[NFC] Add scalable vectorisation tests for int/FP <> int/FP conversions

[david-arm]

Summary

We can already vectorise loops that involve int<>int, fp<>fp, int<>fp
and fp<>int conversions, however we didn't previously have any tests
for them. This patch adds some tests for each conversion type.

[fhahn]

Would it make sense to have all the conversion tests in a single file, as the tests are quite compact?

llvm/test/Transforms/LoopVectorize/AArch64/sve-conv-int-to-int.ll

6 ↗

(On Diff #335497)

Do we still need this? IIRC most those checks have been removed by now?

[david-arm]

Sure, I can combine them into a single test if you think that's better? I wasn't sure whether in general we prefer to have more concise test files or to have fewer, larger files. I could create a file with a name like sve-type-conv.ll.

llvm/test/Transforms/LoopVectorize/AArch64/sve-conv-int-to-int.ll

6 ↗

(On Diff #335497)

Yes you're right. When I created the patch originally I think we still needed them, but I'll rebase and remove them.

[fhahn]

In D99951#2679444, @david-arm wrote:

Sure, I can combine them into a single test if you think that's better? I wasn't sure whether in general we prefer to have more concise test files or to have fewer, larger files. I could create a file with a name like sve-type-conv.ll.

Sounds good to me. I don't think there's a general answer really. In this case it seems like they all test the same/similar code path and there might be less processing overhead with fewer, larger files than with many small files.

[david-arm]

• Removed warnings check line.
• Combined all tests into one file.

[kmclaughlin]

llvm/test/Transforms/LoopVectorize/AArch64/sve-type-conv.ll

157–158

I'm not sure why this test and the one below are using both a [u|s]itofp and an fptrunc, is it possible to just use uitofp i64 %0 to half here?

218

nit: Should this test be called @u8_to_u16?

[david-arm]

• Removed unnecessary fptrunc from tests

[david-arm]

llvm/test/Transforms/LoopVectorize/AArch64/sve-type-conv.ll

157–158

Yeah, this is an artefact of clang. When compiling a loop such as this:

void foo(float16_t *dst, uint64_t *src, long long n) {
  for (long long i = 0; i < n; i++)
    dst[i] = src[i];
}

it generates IR with two-part conversions, i.e. uitfop + fptrunc. Not sure why, but I've changed the test anyway.

[kmclaughlin]

Thanks for updating the tests, LGTM!

