[llvm-branch-commits] [llvm] [LoongArch] Load suitable immediates with LU12I.W+SLLI.D (PR #242357)

Rong Mantle Bao via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Sat Nov 14 00:57:40 PST 2026


https://github.com/CSharperMantle updated https://github.com/llvm/llvm-project/pull/242357

>From 3fbedbb7eb5d0caf2851834197054f3510495b3a Mon Sep 17 00:00:00 2001
From: Rong Bao <rong.bao at csmantle.top>
Date: Fri, 13 Nov 2026 20:45:31 +0800
Subject: [PATCH 1/5] [LoongArch][NFC] Add baseline test cases for immediates
 loadable with LU12I.W+SLLI.D

---
 llvm/test/CodeGen/LoongArch/imm.ll | 59 ++++++++++++++++++++++++++++++
 1 file changed, 59 insertions(+)

diff --git a/llvm/test/CodeGen/LoongArch/imm.ll b/llvm/test/CodeGen/LoongArch/imm.ll
index aca508e99fb960..4e023e175beaaf 100644
--- a/llvm/test/CodeGen/LoongArch/imm.ll
+++ b/llvm/test/CodeGen/LoongArch/imm.ll
@@ -219,3 +219,62 @@ define i64 @imm110000014000000a() {
 ; CHECK-NEXT:    ret
   ret i64 1224979104013484042
 }
+
+define i64 @imm0008000000000000() {
+; CHECK-LABEL: imm0008000000000000:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ori $a0, $zero, 0
+; CHECK-NEXT:    lu32i.d $a0, -524288
+; CHECK-NEXT:    lu52i.d $a0, $a0, 0
+; CHECK-NEXT:    ret
+  ret i64 u0x0008000000000000
+}
+
+define i64 @imm000f800000000000() {
+; CHECK-LABEL: imm000f800000000000:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ori $a0, $zero, 0
+; CHECK-NEXT:    lu32i.d $a0, -32768
+; CHECK-NEXT:    lu52i.d $a0, $a0, 0
+; CHECK-NEXT:    ret
+  ret i64 u0x000f800000000000
+}
+
+define i64 @imm0000000100000000() {
+; CHECK-LABEL: imm0000000100000000:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ori $a0, $zero, 0
+; CHECK-NEXT:    lu32i.d $a0, 1
+; CHECK-NEXT:    ret
+  ret i64 u0x0000000100000000
+}
+
+define i64 @imm7ff8000000000000() {
+; CHECK-LABEL: imm7ff8000000000000:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ori $a0, $zero, 0
+; CHECK-NEXT:    lu32i.d $a0, -524288
+; CHECK-NEXT:    lu52i.d $a0, $a0, 2047
+; CHECK-NEXT:    ret
+  ret i64 u0x7ff8000000000000
+}
+
+define i64 @imm8000100000000000() {
+; CHECK-LABEL: imm8000100000000000:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ori $a0, $zero, 0
+; CHECK-NEXT:    lu32i.d $a0, 4096
+; CHECK-NEXT:    lu52i.d $a0, $a0, -2048
+; CHECK-NEXT:    ret
+  ret i64 u0x8000100000000000
+}
+
+define i64 @imm000fffff00000000() {
+; CHECK-LABEL: imm000fffff00000000:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ori $a0, $zero, 0
+; CHECK-NEXT:    lu32i.d $a0, -1
+; CHECK-NEXT:    lu52i.d $a0, $a0, 0
+; CHECK-NEXT:    ret
+  ret i64 u0x000fffff00000000
+}

>From f645d006c0c6349f10c79c62e74b96df77ddc822 Mon Sep 17 00:00:00 2001
From: Rong Bao <rong.bao at csmantle.top>
Date: Fri, 13 Nov 2026 20:41:27 +0800
Subject: [PATCH 2/5] [LoongArch] Load suitable immediates with LU12I.W+SLLI.D

---
 .../AsmParser/LoongArchAsmParser.cpp          |  1 +
 .../LoongArch/LoongArchISelDAGToDAG.cpp       |  1 +
 .../Target/LoongArch/LoongArchInstrInfo.cpp   |  1 +
 .../MCTargetDesc/LoongArchMatInt.cpp          | 15 ++++
 llvm/test/CodeGen/LoongArch/bittest.ll        |  4 +-
 .../CodeGen/LoongArch/calling-conv-lp64d.ll   |  5 +-
 .../CodeGen/LoongArch/fp-minimum-maximum.ll   | 14 ++--
 llvm/test/CodeGen/LoongArch/fptosi-sat.ll     | 14 ++--
 .../LoongArch/fsqrt-reciprocal-estimate.ll    | 60 ++++++++------
 llvm/test/CodeGen/LoongArch/imm.ll            | 24 +++---
 .../LoongArch/merge-base-offset-tlsle.ll      |  4 +-
 .../CodeGen/LoongArch/merge-base-offset.ll    | 15 ++--
 .../LoongArch/musttail-indirect-args.ll       |  5 +-
 llvm/test/CodeGen/LoongArch/vector-fp-imm.ll  | 79 +++++++++----------
 14 files changed, 134 insertions(+), 108 deletions(-)

diff --git a/llvm/lib/Target/LoongArch/AsmParser/LoongArchAsmParser.cpp b/llvm/lib/Target/LoongArch/AsmParser/LoongArchAsmParser.cpp
index 8041cbc0e707dc..a98b96ac590f11 100644
--- a/llvm/lib/Target/LoongArch/AsmParser/LoongArchAsmParser.cpp
+++ b/llvm/lib/Target/LoongArch/AsmParser/LoongArchAsmParser.cpp
@@ -1546,6 +1546,7 @@ void LoongArchAsmParser::emitLoadImm(MCInst &Inst, SMLoc IDLoc,
     case LoongArch::ORI:
     case LoongArch::LU32I_D:
     case LoongArch::LU52I_D:
+    case LoongArch::SLLI_D:
       Out.emitInstruction(
           MCInstBuilder(Inst.Opc).addReg(DestReg).addReg(SrcReg).addImm(
               Inst.Imm),
diff --git a/llvm/lib/Target/LoongArch/LoongArchISelDAGToDAG.cpp b/llvm/lib/Target/LoongArch/LoongArchISelDAGToDAG.cpp
index 924f360a3ea4db..bc4004afc18e36 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelDAGToDAG.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelDAGToDAG.cpp
@@ -71,6 +71,7 @@ void LoongArchDAGToDAGISel::Select(SDNode *Node) {
       case LoongArch::ORI:
       case LoongArch::LU32I_D:
       case LoongArch::LU52I_D:
+      case LoongArch::SLLI_D:
         Result = CurDAG->getMachineNode(Inst.Opc, DL, GRLenVT, SrcReg, SDImm);
         break;
       case LoongArch::BSTRINS_D:
diff --git a/llvm/lib/Target/LoongArch/LoongArchInstrInfo.cpp b/llvm/lib/Target/LoongArch/LoongArchInstrInfo.cpp
index a4946e1bac9fb2..45bb7759b30b86 100644
--- a/llvm/lib/Target/LoongArch/LoongArchInstrInfo.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchInstrInfo.cpp
@@ -291,6 +291,7 @@ void LoongArchInstrInfo::movImm(MachineBasicBlock &MBB,
     case LoongArch::ORI:
     case LoongArch::LU32I_D: // "rj" is needed due to InstrInfo pattern
     case LoongArch::LU52I_D:
+    case LoongArch::SLLI_D:
       BuildMI(MBB, MBBI, DL, get(Inst.Opc), DstReg)
           .addReg(SrcReg, RegState::Kill)
           .addImm(Inst.Imm)
diff --git a/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchMatInt.cpp b/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchMatInt.cpp
index 6343817a0616d1..f041326e65d8d0 100644
--- a/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchMatInt.cpp
+++ b/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchMatInt.cpp
@@ -32,6 +32,21 @@ LoongArchMatInt::InstSeq LoongArchMatInt::generateInstSeq(int64_t Val) {
     return Insts;
   }
 
