From bab04bf72b531ae78e5ef2a62f9baa517c93ea9a Mon Sep 17 00:00:00 2001
From: gonglingqin <gonglingqin@loongson.cn>
Date: Mon, 29 Aug 2022 09:57:18 +0800
Subject: [PATCH] [LoongArch] Combine add + shl to alsl.[w/d/wu]

Differential Revision: https://reviews.llvm.org/D132718
---
 llvm/lib/Target/LoongArch/LoongArchInstrInfo.td    |  16 +-
 llvm/test/CodeGen/LoongArch/alsl.ll                | 286 +++++++++++++++++++++
 .../LoongArch/ir-instruction/load-store-fp.ll      |   6 +-
 .../CodeGen/LoongArch/ir-instruction/load-store.ll |  24 +-
 4 files changed, 311 insertions(+), 21 deletions(-)
 create mode 100644 llvm/test/CodeGen/LoongArch/alsl.ll

diff --git a/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td
index 8566c63..4038636 100644
--- a/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td
+++ b/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td
@@ -87,7 +87,8 @@ def uimm2 : Operand<GRLenVT> {
   let ParserMatchClass = UImmAsmOperand<2>;
 }
 
-def uimm2_plus1 : Operand<GRLenVT> {
+def uimm2_plus1 : Operand<GRLenVT>,
+    ImmLeaf<GRLenVT, [{return isUInt<2>(Imm - 1);}]> {
   let ParserMatchClass = UImmAsmOperand<2, "plus1">;
   let EncoderMethod = "getImmOpValueSub1";
   let DecoderMethod = "decodeUImmOperand<2, 1>";
@@ -733,6 +734,19 @@ def : Pat<(AddLike (i64 BaseAddr:$rj), simm12:$imm12),
           (ADDI_D (i64 BaseAddr:$rj), simm12:$imm12)>;
 } // Predicates = [IsLA64]
 
+/// Shifted addition
+let Predicates = [IsLA32] in {
+def : Pat<(add GPR:$rk, (shl GPR:$rj, uimm2_plus1:$imm2)),
+          (ALSL_W GPR:$rj, GPR:$rk, uimm2_plus1:$imm2)>;
+} // Predicates = [IsLA32]
+let Predicates = [IsLA64] in {
+def : Pat<(add GPR:$rk, (shl GPR:$rj, uimm2_plus1:$imm2)),
+          (ALSL_D GPR:$rj, GPR:$rk, uimm2_plus1:$imm2)>;
+def : Pat<(loongarch_bstrpick (add GPR:$rk, (shl GPR:$rj, uimm2_plus1:$imm2)),
+                              (i64 31), (i64 0)),
+          (ALSL_WU GPR:$rj, GPR:$rk, uimm2_plus1:$imm2)>;
+} // Predicates = [IsLA64]
+
 /// Shift
 
 let Predicates = [IsLA32] in {
diff --git a/llvm/test/CodeGen/LoongArch/alsl.ll b/llvm/test/CodeGen/LoongArch/alsl.ll
new file mode 100644
index 0000000..cd6cd4e
--- /dev/null
+++ b/llvm/test/CodeGen/LoongArch/alsl.ll
@@ -0,0 +1,286 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc --mtriple=loongarch32 < %s | FileCheck %s --check-prefix=LA32
+; RUN: llc --mtriple=loongarch64 < %s | FileCheck %s --check-prefix=LA64
+
+define i8 @alsl_i8(i8 signext %a, i8 signext %b) nounwind {
+; LA32-LABEL: alsl_i8:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    alsl.w $a0, $a0, $a1, 1
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: alsl_i8:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    alsl.d $a0, $a0, $a1, 1
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i8 %a, 2
+  %add = add nsw i8 %b, %mul
+  ret i8 %add
+}
+
+define i16 @alsl_i16(i16 signext %a, i16 signext %b) nounwind {
+; LA32-LABEL: alsl_i16:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    alsl.w $a0, $a0, $a1, 2
