Don't use runtime bounds check between address spaces.

Don't vectorize with a runtime check if it requires a
comparison between pointers with different address spaces.
The values can't be assumed to be directly comparable.
Previously it would create an illegal bitcast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@191862 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 377 insertions, 0 deletions

diff --git a/test/Transforms/LoopVectorize/runtime-check-address-space.ll b/test/Transforms/LoopVectorize/runtime-check-address-space.ll
new file mode 100644
index 0000000..6c86561
--- /dev/null
+++ b/test/Transforms/LoopVectorize/runtime-check-address-space.ll
@@ -0,0 +1,235 @@
+; RUN: opt -S -march=r600 -mcpu=cayman -loop-vectorize -force-vector-unroll=1 -force-vector-width=4 -dce -instcombine < %s | FileCheck %s
+
+; Check vectorization that would ordinarily require a runtime bounds
+; check on the pointers when mixing address spaces. For now we cannot
+; assume address spaces do not alias, and we can't assume that
+; different pointers are directly comparable.
+;
+; These all test this basic loop for different combinations of address
+; spaces, and swapping in globals or adding noalias.
+;
+;void foo(int addrspace(N)* [noalias] a, int addrspace(M)* [noalias] b, int n)
+;{
+;    for (int i = 0; i < n; ++i)
+;    {
+;        a[i] = 3 * b[i];
+;    }
+;}
+
+; Artificial datalayout
+target datalayout = "e-p:32:32:32-p1:16:16:16-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024-v2048:2048:2048-n32:64"
+
+
+@g_as1 = common addrspace(1) global [1024 x i32] zeroinitializer, align 16
+@q_as2 = common addrspace(2) global [1024 x i32] zeroinitializer, align 16
+
+; Both parameters are unidentified objects with the same address
+; space, so this should vectorize normally.
+define void @foo(i32 addrspace(1)* %a, i32 addrspace(1)* %b, i32 %n) #0 {
+; CHECK-LABEL: @foo(
+; CHECK: <4 x i32>
+; CHECK: ret
+
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp slt i32 %i.0, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %idxprom = sext i32 %i.0 to i64
+  %arrayidx = getelementptr inbounds i32 addrspace(1)* %b, i64 %idxprom
+  %0 = load i32 addrspace(1)* %arrayidx, align 4
+  %mul = mul nsw i32 %0, 3
+  %idxprom1 = sext i32 %i.0 to i64
+  %arrayidx2 = getelementptr inbounds i32 addrspace(1)* %a, i64 %idxprom1
+  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+; Parameters are unidentified and different address spaces, so cannot vectorize.
+define void @bar0(i32* %a, i32 addrspace(1)* %b, i32 %n) #0 {
+; CHECK-LABEL: @bar0(
+; CHECK-NOT: <4 x i32>
+; CHECK: ret
+
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp slt i32 %i.0, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %idxprom = sext i32 %i.0 to i64
+  %arrayidx = getelementptr inbounds i32 addrspace(1)* %b, i64 %idxprom
+  %0 = load i32 addrspace(1)* %arrayidx, align 4
+  %mul = mul nsw i32 %0, 3
+  %idxprom1 = sext i32 %i.0 to i64
+  %arrayidx2 = getelementptr inbounds i32* %a, i64 %idxprom1
+  store i32 %mul, i32* %arrayidx2, align 4
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+; Swapped arguments should be the same
+define void @bar1(i32 addrspace(1)* %a, i32* %b, i32 %n) #0 {
+; CHECK-LABEL: @bar1(
+; CHECK-NOT: <4 x i32>
+; CHECK: ret
+
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp slt i32 %i.0, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %idxprom = sext i32 %i.0 to i64
+  %arrayidx = getelementptr inbounds i32* %b, i64 %idxprom
+  %0 = load i32* %arrayidx, align 4
+  %mul = mul nsw i32 %0, 3
+  %idxprom1 = sext i32 %i.0 to i64
+  %arrayidx2 = getelementptr inbounds i32 addrspace(1)* %a, i64 %idxprom1
+  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+; We should still be able to vectorize with noalias even if the
+; address spaces are different.
