[SDag][ARM][RISCV] Allow lowering CTPOP into a libcall (#101786)

This is a reland of #99752 with the bug fixed (see test diff in the
third commit in this PR).
All `popcount` libcalls return `int`, but `ISD::CTPOP` returns the type
of the argument, which can be wider than `int`. The fix is to make DAG
legalizer pass the correct return type to `makeLibCall` and sign-extend
the result afterwards.

Original commit message:
The main change is adding CTPOP to `RuntimeLibcalls.def` to allow
targets to use LibCall action for CTPOP. DAG legalizers are changed
accordingly.

Pull Request: llvm/llvm-project#101786

Author: s-barannikov

llvm/include/llvm/IR/RuntimeLibcalls.def

@@ -136,6 +136,9 @@ HANDLE_LIBCALL(CTLZ_I16, "__clzhi2")
 HANDLE_LIBCALL(CTLZ_I32, "__clzsi2")
 HANDLE_LIBCALL(CTLZ_I64, "__clzdi2")
 HANDLE_LIBCALL(CTLZ_I128, "__clzti2")
+HANDLE_LIBCALL(CTPOP_I32, "__popcountsi2")
+HANDLE_LIBCALL(CTPOP_I64, "__popcountdi2")
+HANDLE_LIBCALL(CTPOP_I128, "__popcountti2")
 HANDLE_LIBCALL(BSWAP_I8, "__bswapqi2")
 HANDLE_LIBCALL(BSWAP_I16, "__bswaphi2")
 HANDLE_LIBCALL(BSWAP_I32, "__bswapsi2")

llvm/lib/Target/ARM/ARMISelLowering.cpp

@@ -1221,10 +1221,8 @@ ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::ROTR, VT, Expand);
   }
   setOperationAction(ISD::CTTZ,  MVT::i32, Custom);
-  // TODO: These two should be set to LibCall, but this currently breaks
-  //   the Linux kernel build. See #101786.
-  setOperationAction(ISD::CTPOP, MVT::i32, Expand);
-  setOperationAction(ISD::CTPOP, MVT::i64, Expand);
+  setOperationAction(ISD::CTPOP, MVT::i32, LibCall);
+  setOperationAction(ISD::CTPOP, MVT::i64, LibCall);
   if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only()) {
     setOperationAction(ISD::CTLZ, MVT::i32, Expand);
     setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, LibCall);

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -396,13 +396,11 @@ RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
       setOperationAction({ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF}, MVT::i32, Custom);
   } else {
     setOperationAction(ISD::CTTZ, XLenVT, Expand);
-    // TODO: These should be set to LibCall, but this currently breaks
-    //   the Linux kernel build. See #101786. Lacks i128 tests, too.
     if (Subtarget.is64Bit())
-      setOperationAction(ISD::CTPOP, MVT::i128, Expand);
+      setOperationAction(ISD::CTPOP, MVT::i128, LibCall);
     else
-      setOperationAction(ISD::CTPOP, MVT::i32, Expand);
-    setOperationAction(ISD::CTPOP, MVT::i64, Expand);
+      setOperationAction(ISD::CTPOP, MVT::i32, LibCall);
+    setOperationAction(ISD::CTPOP, MVT::i64, LibCall);
   }
 
