[Mips] Fix fast isel for i16 bswap. (llvm#103398)

We need to mask the SRL result to 8 bits before ORing in the SLL. This
is needed in case bits 23:16 of the input aren't zero. They will have
been shifted into bits 15:8.

We don't need to AND the result with 0xffff. It's ok if the upper 16
bits of the register are garbage.

Fixes llvm#103035.

Author: topperc

llvm/lib/Target/Mips/MipsFastISel.cpp

@@ -1608,8 +1608,8 @@ bool MipsFastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
         }
         emitInst(Mips::SLL, TempReg[0]).addReg(SrcReg).addImm(8);
         emitInst(Mips::SRL, TempReg[1]).addReg(SrcReg).addImm(8);
-        emitInst(Mips::OR, TempReg[2]).addReg(TempReg[0]).addReg(TempReg[1]);
-        emitInst(Mips::ANDi, DestReg).addReg(TempReg[2]).addImm(0xFFFF);
+        emitInst(Mips::ANDi, TempReg[2]).addReg(TempReg[1]).addImm(0xFF);
+        emitInst(Mips::OR, DestReg).addReg(TempReg[0]).addReg(TempReg[2]);
         updateValueMap(II, DestReg);
         return true;
       }

llvm/test/CodeGen/Mips/Fast-ISel/bswap1.ll

@@ -21,8 +21,8 @@ define void @b16() {
 
   ; 32R1:           sll   $[[TMP1:[0-9]+]], $[[A_VAL]], 8
   ; 32R1:           srl   $[[TMP2:[0-9]+]], $[[A_VAL]], 8
-  ; 32R1:           or    $[[TMP3:[0-9]+]], $[[TMP1]], $[[TMP2]]
-  ; 32R1:           andi  $[[TMP4:[0-9]+]], $[[TMP3]], 65535
+  ; 32R1:           andi  $[[TMP3:[0-9]+]], $[[TMP2]], 255
+  ; 32R1:           or    $[[RESULT:[0-9]+]], $[[TMP1]], $[[TMP3]]
 
   ; 32R2:           wsbh  $[[RESULT:[0-9]+]], $[[A_VAL]]