c4c2d777f4
Update handling of ReductionStartVector in VPlanUnroll for partial reductions. The new code makes sure all parts are properly set to the cloned ReductionStartVector. Fixes a mis-compile reported for https://github.com/llvm/llvm-project/pull/142290.
155 lines
10 KiB
LLVM
155 lines
10 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
|
|
; RUN: opt -p loop-vectorize -force-vector-width=16 -force-vector-interleave=2 -mattr=+dotprod -S %s | FileCheck --check-prefix=IC2 %s
|
|
; RUN: opt -p loop-vectorize -force-vector-width=16 -force-vector-interleave=4 -mattr=+dotprod -S %s | FileCheck --check-prefix=IC4 %s
|
|
|
|
target triple = "arm64-apple-macosx"
|
|
|
|
define i32 @partial_reduce_with_non_constant_start_value(ptr %src, i32 %rdx.start, i64 %n) {
|
|
; IC2-LABEL: define i32 @partial_reduce_with_non_constant_start_value(
|
|
; IC2-SAME: ptr [[SRC:%.*]], i32 [[RDX_START:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
|
|
; IC2-NEXT: [[ENTRY:.*]]:
|
|
; IC2-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 32
|
|
; IC2-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
|
|
; IC2: [[VECTOR_PH]]:
|
|
; IC2-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], 32
|
|
; IC2-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
|
|
; IC2-NEXT: [[TMP0:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[RDX_START]], i32 0
|
|
; IC2-NEXT: br label %[[VECTOR_BODY:.*]]
|
|
; IC2: [[VECTOR_BODY]]:
|
|
; IC2-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
|
|
; IC2-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ [[TMP0]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE:%.*]], %[[VECTOR_BODY]] ]
|
|
; IC2-NEXT: [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE3:%.*]], %[[VECTOR_BODY]] ]
|
|
; IC2-NEXT: [[TMP1:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 [[INDEX]]
|
|
; IC2-NEXT: [[TMP2:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 0
|
|
; IC2-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 16
|
|
; IC2-NEXT: [[WIDE_LOAD:%.*]] = load <16 x i8>, ptr [[TMP2]], align 1
|
|
; IC2-NEXT: [[WIDE_LOAD2:%.*]] = load <16 x i8>, ptr [[TMP3]], align 1
|
|
; IC2-NEXT: [[TMP4:%.*]] = zext <16 x i8> [[WIDE_LOAD]] to <16 x i32>
|
|
; IC2-NEXT: [[TMP5:%.*]] = zext <16 x i8> [[WIDE_LOAD2]] to <16 x i32>
|
|
; IC2-NEXT: [[TMP6:%.*]] = mul nuw nsw <16 x i32> [[TMP4]], [[TMP4]]
|
|
; IC2-NEXT: [[TMP7:%.*]] = mul nuw nsw <16 x i32> [[TMP5]], [[TMP5]]
|
|
; IC2-NEXT: [[PARTIAL_REDUCE]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI]], <16 x i32> [[TMP6]])
|
|
; IC2-NEXT: [[PARTIAL_REDUCE3]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI1]], <16 x i32> [[TMP7]])
|
|
; IC2-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 32
|
|
; IC2-NEXT: [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
|
|
; IC2-NEXT: br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
|
|
; IC2: [[MIDDLE_BLOCK]]:
|
|
; IC2-NEXT: [[BIN_RDX:%.*]] = add <4 x i32> [[PARTIAL_REDUCE3]], [[PARTIAL_REDUCE]]
|
|
; IC2-NEXT: [[TMP9:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX]])
|
|
; IC2-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
|
|
; IC2-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
|
|
; IC2: [[SCALAR_PH]]:
|
|
; IC2-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
|
|
; IC2-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP9]], %[[MIDDLE_BLOCK]] ], [ [[RDX_START]], %[[ENTRY]] ]
