Set alignment operand for NEON VST instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@114709 91177308-0d34-0410-b5e6-96231b3b80d8

6 files changed, 86 insertions, 56 deletions

diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index c81f950..bcaf662 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -1002,6 +1002,24 @@ SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
 }
 
+/// GetVLDSTAlign - Get the alignment (in bytes) for the alignment operand
+/// of a NEON VLD or VST instruction.  The supported values depend on the
+/// number of registers being loaded.
+static unsigned GetVLDSTAlign(SDNode *N, unsigned NumVecs, bool is64BitVector) {
+  unsigned NumRegs = NumVecs;
+  if (!is64BitVector && NumVecs < 3)
+    NumRegs *= 2;
+
+  unsigned Alignment = cast<MemIntrinsicSDNode>(N)->getAlignment();
+  if (Alignment >= 32 && NumRegs == 4)
+    return 32;
+  if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4))
+    return 16;
+  if (Alignment >= 8)
+    return 8;
+  return 0;
+}
+
 SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
                                    unsigned *DOpcodes, unsigned *QOpcodes0,
                                    unsigned *QOpcodes1) {
@@ -1016,20 +1034,7 @@ SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
   EVT VT = N->getValueType(0);
   bool is64BitVector = VT.is64BitVector();
 
-  // Set the alignment.  The supported values depend on the number of
-  // registers being loaded.
-  unsigned NumRegs = NumVecs;
-  if (!is64BitVector && NumVecs < 3)
-    NumRegs *= 2;
-  unsigned Alignment = cast<MemIntrinsicSDNode>(N)->getAlignment();
-  if (Alignment >= 32 && NumRegs == 4)
-    Alignment = 32;
-  else if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4))
-    Alignment = 16;
-  else if (Alignment >= 8)
-    Alignment = 8;
-  else
-    Alignment = 0;
+  unsigned Alignment = GetVLDSTAlign(N, NumVecs, is64BitVector);
   Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
 
   unsigned OpcodeIndex;
@@ -1143,6 +1148,9 @@ SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
   EVT VT = N->getOperand(3).getValueType();
   bool is64BitVector = VT.is64BitVector();
 
+  unsigned Alignment = GetVLDSTAlign(N, NumVecs, is64BitVector);
+  Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
+
   unsigned OpcodeIndex;
   switch (VT.getSimpleVT().SimpleTy) {
   default: llvm_unreachable("unhandled vst type");
diff --git a/test/CodeGen/ARM/vld1.ll b/test/CodeGen/ARM/vld1.ll
index 2654326..2800475 100644
--- a/test/CodeGen/ARM/vld1.ll
+++ b/test/CodeGen/ARM/vld1.ll
@@ -43,16 +43,17 @@ define <1 x i64> @vld1i64(i64* %A) nounwind {
 define <16 x i8> @vld1Qi8(i8* %A) nounwind {
 ;CHECK: vld1Qi8:
 ;Check the alignment value.  Max for this instruction is 128 bits:
-;CHECK: vld1.8 {d0, d1}, [r0, :128]
-	%tmp1 = call <16 x i8> @llvm.arm.neon.vld1.v16i8(i8* %A, i32 32)
+;CHECK: vld1.8 {d0, d1}, [r0, :64]
+	%tmp1 = call <16 x i8> @llvm.arm.neon.vld1.v16i8(i8* %A, i32 8)
 	ret <16 x i8> %tmp1
 }
 
 define <8 x i16> @vld1Qi16(i16* %A) nounwind {
 ;CHECK: vld1Qi16:
-;CHECK: vld1.16
+;Check the alignment value.  Max for this instruction is 128 bits:
+;CHECK: vld1.16 {d0, d1}, [r0, :128]
 	%tmp0 = bitcast i16* %A to i8*
-	%tmp1 = call <8 x i16> @llvm.arm.neon.vld1.v8i16(i8* %tmp0, i32 1)
+	%tmp1 = call <8 x i16> @llvm.arm.neon.vld1.v8i16(i8* %tmp0, i32 32)
 	ret <8 x i16> %tmp1
 }
 
diff --git a/test/CodeGen/ARM/vst1.ll b/test/CodeGen/ARM/vst1.ll
index 2b535ad..c43b8aa 100644
--- a/test/CodeGen/ARM/vst1.ll
+++ b/test/CodeGen/ARM/vst1.ll
@@ -2,9 +2,10 @@
 
 define void @vst1i8(i8* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vst1i8:
-;CHECK: vst1.8
+;Check the alignment value.  Max for this instruction is 64 bits:
+;CHECK: vst1.8 {d0}, [r0, :64]
 	%tmp1 = load <8 x i8>* %B
-	call void @llvm.arm.neon.vst1.v8i8(i8* %A, <8 x i8> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst1.v8i8(i8* %A, <8 x i8> %tmp1, i32 16)
 	ret void
 }
 