+  // Values with lo32==0 and hi32 containing <=19 significant bits
+  // ("significant" as in "significant digits") can be loaded by SLLI.D-ing them
+  // into place.
+  // https://github.com/v8/v8/blob/cd5440417ebc15558b3fc626fb7c0cf1b6f97c66/src/codegen/loong64/macro-assembler-loong64.cc#L2083-L2085
+  if ((Val & 0xFFFFFFFF) == 0) {
+    const int32_t Hi32 = Val >> 32;
+    const int TZC = llvm::countr_zero(static_cast<uint32_t>(Hi32));
+    const int LZC = llvm::countl_zero(static_cast<uint32_t>(Hi32));
+    if (TZC < 20 && ((TZC + LZC > 12) || (LZC == 0 && TZC + LZC >= 12))) {
+      Insts.push_back(Inst(LoongArch::LU12I_W, Hi32 >> TZC));
+      Insts.push_back(Inst(LoongArch::SLLI_D, TZC + 20));
+      return Insts;
+    }
+  }
+
   // lo32
   if (Hi20 == 0)
     Insts.push_back(Inst(LoongArch::ORI, Lo12));
diff --git a/llvm/test/CodeGen/LoongArch/bittest.ll b/llvm/test/CodeGen/LoongArch/bittest.ll
index 6d7c05a16b5a0e..2b16330d495b4d 100644
--- a/llvm/test/CodeGen/LoongArch/bittest.ll
+++ b/llvm/test/CodeGen/LoongArch/bittest.ll
@@ -749,8 +749,8 @@ define i64 @bit_32_z_select_i64(i64 %a, i64 %b, i64 %c) {
 ;
 ; LA64-LABEL: bit_32_z_select_i64:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    ori $a3, $zero, 0
-; LA64-NEXT:    lu32i.d $a3, 1
+; LA64-NEXT:    lu12i.w $a3, 1
+; LA64-NEXT:    slli.d $a3, $a3, 20
 ; LA64-NEXT:    and $a0, $a0, $a3
 ; LA64-NEXT:    maskeqz $a2, $a2, $a0
 ; LA64-NEXT:    masknez $a0, $a1, $a0
diff --git a/llvm/test/CodeGen/LoongArch/calling-conv-lp64d.ll b/llvm/test/CodeGen/LoongArch/calling-conv-lp64d.ll
index 733c53697b987d..64baa8e9cfbe74 100644
--- a/llvm/test/CodeGen/LoongArch/calling-conv-lp64d.ll
+++ b/llvm/test/CodeGen/LoongArch/calling-conv-lp64d.ll
@@ -64,9 +64,8 @@ define i64 @caller_double_in_gpr_exhausted_fprs() nounwind {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    addi.d $sp, $sp, -16
 ; CHECK-NEXT:    st.d $ra, $sp, 8 # 8-byte Folded Spill
-; CHECK-NEXT:    ori $a0, $zero, 0
-; CHECK-NEXT:    lu32i.d $a0, 131072
-; CHECK-NEXT:    lu52i.d $a0, $a0, 1026
+; CHECK-NEXT:    lu12i.w $a0, 8209
+; CHECK-NEXT:    slli.d $a0, $a0, 37
 ; CHECK-NEXT:    vldi $vr0, -912
 ; CHECK-NEXT:    vldi $vr1, -1024
 ; CHECK-NEXT:    vldi $vr2, -1016
diff --git a/llvm/test/CodeGen/LoongArch/fp-minimum-maximum.ll b/llvm/test/CodeGen/LoongArch/fp-minimum-maximum.ll
index fce256d4615d65..a08d2264123ac0 100644
--- a/llvm/test/CodeGen/LoongArch/fp-minimum-maximum.ll
+++ b/llvm/test/CodeGen/LoongArch/fp-minimum-maximum.ll
@@ -177,11 +177,12 @@ define double @minimum_f64(double %a, double %b) {
 ;
 ; LA64D-LABEL: minimum_f64:
 ; LA64D:       # %bb.0:
-; LA64D-NEXT:    pcalau12i $a0, %pc_hi20(.LCPI2_0)
-; LA64D-NEXT:    fld.d $fa2, $a0, %pc_lo12(.LCPI2_0)
 ; LA64D-NEXT:    fcmp.cun.d $fcc0, $fa0, $fa1
 ; LA64D-NEXT:    fmin.d $fa0, $fa0, $fa1
-; LA64D-NEXT:    fsel $fa0, $fa0, $fa2, $fcc0
+; LA64D-NEXT:    lu12i.w $a0, 4095
+; LA64D-NEXT:    slli.d $a0, $a0, 39
+; LA64D-NEXT:    movgr2fr.d $fa1, $a0
+; LA64D-NEXT:    fsel $fa0, $fa0, $fa1, $fcc0
 ; LA64D-NEXT:    ret
 ;
 ; LA32-SOFT-LABEL: minimum_f64:
@@ -246,11 +247,12 @@ define double @maximum_f64(double %a, double %b) {
 ;
 ; LA64D-LABEL: maximum_f64:
 ; LA64D:       # %bb.0:
-; LA64D-NEXT:    pcalau12i $a0, %pc_hi20(.LCPI3_0)
-; LA64D-NEXT:    fld.d $fa2, $a0, %pc_lo12(.LCPI3_0)
 ; LA64D-NEXT:    fcmp.cun.d $fcc0, $fa0, $fa1
 ; LA64D-NEXT:    fmax.d $fa0, $fa0, $fa1
-; LA64D-NEXT:    fsel $fa0, $fa0, $fa2, $fcc0
+; LA64D-NEXT:    lu12i.w $a0, 4095
+; LA64D-NEXT:    slli.d $a0, $a0, 39
+; LA64D-NEXT:    movgr2fr.d $fa1, $a0
+; LA64D-NEXT:    fsel $fa0, $fa0, $fa1, $fcc0
 ; LA64D-NEXT:    ret
 ;
 ; LA32-SOFT-LABEL: maximum_f64:
diff --git a/llvm/test/CodeGen/LoongArch/fptosi-sat.ll b/llvm/test/CodeGen/LoongArch/fptosi-sat.ll
index 382dae0f88daff..5c9bf0fc48cee2 100644
--- a/llvm/test/CodeGen/LoongArch/fptosi-sat.ll
+++ b/llvm/test/CodeGen/LoongArch/fptosi-sat.ll
@@ -379,9 +379,8 @@ define i8 @fptosi_sat_f64_to_i8(double %a) nounwind {
 ; LA64F-NEXT:    addi.w $a1, $zero, -128
 ; LA64F-NEXT:    maskeqz $a1, $a1, $s0
 ; LA64F-NEXT:    or $s0, $a1, $a0
-; LA64F-NEXT:    ori $a0, $zero, 0
-; LA64F-NEXT:    lu32i.d $a0, -16384
-; LA64F-NEXT:    lu52i.d $a1, $a0, 1029
+; LA64F-NEXT:    lu12i.w $a0, 65919
+; LA64F-NEXT:    slli.d $a1, $a0, 34
 ; LA64F-NEXT:    move $a0, $fp
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__gtdf2)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
@@ -403,12 +402,13 @@ define i8 @fptosi_sat_f64_to_i8(double %a) nounwind {
 ;
 ; LA64D-LABEL: fptosi_sat_f64_to_i8:
 ; LA64D:       # %bb.0: # %entry
-; LA64D-NEXT:    pcalau12i $a0, %pc_hi20(.LCPI4_0)
-; LA64D-NEXT:    fld.d $fa1, $a0, %pc_lo12(.LCPI4_0)
 ; LA64D-NEXT:    lu52i.d $a0, $zero, -1018
+; LA64D-NEXT:    movgr2fr.d $fa1, $a0
+; LA64D-NEXT:    fmax.d $fa1, $fa0, $fa1
+; LA64D-NEXT:    lu12i.w $a0, 65919
+; LA64D-NEXT:    slli.d $a0, $a0, 34
 ; LA64D-NEXT:    movgr2fr.d $fa2, $a0
-; LA64D-NEXT:    fmax.d $fa2, $fa0, $fa2
-; LA64D-NEXT:    fmin.d $fa1, $fa2, $fa1
+; LA64D-NEXT:    fmin.d $fa1, $fa1, $fa2
 ; LA64D-NEXT:    ftintrz.l.d $fa1, $fa1
 ; LA64D-NEXT:    fcmp.cun.d $fcc0, $fa0, $fa0
 ; LA64D-NEXT:    movfr2gr.d $a0, $fa1
diff --git a/llvm/test/CodeGen/LoongArch/fsqrt-reciprocal-estimate.ll b/llvm/test/CodeGen/LoongArch/fsqrt-reciprocal-estimate.ll
index df0387b87fe134..1d2fd0c5fd48b7 100644
--- a/llvm/test/CodeGen/LoongArch/fsqrt-reciprocal-estimate.ll
+++ b/llvm/test/CodeGen/LoongArch/fsqrt-reciprocal-estimate.ll
@@ -190,10 +190,11 @@ define double @sqrt_simplify_before_recip_3_uses_f64(double %x, ptr %p1, ptr %p2
 ;
 ; LA64D-LABEL: sqrt_simplify_before_recip_3_uses_f64:
 ; LA64D:       # %bb.0:
-; LA64D-NEXT:    pcalau12i $a2, %pc_hi20(.LCPI2_0)
-; LA64D-NEXT:    fld.d $fa2, $a2, %pc_lo12(.LCPI2_0)
 ; LA64D-NEXT:    fsqrt.d $fa1, $fa0
 ; LA64D-NEXT:    frsqrt.d $fa0, $fa0
+; LA64D-NEXT:    lu12i.w $a2, 16453
+; LA64D-NEXT:    slli.d $a2, $a2, 36
+; LA64D-NEXT:    movgr2fr.d $fa2, $a2
 ; LA64D-NEXT:    fdiv.d $fa2, $fa2, $fa1
 ; LA64D-NEXT:    fst.d $fa0, $a0, 0
 ; LA64D-NEXT:    fst.d $fa2, $a1, 0
@@ -210,12 +211,13 @@ define double @sqrt_simplify_before_recip_3_uses_f64(double %x, ptr %p1, ptr %p2
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa4
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa2
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa2, $fa0, $fa1
-; LA64D-FRECIPE-NEXT:    pcalau12i $a2, %pc_hi20(.LCPI2_0)
-; LA64D-FRECIPE-NEXT:    fld.d $fa5, $a2, %pc_lo12(.LCPI2_0)
 ; LA64D-FRECIPE-NEXT:    fmadd.d $fa2, $fa2, $fa1, $fa3
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa4
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa2
-; LA64D-FRECIPE-NEXT:    fmul.d $fa2, $fa1, $fa5
+; LA64D-FRECIPE-NEXT:    lu12i.w $a2, 16453
+; LA64D-FRECIPE-NEXT:    slli.d $a2, $a2, 36
+; LA64D-FRECIPE-NEXT:    movgr2fr.d $fa2, $a2
+; LA64D-FRECIPE-NEXT:    fmul.d $fa2, $fa1, $fa2
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa0, $fa0, $fa1
 ; LA64D-FRECIPE-NEXT:    fst.d $fa1, $a0, 0
 ; LA64D-FRECIPE-NEXT:    fst.d $fa2, $a1, 0
@@ -319,12 +321,14 @@ define double @sqrt_simplify_before_recip_3_uses_order_f64(double %x, ptr %p1, p
 ;
 ; LA64D-LABEL: sqrt_simplify_before_recip_3_uses_order_f64:
 ; LA64D:       # %bb.0:
-; LA64D-NEXT:    pcalau12i $a2, %pc_hi20(.LCPI3_0)
-; LA64D-NEXT:    fld.d $fa1, $a2, %pc_lo12(.LCPI3_0)
-; LA64D-NEXT:    pcalau12i $a2, %pc_hi20(.LCPI3_1)
