[LoongArch] Implement isLegalAddressingMode for lsx/lasx

[zhaoqi5]

TODO: Only add tests for general vector load and store. Maybe consider adding more cases?

[llvmbot]

@llvm/pr-subscribers-backend-loongarch

Author: ZhaoQi (zhaoqi5)

Changes

TODO: Only add tests for general vector load and store. Maybe consider adding more cases?

---

Full diff: https://github.com/llvm/llvm-project/pull/151917.diff

3 Files Affected:

• (modified) llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp (+19-3)
• (modified) llvm/test/CodeGen/LoongArch/lasx/loop-reduce.ll (+9-10)
• (modified) llvm/test/CodeGen/LoongArch/lsx/loop-reduce.ll (+5-6)

diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index a5bf0e57e3053..529a6d7063106 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -8385,7 +8385,14 @@ bool LoongArchTargetLowering::isLegalAddressingMode(const DataLayout &DL,
   //  2. reg + 12-bit signed offset
   //  3. reg + 14-bit signed offset left-shifted by 2
   //  4. reg1 + reg2
-  // TODO: Add more checks after support vector extension.
+  //
+  // LoongArch LSX/LASX vector extension has three basic addressing modes:
+  //  1. reg + 12-bit signed offset
+  //  2. reg + {12/11/10/9}-bit or 8-bit signed offset multiplied by element
+  //  width. ie.
+  //     si8,  si8<<1,  si8<<2, si8<<3
+  //    si12, si11<<1, si10<<2, si9<<3
+  //  3. reg1 + reg2
 
   // No global is ever allowed as a base.
   if (AM.BaseGV)
@@ -8393,8 +8400,17 @@ bool LoongArchTargetLowering::isLegalAddressingMode(const DataLayout &DL,
 
   // Require a 12-bit signed offset or 14-bit signed offset left-shifted by 2
   // with `UAL` feature.
-  if (!isInt<12>(AM.BaseOffs) &&
-      !(isShiftedInt<14, 2>(AM.BaseOffs) && Subtarget.hasUAL()))
+  if (!isInt<12>(AM.BaseOffs) && !(isShiftedInt<14, 2>(AM.BaseOffs) &&
+                                   Subtarget.hasUAL() && !isa<VectorType>(Ty)))
+    return false;
+
+  // FIXME: Is it necessary and possible to perform fine-grained processing
+  // according to vector element types?
+  if (Subtarget.hasExtLSX() && isa<VectorType>(Ty) &&
+      !(isInt<8>(AM.BaseOffs) || isShiftedInt<8, 1>(AM.BaseOffs) ||
+        isShiftedInt<8, 2>(AM.BaseOffs) || isShiftedInt<8, 3>(AM.BaseOffs) ||
+        isShiftedInt<11, 1>(AM.BaseOffs) || isShiftedInt<10, 2>(AM.BaseOffs) ||
+        isShiftedInt<9, 3>(AM.BaseOffs)))
     return false;
 