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: alsl_i16:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    alsl.d $a0, $a0, $a1, 2
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i16 %a, 4
+  %add = add nsw i16 %b, %mul
+  ret i16 %add
+}
+
+define i32 @alsl_i32(i32 signext %a, i32 signext %b) nounwind {
+; LA32-LABEL: alsl_i32:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    alsl.w $a0, $a0, $a1, 3
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: alsl_i32:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    alsl.d $a0, $a0, $a1, 3
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i32 %a, 8
+  %add = add nsw i32 %b, %mul
+  ret i32 %add
+}
+
+define i64 @alsl_i64(i64 signext %a, i64 signext %b) nounwind {
+; LA32-LABEL: alsl_i64:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    slli.w $a1, $a1, 4
+; LA32-NEXT:    srli.w $a4, $a0, 28
+; LA32-NEXT:    or $a1, $a1, $a4
+; LA32-NEXT:    add.w $a1, $a3, $a1
+; LA32-NEXT:    alsl.w $a0, $a0, $a2, 4
+; LA32-NEXT:    sltu $a2, $a0, $a2
+; LA32-NEXT:    add.w $a1, $a1, $a2
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: alsl_i64:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    alsl.d $a0, $a0, $a1, 4
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i64 %a, 16
+  %add = add nsw i64 %b, %mul
+  ret i64 %add
+}
+
+define i32 @alsl_zext_i8(i8 signext %a, i8 signext %b) nounwind {
+; LA32-LABEL: alsl_zext_i8:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    alsl.w $a0, $a0, $a1, 1
+; LA32-NEXT:    andi $a0, $a0, 255
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: alsl_zext_i8:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    alsl.d $a0, $a0, $a1, 1
+; LA64-NEXT:    andi $a0, $a0, 255
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i8 %a, 2
+  %add = add nsw i8 %b, %mul
+  %zext = zext i8 %add to i32
+  ret i32 %zext
+}
+
+define i32 @alsl_zext_i16(i16 signext %a, i16 signext %b) nounwind {
+; LA32-LABEL: alsl_zext_i16:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    alsl.w $a0, $a0, $a1, 2
+; LA32-NEXT:    bstrpick.w $a0, $a0, 15, 0
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: alsl_zext_i16:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    alsl.d $a0, $a0, $a1, 2
+; LA64-NEXT:    bstrpick.d $a0, $a0, 15, 0
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i16 %a, 4
+  %add = add nsw i16 %b, %mul
+  %zext = zext i16 %add to i32
+  ret i32 %zext
+}
+
+define i64 @alsl_zext_i32(i32 signext %a, i32 signext %b) nounwind {
+; LA32-LABEL: alsl_zext_i32:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    alsl.w $a0, $a0, $a1, 3
+; LA32-NEXT:    move $a1, $zero
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: alsl_zext_i32:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    alsl.wu $a0, $a0, $a1, 3
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i32 %a, 8
+  %add = add nsw i32 %b, %mul
+  %zext = zext i32 %add to i64
+  ret i64 %zext
+}
+
+;; Check that alsl.w or alsl.d is not emitted.