+define void @bar2(i32* noalias %a, i32 addrspace(1)* noalias %b, i32 %n) #0 {
+; CHECK-LABEL: @bar2(
+; CHECK: <4 x i32>
+; CHECK: ret
+
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp slt i32 %i.0, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %idxprom = sext i32 %i.0 to i64
+  %arrayidx = getelementptr inbounds i32 addrspace(1)* %b, i64 %idxprom
+  %0 = load i32 addrspace(1)* %arrayidx, align 4
+  %mul = mul nsw i32 %0, 3
+  %idxprom1 = sext i32 %i.0 to i64
+  %arrayidx2 = getelementptr inbounds i32* %a, i64 %idxprom1
+  store i32 %mul, i32* %arrayidx2, align 4
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+; Store to identified global with different address space. This isn't
+; generally safe and shouldn't be vectorized.
+define void @arst0(i32* %b, i32 %n) #0 {
+; CHECK-LABEL: @arst0(
+; CHECK-NOT: <4 x i32>
+; CHECK: ret
+
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp slt i32 %i.0, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %idxprom = sext i32 %i.0 to i64
+  %arrayidx = getelementptr inbounds i32* %b, i64 %idxprom
+  %0 = load i32* %arrayidx, align 4
+  %mul = mul nsw i32 %0, 3
+  %idxprom1 = sext i32 %i.0 to i64
+  %arrayidx2 = getelementptr inbounds [1024 x i32] addrspace(1)* @g_as1, i64 0, i64 %idxprom1
+  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+
+; Load from identified global with different address space.
+; This isn't generally safe and shouldn't be vectorized.
+define void @arst1(i32* %b, i32 %n) #0 {
+; CHECK-LABEL: @arst1(
+; CHECK-NOT: <4 x i32>
+; CHECK: ret
+
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp slt i32 %i.0, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %idxprom = sext i32 %i.0 to i64
+  %arrayidx = getelementptr inbounds [1024 x i32] addrspace(1)* @g_as1, i64 0, i64 %idxprom
+  %0 = load i32 addrspace(1)* %arrayidx, align 4
+  %mul = mul nsw i32 %0, 3
+  %idxprom1 = sext i32 %i.0 to i64
+  %arrayidx2 = getelementptr inbounds i32* %b, i64 %idxprom1
+  store i32 %mul, i32* %arrayidx2, align 4
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+; Read and write to 2 identified globals in different address
+; spaces. This should be vectorized.
+define void @aoeu(i32 %n) #0 {
+; CHECK-LABEL: @aoeu(
+; CHECK: <4 x i32>
+; CHECK: ret
+
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp slt i32 %i.0, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %idxprom = sext i32 %i.0 to i64
+  %arrayidx = getelementptr inbounds [1024 x i32] addrspace(2)* @q_as2, i64 0, i64 %idxprom
+  %0 = load i32 addrspace(2)* %arrayidx, align 4
+  %mul = mul nsw i32 %0, 3
+  %idxprom1 = sext i32 %i.0 to i64
+  %arrayidx2 = getelementptr inbounds [1024 x i32] addrspace(1)* @g_as1, i64 0, i64 %idxprom1
+  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+attributes #0 = { nounwind uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
diff --git a/test/Transforms/LoopVectorize/runtime-check-readonly-address-space.ll b/test/Transforms/LoopVectorize/runtime-check-readonly-address-space.ll
new file mode 100644
index 0000000..212b37c
--- /dev/null
+++ b/test/Transforms/LoopVectorize/runtime-check-readonly-address-space.ll
@@ -0,0 +1,142 @@
+; RUN: opt -S -march=r600 -mcpu=cayman -loop-vectorize -force-vector-unroll=1 -force-vector-width=4 -dce -instcombine < %s | FileCheck %s
+
+; Artificial datalayout
+target datalayout = "e-p:32:32:32-p1:16:16:16-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024-v2048:2048:2048-n32:64"
+
+
+define void @add_ints_1_1_1(i32 addrspace(1)* %a, i32 addrspace(1)* %b, i32 addrspace(1)* %c) #0 {
