PPC: Don't restrict lvsl generation to after type legalization

This is a re-commit of r190764, with an extra check to make sure that we're not
performing the transformation on illegal types (a small test case has been
added for this as well).

Original commit message:

The PPC backend uses a target-specific DAG combine to turn unaligned Altivec
loads into a permutation-based sequence when possible. Unfortunately, the
target-specific DAG combine is not always called on all loads of interest
(sometimes the routines in DAGCombine call CombineTo such that the new node and
users are not added to the worklist); allowing the combine to trigger early
(before type legalization) mitigates this problem. Because the autovectorizers
only create legal vector types, I don't expect a lot of cases where this
optimization is enabled by type legalization in practice.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@190771 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/PowerPC/PPCISelLowering.cpp b/lib/Target/PowerPC/PPCISelLowering.cpp
index ae046f0b7af033d03cb1fbbb455a6729f73a9799..83bd1abddedb98ed38596f153e8bc0ac9fb7b133 100644 (file)
--- a/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -7177,7 +7177,8 @@ SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
     unsigned ABIAlignment = getDataLayout()->getABITypeAlignment(Ty);
     if (ISD::isNON_EXTLoad(N) && VT.isVector() &&
         TM.getSubtarget<PPCSubtarget>().hasAltivec() &&
-        DCI.getDAGCombineLevel() == AfterLegalizeTypes &&
+        (VT == MVT::v16i8 || VT == MVT::v8i16 ||
+         VT == MVT::v4i32 || VT == MVT::v4f32) &&
         LD->getAlignment() < ABIAlignment) {
       // This is a type-legal unaligned Altivec load.
       SDValue Chain = LD->getChain();