llvm/test/Transforms/LoopVectorize/AArch64/sve-type-conv.ll

+; RUN: opt -loop-vectorize -dce -instcombine < %s -S | FileCheck %s
+
+target triple = "aarch64-unknown-linux-gnu"
+
+
+define void @f16_to_f32(float* noalias nocapture %dst, half* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @f16_to_f32(
+; CHECK: vector.body
+; CHECK:   %{{.*}} = fpext <vscale x 8 x half> %{{.*}} to <vscale x 8 x float>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds half, half* %src, i64 %i.07
+  %0 = load half, half* %arrayidx, align 2
+  %conv = fpext half %0 to float
+  %arrayidx1 = getelementptr inbounds float, float* %dst, i64 %i.07
+  store float %conv, float* %arrayidx1, align 4
+  %inc = add nuw nsw i64 %i.07, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+define void @f64_to_f32(float* noalias nocapture %dst, double* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @f64_to_f32(
+; CHECK: vector.body
+; CHECK:   %{{.*}} = fptrunc <vscale x 8 x double> %{{.*}} to <vscale x 8 x float>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds double, double* %src, i64 %i.07
+  %0 = load double, double* %arrayidx, align 8
+  %conv = fptrunc double %0 to float
+  %arrayidx1 = getelementptr inbounds float, float* %dst, i64 %i.07
+  store float %conv, float* %arrayidx1, align 4
+  %inc = add nuw nsw i64 %i.07, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+define void @f16_to_s8(i8* noalias nocapture %dst, half* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @f16_to_s8(
+; CHECK: vector.body
+; CHECK:   %{{.*}} = fptosi <vscale x 8 x half> %{{.*}} to <vscale x 8 x i8>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.08 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds half, half* %src, i64 %i.08
+  %0 = load half, half* %arrayidx, align 2
+  %conv1 = fptosi half %0 to i8
+  %arrayidx2 = getelementptr inbounds i8, i8* %dst, i64 %i.08
+  store i8 %conv1, i8* %arrayidx2, align 1
+  %inc = add nuw nsw i64 %i.08, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+define void @f32_to_u64(i64* noalias nocapture %dst, float* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @f32_to_u64(
+; CHECK: vector.body
+; CHECK:   %{{.*}} = fptoui <vscale x 8 x float> %{{.*}} to <vscale x 8 x i64>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds float, float* %src, i64 %i.07
+  %0 = load float, float* %arrayidx, align 4
+  %conv = fptoui float %0 to i64
+  %arrayidx1 = getelementptr inbounds i64, i64* %dst, i64 %i.07
+  store i64 %conv, i64* %arrayidx1, align 8
+  %inc = add nuw nsw i64 %i.07, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+define void @s8_to_f32(float* noalias nocapture %dst, i8* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @s8_to_f32(
+; CHECK: vector.body
+; CHECK:   %{{.*}} = sitofp <vscale x 8 x i8> %{{.*}} to <vscale x 8 x float>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i8, i8* %src, i64 %i.07
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = sitofp i8 %0 to float
+  %arrayidx1 = getelementptr inbounds float, float* %dst, i64 %i.07
+  store float %conv, float* %arrayidx1, align 4
+  %inc = add nuw nsw i64 %i.07, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+define void @u16_to_f32(float* noalias nocapture %dst, i16* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @u16_to_f32(
+; CHECK: vector.body
+; CHECK:   %{{.*}} = uitofp <vscale x 8 x i16> %{{.*}} to <vscale x 8 x float>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i16, i16* %src, i64 %i.07
+  %0 = load i16, i16* %arrayidx, align 2
+  %conv = uitofp i16 %0 to float
+  %arrayidx1 = getelementptr inbounds float, float* %dst, i64 %i.07
+  store float %conv, float* %arrayidx1, align 4
+  %inc = add nuw nsw i64 %i.07, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+define void @u64_to_f16(half* noalias nocapture %dst, i64* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @u64_to_f16(
+; CHECK:      vector.body
+; CHECK:        %{{.*}} = uitofp <vscale x 8 x i64> %{{.*}} to <vscale x 8 x half>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.08 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i64, i64* %src, i64 %i.08
+  %0 = load i64, i64* %arrayidx, align 8
+  %conv1 = uitofp i64 %0 to half
+  %arrayidx2 = getelementptr inbounds half, half* %dst, i64 %i.08
+  store half %conv1, half* %arrayidx2, align 2
+  %inc = add nuw nsw i64 %i.08, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+define void @s64_to_f16(half* noalias nocapture %dst, i64* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @s64_to_f16(
+; CHECK:      vector.body
+; CHECK:        %{{.*}} = sitofp <vscale x 8 x i64> %{{.*}} to <vscale x 8 x half>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.08 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i64, i64* %src, i64 %i.08
+  %0 = load i64, i64* %arrayidx, align 8
+  %conv1 = sitofp i64 %0 to half
+  %arrayidx2 = getelementptr inbounds half, half* %dst, i64 %i.08
+  store half %conv1, half* %arrayidx2, align 2
+  %inc = add nuw nsw i64 %i.08, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+define void @s8_to_s32(i32* noalias nocapture %dst, i8* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @s8_to_s32(
+; CHECK: vector.body
+; CHECK:   %{{.*}} = sext <vscale x 8 x i8> %{{.*}} to <vscale x 8 x i32>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i8, i8* %src, i64 %i.07
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = sext i8 %0 to i32
+  %arrayidx1 = getelementptr inbounds i32, i32* %dst, i64 %i.07
+  store i32 %conv, i32* %arrayidx1, align 4
+  %inc = add nuw nsw i64 %i.07, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+define void @u8_to_u16(i16* noalias nocapture %dst, i8* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @u8_to_u16(
+; CHECK: vector.body
+; CHECK:   %{{.*}} = zext <vscale x 8 x i8> %{{.*}} to <vscale x 8 x i16>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i8, i8* %src, i64 %i.07
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = zext i8 %0 to i16
+  %arrayidx1 = getelementptr inbounds i16, i16* %dst, i64 %i.07
+  store i16 %conv, i16* %arrayidx1, align 2
+  %inc = add nuw nsw i64 %i.07, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+define void @s64_to_s8(i8* noalias nocapture %dst, i64* noalias nocapture readonly %src, i64 %N) #0 {
+; CHECK-LABEL: @s64_to_s8(
+; CHECK: vector.body
+; CHECK:   %{{.*}} = trunc <vscale x 8 x i64> %{{.*}} to <vscale x 8 x i8>
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i64, i64* %src, i64 %i.07
+  %0 = load i64, i64* %arrayidx, align 8
+  %conv = trunc i64 %0 to i8
+  %arrayidx1 = getelementptr inbounds i8, i8* %dst, i64 %i.07
+  store i8 %conv, i8* %arrayidx1, align 1
+  %inc = add nuw nsw i64 %i.07, 1
+  %exitcond.not = icmp eq i64 %inc, %N
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+attributes #0 = { "target-features"="+sve" }
+
+!0 = distinct !{!0, !1, !2, !3, !4, !5}
+!1 = !{!"llvm.loop.mustprogress"}
+!2 = !{!"llvm.loop.vectorize.width", i32 8}
+!3 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
+!4 = !{!"llvm.loop.interleave.count", i32 1}
+!5 = !{!"llvm.loop.vectorize.enable", i1 true}