[llvm] [LoongArch] Optimize inserting extracted element for v4i64/v8i32 (PR #152629)

[via llvm-commits]

https://github.com/tangaac updated https://github.com/llvm/llvm-project/pull/152629

>From 7132dbf696a6df1526ec2830fda69f07e6fd0dc9 Mon Sep 17 00:00:00 2001
From: tangaac <tangyan01 at loongson.cn>
Date: Fri, 8 Aug 2025 10:34:59 +0800
Subject: [PATCH 1/2] Optimize inserting extracted element for v4i64/v8i32

---
 .../Target/LoongArch/LoongArchLASXInstrInfo.td   | 16 ++++++++++++++++
 .../ir-instruction/insert-extract-element.ll     |  8 ++++----
 2 files changed, 20 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
index d8bb16fe9b94d..37d665010aa8c 100644
--- a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
+++ b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
@@ -1640,6 +1640,22 @@ defm : PairInsertExtractPatV8<v8f32, f32>;
 defm : PairInsertExtractPatV4<v4i64, GRLenVT>;
 defm : PairInsertExtractPatV4<v4f64, f64>;
 
+foreach imm1 = 0...7 in {
+  foreach imm2 = 0...7 in {
+    def : Pat<(vector_insert v8i32:$xd,
+                  (GRLenVT(vector_extract v8i32:$xj, imm1)), imm2),
+              (XVINSVE0_W v8i32:$xd, (XVPICKVE_W v8i32:$xj, imm1), imm2)>;
+  }
+}
+
+foreach imm1 = 0...3 in {
+  foreach imm2 = 0...3 in {
+    def : Pat<(vector_insert v4i64:$xd,
+                  (GRLenVT(vector_extract v4i64:$xj, imm1)), imm2),
+              (XVINSVE0_D v4i64:$xd, (XVPICKVE_D v4i64:$xj, imm1), imm2)>;
+  }
+}
+
 // PseudoXVINSGR2VR_{B/H}
 def : Pat<(vector_insert v32i8:$xd, GRLenVT:$rj, uimm5:$imm),
           (PseudoXVINSGR2VR_B v32i8:$xd, GRLenVT:$rj, uimm5:$imm)>;
diff --git a/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insert-extract-element.ll b/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insert-extract-element.ll
index ac5a2143451d0..91a1247cab89e 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insert-extract-element.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insert-extract-element.ll
@@ -30,8 +30,8 @@ entry:
 define <8 x i32> @insert_extract_v8i32(<8 x i32> %a) nounwind {
 ; CHECK-LABEL: insert_extract_v8i32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    xvpickve2gr.w $a0, $xr0, 7
-; CHECK-NEXT:    xvinsgr2vr.w $xr0, $a0, 1
+; CHECK-NEXT:    xvpickve.w $xr1, $xr0, 7
+; CHECK-NEXT:    xvinsve0.w $xr0, $xr1, 1
 ; CHECK-NEXT:    ret
 entry:
   %b = extractelement <8 x i32> %a, i32 7
@@ -54,8 +54,8 @@ entry:
 define <4 x i64> @insert_extract_v4i64(<4 x i64> %a) nounwind {
 ; CHECK-LABEL: insert_extract_v4i64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 3
-; CHECK-NEXT:    xvinsgr2vr.d $xr0, $a0, 1
+; CHECK-NEXT:    xvpickve.d $xr1, $xr0, 3
+; CHECK-NEXT:    xvinsve0.d $xr0, $xr1, 1
 ; CHECK-NEXT:    ret
 entry:
   %b = extractelement <4 x i64> %a, i32 3

>From b1b65af2d30b0e2ab74fb4ad7804f4e760b29e6c Mon Sep 17 00:00:00 2001
From: tangaac <tangyan01 at loongson.cn>
Date: Fri, 8 Aug 2025 11:48:05 +0800
Subject: [PATCH 2/2] remove uncessary imm

---
 .../LoongArch/LoongArchLASXInstrInfo.td       | 24 +++++++------------
 1 file changed, 9 insertions(+), 15 deletions(-)

diff --git a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
index 37d665010aa8c..1fcbbd62edbc5 100644
--- a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
+++ b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
@@ -1640,21 +1640,15 @@ defm : PairInsertExtractPatV8<v8f32, f32>;
 defm : PairInsertExtractPatV4<v4i64, GRLenVT>;
 defm : PairInsertExtractPatV4<v4f64, f64>;
 
-foreach imm1 = 0...7 in {
-  foreach imm2 = 0...7 in {
-    def : Pat<(vector_insert v8i32:$xd,
-                  (GRLenVT(vector_extract v8i32:$xj, imm1)), imm2),
-              (XVINSVE0_W v8i32:$xd, (XVPICKVE_W v8i32:$xj, imm1), imm2)>;
-  }
-}
-
-foreach imm1 = 0...3 in {
-  foreach imm2 = 0...3 in {
-    def : Pat<(vector_insert v4i64:$xd,
-                  (GRLenVT(vector_extract v4i64:$xj, imm1)), imm2),
-              (XVINSVE0_D v4i64:$xd, (XVPICKVE_D v4i64:$xj, imm1), imm2)>;
-  }
-}
+def : Pat<(vector_insert v8i32:$xd,
+              (GRLenVT(vector_extract v8i32:$xj, uimm3:$imm1)), uimm3:$imm2),
+          (XVINSVE0_W v8i32:$xd, (XVPICKVE_W v8i32:$xj, uimm3:$imm1),
+              uimm3:$imm2)>;
+
+def : Pat<(vector_insert v4i64:$xd,
+              (GRLenVT(vector_extract v4i64:$xj, uimm2:$imm1)), uimm2:$imm2),
+          (XVINSVE0_D v4i64:$xd, (XVPICKVE_D v4i64:$xj, uimm2:$imm1),
+              uimm2:$imm2)>;
 
 // PseudoXVINSGR2VR_{B/H}
 def : Pat<(vector_insert v32i8:$xd, GRLenVT:$rj, uimm5:$imm),