-; LA64D-NEXT:    fld.d $fa2, $a2, %pc_lo12(.LCPI3_1)
 ; LA64D-NEXT:    fsqrt.d $fa0, $fa0
+; LA64D-NEXT:    lu12i.w $a2, 16453
+; LA64D-NEXT:    slli.d $a2, $a2, 36
+; LA64D-NEXT:    movgr2fr.d $fa1, $a2
 ; LA64D-NEXT:    fdiv.d $fa1, $fa1, $fa0
+; LA64D-NEXT:    lu12i.w $a2, 32907
+; LA64D-NEXT:    slli.d $a2, $a2, 35
+; LA64D-NEXT:    movgr2fr.d $fa2, $a2
 ; LA64D-NEXT:    fdiv.d $fa2, $fa2, $fa0
 ; LA64D-NEXT:    fst.d $fa1, $a0, 0
 ; LA64D-NEXT:    fst.d $fa2, $a1, 0
@@ -342,14 +346,16 @@ define double @sqrt_simplify_before_recip_3_uses_order_f64(double %x, ptr %p1, p
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa2, $fa0, $fa1
 ; LA64D-FRECIPE-NEXT:    fmadd.d $fa2, $fa2, $fa1, $fa3
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa4
-; LA64D-FRECIPE-NEXT:    pcalau12i $a2, %pc_hi20(.LCPI3_0)
-; LA64D-FRECIPE-NEXT:    fld.d $fa3, $a2, %pc_lo12(.LCPI3_0)
-; LA64D-FRECIPE-NEXT:    pcalau12i $a2, %pc_hi20(.LCPI3_1)
-; LA64D-FRECIPE-NEXT:    fld.d $fa4, $a2, %pc_lo12(.LCPI3_1)
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa2
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa0, $fa0, $fa1
-; LA64D-FRECIPE-NEXT:    fmul.d $fa2, $fa1, $fa3
-; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa4
+; LA64D-FRECIPE-NEXT:    lu12i.w $a2, 16453
+; LA64D-FRECIPE-NEXT:    slli.d $a2, $a2, 36
+; LA64D-FRECIPE-NEXT:    movgr2fr.d $fa2, $a2
+; LA64D-FRECIPE-NEXT:    fmul.d $fa2, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    lu12i.w $a2, 32907
+; LA64D-FRECIPE-NEXT:    slli.d $a2, $a2, 35
+; LA64D-FRECIPE-NEXT:    movgr2fr.d $fa3, $a2
+; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa3
 ; LA64D-FRECIPE-NEXT:    fst.d $fa2, $a0, 0
 ; LA64D-FRECIPE-NEXT:    fst.d $fa1, $a1, 0
 ; LA64D-FRECIPE-NEXT:    ret
@@ -483,13 +489,15 @@ define double @sqrt_simplify_before_recip_4_uses_f64(double %x, ptr %p1, ptr %p2
 ;
 ; LA64D-LABEL: sqrt_simplify_before_recip_4_uses_f64:
 ; LA64D:       # %bb.0:
-; LA64D-NEXT:    pcalau12i $a3, %pc_hi20(.LCPI4_0)
-; LA64D-NEXT:    fld.d $fa2, $a3, %pc_lo12(.LCPI4_0)
-; LA64D-NEXT:    pcalau12i $a3, %pc_hi20(.LCPI4_1)
-; LA64D-NEXT:    fld.d $fa3, $a3, %pc_lo12(.LCPI4_1)
 ; LA64D-NEXT:    fsqrt.d $fa1, $fa0
 ; LA64D-NEXT:    frsqrt.d $fa0, $fa0
+; LA64D-NEXT:    lu12i.w $a3, 16453
+; LA64D-NEXT:    slli.d $a3, $a3, 36
+; LA64D-NEXT:    movgr2fr.d $fa2, $a3
 ; LA64D-NEXT:    fdiv.d $fa2, $fa2, $fa1
+; LA64D-NEXT:    lu12i.w $a3, 32907
+; LA64D-NEXT:    slli.d $a3, $a3, 35
+; LA64D-NEXT:    movgr2fr.d $fa3, $a3
 ; LA64D-NEXT:    fdiv.d $fa3, $fa3, $fa1
 ; LA64D-NEXT:    fst.d $fa0, $a0, 0
 ; LA64D-NEXT:    fst.d $fa2, $a1, 0
@@ -508,14 +516,16 @@ define double @sqrt_simplify_before_recip_4_uses_f64(double %x, ptr %p1, ptr %p2
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa2
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa2, $fa0, $fa1
 ; LA64D-FRECIPE-NEXT:    fmadd.d $fa2, $fa2, $fa1, $fa3
-; LA64D-FRECIPE-NEXT:    pcalau12i $a3, %pc_hi20(.LCPI4_0)
-; LA64D-FRECIPE-NEXT:    fld.d $fa3, $a3, %pc_lo12(.LCPI4_0)
-; LA64D-FRECIPE-NEXT:    pcalau12i $a3, %pc_hi20(.LCPI4_1)
-; LA64D-FRECIPE-NEXT:    fld.d $fa5, $a3, %pc_lo12(.LCPI4_1)
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa4
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa1, $fa1, $fa2
-; LA64D-FRECIPE-NEXT:    fmul.d $fa2, $fa1, $fa3
-; LA64D-FRECIPE-NEXT:    fmul.d $fa3, $fa1, $fa5
+; LA64D-FRECIPE-NEXT:    lu12i.w $a3, 16453
+; LA64D-FRECIPE-NEXT:    slli.d $a3, $a3, 36
+; LA64D-FRECIPE-NEXT:    movgr2fr.d $fa2, $a3
+; LA64D-FRECIPE-NEXT:    fmul.d $fa2, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    lu12i.w $a3, 32907
+; LA64D-FRECIPE-NEXT:    slli.d $a3, $a3, 35
+; LA64D-FRECIPE-NEXT:    movgr2fr.d $fa3, $a3
+; LA64D-FRECIPE-NEXT:    fmul.d $fa3, $fa1, $fa3
 ; LA64D-FRECIPE-NEXT:    fmul.d $fa0, $fa0, $fa1
 ; LA64D-FRECIPE-NEXT:    fst.d $fa1, $a0, 0
 ; LA64D-FRECIPE-NEXT:    fst.d $fa2, $a1, 0
diff --git a/llvm/test/CodeGen/LoongArch/imm.ll b/llvm/test/CodeGen/LoongArch/imm.ll
index 4e023e175beaaf..06a61b58da312d 100644
--- a/llvm/test/CodeGen/LoongArch/imm.ll
+++ b/llvm/test/CodeGen/LoongArch/imm.ll
@@ -223,9 +223,8 @@ define i64 @imm110000014000000a() {
 define i64 @imm0008000000000000() {
 ; CHECK-LABEL: imm0008000000000000:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    ori $a0, $zero, 0
-; CHECK-NEXT:    lu32i.d $a0, -524288
-; CHECK-NEXT:    lu52i.d $a0, $a0, 0
+; CHECK-NEXT:    lu12i.w $a0, 1
+; CHECK-NEXT:    slli.d $a0, $a0, 39
 ; CHECK-NEXT:    ret
   ret i64 u0x0008000000000000
 }
@@ -233,9 +232,8 @@ define i64 @imm0008000000000000() {
 define i64 @imm000f800000000000() {
 ; CHECK-LABEL: imm000f800000000000:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    ori $a0, $zero, 0
-; CHECK-NEXT:    lu32i.d $a0, -32768
-; CHECK-NEXT:    lu52i.d $a0, $a0, 0
+; CHECK-NEXT:    lu12i.w $a0, 31
+; CHECK-NEXT:    slli.d $a0, $a0, 35
 ; CHECK-NEXT:    ret
   ret i64 u0x000f800000000000
 }
@@ -243,8 +241,8 @@ define i64 @imm000f800000000000() {
 define i64 @imm0000000100000000() {
 ; CHECK-LABEL: imm0000000100000000:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    ori $a0, $zero, 0
-; CHECK-NEXT:    lu32i.d $a0, 1
+; CHECK-NEXT:    lu12i.w $a0, 1
+; CHECK-NEXT:    slli.d $a0, $a0, 20
 ; CHECK-NEXT:    ret
   ret i64 u0x0000000100000000
 }
@@ -252,9 +250,8 @@ define i64 @imm0000000100000000() {
 define i64 @imm7ff8000000000000() {
 ; CHECK-LABEL: imm7ff8000000000000:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    ori $a0, $zero, 0
-; CHECK-NEXT:    lu32i.d $a0, -524288
-; CHECK-NEXT:    lu52i.d $a0, $a0, 2047
+; CHECK-NEXT:    lu12i.w $a0, 4095
+; CHECK-NEXT:    slli.d $a0, $a0, 39
 ; CHECK-NEXT:    ret
   ret i64 u0x7ff8000000000000
 }
@@ -262,9 +259,8 @@ define i64 @imm7ff8000000000000() {
 define i64 @imm8000100000000000() {
 ; CHECK-LABEL: imm8000100000000000:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    ori $a0, $zero, 0
-; CHECK-NEXT:    lu32i.d $a0, 4096
-; CHECK-NEXT:    lu52i.d $a0, $a0, -2048
+; CHECK-NEXT:    lu12i.w $a0, -524287
+; CHECK-NEXT:    slli.d $a0, $a0, 32
 ; CHECK-NEXT:    ret
   ret i64 u0x8000100000000000
 }
diff --git a/llvm/test/CodeGen/LoongArch/merge-base-offset-tlsle.ll b/llvm/test/CodeGen/LoongArch/merge-base-offset-tlsle.ll
index fd1468308ec626..c1283f7fd3f9bd 100644
--- a/llvm/test/CodeGen/LoongArch/merge-base-offset-tlsle.ll
+++ b/llvm/test/CodeGen/LoongArch/merge-base-offset-tlsle.ll
@@ -784,8 +784,8 @@ define dso_local ptr @tlsle_load_addr_offset_281474439839744() nounwind {
 ; LA64-NEXT:    lu12i.w $a0, %le_hi20_r(g_a64)
 ; LA64-NEXT:    add.d $a0, $a0, $tp, %le_add_r(g_a64)
 ; LA64-NEXT:    addi.d $a0, $a0, %le_lo12_r(g_a64)
-; LA64-NEXT:    ori $a1, $zero, 0
-; LA64-NEXT:    lu32i.d $a1, 524287
+; LA64-NEXT:    lu12i.w $a1, 524287
+; LA64-NEXT:    slli.d $a1, $a1, 20
 ; LA64-NEXT:    add.d $a0, $a0, $a1
 ; LA64-NEXT:    ret
 entry:
diff --git a/llvm/test/CodeGen/LoongArch/merge-base-offset.ll b/llvm/test/CodeGen/LoongArch/merge-base-offset.ll
index 9a2a6abcdae9a8..2408c3e5328c0c 100644
--- a/llvm/test/CodeGen/LoongArch/merge-base-offset.ll
+++ b/llvm/test/CodeGen/LoongArch/merge-base-offset.ll
@@ -1129,18 +1129,21 @@ define dso_local ptr @load_addr_offset_281474439839744() nounwind {
 ; LA64:       # %bb.0: # %entry
 ; LA64-NEXT:    pcalau12i $a0, %pc_hi20(g_a64)
 ; LA64-NEXT:    addi.d $a0, $a0, %pc_lo12(g_a64)
-; LA64-NEXT:    ori $a1, $zero, 0