+define i8 @mul_add_i8(i8 signext %a, i8 signext %b) nounwind {
+; LA32-LABEL: mul_add_i8:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    ori $a2, $zero, 3
+; LA32-NEXT:    mul.w $a0, $a0, $a2
+; LA32-NEXT:    add.w $a0, $a1, $a0
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: mul_add_i8:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    ori $a2, $zero, 3
+; LA64-NEXT:    mul.d $a0, $a0, $a2
+; LA64-NEXT:    add.d $a0, $a1, $a0
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i8 %a, 3
+  %add = add nsw i8 %b, %mul
+  ret i8 %add
+}
+
+define i16 @mul_add_i16(i16 signext %a, i16 signext %b) nounwind {
+; LA32-LABEL: mul_add_i16:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    ori $a2, $zero, 10
+; LA32-NEXT:    mul.w $a0, $a0, $a2
+; LA32-NEXT:    add.w $a0, $a1, $a0
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: mul_add_i16:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    ori $a2, $zero, 10
+; LA64-NEXT:    mul.d $a0, $a0, $a2
+; LA64-NEXT:    add.d $a0, $a1, $a0
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i16 %a, 10
+  %add = add nsw i16 %b, %mul
+  ret i16 %add
+}
+
+define i32 @mul_add_i32(i32 signext %a, i32 signext %b) nounwind {
+; LA32-LABEL: mul_add_i32:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    ori $a2, $zero, 12
+; LA32-NEXT:    mul.w $a0, $a0, $a2
+; LA32-NEXT:    add.w $a0, $a1, $a0
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: mul_add_i32:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    ori $a2, $zero, 12
+; LA64-NEXT:    mul.d $a0, $a0, $a2
+; LA64-NEXT:    add.d $a0, $a1, $a0
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i32 %a, 12
+  %add = add nsw i32 %b, %mul
+  ret i32 %add
+}
+
+define i64 @mul_add_i64(i64 signext %a, i64 signext %b) nounwind {
+; LA32-LABEL: mul_add_i64:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    ori $a4, $zero, 15
+; LA32-NEXT:    mul.w $a1, $a1, $a4
+; LA32-NEXT:    mulh.wu $a5, $a0, $a4
+; LA32-NEXT:    add.w $a1, $a5, $a1
+; LA32-NEXT:    add.w $a1, $a3, $a1
+; LA32-NEXT:    mul.w $a0, $a0, $a4
+; LA32-NEXT:    add.w $a0, $a2, $a0
+; LA32-NEXT:    sltu $a2, $a0, $a2
+; LA32-NEXT:    add.w $a1, $a1, $a2
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: mul_add_i64:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    ori $a2, $zero, 15
+; LA64-NEXT:    mul.d $a0, $a0, $a2
+; LA64-NEXT:    add.d $a0, $a1, $a0
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i64 %a, 15
+  %add = add nsw i64 %b, %mul
+  ret i64 %add
+}
+
+define i32 @mul_add_zext_i8(i8 signext %a, i8 signext %b) nounwind {
+; LA32-LABEL: mul_add_zext_i8:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    ori $a2, $zero, 5
+; LA32-NEXT:    mul.w $a0, $a0, $a2
+; LA32-NEXT:    add.w $a0, $a1, $a0
+; LA32-NEXT:    andi $a0, $a0, 255
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: mul_add_zext_i8:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    ori $a2, $zero, 5
+; LA64-NEXT:    mul.d $a0, $a0, $a2
+; LA64-NEXT:    add.d $a0, $a1, $a0
+; LA64-NEXT:    andi $a0, $a0, 255
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i8 %a, 5
+  %add = add nsw i8 %b, %mul
+  %zext = zext i8 %add to i32
+  ret i32 %zext
+}
+
+define i32 @mul_add_zext_i16(i16 signext %a, i16 signext %b) nounwind {
+; LA32-LABEL: mul_add_zext_i16:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    ori $a2, $zero, 15
+; LA32-NEXT:    mul.w $a0, $a0, $a2
+; LA32-NEXT:    add.w $a0, $a1, $a0
+; LA32-NEXT:    bstrpick.w $a0, $a0, 15, 0
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: mul_add_zext_i16:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    ori $a2, $zero, 15
+; LA64-NEXT:    mul.d $a0, $a0, $a2
+; LA64-NEXT:    add.d $a0, $a1, $a0
+; LA64-NEXT:    bstrpick.d $a0, $a0, 15, 0
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i16 %a, 15
+  %add = add nsw i16 %b, %mul
+  %zext = zext i16 %add to i32
+  ret i32 %zext
+}
+
+;; Check that alsl.wu is not emitted.