|
|
; IC2-NEXT: br label %[[LOOP:.*]]
|
|
; IC2: [[LOOP]]:
|
|
; IC2-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
|
|
; IC2-NEXT: [[RDX:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[RDX_NEXT:%.*]], %[[LOOP]] ]
|
|
; IC2-NEXT: [[GEP_SRC:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 [[IV]]
|
|
; IC2-NEXT: [[L:%.*]] = load i8, ptr [[GEP_SRC]], align 1
|
|
; IC2-NEXT: [[CONV:%.*]] = zext i8 [[L]] to i32
|
|
; IC2-NEXT: [[MUL:%.*]] = mul nuw nsw i32 [[CONV]], [[CONV]]
|
|
; IC2-NEXT: [[RDX_NEXT]] = add nsw i32 [[MUL]], [[RDX]]
|
|
; IC2-NEXT: [[IV_NEXT]] = add nsw i64 [[IV]], 1
|
|
; IC2-NEXT: [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
|
|
; IC2-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
|
|
; IC2: [[EXIT]]:
|
|
; IC2-NEXT: [[RDX_NEXT_LCSSA:%.*]] = phi i32 [ [[RDX_NEXT]], %[[LOOP]] ], [ [[TMP9]], %[[MIDDLE_BLOCK]] ]
|
|
; IC2-NEXT: ret i32 [[RDX_NEXT_LCSSA]]
|
|
;
|
|
; IC4-LABEL: define i32 @partial_reduce_with_non_constant_start_value(
|
|
; IC4-SAME: ptr [[SRC:%.*]], i32 [[RDX_START:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
|
|
; IC4-NEXT: [[ENTRY:.*]]:
|
|
; IC4-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 64
|
|
; IC4-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
|
|
; IC4: [[VECTOR_PH]]:
|
|
; IC4-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], 64
|
|
; IC4-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
|
|
; IC4-NEXT: [[TMP0:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[RDX_START]], i32 0
|
|
; IC4-NEXT: br label %[[VECTOR_BODY:.*]]
|
|
; IC4: [[VECTOR_BODY]]:
|
|
; IC4-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
|
|
; IC4-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ [[TMP0]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE:%.*]], %[[VECTOR_BODY]] ]
|
|
; IC4-NEXT: [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE7:%.*]], %[[VECTOR_BODY]] ]
|
|
; IC4-NEXT: [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE8:%.*]], %[[VECTOR_BODY]] ]
|
|
; IC4-NEXT: [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE9:%.*]], %[[VECTOR_BODY]] ]
|
|
; IC4-NEXT: [[TMP1:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 [[INDEX]]
|
|
; IC4-NEXT: [[TMP2:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 0
|
|
; IC4-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 16
|
|
; IC4-NEXT: [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 32
|
|
; IC4-NEXT: [[TMP5:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 48
|
|
; IC4-NEXT: [[WIDE_LOAD:%.*]] = load <16 x i8>, ptr [[TMP2]], align 1
|
|
; IC4-NEXT: [[WIDE_LOAD4:%.*]] = load <16 x i8>, ptr [[TMP3]], align 1
|
|
; IC4-NEXT: [[WIDE_LOAD5:%.*]] = load <16 x i8>, ptr [[TMP4]], align 1
|
|
; IC4-NEXT: [[WIDE_LOAD6:%.*]] = load <16 x i8>, ptr [[TMP5]], align 1
|
|
; IC4-NEXT: [[TMP6:%.*]] = zext <16 x i8> [[WIDE_LOAD]] to <16 x i32>
|
|
; IC4-NEXT: [[TMP7:%.*]] = zext <16 x i8> [[WIDE_LOAD4]] to <16 x i32>
|
|
; IC4-NEXT: [[TMP8:%.*]] = zext <16 x i8> [[WIDE_LOAD5]] to <16 x i32>
|
|
; IC4-NEXT: [[TMP9:%.*]] = zext <16 x i8> [[WIDE_LOAD6]] to <16 x i32>
|
|
; IC4-NEXT: [[TMP10:%.*]] = mul nuw nsw <16 x i32> [[TMP6]], [[TMP6]]
|
|
; IC4-NEXT: [[TMP11:%.*]] = mul nuw nsw <16 x i32> [[TMP7]], [[TMP7]]
|
|
; IC4-NEXT: [[TMP12:%.*]] = mul nuw nsw <16 x i32> [[TMP8]], [[TMP8]]
|
|
; IC4-NEXT: [[TMP13:%.*]] = mul nuw nsw <16 x i32> [[TMP9]], [[TMP9]]