-; LA64-NEXT:    lu32i.d $a1, 524287
+; LA64-NEXT:    lu12i.w $a1, 524287
+; LA64-NEXT:    slli.d $a1, $a1, 20
 ; LA64-NEXT:    add.d $a0, $a0, $a1
 ; LA64-NEXT:    ret
 ;
 ; LA64-LARGE-LABEL: load_addr_offset_281474439839744:
 ; LA64-LARGE:       # %bb.0: # %entry
-; LA64-LARGE-NEXT:    pcalau12i $a0, %pc_hi20(g_a64+2251795518717952)
-; LA64-LARGE-NEXT:    addi.d $a1, $zero, %pc_lo12(g_a64+2251795518717952)
-; LA64-LARGE-NEXT:    lu32i.d $a1, %pc64_lo20(g_a64+2251795518717952)
-; LA64-LARGE-NEXT:    lu52i.d $a1, $a1, %pc64_hi12(g_a64+2251795518717952)
+; LA64-LARGE-NEXT:    pcalau12i $a0, %pc_hi20(g_a64)
+; LA64-LARGE-NEXT:    addi.d $a1, $zero, %pc_lo12(g_a64)
+; LA64-LARGE-NEXT:    lu32i.d $a1, %pc64_lo20(g_a64)
+; LA64-LARGE-NEXT:    lu52i.d $a1, $a1, %pc64_hi12(g_a64)
 ; LA64-LARGE-NEXT:    add.d $a0, $a1, $a0
+; LA64-LARGE-NEXT:    lu12i.w $a1, 524287
+; LA64-LARGE-NEXT:    slli.d $a1, $a1, 20
+; LA64-LARGE-NEXT:    add.d $a0, $a0, $a1
 ; LA64-LARGE-NEXT:    ret
 entry:
   ret ptr getelementptr inbounds ([1 x i64], ptr @g_a64, i64 281474439839744)
diff --git a/llvm/test/CodeGen/LoongArch/musttail-indirect-args.ll b/llvm/test/CodeGen/LoongArch/musttail-indirect-args.ll
index d088d6065aa07b..dcdedd33f409e6 100644
--- a/llvm/test/CodeGen/LoongArch/musttail-indirect-args.ll
+++ b/llvm/test/CodeGen/LoongArch/musttail-indirect-args.ll
@@ -43,9 +43,8 @@ define void @caller_no_musttail_indirect() nounwind {
 ;
 ; LA64-LABEL: caller_no_musttail_indirect:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    ori $a0, $zero, 0
-; LA64-NEXT:    lu32i.d $a0, -65536
-; LA64-NEXT:    lu52i.d $a1, $a0, 1023
+; LA64-NEXT:    lu12i.w $a0, 16383
+; LA64-NEXT:    slli.d $a1, $a0, 36
 ; LA64-NEXT:    move $a0, $zero
 ; LA64-NEXT:    pcaddu18i $t8, %call36(callee_musttail_indirect)
 ; LA64-NEXT:    jr $t8
diff --git a/llvm/test/CodeGen/LoongArch/vector-fp-imm.ll b/llvm/test/CodeGen/LoongArch/vector-fp-imm.ll
index a16af2bbb6596e..f29b5ace08e56d 100644
--- a/llvm/test/CodeGen/LoongArch/vector-fp-imm.ll
+++ b/llvm/test/CodeGen/LoongArch/vector-fp-imm.ll
@@ -554,22 +554,21 @@ define void @test_d4(ptr %P, ptr %S) nounwind {
 ; LA64F-NEXT:    st.d $s2, $sp, 8 # 8-byte Folded Spill
 ; LA64F-NEXT:    st.d $s3, $sp, 0 # 8-byte Folded Spill
 ; LA64F-NEXT:    ld.d $fp, $a0, 24
-; LA64F-NEXT:    ld.d $s0, $a0, 8
-; LA64F-NEXT:    ld.d $s1, $a0, 0
-; LA64F-NEXT:    ld.d $a0, $a0, 16
+; LA64F-NEXT:    ld.d $s0, $a0, 16
+; LA64F-NEXT:    ld.d $s1, $a0, 8
+; LA64F-NEXT:    ld.d $a0, $a0, 0
 ; LA64F-NEXT:    move $s2, $a1
-; LA64F-NEXT:    ori $a1, $zero, 0
-; LA64F-NEXT:    lu32i.d $a1, -524288
-; LA64F-NEXT:    lu52i.d $a1, $a1, 1024
+; LA64F-NEXT:    lu52i.d $a1, $zero, 1023
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
 ; LA64F-NEXT:    move $s3, $a0
-; LA64F-NEXT:    lu52i.d $a1, $zero, 1023
+; LA64F-NEXT:    lu52i.d $a1, $zero, 1024
 ; LA64F-NEXT:    move $a0, $s1
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
 ; LA64F-NEXT:    move $s1, $a0
-; LA64F-NEXT:    lu52i.d $a1, $zero, 1024
+; LA64F-NEXT:    lu12i.w $a0, 2049
+; LA64F-NEXT:    slli.d $a1, $a0, 39
 ; LA64F-NEXT:    move $a0, $s0
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
@@ -579,9 +578,9 @@ define void @test_d4(ptr %P, ptr %S) nounwind {
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
 ; LA64F-NEXT:    st.d $a0, $s2, 24
-; LA64F-NEXT:    st.d $s0, $s2, 8
-; LA64F-NEXT:    st.d $s1, $s2, 0
-; LA64F-NEXT:    st.d $s3, $s2, 16
+; LA64F-NEXT:    st.d $s0, $s2, 16
+; LA64F-NEXT:    st.d $s1, $s2, 8
+; LA64F-NEXT:    st.d $s3, $s2, 0
 ; LA64F-NEXT:    ld.d $s3, $sp, 0 # 8-byte Folded Reload
 ; LA64F-NEXT:    ld.d $s2, $sp, 8 # 8-byte Folded Reload
 ; LA64F-NEXT:    ld.d $s1, $sp, 16 # 8-byte Folded Reload
@@ -794,67 +793,67 @@ define void @test_d8(ptr %P, ptr %S) nounwind {
 ; LA64F-NEXT:    st.d $s8, $sp, 24 # 8-byte Folded Spill
 ; LA64F-NEXT:    ld.d $a2, $a0, 56
 ; LA64F-NEXT:    st.d $a2, $sp, 16 # 8-byte Folded Spill
-; LA64F-NEXT:    ld.d $s1, $a0, 40
-; LA64F-NEXT:    ld.d $s2, $a0, 32
-; LA64F-NEXT:    ld.d $s3, $a0, 24
-; LA64F-NEXT:    ld.d $s4, $a0, 8
-; LA64F-NEXT:    ld.d $s5, $a0, 0
-; LA64F-NEXT:    ld.d $s6, $a0, 48
-; LA64F-NEXT:    ld.d $a0, $a0, 16
+; LA64F-NEXT:    ld.d $a2, $a0, 48
+; LA64F-NEXT:    st.d $a2, $sp, 0 # 8-byte Folded Spill
+; LA64F-NEXT:    ld.d $s2, $a0, 40
+; LA64F-NEXT:    ld.d $s3, $a0, 32
+; LA64F-NEXT:    ld.d $s4, $a0, 24
+; LA64F-NEXT:    ld.d $s5, $a0, 16
+; LA64F-NEXT:    ld.d $s6, $a0, 8
+; LA64F-NEXT:    ld.d $a0, $a0, 0
 ; LA64F-NEXT:    move $fp, $a1
-; LA64F-NEXT:    ori $a1, $zero, 0
-; LA64F-NEXT:    lu32i.d $a1, -524288
-; LA64F-NEXT:    lu52i.d $s7, $a1, 1024
+; LA64F-NEXT:    lu52i.d $s7, $zero, 1023
 ; LA64F-NEXT:    move $a1, $s7
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
 ; LA64F-NEXT:    st.d $a0, $sp, 8 # 8-byte Folded Spill
+; LA64F-NEXT:    lu52i.d $s0, $zero, 1024
 ; LA64F-NEXT:    move $a0, $s6
-; LA64F-NEXT:    move $a1, $s7
+; LA64F-NEXT:    move $a1, $s0
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
 ; LA64F-NEXT:    move $s6, $a0
-; LA64F-NEXT:    lu52i.d $s7, $zero, 1023
+; LA64F-NEXT:    lu12i.w $a0, 2049
+; LA64F-NEXT:    slli.d $s8, $a0, 39
 ; LA64F-NEXT:    move $a0, $s5
-; LA64F-NEXT:    move $a1, $s7
+; LA64F-NEXT:    move $a1, $s8
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
 ; LA64F-NEXT:    move $s5, $a0
-; LA64F-NEXT:    lu52i.d $s0, $zero, 1024
+; LA64F-NEXT:    lu52i.d $s1, $zero, 1025
 ; LA64F-NEXT:    move $a0, $s4
-; LA64F-NEXT:    move $a1, $s0
+; LA64F-NEXT:    move $a1, $s1
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
 ; LA64F-NEXT:    move $s4, $a0
-; LA64F-NEXT:    lu52i.d $s8, $zero, 1025
 ; LA64F-NEXT:    move $a0, $s3
-; LA64F-NEXT:    move $a1, $s8
+; LA64F-NEXT:    move $a1, $s7
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
 ; LA64F-NEXT:    move $s3, $a0
 ; LA64F-NEXT:    move $a0, $s2
-; LA64F-NEXT:    move $a1, $s7
-; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
-; LA64F-NEXT:    jirl $ra, $ra, 0
-; LA64F-NEXT:    move $s2, $a0
-; LA64F-NEXT:    move $a0, $s1
 ; LA64F-NEXT:    move $a1, $s0
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
 ; LA64F-NEXT:    move $s0, $a0
-; LA64F-NEXT:    ld.d $a0, $sp, 16 # 8-byte Folded Reload
+; LA64F-NEXT:    ld.d $a0, $sp, 0 # 8-byte Folded Reload
 ; LA64F-NEXT:    move $a1, $s8
 ; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
 ; LA64F-NEXT:    jirl $ra, $ra, 0
+; LA64F-NEXT:    move $s2, $a0
+; LA64F-NEXT:    ld.d $a0, $sp, 16 # 8-byte Folded Reload
+; LA64F-NEXT:    move $a1, $s1
+; LA64F-NEXT:    pcaddu18i $ra, %call36(__adddf3)
+; LA64F-NEXT:    jirl $ra, $ra, 0
 ; LA64F-NEXT:    st.d $a0, $fp, 56
+; LA64F-NEXT:    st.d $s2, $fp, 48
 ; LA64F-NEXT:    st.d $s0, $fp, 40
-; LA64F-NEXT:    st.d $s2, $fp, 32
-; LA64F-NEXT:    st.d $s3, $fp, 24
-; LA64F-NEXT:    st.d $s4, $fp, 8
-; LA64F-NEXT:    st.d $s5, $fp, 0
-; LA64F-NEXT:    st.d $s6, $fp, 48
+; LA64F-NEXT:    st.d $s3, $fp, 32
+; LA64F-NEXT:    st.d $s4, $fp, 24
+; LA64F-NEXT:    st.d $s5, $fp, 16
+; LA64F-NEXT:    st.d $s6, $fp, 8
 ; LA64F-NEXT:    ld.d $a0, $sp, 8 # 8-byte Folded Reload
-; LA64F-NEXT:    st.d $a0, $fp, 16
+; LA64F-NEXT:    st.d $a0, $fp, 0
 ; LA64F-NEXT:    ld.d $s8, $sp, 24 # 8-byte Folded Reload
 ; LA64F-NEXT:    ld.d $s7, $sp, 32 # 8-byte Folded Reload
 ; LA64F-NEXT:    ld.d $s6, $sp, 40 # 8-byte Folded Reload

