[PATCH] D158495: [PowerPC] Try to use less instructions to materialize 64-bit constant when High32=Low32

[Esme Yi via Phabricator via llvm-commits]

Esme created this revision.
Esme added reviewers: qiongsiwu, shchenz, stefanp, PowerPC.
Herald added subscribers: kbarton, hiraditya, nemanjai.
Herald added a project: All.
Esme requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Materialization a 64-bit constant with High32=Low32 only requires 2 instructions instead of 3 when Low32 can be materialized in 1 instruction.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D158495

Files:
  llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
  llvm/test/CodeGen/PowerPC/constants-i64.ll


Index: llvm/test/CodeGen/PowerPC/constants-i64.ll
===================================================================
--- llvm/test/CodeGen/PowerPC/constants-i64.ll
+++ llvm/test/CodeGen/PowerPC/constants-i64.ll
@@ -424,4 +424,24 @@
   ret i64 82134617250843903 ;0x0123CCFFCCFFCCFF
 }
 
+define i64 @imm23() {
+; CHECK-LABEL: imm23:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    li 3, 32767
+; CHECK-NEXT:    rldimi 3, 3, 32, 0
+; CHECK-NEXT:    blr
+entry:
+  ret i64 140733193420799 ;0x00007FFF00007FFF
+}
+
+define i64 @imm24() {
+; CHECK-LABEL: imm24:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    lis 3, -9
+; CHECK-NEXT:    rldimi 3, 3, 32, 0
+; CHECK-NEXT:    blr
+entry:
+  ret i64 18444210803213533184 ;0xFFF70000FFF70000
+}
+
 attributes #0 = { nounwind readnone }
Index: llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
===================================================================
--- llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
+++ llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
@@ -1165,6 +1165,31 @@
     return CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, SDValue(Result, 0),
                                   getI32Imm(Shift), getI32Imm(0));
   }
+  // 2-7) Patterns : High word == Low word
+  // This may require 2 to 3 instructions, depending on whether Lo32 can be
+  // materialized in 1 instruction.
+  if (Hi32 == Lo32) {
+    // Handle the first 32 bits.
+    uint64_t ImmHi16 = (Lo32 >> 16) & 0xffff;
+    uint64_t ImmLo16 = Lo32 & 0xffff;
+    if (isInt<16>(Lo32))
+      Result =
+          CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(ImmLo16));
+    else if (!ImmLo16)
+      Result =
+          CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(ImmHi16));
+    else {
+      InstCnt = 3;
+      Result =
+          CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(ImmHi16));
+      Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
+                                      SDValue(Result, 0), getI32Imm(ImmLo16));
+    }
+    // Use rldimi to insert the Low word into High word.
+    SDValue Ops[] = {SDValue(Result, 0), SDValue(Result, 0), getI32Imm(32),
+                     getI32Imm(0)};
+    return CurDAG->getMachineNode(PPC::RLDIMI, dl, MVT::i64, Ops);
+  }
 
   // Following patterns use 3 instructions to materialize the Imm.
   InstCnt = 3;
@@ -1215,20 +1240,7 @@
     return CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, SDValue(Result, 0),
                                   getI32Imm(TO), getI32Imm(LZ));
   }
-  // 3-4) Patterns : High word == Low word
-  if (Hi32 == Lo32) {
-    // Handle the first 32 bits.
-    uint64_t ImmHi16 = (Lo32 >> 16) & 0xffff;
-    unsigned Opcode = ImmHi16 ? PPC::LIS8 : PPC::LI8;
-    Result = CurDAG->getMachineNode(Opcode, dl, MVT::i64, getI32Imm(ImmHi16));
-    Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, SDValue(Result, 0),
-                                    getI32Imm(Lo32 & 0xffff));
-    // Use rldimi to insert the Low word into High word.
-    SDValue Ops[] = {SDValue(Result, 0), SDValue(Result, 0), getI32Imm(32),
-                     getI32Imm(0)};
-    return CurDAG->getMachineNode(PPC::RLDIMI, dl, MVT::i64, Ops);
-  }
-  // 3-5) Patterns : {******}{33 zeros}{******}
+  // 3-4) Patterns : {******}{33 zeros}{******}
   //                 {******}{33 ones}{******}
   // If the Imm contains 33 consecutive zeros/ones, it means that a total of 31
   // bits remain on both sides. Rotate right the Imm to construct an int<32>


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D158495.552264.patch
Type: text/x-patch
Size: 3487 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230822/93092681/attachment.bin>