[SLP] Vectorize non-power-of-2 ops with padding.

This patch introduces a new VectorizeWithPadding node type for root and
leave nodes to allow vectorizing loads/stores with non-power-of-2 number
of elements.

VectorizeWithPadding load nodes will pad the result to the next power of 2
with poison elements.

Non-leaf nodes will operate on normal power-of-2 vectors. For those
non-leaf nodes, we still track the number of padding elements needed to
go to the next power-of-2, to be used in various places, like cost
computation.

VectorizeWithPadding store nodes strip away the padding elements and
store the non-power-of-2 number of data elements.

Note that re-ordering and shuffling is not implemented for nodes
requiring padding yet to keep the initial implementation simpler.

The initial implementation also only tries to vectorize with padding if
original number of elements + 1 is a power-of-2, i.e. if only a single
padding element is needed.

The feature is guarded by a new flag, off by defaul for now.

Author: fhahn

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -1,35 +1,65 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
-; RUN: opt -passes=slp-vectorizer -mtriple=arm64-apple-ios -S %s | FileCheck %s
+; RUN: opt -passes=slp-vectorizer -slp-vectorize-with-padding -mtriple=arm64-apple-ios -S %s | FileCheck --check-prefixes=PADDING %s
+; RUN: opt -passes=slp-vectorizer -slp-vectorize-with-padding=false -mtriple=arm64-apple-ios -S %s | FileCheck --check-prefixes=NO-PADDING %s
 
