[AArch64][NFC] NFC for const vector as Instruction operand

[sushgokh]

Current cost-modelling does not take into account cost of materializing const vector. This results in some cases, as the test shows, being vectorized but this may not always be profitable. Future patch will try to address this issue.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Sushant Gokhale (sushgokh)

Changes

Current cost-modelling does not take into account cost of materializing const vector. This results in some cases, as the test shows, being vectorized but this may not always be profitable. Future patch will try to address this issue.

---

Full diff: https://github.com/llvm/llvm-project/pull/116790.diff

1 Files Affected:

• (added) llvm/test/Transforms/SLPVectorizer/materialize-vector-of-consts.ll (+100)

diff --git a/llvm/test/Transforms/SLPVectorizer/materialize-vector-of-consts.ll b/llvm/test/Transforms/SLPVectorizer/materialize-vector-of-consts.ll
new file mode 100644
index 00000000000000..2f58bd25b75647
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/materialize-vector-of-consts.ll
@@ -0,0 +1,100 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: %if aarch64-registered-target %{ opt -passes=slp-vectorizer -mtriple=aarch64 -S %s | FileCheck %s %}
+
+define <2 x float> @v2f32_diff_consts(float %a, float %b)
+; CHECK-LABEL: define <2 x float> @v2f32_diff_consts(
+; CHECK-SAME: float [[A:%.*]], float [[B:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x float> poison, float [[A]], i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <2 x float> [[TMP1]], float [[B]], i32 1
+; CHECK-NEXT:    [[TMP3:%.*]] = fmul <2 x float> [[TMP2]], <float 2.200000e+01, float 2.300000e+01>
+; CHECK-NEXT:    ret <2 x float> [[TMP3]]
+;
+{
+  %1 = fmul float %a, 22.0
+  %2 = fmul float %b, 23.0
+  %3 = insertelement <2 x float> poison, float %1, i32 0
+  %4 = insertelement <2 x float> %3, float %2, i32 1
+  ret <2 x float> %4
+}
+
+define <2 x float> @v2f32_const_splat(float %a, float %b)
+; CHECK-LABEL: define <2 x float> @v2f32_const_splat(
+; CHECK-SAME: float [[A:%.*]], float [[B:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x float> poison, float [[A]], i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <2 x float> [[TMP1]], float [[B]], i32 1
+; CHECK-NEXT:    [[TMP3:%.*]] = fmul <2 x float> [[TMP2]], splat (float 2.200000e+01)
+; CHECK-NEXT:    ret <2 x float> [[TMP3]]
+;
+{
+  %1 = fmul float %a, 22.0
+  %2 = fmul float %b, 22.0
+  %3 = insertelement <2 x float> poison, float %1, i32 0
+  %4 = insertelement <2 x float> %3, float %2, i32 1
+  ret <2 x float> %4
+}
+
+define <4 x double> @v4f64_illegal_type(double %a, double %b, double %c, double %d)
+; CHECK-LABEL: define <4 x double> @v4f64_illegal_type(
+; CHECK-SAME: double [[A:%.*]], double [[B:%.*]], double [[C:%.*]], double [[D:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <4 x double> poison, double [[A]], i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <4 x double> [[TMP1]], double [[B]], i32 1
+; CHECK-NEXT:    [[TMP3:%.*]] = insertelement <4 x double> [[TMP2]], double [[C]], i32 2
+; CHECK-NEXT:    [[TMP4:%.*]] = insertelement <4 x double> [[TMP3]], double [[D]], i32 3
+; CHECK-NEXT:    [[TMP5:%.*]] = fmul <4 x double> [[TMP4]], <double 2.100000e+01, double 2.200000e+01, double 2.300000e+01, double 2.400000e+01>
+; CHECK-NEXT:    ret <4 x double> [[TMP5]]
+;
+{
+  %1 = fmul double %a, 21.0
+  %2 = fmul double %b, 22.0
+  %3 = fmul double %c, 23.0
+  %4 = fmul double %d, 24.0
+  %5 = insertelement <4 x double> poison, double %1, i32 0
+  %6 = insertelement <4 x double> %5, double %2, i32 1
+  %7 = insertelement <4 x double> %6, double %3, i32 2
+  %8 = insertelement <4 x double> %7, double %4, i32 3
+  ret <4 x double> %8
+}
+
+define <2 x double> @v2f64_dup_const_vector_case1(double %a, double %b, double %c, double %d)
+; CHECK-LABEL: define <2 x double> @v2f64_dup_const_vector_case1(
+; CHECK-SAME: double [[A:%.*]], double [[B:%.*]], double [[C:%.*]], double [[D:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x double> poison, double [[A]], i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <2 x double> [[TMP1]], double [[B]], i32 1
+; CHECK-NEXT:    [[TMP3:%.*]] = fmul <2 x double> [[TMP2]], <double 2.100000e+01, double 2.200000e+01>
+; CHECK-NEXT:    [[TMP4:%.*]] = insertelement <2 x double> poison, double [[C]], i32 0
+; CHECK-NEXT:    [[TMP5:%.*]] = insertelement <2 x double> [[TMP4]], double [[D]], i32 1
+; CHECK-NEXT:    [[TMP6:%.*]] = fmul <2 x double> [[TMP5]], <double 2.100000e+01, double 2.200000e+01>
+; CHECK-NEXT:    [[TMP7:%.*]] = fadd <2 x double> [[TMP3]], [[TMP6]]
+; CHECK-NEXT:    ret <2 x double> [[TMP7]]
+;
+{
+  %1 = fmul double %a, 21.0
+  %2 = fmul double %b, 22.0
+  %3 = fmul double %c, 21.0
+  %4 = fmul double %d, 22.0
+  %5 = insertelement <2 x double> poison, double %1, i32 0
+  %6 = insertelement <2 x double> %5, double %2, i32 1
+  %7 = insertelement <2 x double> poison, double %3, i32 0
+  %8 = insertelement <2 x double> %7, double %4, i32 1
+  %9 = fadd <2 x double> %6, %8
+  ret <2 x double> %9
+}
+
+define <2 x double> @v2f64_dup_const_vector_case2(double %a, double %b, double %c, double %d)
+; CHECK-LABEL: define <2 x double> @v2f64_dup_const_vector_case2(
+; CHECK-SAME: double [[A:%.*]], double [[B:%.*]], double [[C:%.*]], double [[D:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x double> poison, double [[A]], i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <2 x double> [[TMP1]], double [[B]], i32 1
+; CHECK-NEXT:    [[TMP3:%.*]] = fmul <2 x double> [[TMP2]], <double 2.100000e+01, double 2.200000e+01>
+; CHECK-NEXT:    [[TMP4:%.*]] = fadd <2 x double> [[TMP3]], <double 2.100000e+01, double 2.200000e+01>
+; CHECK-NEXT:    ret <2 x double> [[TMP4]]
+;
+{
+  %1 = fmul double %a, 21.0
+  %2 = fmul double %b, 22.0
+  %3 = fadd double %1, 21.0
+  %4 = fadd double %2, 22.0
+  %5 = insertelement <2 x double> poison, double %3, i32 0
+  %6 = insertelement <2 x double> %5, double %4, i32 1
+  ret <2 x double> %6
+}

[davemgreen]

I have been a little sceptical about costing constants with how they are currently represented in llvm and how often they can be "free". They can certainly be more expensive at times, looking forward to being proved incorrect about it.

The tests LGTM.