@@ -46,18 +47,20 @@ define void @vst1i64(i64* %A, <1 x i64>* %B) nounwind {
 
 define void @vst1Qi8(i8* %A, <16 x i8>* %B) nounwind {
 ;CHECK: vst1Qi8:
-;CHECK: vst1.8
+;Check the alignment value.  Max for this instruction is 128 bits:
+;CHECK: vst1.8 {d0, d1}, [r0, :64]
 	%tmp1 = load <16 x i8>* %B
-	call void @llvm.arm.neon.vst1.v16i8(i8* %A, <16 x i8> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst1.v16i8(i8* %A, <16 x i8> %tmp1, i32 8)
 	ret void
 }
 
 define void @vst1Qi16(i16* %A, <8 x i16>* %B) nounwind {
 ;CHECK: vst1Qi16:
-;CHECK: vst1.16
+;Check the alignment value.  Max for this instruction is 128 bits:
+;CHECK: vst1.16 {d0, d1}, [r0, :128]
 	%tmp0 = bitcast i16* %A to i8*
 	%tmp1 = load <8 x i16>* %B
-	call void @llvm.arm.neon.vst1.v8i16(i8* %tmp0, <8 x i16> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst1.v8i16(i8* %tmp0, <8 x i16> %tmp1, i32 32)
 	ret void
 }
 
diff --git a/test/CodeGen/ARM/vst2.ll b/test/CodeGen/ARM/vst2.ll
index aed15fd..1124640 100644
--- a/test/CodeGen/ARM/vst2.ll
+++ b/test/CodeGen/ARM/vst2.ll
@@ -2,18 +2,20 @@
 
 define void @vst2i8(i8* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vst2i8:
-;CHECK: vst2.8
+;Check the alignment value.  Max for this instruction is 128 bits:
+;CHECK: vst2.8 {d0, d1}, [r0, :64]
 	%tmp1 = load <8 x i8>* %B
-	call void @llvm.arm.neon.vst2.v8i8(i8* %A, <8 x i8> %tmp1, <8 x i8> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst2.v8i8(i8* %A, <8 x i8> %tmp1, <8 x i8> %tmp1, i32 8)
 	ret void
 }
 
 define void @vst2i16(i16* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vst2i16:
-;CHECK: vst2.16
+;Check the alignment value.  Max for this instruction is 128 bits:
+;CHECK: vst2.16 {d0, d1}, [r0, :128]
 	%tmp0 = bitcast i16* %A to i8*
 	%tmp1 = load <4 x i16>* %B
-	call void @llvm.arm.neon.vst2.v4i16(i8* %tmp0, <4 x i16> %tmp1, <4 x i16> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst2.v4i16(i8* %tmp0, <4 x i16> %tmp1, <4 x i16> %tmp1, i32 32)
 	ret void
 }
 
@@ -37,36 +39,40 @@ define void @vst2f(float* %A, <2 x float>* %B) nounwind {
 
 define void @vst2i64(i64* %A, <1 x i64>* %B) nounwind {
 ;CHECK: vst2i64:
-;CHECK: vst1.64
+;Check the alignment value.  Max for this instruction is 128 bits:
+;CHECK: vst1.64 {d0, d1}, [r0, :128]
 	%tmp0 = bitcast i64* %A to i8*
 	%tmp1 = load <1 x i64>* %B
-	call void @llvm.arm.neon.vst2.v1i64(i8* %tmp0, <1 x i64> %tmp1, <1 x i64> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst2.v1i64(i8* %tmp0, <1 x i64> %tmp1, <1 x i64> %tmp1, i32 32)
 	ret void
 }
 
 define void @vst2Qi8(i8* %A, <16 x i8>* %B) nounwind {
 ;CHECK: vst2Qi8:
-;CHECK: vst2.8
+;Check the alignment value.  Max for this instruction is 256 bits:
+;CHECK: vst2.8 {d0, d1, d2, d3}, [r0, :64]
 	%tmp1 = load <16 x i8>* %B
-	call void @llvm.arm.neon.vst2.v16i8(i8* %A, <16 x i8> %tmp1, <16 x i8> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst2.v16i8(i8* %A, <16 x i8> %tmp1, <16 x i8> %tmp1, i32 8)
 	ret void
 }
 
