[PowerPC] Do not convert loop to HW loop if the body contains calls to lrint/lround

[nemanjai]

Summary

These two intrinsics are lowered to calls so should prevent the formation of CTR loops.

This patch rather aggressively disables CTR loops if there are calls to unknown intrinsics so that we don't keep hitting similar issues in the future. I will certainly collect some data about the number of CTR loops we emit before and after the patch before committing such an "aggressively pessimistic" patch. Posting this early just to gauge what others think about this conservative behaviour.

[nemanjai]

lib/Target/PowerPC/PPCTargetTransformInfo.cpp

281

@hfinkel What do you think about this conservative approach to prevent similar issues in the future? This is I think a third time I address a similar bug.

[hfinkel]

lib/Target/PowerPC/PPCTargetTransformInfo.cpp

281

I really thought that I had replied to this already, but apparently not. I'm okay with that, but I'd like to do it in a separate commit, and we need to audit the existing intrinsics and add them as necessary (both ones that have straightforward codegen and ones like llvm.assume which have no codegen).

[joerg]

Can we please get this or the other fix in RSN? Mesa hits this.

[nemanjai]

Remove the change to make this conservatively correct until we can have a look at existing intrinsics that could/should prevent HW loop generation.

[nemanjai]

@hfinkel @joerg @vddvss What do you guys think? Should we get this fix committed and I'll follow-up with a look at all the other intrinsics that we currently have?

[hfinkel]

LGTM

lib/Target/PowerPC/PPCTargetTransformInfo.cpp

if (CallInst *CI = dyn_cast<CallInst>(J)) {
       }
 
       if (Function *F = CI->getCalledFunction()) {
         // Most intrinsics don't become function calls, but some might.
         // sin, cos, exp and log are always calls.
         unsigned Opcode = 0;
         if (F->getIntrinsicID() != Intrinsic::not_intrinsic) {
           switch (F->getIntrinsicID()) {
-          default: continue;
+          // If it is some intrinsic we don't understand, assume it may lower to
+          // a call for safety.
+          default: return true;
           // If we have a call to ppc_is_decremented_ctr_nonzero, or ppc_mtctr
           // we're definitely using CTR.
           case Intrinsic::set_loop_iterations:
           case Intrinsic::loop_decrement:
             return true;
 
 // VisualStudio defines setjmp as _setjmp
 #if defined(_MSC_VER) && defined(setjmp) && \
#endif
               return true;
             else
               continue; // ISD::FCOPYSIGN is never a library call.
           case Intrinsic::sqrt:               Opcode = ISD::FSQRT;      break;
           case Intrinsic::floor:              Opcode = ISD::FFLOOR;     break;
           case Intrinsic::ceil:               Opcode = ISD::FCEIL;      break;
           case Intrinsic::trunc:              Opcode = ISD::FTRUNC;     break;
           case Intrinsic::rint:               Opcode = ISD::FRINT;      break;
+          case Intrinsic::lrint:              Opcode = ISD::LRINT;      break;
+          case Intrinsic::llrint:             Opcode = ISD::LLRINT;     break;
           case Intrinsic::nearbyint:          Opcode = ISD::FNEARBYINT; break;
           case Intrinsic::round:              Opcode = ISD::FROUND;     break;
+          case Intrinsic::lround:             Opcode = ISD::LROUND;     break;
+          case Intrinsic::llround:            Opcode = ISD::LLROUND;    break;
           case Intrinsic::minnum:             Opcode = ISD::FMINNUM;    break;
           case Intrinsic::maxnum:             Opcode = ISD::FMAXNUM;    break;
           case Intrinsic::umul_with_overflow: Opcode = ISD::UMULO;      break;
           case Intrinsic::smul_with_overflow: Opcode = ISD::SMULO;      break;
           }
         }
 