>From 006399177b6c06276e7b03d3c1053c637a36b651 Mon Sep 17 00:00:00 2001
From: Rong Bao <rong.bao at csmantle.top>
Date: Fri, 13 Nov 2026 21:26:15 +0800
Subject: [PATCH 3/5] [LoongArch] Teach
 LoongArchMergeBaseOffsetOpt::foldLargeOffset() about SLLI.D

---
 .../LoongArch/LoongArchMergeBaseOffset.cpp    | 45 ++++++++++++++-----
 .../CodeGen/LoongArch/merge-base-offset.ll    | 11 ++---
 2 files changed, 38 insertions(+), 18 deletions(-)

diff --git a/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp b/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp
index 93f5850bce569f..22c3852e7948d4 100644
--- a/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp
@@ -341,10 +341,13 @@ void LoongArchMergeBaseOffsetOpt::foldOffset(
 //        |     OffsetLo12: ori voff, vreg3, 122 !                    |
 //        |     OffsetLo20: lu32i.d voff, 123       ------------------+
 //        +-> 4) Offset that has non zero bits in Hi12 bits:          |
-//              OffsetHi20: lu12i.w vreg3, 121 !                      |
-//              OffsetLo12: ori voff, vreg3, 122 !                    |
-//              OffsetLo20: lu32i.d vreg3, 123 !                      |
-//              OffsetHi12: lu52i.d voff, vrg3, 124 ------------------+
+//        |     OffsetHi20: lu12i.w vreg3, 121 !                      |
+//        |     OffsetLo12: ori voff, vreg3, 122 !                    |
+//        |     OffsetLo20: lu32i.d vreg3, 123 !                      |
+//        |     OffsetHi12: lu52i.d voff, vreg3, 124 -----------------+
+//        +-> 5) Offset that is a field shifted into place:           |
+//                          lu12i.w voff, 4                           |
+//                          slli.d voff, voff, 20   ------------------+
 //                                                                    |
 //        TailAdd: add.d  vreg4, vreg2, voff       <------------------+
 //
@@ -361,8 +364,9 @@ bool LoongArchMergeBaseOffsetOpt::foldLargeOffset(
   SmallVector<MachineInstr *, 4> Instrs;
   int64_t Offset = 0;
   int64_t Mask = -1;
+  int Shift = 0;
 