 define void @vst2Qi16(i16* %A, <8 x i16>* %B) nounwind {
 ;CHECK: vst2Qi16:
-;CHECK: vst2.16
+;Check the alignment value.  Max for this instruction is 256 bits:
+;CHECK: vst2.16 {d0, d1, d2, d3}, [r0, :128]
 	%tmp0 = bitcast i16* %A to i8*
 	%tmp1 = load <8 x i16>* %B
-	call void @llvm.arm.neon.vst2.v8i16(i8* %tmp0, <8 x i16> %tmp1, <8 x i16> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst2.v8i16(i8* %tmp0, <8 x i16> %tmp1, <8 x i16> %tmp1, i32 16)
 	ret void
 }
 
 define void @vst2Qi32(i32* %A, <4 x i32>* %B) nounwind {
 ;CHECK: vst2Qi32:
-;CHECK: vst2.32
+;Check the alignment value.  Max for this instruction is 256 bits:
+;CHECK: vst2.32 {d0, d1, d2, d3}, [r0, :256]
 	%tmp0 = bitcast i32* %A to i8*
 	%tmp1 = load <4 x i32>* %B
-	call void @llvm.arm.neon.vst2.v4i32(i8* %tmp0, <4 x i32> %tmp1, <4 x i32> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst2.v4i32(i8* %tmp0, <4 x i32> %tmp1, <4 x i32> %tmp1, i32 64)
 	ret void
 }
 
diff --git a/test/CodeGen/ARM/vst3.ll b/test/CodeGen/ARM/vst3.ll
index 1feaed5..0a2df77 100644
--- a/test/CodeGen/ARM/vst3.ll
+++ b/test/CodeGen/ARM/vst3.ll
@@ -2,9 +2,11 @@
 
 define void @vst3i8(i8* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vst3i8:
-;CHECK: vst3.8
+;Check the alignment value.  Max for this instruction is 64 bits:
+;This test runs at -O0 so do not check for specific register numbers.
+;CHECK: vst3.8 {d{{.*}}, d{{.*}}, d{{.*}}}, [r{{.*}}, :64]
 	%tmp1 = load <8 x i8>* %B
-	call void @llvm.arm.neon.vst3.v8i8(i8* %A, <8 x i8> %tmp1, <8 x i8> %tmp1, <8 x i8> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst3.v8i8(i8* %A, <8 x i8> %tmp1, <8 x i8> %tmp1, <8 x i8> %tmp1, i32 32)
 	ret void
 }
 
@@ -37,19 +39,23 @@ define void @vst3f(float* %A, <2 x float>* %B) nounwind {
 
 define void @vst3i64(i64* %A, <1 x i64>* %B) nounwind {
 ;CHECK: vst3i64:
-;CHECK: vst1.64
+;Check the alignment value.  Max for this instruction is 64 bits:
+;This test runs at -O0 so do not check for specific register numbers.
+;CHECK: vst1.64 {d{{.*}}, d{{.*}}, d{{.*}}}, [r{{.*}}, :64]
 	%tmp0 = bitcast i64* %A to i8*
 	%tmp1 = load <1 x i64>* %B
-	call void @llvm.arm.neon.vst3.v1i64(i8* %tmp0, <1 x i64> %tmp1, <1 x i64> %tmp1, <1 x i64> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst3.v1i64(i8* %tmp0, <1 x i64> %tmp1, <1 x i64> %tmp1, <1 x i64> %tmp1, i32 16)
 	ret void
 }
 
 define void @vst3Qi8(i8* %A, <16 x i8>* %B) nounwind {
 ;CHECK: vst3Qi8:
-;CHECK: vst3.8
-;CHECK: vst3.8
+;Check the alignment value.  Max for this instruction is 64 bits:
+;This test runs at -O0 so do not check for specific register numbers.
+;CHECK: vst3.8 {d{{.*}}, d{{.*}}, d{{.*}}}, [r{{.*}}, :64]!
+;CHECK: vst3.8 {d{{.*}}, d{{.*}}, d{{.*}}}, [r{{.*}}, :64]
 	%tmp1 = load <16 x i8>* %B
-	call void @llvm.arm.neon.vst3.v16i8(i8* %A, <16 x i8> %tmp1, <16 x i8> %tmp1, <16 x i8> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst3.v16i8(i8* %A, <16 x i8> %tmp1, <16 x i8> %tmp1, <16 x i8> %tmp1, i32 32)
 	ret void
 }
 
diff --git a/test/CodeGen/ARM/vst4.ll b/test/CodeGen/ARM/vst4.ll
index d302f09..a339192 100644
--- a/test/CodeGen/ARM/vst4.ll
+++ b/test/CodeGen/ARM/vst4.ll
@@ -2,27 +2,30 @@
 