+define i64 @mul_add_zext_i32(i32 signext %a, i32 signext %b) nounwind {
+; LA32-LABEL: mul_add_zext_i32:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    ori $a2, $zero, 5
+; LA32-NEXT:    mul.w $a0, $a0, $a2
+; LA32-NEXT:    add.w $a0, $a1, $a0
+; LA32-NEXT:    move $a1, $zero
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: mul_add_zext_i32:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    ori $a2, $zero, 5
+; LA64-NEXT:    mul.d $a0, $a0, $a2
+; LA64-NEXT:    add.d $a0, $a1, $a0
+; LA64-NEXT:    bstrpick.d $a0, $a0, 31, 0
+; LA64-NEXT:    ret
+entry:
+  %mul = mul nsw i32 %a, 5
+  %add = add nsw i32 %b, %mul
+  %zext = zext i32 %add to i64
+  ret i64 %zext
+}
diff --git a/llvm/test/CodeGen/LoongArch/ir-instruction/load-store-fp.ll b/llvm/test/CodeGen/LoongArch/ir-instruction/load-store-fp.ll
index cba7ed5..3f392c5 100644
--- a/llvm/test/CodeGen/LoongArch/ir-instruction/load-store-fp.ll
+++ b/llvm/test/CodeGen/LoongArch/ir-instruction/load-store-fp.ll
@@ -36,8 +36,7 @@ define float @fldx_s(ptr %a, i64 %idx) nounwind {
 define double @fldx_d(ptr %a, i64 %idx) nounwind {
 ; LA32F-LABEL: fldx_d:
 ; LA32F:       # %bb.0:
-; LA32F-NEXT:    slli.w $a1, $a1, 3
-; LA32F-NEXT:    add.w $a1, $a0, $a1
+; LA32F-NEXT:    alsl.w $a1, $a1, $a0, 3
 ; LA32F-NEXT:    ld.w $a0, $a1, 0
 ; LA32F-NEXT:    ld.w $a1, $a1, 4
 ; LA32F-NEXT:    ret
@@ -96,8 +95,7 @@ define void @fstx_s(ptr %dst, i64 %idx, float %val) nounwind {
 define void @fstx_d(ptr %dst, i64 %idx, double %val) nounwind {
 ; LA32F-LABEL: fstx_d:
 ; LA32F:       # %bb.0:
-; LA32F-NEXT:    slli.w $a1, $a1, 3
-; LA32F-NEXT:    add.w $a0, $a0, $a1
+; LA32F-NEXT:    alsl.w $a0, $a1, $a0, 3
 ; LA32F-NEXT:    st.w $a4, $a0, 4
 ; LA32F-NEXT:    st.w $a3, $a0, 0
 ; LA32F-NEXT:    ret
diff --git a/llvm/test/CodeGen/LoongArch/ir-instruction/load-store.ll b/llvm/test/CodeGen/LoongArch/ir-instruction/load-store.ll
index 37a5e64..387a62bd 100644
--- a/llvm/test/CodeGen/LoongArch/ir-instruction/load-store.ll
+++ b/llvm/test/CodeGen/LoongArch/ir-instruction/load-store.ll
@@ -252,8 +252,7 @@ define i64 @ldx_b(ptr %a, i64 %idx) nounwind {
 define i64 @ldx_h(ptr %a, i64 %idx) nounwind {
 ; LA32-LABEL: ldx_h:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    slli.w $a1, $a1, 1
-; LA32-NEXT:    add.w $a1, $a0, $a1
+; LA32-NEXT:    alsl.w $a1, $a1, $a0, 1
 ; LA32-NEXT:    ld.h $a2, $a1, 0
 ; LA32-NEXT:    ld.h $a0, $a0, 0
 ; LA32-NEXT:    srai.w $a1, $a2, 31
@@ -277,8 +276,7 @@ define i64 @ldx_h(ptr %a, i64 %idx) nounwind {
 define i64 @ldx_w(ptr %a, i64 %idx) nounwind {
