[RISCV][SLP] Add tests for unprofitable SLP vectorization due to GEP. NFC

[ABataev]

llvm/test/Analysis/CostModel/RISCV/gep-zero-indices.ll

7–23 ↗

(On Diff #519599)

Could you add some more tests with other types/vector lengths?

[luke]

Add more zero indices tests

[luke]

Add a comment to describe what the gep-zero-indices tests are for, and add
a few more test cases

[reames]

LGTM to the test additions in test/Analysis/CostModel/RISCV/gep-zero-indices.ll only. The other piece needs a bit of cleanup, and is tied more to the sequence.

llvm/test/Analysis/CostModel/RISCV/gep-zero-indices.ll

7 ↗

(On Diff #533298)

indices

llvm/test/Transforms/SLPVectorizer/RISCV/gep.ll

6

This needs more explanation as to *why* this shouldn't be vectorized.

[luke]

Include new cost model tests for GEPs in this patch

[reames]

LGTM to the CostModel tests only. Please split and land. Please do *not* update this patch again without doing so, and likely rearranging the dependencies on other reviews.

[luke]

Split out the cost model tests

[luke]

Split out non-SLP changes

[luke]

Undo accidental diff update: apologies for the noise.

[ABataev]

Retitle to add SLP again and add NFC

[ABataev]

LG

llvm/test/Transforms/SLPVectorizer/RISCV/gep.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=slp-vectorizer -mtriple=riscv64 -mattr=+v \
+; RUN: -riscv-v-slp-max-vf=0 -S | FileCheck %s
+
+; FIXME: This should not be vectorized, as the cost of computing the
+; offsets nullifies the benefits of vectorizing:
+;
+; copy_with_offset_v2i8:
+;         addi    a0, a0, 8
+;         vsetivli        zero, 2, e8, mf8, ta, ma
+;         vle8.v  v8, (a0)
+;         addi    a1, a1, 16
+;         vse8.v  v8, (a1)
+;         ret
+;
+; Compared to the scalar version where the offsets can be folded into the
+; addressing mode:
+;
+; copy_with_offset_v2i8:
+;         lbu     a2, 8(a0)
+;         lbu     a0, 9(a0)
+;         sb      a2, 16(a1)
+;         sb      a0, 17(a1)
+;   ret
+
+define void @copy_with_offset_v2i8(ptr noalias %p, ptr noalias %q) {
+; CHECK-LABEL: @copy_with_offset_v2i8(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 8
+; CHECK-NEXT:    [[Q1:%.*]] = getelementptr i8, ptr [[Q:%.*]], i32 16
+; CHECK-NEXT:    [[TMP0:%.*]] = load <2 x i8>, ptr [[P1]], align 1
+; CHECK-NEXT:    store <2 x i8> [[TMP0]], ptr [[Q1]], align 1
+; CHECK-NEXT:    ret void
+;
+entry:
+  %p1 = getelementptr i8, ptr %p, i32 8
+  %x1 = load i8, ptr %p1
+  %q1 = getelementptr i8, ptr %q, i32 16
+  store i8 %x1, ptr %q1
+
+  %p2 = getelementptr i8, ptr %p, i32 9
+  %x2 = load i8, ptr %p2
+  %q2 = getelementptr i8, ptr %q, i32 17
+  store i8 %x2, ptr %q2
+
+  ret void
+}
+
+; This on the other hand, should be vectorized as the vector savings outweigh
+; the GEP costs.
+define void @copy_with_offset_v4i8(ptr noalias %p, ptr noalias %q) {
+; CHECK-LABEL: @copy_with_offset_v4i8(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 8
+; CHECK-NEXT:    [[Q1:%.*]] = getelementptr i8, ptr [[Q:%.*]], i32 16
+; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i8>, ptr [[P1]], align 1
+; CHECK-NEXT:    store <4 x i8> [[TMP0]], ptr [[Q1]], align 1
+; CHECK-NEXT:    ret void
+;
+entry:
+  %p1 = getelementptr i8, ptr %p, i32 8
+  %x1 = load i8, ptr %p1
+  %q1 = getelementptr i8, ptr %q, i32 16
+  store i8 %x1, ptr %q1
+
+  %p2 = getelementptr i8, ptr %p, i32 9
+  %x2 = load i8, ptr %p2
+  %q2 = getelementptr i8, ptr %q, i32 17
+  store i8 %x2, ptr %q2
+
+  %p3 = getelementptr i8, ptr %p, i32 10
+  %x3 = load i8, ptr %p3
+  %q3 = getelementptr i8, ptr %q, i32 18
+  store i8 %x3, ptr %q3
+
+  %p4 = getelementptr i8, ptr %p, i32 11
+  %x4 = load i8, ptr %p4
+  %q4 = getelementptr i8, ptr %q, i32 19
+  store i8 %x4, ptr %q4
+
+  ret void
+}