         // PowerPC does not use [US]DIVREM or other library calls for

test/CodeGen/PowerPC/pr43527.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -ppc-asm-full-reg-names -verify-machineinstrs \
+; RUN:   -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr8 < %s | FileCheck %s
+; We don't want to produce a CTR loop due to the call to lrint in the body.
+define dso_local void @test(i64 %arg, i64 %arg1) {
+; CHECK-LABEL: test:
+; CHECK:       # %bb.0: # %bb
+; CHECK-NEXT:    bc 4, 4*cr5+lt, .LBB0_5
+; CHECK-NEXT:  # %bb.1: # %bb3
+; CHECK-NEXT:    bc 12, 4*cr5+lt, .LBB0_6
+; CHECK-NEXT:  # %bb.2: # %bb4
+; CHECK-NEXT:    mflr r0
+; CHECK-NEXT:    .cfi_def_cfa_offset 64
+; CHECK-NEXT:    .cfi_offset lr, 16
+; CHECK-NEXT:    .cfi_offset r29, -24
+; CHECK-NEXT:    .cfi_offset r30, -16
+; CHECK-NEXT:    std r29, -24(r1) # 8-byte Folded Spill
+; CHECK-NEXT:    std r30, -16(r1) # 8-byte Folded Spill
+; CHECK-NEXT:    std r0, 16(r1)
+; CHECK-NEXT:    stdu r1, -64(r1)
+; CHECK-NEXT:    sub r30, r4, r3
+; CHECK-NEXT:    li r29, 0
+; CHECK-NEXT:    .p2align 5
+; CHECK-NEXT:  .LBB0_3: # %bb5
+; CHECK-NEXT:    #
+; CHECK-NEXT:    lfsx f1, 0, r29
+; CHECK-NEXT:    bl lrint
+; CHECK-NEXT:    nop
+; CHECK-NEXT:    addi r30, r30, -1
+; CHECK-NEXT:    addi r29, r29, 4
+; CHECK-NEXT:    cmpldi r30, 0
+; CHECK-NEXT:    bne cr0, .LBB0_3
+; CHECK-NEXT:  # %bb.4: # %bb15
+; CHECK-NEXT:    stb r3, 0(r3)
+; CHECK-NEXT:    addi r1, r1, 64
+; CHECK-NEXT:    ld r0, 16(r1)
+; CHECK-NEXT:    mtlr r0
+; CHECK-NEXT:    ld r30, -16(r1) # 8-byte Folded Reload
+; CHECK-NEXT:    ld r29, -24(r1) # 8-byte Folded Reload
+; CHECK-NEXT:    blr
+; CHECK-NEXT:  .LBB0_5: # %bb2
+; CHECK-NEXT:  .LBB0_6: # %bb14
+bb:
+  br i1 undef, label %bb3, label %bb2
+
+bb2:                                              ; preds = %bb
+  unreachable
+
+bb3:                                              ; preds = %bb
+  %tmp = sub i64 %arg1, %arg
+  br i1 undef, label %bb4, label %bb14
+
+bb4:                                              ; preds = %bb3
+  br label %bb5
+
+bb5:                                              ; preds = %bb5, %bb4
+  %tmp6 = phi i64 [ %tmp12, %bb5 ], [ 0, %bb4 ]
+  %tmp7 = getelementptr inbounds float, float* null, i64 %tmp6
+  %tmp8 = load float, float* %tmp7, align 4
+  %tmp9 = fpext float %tmp8 to double
+  %tmp10 = tail call i64 @llvm.lrint.i64.f64(double %tmp9) #2
+  %tmp11 = trunc i64 %tmp10 to i8
+  store i8 %tmp11, i8* undef, align 1
+  %tmp12 = add nuw i64 %tmp6, 1
+  %tmp13 = icmp eq i64 %tmp12, %tmp
+  br i1 %tmp13, label %bb15, label %bb5
+
+bb14:                                             ; preds = %bb3
+  unreachable
+
+bb15:                                             ; preds = %bb5
+  ret void
+}
+
+declare i64 @llvm.lrint.i64.f64(double)