 ; LA32-LABEL: ldx_w:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    slli.w $a1, $a1, 2
-; LA32-NEXT:    add.w $a1, $a0, $a1
+; LA32-NEXT:    alsl.w $a1, $a1, $a0, 2
 ; LA32-NEXT:    ld.w $a2, $a1, 0
 ; LA32-NEXT:    ld.w $a0, $a0, 0
 ; LA32-NEXT:    srai.w $a1, $a2, 31
@@ -302,8 +300,7 @@ define i64 @ldx_w(ptr %a, i64 %idx) nounwind {
 define i64 @ldx_d(ptr %a, i64 %idx) nounwind {
 ; LA32-LABEL: ldx_d:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    slli.w $a1, $a1, 3
-; LA32-NEXT:    add.w $a1, $a0, $a1
+; LA32-NEXT:    alsl.w $a1, $a1, $a0, 3
 ; LA32-NEXT:    ld.w $a2, $a1, 0
 ; LA32-NEXT:    ld.w $a3, $a0, 0
 ; LA32-NEXT:    ld.w $a1, $a1, 4
@@ -352,8 +349,7 @@ define i64 @ldx_bu(ptr %a, i64 %idx) nounwind {
 define i64 @ldx_hu(ptr %a, i64 %idx) nounwind {
 ; LA32-LABEL: ldx_hu:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    slli.w $a1, $a1, 1
-; LA32-NEXT:    add.w $a1, $a0, $a1
+; LA32-NEXT:    alsl.w $a1, $a1, $a0, 1
 ; LA32-NEXT:    ld.hu $a1, $a1, 0
 ; LA32-NEXT:    ld.hu $a0, $a0, 0
 ; LA32-NEXT:    add.w $a0, $a1, $a0
@@ -379,8 +375,7 @@ define i64 @ldx_hu(ptr %a, i64 %idx) nounwind {
 define i64 @ldx_wu(ptr %a, i64 %idx) nounwind {
 ; LA32-LABEL: ldx_wu:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    slli.w $a1, $a1, 2
-; LA32-NEXT:    add.w $a1, $a0, $a1
+; LA32-NEXT:    alsl.w $a1, $a1, $a0, 2
 ; LA32-NEXT:    ld.w $a1, $a1, 0
 ; LA32-NEXT:    ld.w $a0, $a0, 0
 ; LA32-NEXT:    add.w $a0, $a1, $a0
@@ -480,8 +475,7 @@ define void @stx_b(ptr %dst, i64 %idx, i8 %val) nounwind {
 define void @stx_h(ptr %dst, i64 %idx, i16 %val) nounwind {
 ; LA32-LABEL: stx_h:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    slli.w $a1, $a1, 1
-; LA32-NEXT:    add.w $a0, $a0, $a1
+; LA32-NEXT:    alsl.w $a0, $a1, $a0, 1
 ; LA32-NEXT:    st.h $a3, $a0, 0
 ; LA32-NEXT:    ret
 ;
@@ -498,8 +492,7 @@ define void @stx_h(ptr %dst, i64 %idx, i16 %val) nounwind {
 define void @stx_w(ptr %dst, i64 %idx, i32 %val) nounwind {
 ; LA32-LABEL: stx_w:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    slli.w $a1, $a1, 2
-; LA32-NEXT:    add.w $a0, $a0, $a1
+; LA32-NEXT:    alsl.w $a0, $a1, $a0, 2
 ; LA32-NEXT:    st.w $a3, $a0, 0
 ; LA32-NEXT:    ret
 ;
@@ -516,8 +509,7 @@ define void @stx_w(ptr %dst, i64 %idx, i32 %val) nounwind {
 define void @stx_d(ptr %dst, i64 %idx, i64 %val) nounwind {
 ; LA32-LABEL: stx_d:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    slli.w $a1, $a1, 3
-; LA32-NEXT:    add.w $a0, $a0, $a1
+; LA32-NEXT:    alsl.w $a0, $a1, $a0, 3
 ; LA32-NEXT:    st.w $a4, $a0, 4
 ; LA32-NEXT:    st.w $a3, $a0, 0
 ; LA32-NEXT:    ret
-- 
2.7.4