r0, :upper16:L_temp$non_lazy_ptr
+; THUMB: ldr r0, [r0]
+; THUMB: ldrh r1, [r0, #16]
+; THUMB: strh r1, [r0, #4]
+; THUMB: ldrh r1, [r0, #18]
+; THUMB: strh r1, [r0, #6]
+; THUMB: ldrh r1, [r0, #20]
+; THUMB: strh r1, [r0, #8]
+; THUMB: ldrh r1, [r0, #22]
+; THUMB: strh r1, [r0, #10]
+; THUMB: ldrh r1, [r0, #24]
+; THUMB: strh r1, [r0, #12]
+; THUMB: bx lr
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 4), i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 16), i32 10, i32 2, i1 false)
+  ret void
+}
+
+define void @t6() nounwind ssp {
+; ARM: t6
+; ARM: movw r0, :lower16:L_temp$non_lazy_ptr
+; ARM: movt r0, :upper16:L_temp$non_lazy_ptr
+; ARM: ldr r0, [r0]
+; ARM: ldrb r1, [r0, #16]
+; ARM: strb r1, [r0, #4]
+; ARM: ldrb r1, [r0, #17]
+; ARM: strb r1, [r0, #5]
+; ARM: ldrb r1, [r0, #18]
+; ARM: strb r1, [r0, #6]
+; ARM: ldrb r1, [r0, #19]
+; ARM: strb r1, [r0, #7]
+; ARM: ldrb r1, [r0, #20]
+; ARM: strb r1, [r0, #8]
+; ARM: ldrb r1, [r0, #21]
+; ARM: strb r1, [r0, #9]
+; ARM: ldrb r1, [r0, #22]
+; ARM: strb r1, [r0, #10]
+; ARM: ldrb r1, [r0, #23]
+; ARM: strb r1, [r0, #11]
+; ARM: ldrb r1, [r0, #24]
+; ARM: strb r1, [r0, #12]
+; ARM: ldrb r1, [r0, #25]
+; ARM: strb r1, [r0, #13]
+; ARM: bx lr
+; THUMB: t6
+; THUMB: movw r0, :lower16:L_temp$non_lazy_ptr
+; THUMB: movt r0, :upper16:L_temp$non_lazy_ptr
+; THUMB: ldr r0, [r0]
+; THUMB: ldrb r1, [r0, #16]
+; THUMB: strb r1, [r0, #4]
+; THUMB: ldrb r1, [r0, #17]
+; THUMB: strb r1, [r0, #5]
+; THUMB: ldrb r1, [r0, #18]
+; THUMB: strb r1, [r0, #6]
+; THUMB: ldrb r1, [r0, #19]
+; THUMB: strb r1, [r0, #7]
+; THUMB: ldrb r1, [r0, #20]
+; THUMB: strb r1, [r0, #8]
+; THUMB: ldrb r1, [r0, #21]
+; THUMB: strb r1, [r0, #9]
+; THUMB: ldrb r1, [r0, #22]
+; THUMB: strb r1, [r0, #10]
+; THUMB: ldrb r1, [r0, #23]
+; THUMB: strb r1, [r0, #11]
+; THUMB: ldrb r1, [r0, #24]
+; THUMB: strb r1, [r0, #12]
+; THUMB: ldrb r1, [r0, #25]
+; THUMB: strb r1, [r0, #13]
+; THUMB: bx lr
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 4), i8* getelementptr inbounds ([60 x i8]* @temp, i32 0, i32 16), i32 10, i32 1, i1 false)
+  ret void
+}