|
|
; IC4-NEXT: [[PARTIAL_REDUCE]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI]], <16 x i32> [[TMP10]])
|
|
; IC4-NEXT: [[PARTIAL_REDUCE7]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI1]], <16 x i32> [[TMP11]])
|
|
; IC4-NEXT: [[PARTIAL_REDUCE8]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI2]], <16 x i32> [[TMP12]])
|
|
; IC4-NEXT: [[PARTIAL_REDUCE9]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI3]], <16 x i32> [[TMP13]])
|
|
; IC4-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 64
|
|
; IC4-NEXT: [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
|
|
; IC4-NEXT: br i1 [[TMP18]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
|
|
; IC4: [[MIDDLE_BLOCK]]:
|
|
; IC4-NEXT: [[BIN_RDX:%.*]] = add <4 x i32> [[PARTIAL_REDUCE7]], [[PARTIAL_REDUCE]]
|
|
; IC4-NEXT: [[BIN_RDX10:%.*]] = add <4 x i32> [[PARTIAL_REDUCE8]], [[BIN_RDX]]
|
|
; IC4-NEXT: [[BIN_RDX11:%.*]] = add <4 x i32> [[PARTIAL_REDUCE9]], [[BIN_RDX10]]
|
|
; IC4-NEXT: [[TMP15:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX11]])
|
|
; IC4-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
|
|
; IC4-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
|
|
; IC4: [[SCALAR_PH]]:
|
|
; IC4-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
|
|
; IC4-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP15]], %[[MIDDLE_BLOCK]] ], [ [[RDX_START]], %[[ENTRY]] ]
|
|
; IC4-NEXT: br label %[[LOOP:.*]]
|
|
; IC4: [[LOOP]]:
|
|
; IC4-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
|
|
; IC4-NEXT: [[RDX:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[RDX_NEXT:%.*]], %[[LOOP]] ]
|
|
; IC4-NEXT: [[GEP_SRC:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 [[IV]]
|
|
; IC4-NEXT: [[L:%.*]] = load i8, ptr [[GEP_SRC]], align 1
|
|
; IC4-NEXT: [[CONV:%.*]] = zext i8 [[L]] to i32
|
|
; IC4-NEXT: [[MUL:%.*]] = mul nuw nsw i32 [[CONV]], [[CONV]]
|
|
; IC4-NEXT: [[RDX_NEXT]] = add nsw i32 [[MUL]], [[RDX]]
|
|
; IC4-NEXT: [[IV_NEXT]] = add nsw i64 [[IV]], 1
|
|
; IC4-NEXT: [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
|
|
; IC4-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
|
|
; IC4: [[EXIT]]:
|
|
; IC4-NEXT: [[RDX_NEXT_LCSSA:%.*]] = phi i32 [ [[RDX_NEXT]], %[[LOOP]] ], [ [[TMP15]], %[[MIDDLE_BLOCK]] ]
|
|
; IC4-NEXT: ret i32 [[RDX_NEXT_LCSSA]]
|
|
;
|
|
entry:
|
|
br label %loop
|
|
|
|
loop:
|
|
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
|
|
%rdx = phi i32 [ %rdx.start, %entry ], [ %rdx.next, %loop ]
|
|
%gep.src = getelementptr inbounds i8, ptr %src, i64 %iv
|
|
%l = load i8, ptr %gep.src, align 1
|
|
%conv = zext i8 %l to i32
|
|
%mul = mul nuw nsw i32 %conv, %conv
|
|
%rdx.next = add nsw i32 %mul, %rdx
|
|
%iv.next = add nsw i64 %iv, 1
|
|
%ec = icmp eq i64 %iv.next, %n
|
|
br i1 %ec, label %exit, label %loop
|
|
|
|
exit:
|
|
ret i32 %rdx.next
|
|
}
|
|
;.
|
|
; IC2: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
|
|
; IC2: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
|
|
; IC2: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
|
|
; IC2: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
|
|
;.
|
|
; IC4: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
|
|
; IC4: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
|
|
; IC4: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
|
|
; IC4: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
|
|
;.
|