From 3d69ff4d07cec00180a1069bca5982063996908e Mon Sep 17 00:00:00 2001
From: Hao Liu <Hao.Liu@arm.com>
Date: Tue, 3 Dec 2013 03:39:47 +0000
Subject: [PATCH] AArch64: Add missing scalar pair intrinsics. E.g. "float32_t
 vaddv_f32(float32x2_t a)" to be matched into "faddp s0, v1.2s".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@196198 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/AArch64/AArch64InstrNEON.td        |  22 +++
 .../AArch64/neon-scalar-reduce-pairwise.ll    | 144 ++++++++++++++++++
 2 files changed, 166 insertions(+)

diff --git a/lib/Target/AArch64/AArch64InstrNEON.td b/lib/Target/AArch64/AArch64InstrNEON.td
index 361a790c0d7..6b83d88d7f0 100644
--- a/lib/Target/AArch64/AArch64InstrNEON.td
+++ b/lib/Target/AArch64/AArch64InstrNEON.td
@@ -5194,6 +5194,8 @@ defm ADDPvv : NeonI_ScalarPair_D_sizes<0b0, 0b1, 0b11011, "addp", 0>;
 // Scalar Reduce Addition Pairwise (Integer)
 def : Pat<(v1i64 (int_aarch64_neon_vpadd (v2i64 VPR128:$Rn))),
           (ADDPvv_D_2D VPR128:$Rn)>;
+def : Pat<(v1i64 (int_aarch64_neon_vaddv (v2i64 VPR128:$Rn))),
+          (ADDPvv_D_2D VPR128:$Rn)>;
 
 // Scalar Reduce Addition Pairwise (Floating Point)
 defm FADDPvv : NeonI_ScalarPair_SD_sizes<0b1, 0b0, 0b01101, "faddp", 0>;
@@ -5237,6 +5239,26 @@ defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vpfmaxnm,
 defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vpfminnm, 
   int_aarch64_neon_vpfminnmq, FMINNMPvv_S_2S, FMINNMPvv_D_2D>;
 
+defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vaddv, 
+    int_aarch64_neon_vaddv, FADDPvv_S_2S, FADDPvv_D_2D>;
+
+def : Pat<(v1f32 (int_aarch64_neon_vaddv (v4f32 VPR128:$Rn))),
+          (FADDPvv_S_2S (v2f32
+               (EXTRACT_SUBREG
+                   (v4f32 (FADDP_4S (v4f32 VPR128:$Rn), (v4f32 VPR128:$Rn))),
+                   sub_64)))>;
+
+defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vmaxv, 
+    int_aarch64_neon_vmaxv, FMAXPvv_S_2S, FMAXPvv_D_2D>;
+
+defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vminv, 
+    int_aarch64_neon_vminv, FMINPvv_S_2S, FMINPvv_D_2D>;
+
+defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vmaxnmv, 
+    int_aarch64_neon_vmaxnmv, FMAXNMPvv_S_2S, FMAXNMPvv_D_2D>;
+
+defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vminnmv, 
+    int_aarch64_neon_vminnmv, FMINNMPvv_S_2S, FMINNMPvv_D_2D>;
 
 // Scalar by element Arithmetic
 
diff --git a/test/CodeGen/AArch64/neon-scalar-reduce-pairwise.ll b/test/CodeGen/AArch64/neon-scalar-reduce-pairwise.ll
index 1bb3b40440a..80e8dc339d6 100644
--- a/test/CodeGen/AArch64/neon-scalar-reduce-pairwise.ll
+++ b/test/CodeGen/AArch64/neon-scalar-reduce-pairwise.ll
@@ -101,3 +101,147 @@ define <1 x double> @test_fminnmp_v1f64(<2 x double> %a) {
         ret <1 x double> %val
 }
 