-  // This can point to one of [ORI, LU12I.W, LU32I.D, LU52I.D]:
+  // This can point to one of [ORI, LU12I.W, LU32I.D, LU52I.D, SLLI.D]:
   for (int i = 0; i < 4; i++) {
     // Handle Reg is R0.
     if (Reg == LoongArch::R0)
@@ -384,7 +388,9 @@ bool LoongArchMergeBaseOffsetOpt::foldLargeOffset(
       MachineOperand ImmOp = Curr->getOperand(2);
       if (ImmOp.getTargetFlags() != LoongArchII::MO_None)
         return false;
-      Offset += ImmOp.getImm();
+      const int64_t Imm = ImmOp.getImm();
+      Offset += (static_cast<uint64_t>(Imm) << Shift) & Mask;
+      Mask &= ~(0x0000000000000FFFULL << Shift);
       Reg = Curr->getOperand(1).getReg();
       Instrs.push_back(Curr);
       break;
@@ -393,7 +399,8 @@ bool LoongArchMergeBaseOffsetOpt::foldLargeOffset(
       MachineOperand ImmOp = Curr->getOperand(1);
       if (ImmOp.getTargetFlags() != LoongArchII::MO_None)
         return false;
-      Offset += SignExtend64<32>(ImmOp.getImm() << 12) & Mask;
+      const int64_t Imm = SignExtend64<32>(ImmOp.getImm() << 12);
+      Offset += (static_cast<uint64_t>(Imm) << Shift) & Mask;
       Reg = LoongArch::R0;
       Instrs.push_back(Curr);
       break;
@@ -402,8 +409,9 @@ bool LoongArchMergeBaseOffsetOpt::foldLargeOffset(
       MachineOperand ImmOp = Curr->getOperand(2);
       if (ImmOp.getTargetFlags() != LoongArchII::MO_None || !Lo20)
         return false;
-      Offset += SignExtend64<52>(ImmOp.getImm() << 32) & Mask;
-      Mask &= ~0xFFFFFFFF00000000ULL;
+      const int64_t Imm = SignExtend64<52>(ImmOp.getImm() << 32);
+      Offset += (static_cast<uint64_t>(Imm) << Shift) & Mask;
+      Mask &= ~(0xFFFFFFFF00000000ULL << Shift);
       Reg = Curr->getOperand(1).getReg();
       Instrs.push_back(Curr);
       break;
@@ -412,8 +420,20 @@ bool LoongArchMergeBaseOffsetOpt::foldLargeOffset(
       MachineOperand ImmOp = Curr->getOperand(2);
       if (ImmOp.getTargetFlags() != LoongArchII::MO_None || !Hi12)
         return false;
-      Offset += ImmOp.getImm() << 52;
-      Mask &= ~0xFFF0000000000000ULL;
+      const int64_t Imm = ImmOp.getImm();
+      Offset += (static_cast<uint64_t>(Imm) << 52) << Shift;
+      Mask &= ~(0xFFF0000000000000ULL << Shift);
+      Reg = Curr->getOperand(1).getReg();
+      Instrs.push_back(Curr);
+      break;
+    }
+    case LoongArch::SLLI_D: {
+      MachineOperand ImmOp = Curr->getOperand(2);
+      if (ImmOp.getTargetFlags() != LoongArchII::MO_None || !(Lo20 && Hi12))
+        return false;
+      Shift += ImmOp.getImm();
+      if (Shift >= 64)
+        return false;
       Reg = Curr->getOperand(1).getReg();
       Instrs.push_back(Curr);
       break;
@@ -424,6 +444,9 @@ bool LoongArchMergeBaseOffsetOpt::foldLargeOffset(
   // Can't fold if the offset is not extracted.
   if (!Offset)
     return false;
+  // Can't fold if the current register does not reach R0.
+  if (Reg != LoongArch::R0)
+    return false;
 
