[PowerPC] mark r+i as legal address mode for vector type after pwr9

Reviewed By: steven.zhang

Differential Revision: https://reviews.llvm.org/D84735

diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 85f1630..dfe0c2d 100644 (file)
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -15128,9 +15128,15 @@ void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
 // by AM is legal for this target, for a load/store of the specified type.
 bool PPCTargetLowering::isLegalAddressingMode(const DataLayout &DL,
                                               const AddrMode &AM, Type *Ty,
-                                              unsigned AS, Instruction *I) const {
-  // PPC does not allow r+i addressing modes for vectors!
-  if (Ty->isVectorTy() && AM.BaseOffs != 0)
+                                              unsigned AS,
+                                              Instruction *I) const {
+  // Vector type r+i form is supported since power9 as DQ form. We don't check
+  // the offset matching DQ form requirement(off % 16 == 0), because on PowerPC,
+  // imm form is preferred and the offset can be adjusted to use imm form later
+  // in pass PPCLoopInstrFormPrep. Also in LSR, for one LSRUse, it uses min and
+  // max offset to check legal addressing mode, we should be a little aggressive
+  // to contain other offsets for that LSRUse.
+  if (Ty->isVectorTy() && AM.BaseOffs != 0 && !Subtarget.hasP9Vector())
     return false;
 
   // PPC allows a sign-extended 16-bit immediate field.

diff --git a/llvm/test/CodeGen/PowerPC/prefer-dqform.ll b/llvm/test/CodeGen/PowerPC/prefer-dqform.ll
index 0d19927..79e6026 100644 (file)
--- a/llvm/test/CodeGen/PowerPC/prefer-dqform.ll
+++ b/llvm/test/CodeGen/PowerPC/prefer-dqform.ll
@@ -12,26 +12,27 @@ target triple = "powerpc64le-unknown-linux-gnu"
 define void @test(i32* dereferenceable(4) %.ial, i32* noalias dereferenceable(4) %.m, i32* noalias dereferenceable(4) %.n, [0 x %_elem_type_of_a]*  %.a, i32* noalias dereferenceable(4) %.lda, [0 x %_elem_type_of_x]* noalias %.x, [0 x %_elem_type_of_y]* noalias %.y) {
 ; CHECK-P9-LABEL: test:
 ; CHECK-P9:       .LBB0_2: # %_loop_2_do_
-; CHECK-P9:         lxvx
-; CHECK-P9:         lxvx
-; CHECK-P9-DAG:     lxvx
-; CHECK-P9-DAG:     lxvx
-; CHECK-P9-DAG:     xvmaddadp
-; CHECK-P9-DAG:     xvmaddadp
-; CHECK-P9-DAG:     stxvx
-; CHECK-P9:         stxvx
+; CHECK-P9:         lxv vs1, -16(r4)
+; CHECK-P9:         lxv vs2, 0(r4)
+; CHECK-P9-DAG:     lxv vs3, -16(r3)
+; CHECK-P9-DAG:     lxv vs4, 0(r3)
+; CHECK-P9-DAG:     xvmaddadp vs1, vs3, vs1
+; CHECK-P9-DAG:     stxv vs1, -16(r4)
+; CHECK-P9-DAG:     xvmaddadp vs2, vs4, vs0
+; CHECK-P9:         stxv vs2, 0(r4)
 ; CHECK-P9:         bdnz .LBB0_2
 ;
+; FIXME: use pair load/store instructions lxvp/stxvp
 ; CHECK-P10-LABEL: test:
 ; CHECK-P10:       .LBB0_2: # %_loop_2_do_
-; CHECK-P10:         lxvx
-; CHECK-P10:         lxvx
-; CHECK-P10-DAG:     lxvx
-; CHECK-P10-DAG:     lxvx
-; CHECK-P10-DAG:     xvmaddadp
-; CHECK-P10-DAG:     xvmaddadp
-; CHECK-P10-DAG:     stxvx
-; CHECK-P10:         stxvx
+; CHECK-P10:         lxv vs1, -16(r4)
+; CHECK-P10:         lxv vs2, 0(r4)
+; CHECK-P10-DAG:     lxv vs3, -16(r3)
+; CHECK-P10-DAG:     lxv vs4, 0(r3)
+; CHECK-P10-DAG:     xvmaddadp vs1, vs3, vs1
+; CHECK-P10-DAG:     xvmaddadp vs2, vs4, vs0
+; CHECK-P10-DAG:     stxv vs1, -16(r4)
+; CHECK-P10:         stxv vs2, 0(r4)
 ; CHECK-P10:         bdnz .LBB0_2
 test_entry:
   %_conv5 = ptrtoint [0 x %_elem_type_of_a]* %.a to i64