 define void @v15_load_i8_mul_by_constant_store(ptr %src, ptr noalias %dst) {
-; CHECK-LABEL: define void @v15_load_i8_mul_by_constant_store(
-; CHECK-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]]) {
-; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[GEP_SRC_0:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 0
-; CHECK-NEXT:    [[TMP0:%.*]] = load <8 x i8>, ptr [[GEP_SRC_0]], align 4
-; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <8 x i8> [[TMP0]], <i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10>
-; CHECK-NEXT:    store <8 x i8> [[TMP1]], ptr [[DST]], align 1
-; CHECK-NEXT:    [[GEP_SRC_8:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 8
-; CHECK-NEXT:    [[DST_8:%.*]] = getelementptr i8, ptr [[DST]], i8 8
-; CHECK-NEXT:    [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_SRC_8]], align 4
-; CHECK-NEXT:    [[TMP3:%.*]] = mul nsw <4 x i8> [[TMP2]], <i8 10, i8 10, i8 10, i8 10>
-; CHECK-NEXT:    store <4 x i8> [[TMP3]], ptr [[DST_8]], align 1
-; CHECK-NEXT:    [[GEP_SRC_12:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 12
-; CHECK-NEXT:    [[L_SRC_12:%.*]] = load i8, ptr [[GEP_SRC_12]], align 4
-; CHECK-NEXT:    [[MUL_12:%.*]] = mul nsw i8 [[L_SRC_12]], 10
-; CHECK-NEXT:    [[DST_12:%.*]] = getelementptr i8, ptr [[DST]], i8 12
-; CHECK-NEXT:    store i8 [[MUL_12]], ptr [[DST_12]], align 1
-; CHECK-NEXT:    [[GEP_SRC_13:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 13
-; CHECK-NEXT:    [[L_SRC_13:%.*]] = load i8, ptr [[GEP_SRC_13]], align 4
-; CHECK-NEXT:    [[MUL_13:%.*]] = mul nsw i8 [[L_SRC_13]], 10
-; CHECK-NEXT:    [[DST_13:%.*]] = getelementptr i8, ptr [[DST]], i8 13
-; CHECK-NEXT:    store i8 [[MUL_13]], ptr [[DST_13]], align 1
-; CHECK-NEXT:    [[GEP_SRC_14:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 14
-; CHECK-NEXT:    [[L_SRC_14:%.*]] = load i8, ptr [[GEP_SRC_14]], align 4
-; CHECK-NEXT:    [[MUL_14:%.*]] = mul nsw i8 [[L_SRC_14]], 10
-; CHECK-NEXT:    [[DST_14:%.*]] = getelementptr i8, ptr [[DST]], i8 14
-; CHECK-NEXT:    store i8 [[MUL_14]], ptr [[DST_14]], align 1
-; CHECK-NEXT:    ret void
+; PADDING-LABEL: define void @v15_load_i8_mul_by_constant_store(
+; PADDING-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]]) {
+; PADDING-NEXT:  entry:
+; PADDING-NEXT:    [[GEP_SRC_0:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 0
+; PADDING-NEXT:    [[TMP0:%.*]] = load <8 x i8>, ptr [[GEP_SRC_0]], align 4
+; PADDING-NEXT:    [[TMP1:%.*]] = mul nsw <8 x i8> [[TMP0]], <i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10>
+; PADDING-NEXT:    store <8 x i8> [[TMP1]], ptr [[DST]], align 1
+; PADDING-NEXT:    [[GEP_SRC_8:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 8
+; PADDING-NEXT:    [[DST_8:%.*]] = getelementptr i8, ptr [[DST]], i8 8
+; PADDING-NEXT:    [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_SRC_8]], align 4
+; PADDING-NEXT:    [[TMP3:%.*]] = mul nsw <4 x i8> [[TMP2]], <i8 10, i8 10, i8 10, i8 10>
+; PADDING-NEXT:    store <4 x i8> [[TMP3]], ptr [[DST_8]], align 1
+; PADDING-NEXT:    [[GEP_SRC_12:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 12
+; PADDING-NEXT:    [[L_SRC_12:%.*]] = load i8, ptr [[GEP_SRC_12]], align 4
+; PADDING-NEXT:    [[MUL_12:%.*]] = mul nsw i8 [[L_SRC_12]], 10
+; PADDING-NEXT:    [[DST_12:%.*]] = getelementptr i8, ptr [[DST]], i8 12
+; PADDING-NEXT:    store i8 [[MUL_12]], ptr [[DST_12]], align 1
+; PADDING-NEXT:    [[GEP_SRC_13:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 13
+; PADDING-NEXT:    [[L_SRC_13:%.*]] = load i8, ptr [[GEP_SRC_13]], align 4
+; PADDING-NEXT:    [[MUL_13:%.*]] = mul nsw i8 [[L_SRC_13]], 10
+; PADDING-NEXT:    [[DST_13:%.*]] = getelementptr i8, ptr [[DST]], i8 13
+; PADDING-NEXT:    store i8 [[MUL_13]], ptr [[DST_13]], align 1
+; PADDING-NEXT:    [[GEP_SRC_14:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 14
+; PADDING-NEXT:    [[L_SRC_14:%.*]] = load i8, ptr [[GEP_SRC_14]], align 4
+; PADDING-NEXT:    [[MUL_14:%.*]] = mul nsw i8 [[L_SRC_14]], 10
+; PADDING-NEXT:    [[DST_14:%.*]] = getelementptr i8, ptr [[DST]], i8 14
+; PADDING-NEXT:    store i8 [[MUL_14]], ptr [[DST_14]], align 1
+; PADDING-NEXT:    ret void
+;
+; NO-PADDING-LABEL: define void @v15_load_i8_mul_by_constant_store(
+; NO-PADDING-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]]) {
+; NO-PADDING-NEXT:  entry:
+; NO-PADDING-NEXT:    [[GEP_SRC_0:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 0
+; NO-PADDING-NEXT:    [[TMP0:%.*]] = load <8 x i8>, ptr [[GEP_SRC_0]], align 4
+; NO-PADDING-NEXT:    [[TMP1:%.*]] = mul nsw <8 x i8> [[TMP0]], <i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10>
+; NO-PADDING-NEXT:    store <8 x i8> [[TMP1]], ptr [[DST]], align 1
+; NO-PADDING-NEXT:    [[GEP_SRC_8:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 8
+; NO-PADDING-NEXT:    [[DST_8:%.*]] = getelementptr i8, ptr [[DST]], i8 8
+; NO-PADDING-NEXT:    [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_SRC_8]], align 4
+; NO-PADDING-NEXT:    [[TMP3:%.*]] = mul nsw <4 x i8> [[TMP2]], <i8 10, i8 10, i8 10, i8 10>
+; NO-PADDING-NEXT:    store <4 x i8> [[TMP3]], ptr [[DST_8]], align 1
+; NO-PADDING-NEXT:    [[GEP_SRC_12:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 12
+; NO-PADDING-NEXT:    [[L_SRC_12:%.*]] = load i8, ptr [[GEP_SRC_12]], align 4
+; NO-PADDING-NEXT:    [[MUL_12:%.*]] = mul nsw i8 [[L_SRC_12]], 10
+; NO-PADDING-NEXT:    [[DST_12:%.*]] = getelementptr i8, ptr [[DST]], i8 12
+; NO-PADDING-NEXT:    store i8 [[MUL_12]], ptr [[DST_12]], align 1
+; NO-PADDING-NEXT:    [[GEP_SRC_13:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 13
+; NO-PADDING-NEXT:    [[L_SRC_13:%.*]] = load i8, ptr [[GEP_SRC_13]], align 4
+; NO-PADDING-NEXT:    [[MUL_13:%.*]] = mul nsw i8 [[L_SRC_13]], 10
+; NO-PADDING-NEXT:    [[DST_13:%.*]] = getelementptr i8, ptr [[DST]], i8 13
+; NO-PADDING-NEXT:    store i8 [[MUL_13]], ptr [[DST_13]], align 1
+; NO-PADDING-NEXT:    [[GEP_SRC_14:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 14
+; NO-PADDING-NEXT:    [[L_SRC_14:%.*]] = load i8, ptr [[GEP_SRC_14]], align 4
+; NO-PADDING-NEXT:    [[MUL_14:%.*]] = mul nsw i8 [[L_SRC_14]], 10
+; NO-PADDING-NEXT:    [[DST_14:%.*]] = getelementptr i8, ptr [[DST]], i8 14
+; NO-PADDING-NEXT:    store i8 [[MUL_14]], ptr [[DST_14]], align 1
+; NO-PADDING-NEXT:    ret void
 ;
 entry:
   %gep.src.0 = getelementptr inbounds i8, ptr %src, i8 0
@@ -123,5 +153,3 @@ entry:
 
   ret void
 }
-
-