   foldOffset(Hi20, Lo12, Lo20, Hi12, Last, TailAdd, Offset);
   LLVM_DEBUG(dbgs() << "  Offset Instrs:\n");
diff --git a/llvm/test/CodeGen/LoongArch/merge-base-offset.ll b/llvm/test/CodeGen/LoongArch/merge-base-offset.ll
index 2408c3e5328c0c..a958f5026c70aa 100644
--- a/llvm/test/CodeGen/LoongArch/merge-base-offset.ll
+++ b/llvm/test/CodeGen/LoongArch/merge-base-offset.ll
@@ -1136,14 +1136,11 @@ define dso_local ptr @load_addr_offset_281474439839744() nounwind {
 ;
 ; LA64-LARGE-LABEL: load_addr_offset_281474439839744:
 ; LA64-LARGE:       # %bb.0: # %entry
-; LA64-LARGE-NEXT:    pcalau12i $a0, %pc_hi20(g_a64)
-; LA64-LARGE-NEXT:    addi.d $a1, $zero, %pc_lo12(g_a64)
-; LA64-LARGE-NEXT:    lu32i.d $a1, %pc64_lo20(g_a64)
-; LA64-LARGE-NEXT:    lu52i.d $a1, $a1, %pc64_hi12(g_a64)
+; LA64-LARGE-NEXT:    pcalau12i $a0, %pc_hi20(g_a64+2251795518717952)
+; LA64-LARGE-NEXT:    addi.d $a1, $zero, %pc_lo12(g_a64+2251795518717952)
+; LA64-LARGE-NEXT:    lu32i.d $a1, %pc64_lo20(g_a64+2251795518717952)
+; LA64-LARGE-NEXT:    lu52i.d $a1, $a1, %pc64_hi12(g_a64+2251795518717952)
 ; LA64-LARGE-NEXT:    add.d $a0, $a1, $a0
-; LA64-LARGE-NEXT:    lu12i.w $a1, 524287
-; LA64-LARGE-NEXT:    slli.d $a1, $a1, 20
-; LA64-LARGE-NEXT:    add.d $a0, $a0, $a1
 ; LA64-LARGE-NEXT:    ret
 entry:
   ret ptr getelementptr inbounds ([1 x i64], ptr @g_a64, i64 281474439839744)