diff --git a/test/CodeGen/PowerPC/unal-altivec2.ll b/test/CodeGen/PowerPC/unal-altivec2.ll
new file mode 100644 (file)
index 0000000..7464675
--- /dev/null
+++ b/test/CodeGen/PowerPC/unal-altivec2.ll
@@ -0,0 +1,166 @@
+; RUN: llc -mtriple=powerpc64-unknown-linux-gnu -mcpu=pwr7 < %s | FileCheck %s
+target datalayout = "E-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f128:128:128-v128:128:128-n32:64"
+target triple = "powerpc64-unknown-linux-gnu"
+
+; Function Attrs: nounwind
+define void @foo(float* noalias nocapture %x, float* noalias nocapture readonly %y) #0 {
+entry:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %entry
+; CHECK-LABEL: @foo
+; CHECK: lvsl
+; CHECK: blr
+  %index = phi i64 [ 0, %entry ], [ %index.next.15, %vector.body ]
+  %0 = getelementptr inbounds float* %y, i64 %index
+  %1 = bitcast float* %0 to <4 x float>*
+  %wide.load = load <4 x float>* %1, align 4
+  %2 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load)
+  %3 = getelementptr inbounds float* %x, i64 %index
+  %4 = bitcast float* %3 to <4 x float>*
+  store <4 x float> %2, <4 x float>* %4, align 4
+  %index.next = add i64 %index, 4
+  %5 = getelementptr inbounds float* %y, i64 %index.next
+  %6 = bitcast float* %5 to <4 x float>*
+  %wide.load.1 = load <4 x float>* %6, align 4
+  %7 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.1)
+  %8 = getelementptr inbounds float* %x, i64 %index.next
+  %9 = bitcast float* %8 to <4 x float>*
+  store <4 x float> %7, <4 x float>* %9, align 4
+  %index.next.1 = add i64 %index.next, 4
+  %10 = getelementptr inbounds float* %y, i64 %index.next.1
+  %11 = bitcast float* %10 to <4 x float>*
+  %wide.load.2 = load <4 x float>* %11, align 4
+  %12 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.2)
+  %13 = getelementptr inbounds float* %x, i64 %index.next.1
+  %14 = bitcast float* %13 to <4 x float>*
+  store <4 x float> %12, <4 x float>* %14, align 4
+  %index.next.2 = add i64 %index.next.1, 4
+  %15 = getelementptr inbounds float* %y, i64 %index.next.2
+  %16 = bitcast float* %15 to <4 x float>*
+  %wide.load.3 = load <4 x float>* %16, align 4
+  %17 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.3)
+  %18 = getelementptr inbounds float* %x, i64 %index.next.2
+  %19 = bitcast float* %18 to <4 x float>*
+  store <4 x float> %17, <4 x float>* %19, align 4
+  %index.next.3 = add i64 %index.next.2, 4
+  %20 = getelementptr inbounds float* %y, i64 %index.next.3
+  %21 = bitcast float* %20 to <4 x float>*
+  %wide.load.4 = load <4 x float>* %21, align 4
+  %22 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.4)
+  %23 = getelementptr inbounds float* %x, i64 %index.next.3
+  %24 = bitcast float* %23 to <4 x float>*
+  store <4 x float> %22, <4 x float>* %24, align 4
+  %index.next.4 = add i64 %index.next.3, 4
+  %25 = getelementptr inbounds float* %y, i64 %index.next.4
+  %26 = bitcast float* %25 to <4 x float>*
+  %wide.load.5 = load <4 x float>* %26, align 4
+  %27 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.5)
+  %28 = getelementptr inbounds float* %x, i64 %index.next.4
+  %29 = bitcast float* %28 to <4 x float>*
+  store <4 x float> %27, <4 x float>* %29, align 4
+  %index.next.5 = add i64 %index.next.4, 4
+  %30 = getelementptr inbounds float* %y, i64 %index.next.5
+  %31 = bitcast float* %30 to <4 x float>*
+  %wide.load.6 = load <4 x float>* %31, align 4
+  %32 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.6)
+  %33 = getelementptr inbounds float* %x, i64 %index.next.5
+  %34 = bitcast float* %33 to <4 x float>*
+  store <4 x float> %32, <4 x float>* %34, align 4
+  %index.next.6 = add i64 %index.next.5, 4
+  %35 = getelementptr inbounds float* %y, i64 %index.next.6
+  %36 = bitcast float* %35 to <4 x float>*
+  %wide.load.7 = load <4 x float>* %36, align 4
+  %37 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.7)
+  %38 = getelementptr inbounds float* %x, i64 %index.next.6
+  %39 = bitcast float* %38 to <4 x float>*
+  store <4 x float> %37, <4 x float>* %39, align 4
+  %index.next.7 = add i64 %index.next.6, 4
+  %40 = getelementptr inbounds float* %y, i64 %index.next.7
+  %41 = bitcast float* %40 to <4 x float>*
+  %wide.load.8 = load <4 x float>* %41, align 4
+  %42 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.8)
+  %43 = getelementptr inbounds float* %x, i64 %index.next.7
+  %44 = bitcast float* %43 to <4 x float>*
+  store <4 x float> %42, <4 x float>* %44, align 4
+  %index.next.8 = add i64 %index.next.7, 4
+  %45 = getelementptr inbounds float* %y, i64 %index.next.8
+  %46 = bitcast float* %45 to <4 x float>*
+  %wide.load.9 = load <4 x float>* %46, align 4
+  %47 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.9)
+  %48 = getelementptr inbounds float* %x, i64 %index.next.8
+  %49 = bitcast float* %48 to <4 x float>*
+  store <4 x float> %47, <4 x float>* %49, align 4
+  %index.next.9 = add i64 %index.next.8, 4
+  %50 = getelementptr inbounds float* %y, i64 %index.next.9
+  %51 = bitcast float* %50 to <4 x float>*
+  %wide.load.10 = load <4 x float>* %51, align 4
+  %52 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.10)
+  %53 = getelementptr inbounds float* %x, i64 %index.next.9
+  %54 = bitcast float* %53 to <4 x float>*
+  store <4 x float> %52, <4 x float>* %54, align 4
+  %index.next.10 = add i64 %index.next.9, 4
+  %55 = getelementptr inbounds float* %y, i64 %index.next.10
+  %56 = bitcast float* %55 to <4 x float>*
+  %wide.load.11 = load <4 x float>* %56, align 4
+  %57 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.11)
+  %58 = getelementptr inbounds float* %x, i64 %index.next.10
+  %59 = bitcast float* %58 to <4 x float>*
+  store <4 x float> %57, <4 x float>* %59, align 4
+  %index.next.11 = add i64 %index.next.10, 4
+  %60 = getelementptr inbounds float* %y, i64 %index.next.11
+  %61 = bitcast float* %60 to <4 x float>*
+  %wide.load.12 = load <4 x float>* %61, align 4
+  %62 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.12)
+  %63 = getelementptr inbounds float* %x, i64 %index.next.11
+  %64 = bitcast float* %63 to <4 x float>*
+  store <4 x float> %62, <4 x float>* %64, align 4
+  %index.next.12 = add i64 %index.next.11, 4
+  %65 = getelementptr inbounds float* %y, i64 %index.next.12
+  %66 = bitcast float* %65 to <4 x float>*
+  %wide.load.13 = load <4 x float>* %66, align 4
+  %67 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.13)
+  %68 = getelementptr inbounds float* %x, i64 %index.next.12
+  %69 = bitcast float* %68 to <4 x float>*
+  store <4 x float> %67, <4 x float>* %69, align 4
+  %index.next.13 = add i64 %index.next.12, 4
+  %70 = getelementptr inbounds float* %y, i64 %index.next.13
+  %71 = bitcast float* %70 to <4 x float>*
+  %wide.load.14 = load <4 x float>* %71, align 4
+  %72 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.14)
+  %73 = getelementptr inbounds float* %x, i64 %index.next.13
+  %74 = bitcast float* %73 to <4 x float>*
+  store <4 x float> %72, <4 x float>* %74, align 4
+  %index.next.14 = add i64 %index.next.13, 4
+  %75 = getelementptr inbounds float* %y, i64 %index.next.14
+  %76 = bitcast float* %75 to <4 x float>*
+  %wide.load.15 = load <4 x float>* %76, align 4
+  %77 = call <4 x float> @llvm_cos_v4f32(<4 x float> %wide.load.15)
+  %78 = getelementptr inbounds float* %x, i64 %index.next.14
+  %79 = bitcast float* %78 to <4 x float>*
+  store <4 x float> %77, <4 x float>* %79, align 4
+  %index.next.15 = add i64 %index.next.14, 4
+  %80 = icmp eq i64 %index.next.15, 2048
+  br i1 %80, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+}
+
+; Function Attrs: nounwind readonly
+declare <4 x float> @llvm_cos_v4f32(<4 x float>) #1
+
+define <2 x double> @bar(double* %x) {
+entry:
+  %p = bitcast double* %x to <2 x double>*
+  %r = load <2 x double>* %p, align 8
+
+; CHECK-LABEL: @bar
+; CHECK-NOT: lvsl
+; CHECK: blr
+
+  ret <2 x double> %r
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readonly }