ARM: add patterns for vqdmlal with separate vqdmull and vqadds

The vqdmlal and vqdmlls instructions are really just a fused pair consisting of
a vqdmull.sN and a vqadd.sN. This adds patterns to LLVM so that we can switch
Clang's CodeGen over to generating these instead of the special vqdmlal
intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189480 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index a5d898540ba87322d350a90ec9ce023ef79da840..cc455adb561c7d9a5eee3f0101a5aac45b996d60 100644 (file)
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -4143,6 +4143,25 @@ defm VQDMLAL  : N3VLInt3_HS<0, 1, 0b1001, 0, IIC_VMACi16D, IIC_VMACi32D,
                             "vqdmlal", "s", int_arm_neon_vqdmlal>;
 defm VQDMLALsl: N3VLInt3SL_HS<0, 0b0011, "vqdmlal", "s", int_arm_neon_vqdmlal>;
 
+def : Pat<(v4i32 (int_arm_neon_vqadds (v4i32 QPR:$src1),
+                     (v4i32 (int_arm_neon_vqdmull (v4i16 DPR:$Vn),
+                                                  (v4i16 DPR:$Vm))))),
+          (VQDMLALv4i32 QPR:$src1, DPR:$Vn, DPR:$Vm)>;
+def : Pat<(v2i64 (int_arm_neon_vqadds (v2i64 QPR:$src1),
+                     (v2i64 (int_arm_neon_vqdmull (v2i32 DPR:$Vn),
+                                                  (v2i32 DPR:$Vm))))),
+          (VQDMLALv2i64 QPR:$src1, DPR:$Vn, DPR:$Vm)>;
+def : Pat<(v4i32 (int_arm_neon_vqadds (v4i32 QPR:$src1),
+                     (v4i32 (int_arm_neon_vqdmull (v4i16 DPR:$Vn),
+                                (v4i16 (NEONvduplane (v4i16 DPR_8:$Vm),
+                                                     imm:$lane)))))),
+          (VQDMLALslv4i16 QPR:$src1, DPR:$Vn, DPR_8:$Vm, imm:$lane)>;
+def : Pat<(v2i64 (int_arm_neon_vqadds (v2i64 QPR:$src1),
+                     (v2i64 (int_arm_neon_vqdmull (v2i32 DPR:$Vn),
+                                (v2i32 (NEONvduplane (v2i32 DPR_VFP2:$Vm),
+                                                     imm:$lane)))))),
+          (VQDMLALslv2i32 QPR:$src1, DPR:$Vn, DPR_VFP2:$Vm, imm:$lane)>;
+
 //   VMLS     : Vector Multiply Subtract (integer and floating-point)
 defm VMLS     : N3VMulOp_QHS<1, 0, 0b1001, 0, IIC_VMACi16D, IIC_VMACi32D,
                              IIC_VMACi16Q, IIC_VMACi32Q, "vmls", "i", sub>;
@@ -4200,6 +4219,25 @@ defm VQDMLSL  : N3VLInt3_HS<0, 1, 0b1011, 0, IIC_VMACi16D, IIC_VMACi32D,
                             "vqdmlsl", "s", int_arm_neon_vqdmlsl>;
 defm VQDMLSLsl: N3VLInt3SL_HS<0, 0b111, "vqdmlsl", "s", int_arm_neon_vqdmlsl>;
 
+def : Pat<(v4i32 (int_arm_neon_vqsubs (v4i32 QPR:$src1),
+                     (v4i32 (int_arm_neon_vqdmull (v4i16 DPR:$Vn),
+                                                  (v4i16 DPR:$Vm))))),
+          (VQDMLSLv4i32 QPR:$src1, DPR:$Vn, DPR:$Vm)>;
+def : Pat<(v2i64 (int_arm_neon_vqsubs (v2i64 QPR:$src1),
+                     (v2i64 (int_arm_neon_vqdmull (v2i32 DPR:$Vn),
+                                                  (v2i32 DPR:$Vm))))),
+          (VQDMLSLv2i64 QPR:$src1, DPR:$Vn, DPR:$Vm)>;
+def : Pat<(v4i32 (int_arm_neon_vqsubs (v4i32 QPR:$src1),
+                     (v4i32 (int_arm_neon_vqdmull (v4i16 DPR:$Vn),
+                                (v4i16 (NEONvduplane (v4i16 DPR_8:$Vm),
+                                                     imm:$lane)))))),
+          (VQDMLSLslv4i16 QPR:$src1, DPR:$Vn, DPR_8:$Vm, imm:$lane)>;
+def : Pat<(v2i64 (int_arm_neon_vqsubs (v2i64 QPR:$src1),
+                     (v2i64 (int_arm_neon_vqdmull (v2i32 DPR:$Vn),
+                                (v2i32 (NEONvduplane (v2i32 DPR_VFP2:$Vm),
+                                                     imm:$lane)))))),
+          (VQDMLSLslv2i32 QPR:$src1, DPR:$Vn, DPR_VFP2:$Vm, imm:$lane)>;
+
 // Fused Vector Multiply-Accumulate and Fused Multiply-Subtract Operations.
 def  VFMAfd   : N3VDMulOp<0, 0, 0b00, 0b1100, 1, IIC_VFMACD, "vfma", "f32",
                           v2f32, fmul_su, fadd_mlx>,