>From f41cd90d68e8b07a97b975122d67b24bed630e83 Mon Sep 17 00:00:00 2001
From: Rong Bao <rong.bao at csmantle.top>
Date: Sat, 14 Nov 2026 10:55:30 +0800
Subject: [PATCH 4/5] [LoongArch][NFC] Fix "i" -> "I" stylistically

---
 llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp b/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp
index 22c3852e7948d4..8cb954bbe44bd8 100644
--- a/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp
@@ -367,7 +367,7 @@ bool LoongArchMergeBaseOffsetOpt::foldLargeOffset(
   int Shift = 0;
 
   // This can point to one of [ORI, LU12I.W, LU32I.D, LU52I.D, SLLI.D]:
-  for (int i = 0; i < 4; i++) {
+  for (int I = 0; I < 4; I++) {
     // Handle Reg is R0.
     if (Reg == LoongArch::R0)
       break;

>From 5f0ab9fa9d53ec9face5794c71022988722f722d Mon Sep 17 00:00:00 2001
From: Rong Bao <rong.bao at csmantle.top>
Date: Sat, 14 Nov 2026 16:53:15 +0800
Subject: [PATCH 5/5] Make the LU52I.D case look like others

---
 llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp b/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp
index 8cb954bbe44bd8..216cf472ad1443 100644
--- a/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchMergeBaseOffset.cpp
@@ -420,8 +420,8 @@ bool LoongArchMergeBaseOffsetOpt::foldLargeOffset(
       MachineOperand ImmOp = Curr->getOperand(2);
       if (ImmOp.getTargetFlags() != LoongArchII::MO_None || !Hi12)
         return false;
-      const int64_t Imm = ImmOp.getImm();
-      Offset += (static_cast<uint64_t>(Imm) << 52) << Shift;
+      const int64_t Imm = ImmOp.getImm() << 52;
+      Offset += (static_cast<uint64_t>(Imm) << Shift) & Mask;
       Mask &= ~(0xFFF0000000000000ULL << Shift);
       Reg = Curr->getOperand(1).getReg();
       Instrs.push_back(Curr);



More information about the llvm-branch-commits mailing list