[Lldb-commits] [clang] [flang] [libc] [libcxx] [lld] [lldb] [llvm] [mlir] [LV] Vectorize maxnum/minnum w/o fast-math flags. (PR #148239)

[Florian Hahn via lldb-commits]

================
@@ -42,18 +42,60 @@ define float @fmaxnum(ptr %src, i64 %n) {
 ; CHECK-LABEL: define float @fmaxnum(
 ; CHECK-SAME: ptr [[SRC:%.*]], i64 [[N:%.*]]) {
 ; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 8
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 8
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ <float -1.000000e+07, float 0xFFF8000000000000, float 0xFFF8000000000000, float 0xFFF8000000000000>, %[[VECTOR_PH]] ], [ [[TMP7:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ splat (float 0xFFF8000000000000), %[[VECTOR_PH]] ], [ [[TMP8:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[GEP_SRC:%.*]] = getelementptr inbounds nuw float, ptr [[SRC]], i64 [[IV]]
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds nuw float, ptr [[GEP_SRC]], i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw float, ptr [[GEP_SRC]], i32 4
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP1]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <4 x float>, ptr [[TMP2]], align 4
+; CHECK-NEXT:    [[TMP7]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI]], <4 x float> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP8]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI1]], <4 x float> [[WIDE_LOAD2]])
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 8
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    [[TMP3:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP4:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD2]], [[WIDE_LOAD2]]
+; CHECK-NEXT:    [[TMP5:%.*]] = or <4 x i1> [[TMP3]], [[TMP4]]
+; CHECK-NEXT:    [[TMP6:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP5]])
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i1> poison, i1 [[TMP6]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i1> [[BROADCAST_SPLATINSERT]], <4 x i1> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP10:%.*]] = or i1 [[TMP6]], [[TMP9]]
+; CHECK-NEXT:    br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[TMP11:%.*]] = select <4 x i1> [[BROADCAST_SPLAT]], <4 x float> [[VEC_PHI]], <4 x float> [[TMP7]]
+; CHECK-NEXT:    [[TMP12:%.*]] = select <4 x i1> [[BROADCAST_SPLAT]], <4 x float> [[VEC_PHI1]], <4 x float> [[TMP8]]
+; CHECK-NEXT:    [[TMP14:%.*]] = select i1 [[TMP6]], i64 [[IV]], i64 [[N_VEC]]
+; CHECK-NEXT:    [[RDX_MINMAX_CMP:%.*]] = fcmp ogt <4 x float> [[TMP11]], [[TMP12]]
+; CHECK-NEXT:    [[RDX_MINMAX_SELECT:%.*]] = select <4 x i1> [[RDX_MINMAX_CMP]], <4 x float> [[TMP11]], <4 x float> [[TMP12]]
+; CHECK-NEXT:    [[TMP13:%.*]] = call float @llvm.vector.reduce.fmax.v4f32(<4 x float> [[RDX_MINMAX_SELECT]])
----------------
fhahn wrote:

Yep, we can also additionally check if the input already has nnan, and avoid adding the extra checks as follow-up

https://github.com/llvm/llvm-project/pull/148239