diff --git a/test/CodeGen/ARM/vqdmul.ll b/test/CodeGen/ARM/vqdmul.ll
index a28cae9aae346884829d69e4e7c6d9c363ea7a2b..01bf1a4a97ffb4bef2dc5c7b92b9d24b543558dd 100644 (file)
--- a/test/CodeGen/ARM/vqdmul.ll
+++ b/test/CodeGen/ARM/vqdmul.ll
@@ -238,6 +238,51 @@ entry:
 declare <4 x i32>  @llvm.arm.neon.vqdmlal.v4i32(<4 x i32>, <4 x i16>, <4 x i16>) nounwind readnone
 declare <2 x i64>  @llvm.arm.neon.vqdmlal.v2i64(<2 x i64>, <2 x i32>, <2 x i32>) nounwind readnone
 
+define <4 x i32> @vqdmlals16_natural(<4 x i32>* %A, <4 x i16>* %B, <4 x i16>* %C) nounwind {
+;CHECK-LABEL: vqdmlals16_natural:
+;CHECK: vqdmlal.s16
+        %tmp1 = load <4 x i32>* %A
+        %tmp2 = load <4 x i16>* %B
+        %tmp3 = load <4 x i16>* %C
+        %tmp4 = call <4 x i32> @llvm.arm.neon.vqdmull.v4i32(<4 x i16> %tmp2, <4 x i16> %tmp3)
+        %tmp5 = call <4 x i32> @llvm.arm.neon.vqadds.v4i32(<4 x i32> %tmp1, <4 x i32> %tmp4)
+        ret <4 x i32> %tmp5
+}
+
+define <2 x i64> @vqdmlals32_natural(<2 x i64>* %A, <2 x i32>* %B, <2 x i32>* %C) nounwind {
+;CHECK-LABEL: vqdmlals32_natural:
+;CHECK: vqdmlal.s32
+        %tmp1 = load <2 x i64>* %A
+        %tmp2 = load <2 x i32>* %B
+        %tmp3 = load <2 x i32>* %C
+        %tmp4 = call <2 x i64> @llvm.arm.neon.vqdmull.v2i64(<2 x i32> %tmp2, <2 x i32> %tmp3)
+        %tmp5 = call <2 x i64> @llvm.arm.neon.vqadds.v2i64(<2 x i64> %tmp1, <2 x i64> %tmp4)
+        ret <2 x i64> %tmp5
+}
+
+define arm_aapcs_vfpcc <4 x i32> @test_vqdmlal_lanes16_natural(<4 x i32> %arg0_int32x4_t, <4 x i16> %arg1_int16x4_t, <4 x i16> %arg2_int16x4_t) nounwind readnone {
+entry:
+; CHECK-LABEL: test_vqdmlal_lanes16_natural:
+; CHECK: vqdmlal.s16 q0, d2, d3[1]
+  %0 = shufflevector <4 x i16> %arg2_int16x4_t, <4 x i16> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1> ; <<4 x i16>> [#uses=1]
+  %1 = tail call <4 x i32> @llvm.arm.neon.vqdmull.v4i32(<4 x i16> %arg1_int16x4_t, <4 x i16> %0)
+  %2 = tail call <4 x i32> @llvm.arm.neon.vqadds.v4i32(<4 x i32> %arg0_int32x4_t, <4 x i32> %1)
+  ret <4 x i32> %2
+}
+
+define arm_aapcs_vfpcc <2 x i64> @test_vqdmlal_lanes32_natural(<2 x i64> %arg0_int64x2_t, <2 x i32> %arg1_int32x2_t, <2 x i32> %arg2_int32x2_t) nounwind readnone {
+entry:
+; CHECK-LABEL: test_vqdmlal_lanes32_natural:
+; CHECK: vqdmlal.s32 q0, d2, d3[1]
+  %0 = shufflevector <2 x i32> %arg2_int32x2_t, <2 x i32> undef, <2 x i32> <i32 1, i32 1> ; <<2 x i32>> [#uses=1]
+  %1 = tail call <2 x i64> @llvm.arm.neon.vqdmull.v2i64(<2 x i32> %arg1_int32x2_t, <2 x i32> %0)
+  %2 = call <2 x i64> @llvm.arm.neon.vqadds.v2i64(<2 x i64> %arg0_int64x2_t, <2 x i64> %1)
+  ret <2 x i64> %2
+}
+
+declare <4 x i32>  @llvm.arm.neon.vqadds.v4i32(<4 x i32>, <4 x i32>) nounwind readnone
+declare <2 x i64>  @llvm.arm.neon.vqadds.v2i64(<2 x i64>, <2 x i64>) nounwind readnone
+
 define <4 x i32> @vqdmlsls16(<4 x i32>* %A, <4 x i16>* %B, <4 x i16>* %C) nounwind {
 ;CHECK-LABEL: vqdmlsls16:
 ;CHECK: vqdmlsl.s16
@@ -278,3 +323,48 @@ entry:
 