+define float @test_vaddv_f32(<2 x float> %a) {
+; CHECK-LABEL: test_vaddv_f32
+; CHECK: faddp {{s[0-9]+}}, {{v[0-9]+}}.2s
+  %1 = tail call <1 x float> @llvm.aarch64.neon.vaddv.v1f32.v2f32(<2 x float> %a)
+  %2 = extractelement <1 x float> %1, i32 0
+  ret float %2
+}
+
+define float @test_vaddvq_f32(<4 x float> %a) {
+; CHECK-LABEL: test_vaddvq_f32
+; CHECK: faddp {{v[0-9]+}}.4s, {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
+; CHECK: faddp {{s[0-9]+}}, {{v[0-9]+}}.2s
+  %1 = tail call <1 x float> @llvm.aarch64.neon.vaddv.v1f32.v4f32(<4 x float> %a)
+  %2 = extractelement <1 x float> %1, i32 0
+  ret float %2
+}
+
+define double @test_vaddvq_f64(<2 x double> %a) {
+; CHECK-LABEL: test_vaddvq_f64
+; CHECK: faddp {{d[0-9]+}}, {{v[0-9]+}}.2d
+  %1 = tail call <1 x double> @llvm.aarch64.neon.vaddv.v1f64.v2f64(<2 x double> %a)
+  %2 = extractelement <1 x double> %1, i32 0
+  ret double %2
+}
+
+define float @test_vmaxv_f32(<2 x float> %a) {
+; CHECK-LABEL: test_vmaxv_f32
+; CHECK: fmaxp {{s[0-9]+}}, {{v[0-9]+}}.2s
+  %1 = tail call <1 x float> @llvm.aarch64.neon.vmaxv.v1f32.v2f32(<2 x float> %a)
+  %2 = extractelement <1 x float> %1, i32 0
+  ret float %2
+}
+
+define double @test_vmaxvq_f64(<2 x double> %a) {
+; CHECK-LABEL: test_vmaxvq_f64
+; CHECK: fmaxp {{d[0-9]+}}, {{v[0-9]+}}.2d
+  %1 = tail call <1 x double> @llvm.aarch64.neon.vmaxv.v1f64.v2f64(<2 x double> %a)
+  %2 = extractelement <1 x double> %1, i32 0
+  ret double %2
+}
+
+define float @test_vminv_f32(<2 x float> %a) {
+; CHECK-LABEL: test_vminv_f32
+; CHECK: fminp {{s[0-9]+}}, {{v[0-9]+}}.2s
+  %1 = tail call <1 x float> @llvm.aarch64.neon.vminv.v1f32.v2f32(<2 x float> %a)
+  %2 = extractelement <1 x float> %1, i32 0
+  ret float %2
+}
+
+define double @test_vminvq_f64(<2 x double> %a) {
+; CHECK-LABEL: test_vminvq_f64
+; CHECK: fminp {{d[0-9]+}}, {{v[0-9]+}}.2d
+  %1 = tail call <1 x double> @llvm.aarch64.neon.vminv.v1f64.v2f64(<2 x double> %a)
+  %2 = extractelement <1 x double> %1, i32 0
+  ret double %2
+}
+
+define double @test_vmaxnmvq_f64(<2 x double> %a) {
+; CHECK-LABEL: test_vmaxnmvq_f64
+; CHECK: fmaxnmp {{d[0-9]+}}, {{v[0-9]+}}.2d
+  %1 = tail call <1 x double> @llvm.aarch64.neon.vmaxnmv.v1f64.v2f64(<2 x double> %a)
+  %2 = extractelement <1 x double> %1, i32 0
+  ret double %2
+}
+
+define float @test_vmaxnmv_f32(<2 x float> %a) {
+; CHECK-LABEL: test_vmaxnmv_f32
+; CHECK: fmaxnmp {{s[0-9]+}}, {{v[0-9]+}}.2s
+  %1 = tail call <1 x float> @llvm.aarch64.neon.vmaxnmv.v1f32.v2f32(<2 x float> %a)
+  %2 = extractelement <1 x float> %1, i32 0
+  ret float %2
+}
+