   switch (AM.Scale) {
diff --git a/llvm/test/CodeGen/LoongArch/lasx/loop-reduce.ll b/llvm/test/CodeGen/LoongArch/lasx/loop-reduce.ll
index 9739d3012bf3c..6c8bfd2e287e7 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/loop-reduce.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/loop-reduce.ll
@@ -10,26 +10,25 @@
 define dso_local void @foo() local_unnamed_addr {
 ; CHECK-LABEL: foo:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    lu12i.w $a0, -2
-; CHECK-NEXT:    pcalau12i $a1, %pc_hi20(.LCPI0_0)
-; CHECK-NEXT:    xvld $xr0, $a1, %pc_lo12(.LCPI0_0)
-; CHECK-NEXT:    pcalau12i $a1, %pc_hi20(.LCPI0_1)
-; CHECK-NEXT:    xvld $xr1, $a1, %pc_lo12(.LCPI0_1)
+; CHECK-NEXT:    pcalau12i $a0, %pc_hi20(.LCPI0_0)
+; CHECK-NEXT:    xvld $xr0, $a0, %pc_lo12(.LCPI0_0)
+; CHECK-NEXT:    pcalau12i $a0, %pc_hi20(.LCPI0_1)
+; CHECK-NEXT:    xvld $xr1, $a0, %pc_lo12(.LCPI0_1)
+; CHECK-NEXT:    ori $a0, $zero, 1024
 ; CHECK-NEXT:    pcalau12i $a1, %pc_hi20(a)
 ; CHECK-NEXT:    addi.d $a1, $a1, %pc_lo12(a)
-; CHECK-NEXT:    lu12i.w $a2, 2
 ; CHECK-NEXT:    .p2align 4, , 16
 ; CHECK-NEXT:  .LBB0_1: # %vector.body
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    add.d $a3, $a1, $a0
-; CHECK-NEXT:    xvldx $xr2, $a3, $a2
+; CHECK-NEXT:    xvld $xr2, $a1, 0
 ; CHECK-NEXT:    xvpermi.d $xr3, $xr2, 78
 ; CHECK-NEXT:    xvori.b $xr4, $xr0, 0
 ; CHECK-NEXT:    xvshuf.d $xr4, $xr0, $xr3
 ; CHECK-NEXT:    xvori.b $xr3, $xr1, 0
 ; CHECK-NEXT:    xvshuf.w $xr3, $xr4, $xr2
-; CHECK-NEXT:    addi.d $a0, $a0, 16
-; CHECK-NEXT:    xvstx $xr3, $a3, $a2
+; CHECK-NEXT:    xvst $xr3, $a1, 0
+; CHECK-NEXT:    addi.d $a0, $a0, -2
+; CHECK-NEXT:    addi.d $a1, $a1, 16
 ; CHECK-NEXT:    bnez $a0, .LBB0_1
 ; CHECK-NEXT:  # %bb.2: # %for.end
 ; CHECK-NEXT:    ret
diff --git a/llvm/test/CodeGen/LoongArch/lsx/loop-reduce.ll b/llvm/test/CodeGen/LoongArch/lsx/loop-reduce.ll
index fbb4d060c9412..9d3be735efa12 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/loop-reduce.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/loop-reduce.ll
@@ -10,18 +10,17 @@
 define dso_local void @foo() local_unnamed_addr {
 ; CHECK-LABEL: foo:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    lu12i.w $a0, -2
+; CHECK-NEXT:    ori $a0, $zero, 1024
 ; CHECK-NEXT:    pcalau12i $a1, %pc_hi20(a)
 ; CHECK-NEXT:    addi.d $a1, $a1, %pc_lo12(a)
-; CHECK-NEXT:    lu12i.w $a2, 2
 ; CHECK-NEXT:    .p2align 4, , 16
 ; CHECK-NEXT:  .LBB0_1: # %vector.body
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    add.d $a3, $a1, $a0
-; CHECK-NEXT:    vldx $vr0, $a3, $a2
+; CHECK-NEXT:    vld $vr0, $a1, 0
 ; CHECK-NEXT:    vshuf4i.w $vr0, $vr0, 9
-; CHECK-NEXT:    addi.d $a0, $a0, 16
-; CHECK-NEXT:    vstx $vr0, $a3, $a2
+; CHECK-NEXT:    vst $vr0, $a1, 0
+; CHECK-NEXT:    addi.d $a0, $a0, -2
+; CHECK-NEXT:    addi.d $a1, $a1, 16
 ; CHECK-NEXT:    bnez $a0, .LBB0_1
 ; CHECK-NEXT:  # %bb.2: # %for.end
 ; CHECK-NEXT:    ret

[zhaoqi5]

Influence to llvm-test-suite compiling using lasx: zhaoqi5/llvm-test-suite@ea9235a.