 declare <4 x i32>  @llvm.arm.neon.vqdmlsl.v4i32(<4 x i32>, <4 x i16>, <4 x i16>) nounwind readnone
 declare <2 x i64>  @llvm.arm.neon.vqdmlsl.v2i64(<2 x i64>, <2 x i32>, <2 x i32>) nounwind readnone
+
+define <4 x i32> @vqdmlsls16_natural(<4 x i32>* %A, <4 x i16>* %B, <4 x i16>* %C) nounwind {
+;CHECK-LABEL: vqdmlsls16_natural:
+;CHECK: vqdmlsl.s16
+        %tmp1 = load <4 x i32>* %A
+        %tmp2 = load <4 x i16>* %B
+        %tmp3 = load <4 x i16>* %C
+        %tmp4 = call <4 x i32> @llvm.arm.neon.vqdmull.v4i32(<4 x i16> %tmp2, <4 x i16> %tmp3)
+        %tmp5 = call <4 x i32> @llvm.arm.neon.vqsubs.v4i32(<4 x i32> %tmp1, <4 x i32> %tmp4)
+        ret <4 x i32> %tmp5
+}
+
+define <2 x i64> @vqdmlsls32_natural(<2 x i64>* %A, <2 x i32>* %B, <2 x i32>* %C) nounwind {
+;CHECK-LABEL: vqdmlsls32_natural:
+;CHECK: vqdmlsl.s32
+        %tmp1 = load <2 x i64>* %A
+        %tmp2 = load <2 x i32>* %B
+        %tmp3 = load <2 x i32>* %C
+        %tmp4 = call <2 x i64> @llvm.arm.neon.vqdmull.v2i64(<2 x i32> %tmp2, <2 x i32> %tmp3)
+        %tmp5 = call <2 x i64> @llvm.arm.neon.vqsubs.v2i64(<2 x i64> %tmp1, <2 x i64> %tmp4)
+        ret <2 x i64> %tmp5
+}
+
+define arm_aapcs_vfpcc <4 x i32> @test_vqdmlsl_lanes16_natural(<4 x i32> %arg0_int32x4_t, <4 x i16> %arg1_int16x4_t, <4 x i16> %arg2_int16x4_t) nounwind readnone {
+entry:
+; CHECK-LABEL: test_vqdmlsl_lanes16_natural:
+; CHECK: vqdmlsl.s16 q0, d2, d3[1]
+  %0 = shufflevector <4 x i16> %arg2_int16x4_t, <4 x i16> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1> ; <<4 x i16>> [#uses=1]
+  %1 = tail call <4 x i32> @llvm.arm.neon.vqdmull.v4i32(<4 x i16> %arg1_int16x4_t, <4 x i16> %0)
+  %2 = tail call <4 x i32> @llvm.arm.neon.vqsubs.v4i32(<4 x i32> %arg0_int32x4_t, <4 x i32> %1)
+  ret <4 x i32> %2
+}
+
+define arm_aapcs_vfpcc <2 x i64> @test_vqdmlsl_lanes32_natural(<2 x i64> %arg0_int64x2_t, <2 x i32> %arg1_int32x2_t, <2 x i32> %arg2_int32x2_t) nounwind readnone {
+entry:
+; CHECK-LABEL: test_vqdmlsl_lanes32_natural:
+; CHECK: vqdmlsl.s32 q0, d2, d3[1]
+  %0 = shufflevector <2 x i32> %arg2_int32x2_t, <2 x i32> undef, <2 x i32> <i32 1, i32 1> ; <<2 x i32>> [#uses=1]
+  %1 = tail call <2 x i64> @llvm.arm.neon.vqdmull.v2i64(<2 x i32> %arg1_int32x2_t, <2 x i32> %0)
+  %2 = call <2 x i64> @llvm.arm.neon.vqsubs.v2i64(<2 x i64> %arg0_int64x2_t, <2 x i64> %1)
+  ret <2 x i64> %2
+}
+
+declare <4 x i32>  @llvm.arm.neon.vqsubs.v4i32(<4 x i32>, <4 x i32>) nounwind readnone
+declare <2 x i64>  @llvm.arm.neon.vqsubs.v2i64(<2 x i64>, <2 x i64>) nounwind readnone