+define double @test_vminnmvq_f64(<2 x double> %a) {
+; CHECK-LABEL: test_vminnmvq_f64
+; CHECK: fminnmp {{d[0-9]+}}, {{v[0-9]+}}.2d
+  %1 = tail call <1 x double> @llvm.aarch64.neon.vminnmv.v1f64.v2f64(<2 x double> %a)
+  %2 = extractelement <1 x double> %1, i32 0
+  ret double %2
+}
+
+define float @test_vminnmv_f32(<2 x float> %a) {
+; CHECK-LABEL: test_vminnmv_f32
+; CHECK: fminnmp {{s[0-9]+}}, {{v[0-9]+}}.2s
+  %1 = tail call <1 x float> @llvm.aarch64.neon.vminnmv.v1f32.v2f32(<2 x float> %a)
+  %2 = extractelement <1 x float> %1, i32 0
+  ret float %2
+}
+
+define <2 x i64> @test_vpaddq_s64(<2 x i64> %a, <2 x i64> %b) {
+; CHECK-LABEL: test_vpaddq_s64
+; CHECK: addp {{v[0-9]+}}.2d, {{v[0-9]+}}.2d, {{v[0-9]+}}.2d
+  %1 = tail call <2 x i64> @llvm.arm.neon.vpadd.v2i64(<2 x i64> %a, <2 x i64> %b)
+  ret <2 x i64> %1
+}
+
+define <2 x i64> @test_vpaddq_u64(<2 x i64> %a, <2 x i64> %b) {
+; CHECK-LABEL: test_vpaddq_u64
+; CHECK: addp {{v[0-9]+}}.2d, {{v[0-9]+}}.2d, {{v[0-9]+}}.2d
+  %1 = tail call <2 x i64> @llvm.arm.neon.vpadd.v2i64(<2 x i64> %a, <2 x i64> %b)
+  ret <2 x i64> %1
+}
+
+define i64 @test_vaddvq_s64(<2 x i64> %a) {
+; CHECK-LABEL: test_vaddvq_s64
+; CHECK: addp {{d[0-9]+}}, {{v[0-9]+}}.2d
+  %1 = tail call <1 x i64> @llvm.aarch64.neon.vaddv.v1i64.v2i64(<2 x i64> %a)
+  %2 = extractelement <1 x i64> %1, i32 0
+  ret i64 %2
+}
+
+define i64 @test_vaddvq_u64(<2 x i64> %a) {
+; CHECK-LABEL: test_vaddvq_u64
+; CHECK: addp {{d[0-9]+}}, {{v[0-9]+}}.2d
+  %1 = tail call <1 x i64> @llvm.aarch64.neon.vaddv.v1i64.v2i64(<2 x i64> %a)
+  %2 = extractelement <1 x i64> %1, i32 0
+  ret i64 %2
+}
+
+declare <1 x i64> @llvm.aarch64.neon.vaddv.v1i64.v2i64(<2 x i64>)
+
+declare <2 x i64> @llvm.arm.neon.vpadd.v2i64(<2 x i64>, <2 x i64>)
+
+declare <1 x float> @llvm.aarch64.neon.vminnmv.v1f32.v2f32(<2 x float>)
+
+declare <1 x double> @llvm.aarch64.neon.vminnmv.v1f64.v2f64(<2 x double>)
+
+declare <1 x float> @llvm.aarch64.neon.vmaxnmv.v1f32.v2f32(<2 x float>)
+
+declare <1 x double> @llvm.aarch64.neon.vmaxnmv.v1f64.v2f64(<2 x double>)
+
+declare <1 x double> @llvm.aarch64.neon.vminv.v1f64.v2f64(<2 x double>)
+
+declare <1 x float> @llvm.aarch64.neon.vminv.v1f32.v2f32(<2 x float>)
+
+declare <1 x double> @llvm.aarch64.neon.vmaxv.v1f64.v2f64(<2 x double>)
+
+declare <1 x float> @llvm.aarch64.neon.vmaxv.v1f32.v2f32(<2 x float>)
+
+declare <1 x double> @llvm.aarch64.neon.vaddv.v1f64.v2f64(<2 x double>)
+
+declare <1 x float> @llvm.aarch64.neon.vaddv.v1f32.v4f32(<4 x float>)
+
+declare <1 x float> @llvm.aarch64.neon.vaddv.v1f32.v2f32(<2 x float>)
\ No newline at end of file