   if (Subtarget.hasStdExtZbb() || Subtarget.hasVendorXTHeadBb() ||

llvm/test/CodeGen/ARM/popcnt.ll

@@ -324,74 +324,20 @@ define i32 @ctpop16(i16 %x) nounwind readnone {
 define i32 @ctpop32(i32 %x) nounwind readnone {
 ; CHECK-LABEL: ctpop32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    ldr r1, .LCPI22_0
-; CHECK-NEXT:    ldr r2, .LCPI22_3
-; CHECK-NEXT:    and r1, r1, r0, lsr #1
-; CHECK-NEXT:    ldr r12, .LCPI22_1
-; CHECK-NEXT:    sub r0, r0, r1
-; CHECK-NEXT:    ldr r3, .LCPI22_2
-; CHECK-NEXT:    and r1, r0, r2
-; CHECK-NEXT:    and r0, r2, r0, lsr #2
-; CHECK-NEXT:    add r0, r1, r0
-; CHECK-NEXT:    add r0, r0, r0, lsr #4
-; CHECK-NEXT:    and r0, r0, r12
-; CHECK-NEXT:    mul r1, r0, r3
-; CHECK-NEXT:    lsr r0, r1, #24
-; CHECK-NEXT:    mov pc, lr
-; CHECK-NEXT:    .p2align 2
-; CHECK-NEXT:  @ %bb.1:
-; CHECK-NEXT:  .LCPI22_0:
-; CHECK-NEXT:    .long 1431655765 @ 0x55555555
-; CHECK-NEXT:  .LCPI22_1:
-; CHECK-NEXT:    .long 252645135 @ 0xf0f0f0f
-; CHECK-NEXT:  .LCPI22_2:
-; CHECK-NEXT:    .long 16843009 @ 0x1010101
-; CHECK-NEXT:  .LCPI22_3:
-; CHECK-NEXT:    .long 858993459 @ 0x33333333
+; CHECK-NEXT:    b __popcountsi2
   %count = tail call i32 @llvm.ctpop.i32(i32 %x)
   ret i32 %count
 }
 
 define i64 @ctpop64(i64 %x) nounwind readnone {
 ; CHECK-LABEL: ctpop64:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    .save {r4, lr}
-; CHECK-NEXT:    push {r4, lr}
-; CHECK-NEXT:    ldr r2, .LCPI23_0
-; CHECK-NEXT:    ldr r3, .LCPI23_3
-; CHECK-NEXT:    and r4, r2, r0, lsr #1
-; CHECK-NEXT:    and r2, r2, r1, lsr #1
-; CHECK-NEXT:    sub r0, r0, r4
-; CHECK-NEXT:    sub r1, r1, r2
-; CHECK-NEXT:    and r4, r0, r3
-; CHECK-NEXT:    and r2, r1, r3
-; CHECK-NEXT:    and r0, r3, r0, lsr #2
-; CHECK-NEXT:    and r1, r3, r1, lsr #2
-; CHECK-NEXT:    add r0, r4, r0
-; CHECK-NEXT:    ldr lr, .LCPI23_1
-; CHECK-NEXT:    add r1, r2, r1
-; CHECK-NEXT:    ldr r12, .LCPI23_2
-; CHECK-NEXT:    add r0, r0, r0, lsr #4
-; CHECK-NEXT:    and r0, r0, lr
-; CHECK-NEXT:    add r1, r1, r1, lsr #4
-; CHECK-NEXT:    mul r2, r0, r12
-; CHECK-NEXT:    and r0, r1, lr
-; CHECK-NEXT:    mul r1, r0, r12
-; CHECK-NEXT:    lsr r0, r2, #24
-; CHECK-NEXT:    add r0, r0, r1, lsr #24
-; CHECK-NEXT:    mov r1, #0
-; CHECK-NEXT:    pop {r4, lr}
+; CHECK-NEXT:    .save {r11, lr}
+; CHECK-NEXT:    push {r11, lr}
+; CHECK-NEXT:    bl __popcountdi2
+; CHECK-NEXT:    asr r1, r0, #31
+; CHECK-NEXT:    pop {r11, lr}
 ; CHECK-NEXT:    mov pc, lr
-; CHECK-NEXT:    .p2align 2
-; CHECK-NEXT:  @ %bb.1:
-; CHECK-NEXT:  .LCPI23_0:
-; CHECK-NEXT:    .long 1431655765 @ 0x55555555
-; CHECK-NEXT:  .LCPI23_1:
-; CHECK-NEXT:    .long 252645135 @ 0xf0f0f0f
-; CHECK-NEXT:  .LCPI23_2:
-; CHECK-NEXT:    .long 16843009 @ 0x1010101
-; CHECK-NEXT:  .LCPI23_3:
-; CHECK-NEXT:    .long 858993459 @ 0x33333333
   %count = tail call i64 @llvm.ctpop.i64(i64 %x)
   ret i64 %count
 }