+; CHECK-LABEL: @add_ints_1_1_1(
+; CHECK: <4 x i32>
+; CHECK: ret
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp ult i64 %i.0, 200
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %arrayidx = getelementptr inbounds i32 addrspace(1)* %b, i64 %i.0
+  %0 = load i32 addrspace(1)* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32 addrspace(1)* %c, i64 %i.0
+  %1 = load i32 addrspace(1)* %arrayidx1, align 4
+  %add = add nsw i32 %0, %1
+  %arrayidx2 = getelementptr inbounds i32 addrspace(1)* %a, i64 %i.0
+  store i32 %add, i32 addrspace(1)* %arrayidx2, align 4
+  %inc = add i64 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+define void @add_ints_as_1_0_0(i32 addrspace(1)* %a, i32* %b, i32* %c) #0 {
+; CHECK-LABEL: @add_ints_as_1_0_0(
+; CHECK-NOT: <4 x i32>
+; CHECK: ret
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp ult i64 %i.0, 200
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %arrayidx = getelementptr inbounds i32* %b, i64 %i.0
+  %0 = load i32* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32* %c, i64 %i.0
+  %1 = load i32* %arrayidx1, align 4
+  %add = add nsw i32 %0, %1
+  %arrayidx2 = getelementptr inbounds i32 addrspace(1)* %a, i64 %i.0
+  store i32 %add, i32 addrspace(1)* %arrayidx2, align 4
+  %inc = add i64 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+define void @add_ints_as_0_1_0(i32* %a, i32 addrspace(1)* %b, i32* %c) #0 {
+; CHECK-LABEL: @add_ints_as_0_1_0(
+; CHECK-NOT: <4 x i32>
+; CHECK: ret
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp ult i64 %i.0, 200
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %arrayidx = getelementptr inbounds i32 addrspace(1)* %b, i64 %i.0
+  %0 = load i32 addrspace(1)* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32* %c, i64 %i.0
+  %1 = load i32* %arrayidx1, align 4
+  %add = add nsw i32 %0, %1
+  %arrayidx2 = getelementptr inbounds i32* %a, i64 %i.0
+  store i32 %add, i32* %arrayidx2, align 4
+  %inc = add i64 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+define void @add_ints_as_0_1_1(i32* %a, i32 addrspace(1)* %b, i32 addrspace(1)* %c) #0 {
+; CHECK-LABEL: @add_ints_as_0_1_1(
+; CHECK-NOT: <4 x i32>
+; CHECK: ret
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp ult i64 %i.0, 200
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %arrayidx = getelementptr inbounds i32 addrspace(1)* %b, i64 %i.0
+  %0 = load i32 addrspace(1)* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32 addrspace(1)* %c, i64 %i.0
+  %1 = load i32 addrspace(1)* %arrayidx1, align 4
+  %add = add nsw i32 %0, %1
+  %arrayidx2 = getelementptr inbounds i32* %a, i64 %i.0
+  store i32 %add, i32* %arrayidx2, align 4
+  %inc = add i64 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+define void @add_ints_as_0_1_2(i32* %a, i32 addrspace(1)* %b, i32 addrspace(2)* %c) #0 {
+; CHECK-LABEL: @add_ints_as_0_1_2(
+; CHECK-NOT: <4 x i32>
+; CHECK: ret
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp ult i64 %i.0, 200
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %arrayidx = getelementptr inbounds i32 addrspace(1)* %b, i64 %i.0
+  %0 = load i32 addrspace(1)* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32 addrspace(2)* %c, i64 %i.0
+  %1 = load i32 addrspace(2)* %arrayidx1, align 4
+  %add = add nsw i32 %0, %1
+  %arrayidx2 = getelementptr inbounds i32* %a, i64 %i.0
+  store i32 %add, i32* %arrayidx2, align 4
+  %inc = add i64 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+attributes #0 = { nounwind uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }