llvm/test/Transforms/LoopUnroll/AMDGPU/unroll-cost-cmp-select.ll

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt -S -mtriple=amdgpu9.0a-amd-amdhsa -passes=loop-unroll < %s | FileCheck %s

; Loop unrolling uses TCK_SizeAndLatency for non-minsize functions. Even though
; AMDGPU reports low size-latency costs for cmp/selects to enable SimplifyCFG
; speculation, this wide loop should still remain too expensive to unroll.

define void @wide_select_loop(ptr addrspace(1) %p, <16 x float> %a, <16 x float> %b) {
; CHECK-LABEL: define void @wide_select_loop(
; CHECK-SAME: ptr addrspace(1) [[P:%.*]], <16 x float> [[A:%.*]], <16 x float> [[B:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*]]:
; CHECK-NEXT:    br label %[[LOOP:.*]]
; CHECK:       [[LOOP]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[INC:%.*]], %[[LOOP]] ]
; CHECK-NEXT:    [[GEP:%.*]] = getelementptr inbounds <16 x float>, ptr addrspace(1) [[P]], i32 [[I]]
; CHECK-NEXT:    [[V:%.*]] = load <16 x float>, ptr addrspace(1) [[GEP]], align 64
; CHECK-NEXT:    [[C1:%.*]] = fcmp ogt <16 x float> [[V]], [[A]]
; CHECK-NEXT:    [[S1:%.*]] = select <16 x i1> [[C1]], <16 x float> [[V]], <16 x float> [[B]]
; CHECK-NEXT:    [[C2:%.*]] = fcmp ogt <16 x float> [[S1]], [[A]]
; CHECK-NEXT:    [[S2:%.*]] = select <16 x i1> [[C2]], <16 x float> [[S1]], <16 x float> [[B]]
; CHECK-NEXT:    [[C3:%.*]] = fcmp ogt <16 x float> [[S2]], [[A]]
; CHECK-NEXT:    [[S3:%.*]] = select <16 x i1> [[C3]], <16 x float> [[S2]], <16 x float> [[B]]
; CHECK-NEXT:    [[C4:%.*]] = fcmp ogt <16 x float> [[S3]], [[A]]
; CHECK-NEXT:    [[S4:%.*]] = select <16 x i1> [[C4]], <16 x float> [[S3]], <16 x float> [[B]]
; CHECK-NEXT:    [[C5:%.*]] = fcmp ogt <16 x float> [[S4]], [[A]]
; CHECK-NEXT:    [[S5:%.*]] = select <16 x i1> [[C5]], <16 x float> [[S4]], <16 x float> [[B]]
; CHECK-NEXT:    store <16 x float> [[S5]], ptr addrspace(1) [[GEP]], align 64
; CHECK-NEXT:    [[INC]] = add nuw nsw i32 [[I]], 1
; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i32 [[INC]], 64
; CHECK-NEXT:    br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %loop

loop:
  %i = phi i32 [ 0, %entry ], [ %inc, %loop ]
  %gep = getelementptr inbounds <16 x float>, ptr addrspace(1) %p, i32 %i
  %v = load <16 x float>, ptr addrspace(1) %gep, align 64
  %c1 = fcmp ogt <16 x float> %v, %a
  %s1 = select <16 x i1> %c1, <16 x float> %v, <16 x float> %b
  %c2 = fcmp ogt <16 x float> %s1, %a
  %s2 = select <16 x i1> %c2, <16 x float> %s1, <16 x float> %b
  %c3 = fcmp ogt <16 x float> %s2, %a
  %s3 = select <16 x i1> %c3, <16 x float> %s2, <16 x float> %b
  %c4 = fcmp ogt <16 x float> %s3, %a
  %s4 = select <16 x i1> %c4, <16 x float> %s3, <16 x float> %b
  %c5 = fcmp ogt <16 x float> %s4, %a
  %s5 = select <16 x i1> %c5, <16 x float> %s4, <16 x float> %b
  store <16 x float> %s5, ptr addrspace(1) %gep, align 64
  %inc = add nuw nsw i32 %i, 1
  %cmp = icmp slt i32 %inc, 64
  br i1 %cmp, label %loop, label %exit

exit:
  ret void
}