 define void @vst4i8(i8* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vst4i8:
-;CHECK: vst4.8
+;Check the alignment value.  Max for this instruction is 256 bits:
+;CHECK: vst4.8 {d0, d1, d2, d3}, [r0, :64]
 	%tmp1 = load <8 x i8>* %B
-	call void @llvm.arm.neon.vst4.v8i8(i8* %A, <8 x i8> %tmp1, <8 x i8> %tmp1, <8 x i8> %tmp1, <8 x i8> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst4.v8i8(i8* %A, <8 x i8> %tmp1, <8 x i8> %tmp1, <8 x i8> %tmp1, <8 x i8> %tmp1, i32 8)
 	ret void
 }
 
 define void @vst4i16(i16* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vst4i16:
-;CHECK: vst4.16
+;Check the alignment value.  Max for this instruction is 256 bits:
+;CHECK: vst4.16 {d0, d1, d2, d3}, [r0, :128]
 	%tmp0 = bitcast i16* %A to i8*
 	%tmp1 = load <4 x i16>* %B
-	call void @llvm.arm.neon.vst4.v4i16(i8* %tmp0, <4 x i16> %tmp1, <4 x i16> %tmp1, <4 x i16> %tmp1, <4 x i16> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst4.v4i16(i8* %tmp0, <4 x i16> %tmp1, <4 x i16> %tmp1, <4 x i16> %tmp1, <4 x i16> %tmp1, i32 16)
 	ret void
 }
 
 define void @vst4i32(i32* %A, <2 x i32>* %B) nounwind {
 ;CHECK: vst4i32:
-;CHECK: vst4.32
+;Check the alignment value.  Max for this instruction is 256 bits:
+;CHECK: vst4.32 {d0, d1, d2, d3}, [r0, :256]
 	%tmp0 = bitcast i32* %A to i8*
 	%tmp1 = load <2 x i32>* %B
-	call void @llvm.arm.neon.vst4.v2i32(i8* %tmp0, <2 x i32> %tmp1, <2 x i32> %tmp1, <2 x i32> %tmp1, <2 x i32> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst4.v2i32(i8* %tmp0, <2 x i32> %tmp1, <2 x i32> %tmp1, <2 x i32> %tmp1, <2 x i32> %tmp1, i32 32)
 	ret void
 }
 
@@ -37,26 +40,29 @@ define void @vst4f(float* %A, <2 x float>* %B) nounwind {
 
 define void @vst4i64(i64* %A, <1 x i64>* %B) nounwind {
 ;CHECK: vst4i64:
-;CHECK: vst1.64
+;Check the alignment value.  Max for this instruction is 256 bits:
+;CHECK: vst1.64 {d0, d1, d2, d3}, [r0, :256]
 	%tmp0 = bitcast i64* %A to i8*
 	%tmp1 = load <1 x i64>* %B
-	call void @llvm.arm.neon.vst4.v1i64(i8* %tmp0, <1 x i64> %tmp1, <1 x i64> %tmp1, <1 x i64> %tmp1, <1 x i64> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst4.v1i64(i8* %tmp0, <1 x i64> %tmp1, <1 x i64> %tmp1, <1 x i64> %tmp1, <1 x i64> %tmp1, i32 64)
 	ret void
 }
 
 define void @vst4Qi8(i8* %A, <16 x i8>* %B) nounwind {
 ;CHECK: vst4Qi8:
-;CHECK: vst4.8
-;CHECK: vst4.8
+;Check the alignment value.  Max for this instruction is 256 bits:
+;CHECK: vst4.8 {d0, d2, d4, d6}, [r0, :256]!
+;CHECK: vst4.8 {d1, d3, d5, d7}, [r0, :256]
 	%tmp1 = load <16 x i8>* %B
-	call void @llvm.arm.neon.vst4.v16i8(i8* %A, <16 x i8> %tmp1, <16 x i8> %tmp1, <16 x i8> %tmp1, <16 x i8> %tmp1, i32 1)
+	call void @llvm.arm.neon.vst4.v16i8(i8* %A, <16 x i8> %tmp1, <16 x i8> %tmp1, <16 x i8> %tmp1, <16 x i8> %tmp1, i32 64)
 	ret void
 }
 
 define void @vst4Qi16(i16* %A, <8 x i16>* %B) nounwind {
 ;CHECK: vst4Qi16:
-;CHECK: vst4.16
-;CHECK: vst4.16
+;Check for no alignment specifier.
+;CHECK: vst4.16 {d0, d2, d4, d6}, [r0]!
+;CHECK: vst4.16 {d1, d3, d5, d7}, [r0]
 	%tmp0 = bitcast i16* %A to i8*
 	%tmp1 = load <8 x i16>* %B
 	call void @llvm.arm.neon.vst4.v8i16(i8* %tmp0, <8 x i16> %tmp1, <8 x i16> %tmp1, <8 x i16> %tmp1, <8 x i16> %tmp1, i32 1)