[LV][AArch64] Don't query registers for illegal scalable vector elts

[huntergr-arm]

When trying to maximize vector bandwidth we ask TTI for the number of
registers required for a given operation. If the type of that operation
happens to be something illegal for scalable vectors (e.g.
<vscale x 4 x fp128>) then we would see a crash.

Instead, just return a default value and let the cost model reject the
invalid operation later.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Graham Hunter (huntergr-arm)

Changes

When trying to maximize vector bandwidth we ask TTI for the number of
registers required for a given operation. If the type of that operation
happens to be something illegal for scalable vectors (e.g.
<vscale x 4 x fp128>) then we would see a crash.

Instead, just return a default value and let the cost model reject the
invalid operation later.

---

Full diff: https://github.com/llvm/llvm-project/pull/109411.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/Vectorize/LoopVectorize.cpp (+3-1)
• (added) llvm/test/Transforms/LoopVectorize/AArch64/scalable-fp-ext-trunc-illegal-type.ll (+81)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 9f554827a8287d..66c2eeed17d38b 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5207,7 +5207,9 @@ LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) {
 
   const auto &TTICapture = TTI;
   auto GetRegUsage = [&TTICapture](Type *Ty, ElementCount VF) -> unsigned {
-    if (Ty->isTokenTy() || !VectorType::isValidElementType(Ty))
+    if (Ty->isTokenTy() || !VectorType::isValidElementType(Ty) ||
+        (VF.isScalable() &&
+         !TTICapture.isElementTypeLegalForScalableVector(Ty)))
       return 0;
     return TTICapture.getRegUsageForType(VectorType::get(Ty, VF));
   };
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/scalable-fp-ext-trunc-illegal-type.ll b/llvm/test/Transforms/LoopVectorize/AArch64/scalable-fp-ext-trunc-illegal-type.ll
new file mode 100644
index 00000000000000..bb89ead3ca303c
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/scalable-fp-ext-trunc-illegal-type.ll
@@ -0,0 +1,81 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 3
+; RUN: opt < %s -passes=loop-vectorize -debug-only=loop-vectorize -vectorizer-maximize-bandwidth -S 2>&1 | FileCheck %s
+; REQUIRES: asserts
+
+target triple = "aarch64-unknown-linux-gnu"
+
+;; Make sure we reject scalable vectors for fp128 types. We were previously
+;; crashing before reaching the cost model when checking for the number of
+;; registers required for a <vscale x 4 x fp128> when trying to maximize
+;; vector bandwidth with SVE.
+
+; CHECK: LV: Found an estimated cost of Invalid for VF vscale x 2 For instruction: %load.ext = fpext double %load.in to fp128
+
+define void @load_ext_trunc_store(ptr readonly %in, ptr noalias %out, i64 %N) #0 {
+; CHECK-LABEL: define void @load_ext_trunc_store(
+; CHECK-SAME: ptr readonly [[IN:%.*]], ptr noalias [[OUT:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N]], i64 1)
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[UMAX]], 4
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[UMAX]], 4
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[UMAX]], [[N_MOD_VF]]
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds double, ptr [[IN]], i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds double, ptr [[TMP2]], i32 0
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x double>, ptr [[TMP4]], align 8
+; CHECK-NEXT:    [[TMP3:%.*]] = fpext <4 x double> [[WIDE_LOAD]] to <4 x fp128>
+; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds float, ptr [[OUT]], i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP5:%.*]] = fptrunc <4 x fp128> [[TMP3]] to <4 x float>
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds float, ptr [[TMP8]], i32 0
+; CHECK-NEXT:    store <4 x float> [[TMP5]], ptr [[TMP12]], align 4
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP14]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[UMAX]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[CMP_N]], label %[[FOR_EXIT:.*]], label %[[SCALAR_PH]]
+; CHECK:       [[SCALAR_PH]]:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT:    br label %[[FOR_BODY:.*]]
+; CHECK:       [[FOR_BODY]]:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[FOR_BODY]] ]
+; CHECK-NEXT:    [[GEP_IN:%.*]] = getelementptr inbounds nuw double, ptr [[IN]], i64 [[IV]]
+; CHECK-NEXT:    [[LOAD_IN:%.*]] = load double, ptr [[GEP_IN]], align 8
+; CHECK-NEXT:    [[LOAD_EXT:%.*]] = fpext double [[LOAD_IN]] to fp128
+; CHECK-NEXT:    [[GEP_OUT:%.*]] = getelementptr inbounds nuw float, ptr [[OUT]], i64 [[IV]]
+; CHECK-NEXT:    [[TRUNC_OUT:%.*]] = fptrunc fp128 [[LOAD_EXT]] to float
+; CHECK-NEXT:    store float [[TRUNC_OUT]], ptr [[GEP_OUT]], align 4
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp ult i64 [[IV_NEXT]], [[N]]
+; CHECK-NEXT:    br i1 [[EXITCOND]], label %[[FOR_BODY]], label %[[FOR_EXIT]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK:       [[FOR_EXIT]]:
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %for.body
+
+for.body:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
+  %gep.in = getelementptr inbounds nuw double, ptr %in, i64 %iv
+  %load.in = load double, ptr %gep.in, align 8
+  %load.ext = fpext double %load.in to fp128
+  %gep.out = getelementptr inbounds nuw float, ptr %out, i64 %iv
+  %trunc.out = fptrunc fp128 %load.ext to float
+  store float %trunc.out, ptr %gep.out, align 4
+  %iv.next = add nuw nsw i64 %iv, 1
+  %exitcond = icmp ult i64 %iv.next, %N
+  br i1 %exitcond, label %for.body, label %for.exit, !llvm.loop !0
+
+for.exit:
+  ret void
+}
+
+attributes #0 = { "target-features"="+sve" vscale_range(1,16) }
+
+!0 = distinct !{!0, !1}
+!1 = !{!"llvm.loop.interleave.count", i32 1}

[fhahn]

LGTM, thanks!

[fhahn]

is this needed?

[huntergr-arm]

I wanted to reduce the number of check lines to review; I've switched to using a flag instead to match the other change.

[fhahn]

Might be simpler/easier to see if using -mattr=+sve in the run line?

[huntergr-arm]

done