2 files changed, 87 insertions, 0 deletions
diff --git a/test/CodeGen/NVPTX/intrin-nocapture.ll b/test/CodeGen/NVPTX/intrin-nocapture.ll
new file mode 100644
index 0000000000..55781bb15a
--- /dev/null
+++ b/test/CodeGen/NVPTX/intrin-nocapture.ll
@@ -0,0 +1,21 @@
+; RUN: opt < %s -O3 -S | FileCheck %s
+
+; Address space intrinsics were erroneously marked NoCapture, leading to bad
+; optimizations (such as the store below being eliminated as dead code). This
+; test makes sure we don't regress.
+
+declare void @foo(i32 addrspace(1)*)
+
+declare i32 addrspace(1)* @llvm.nvvm.ptr.gen.to.global.p1i32.p0i32(i32*)
+
+; CHECK: @bar
+define void @bar() {
+  %t1 = alloca i32
+; CHECK: call i32 addrspace(1)* @llvm.nvvm.ptr.gen.to.global.p1i32.p0i32(i32* %t1)
+; CHECK-NEXT: store i32 10, i32* %t1
+  %t2 = call i32 addrspace(1)* @llvm.nvvm.ptr.gen.to.global.p1i32.p0i32(i32* %t1)
+  store i32 10, i32* %t1
+  call void @foo(i32 addrspace(1)* %t2)
+  ret void
+}
+
diff --git a/test/CodeGen/NVPTX/vector-loads.ll b/test/CodeGen/NVPTX/vector-loads.ll
new file mode 100644
index 0000000000..f5a1795e3c
--- /dev/null
+++ b/test/CodeGen/NVPTX/vector-loads.ll
@@ -0,0 +1,66 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
+
+; Even though general vector types are not supported in PTX, we can still
+; optimize loads/stores with pseudo-vector instructions of the form:
+;
+; ld.v2.f32 {%f0, %f1}, [%r0]
+;
+; which will load two floats at once into scalar registers.
+
+define void @foo(<2 x float>* %a) {
+; CHECK: .func foo
+; CHECK: ld.v2.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}}, [%r{{[0-9]+}}];
+  %t1 = load <2 x float>* %a
+  %t2 = fmul <2 x float> %t1, %t1
+  store <2 x float> %t2, <2 x float>* %a
+  ret void
+}
+
+define void @foo2(<4 x float>* %a) {
+; CHECK: .func foo2
+; CHECK: ld.v4.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}, [%r{{[0-9]+}}];
+  %t1 = load <4 x float>* %a
+  %t2 = fmul <4 x float> %t1, %t1
+  store <4 x float> %t2, <4 x float>* %a
+  ret void
+}
+
+define void @foo3(<8 x float>* %a) {
+; CHECK: .func foo3
+; CHECK: ld.v4.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}, [%r{{[0-9]+}}];
+; CHECK-NEXT: ld.v4.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}, [%r{{[0-9]+}}+16];
+  %t1 = load <8 x float>* %a
+  %t2 = fmul <8 x float> %t1, %t1
+  store <8 x float> %t2, <8 x float>* %a
+  ret void
+}
+
+
+
+define void @foo4(<2 x i32>* %a) {
+; CHECK: .func foo4
+; CHECK: ld.v2.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}}, [%r{{[0-9]+}}];
+  %t1 = load <2 x i32>* %a
+  %t2 = mul <2 x i32> %t1, %t1
+  store <2 x i32> %t2, <2 x i32>* %a
+  ret void
+}
+
+define void @foo5(<4 x i32>* %a) {
+; CHECK: .func foo5
+; CHECK: ld.v4.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}}, [%r{{[0-9]+}}];
+  %t1 = load <4 x i32>* %a
+  %t2 = mul <4 x i32> %t1, %t1
+  store <4 x i32> %t2, <4 x i32>* %a
+  ret void
+}
+
+define void @foo6(<8 x i32>* %a) {
+; CHECK: .func foo6
+; CHECK: ld.v4.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}}, [%r{{[0-9]+}}];
+; CHECK-NEXT: ld.v4.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}}, [%r{{[0-9]+}}+16];
+  %t1 = load <8 x i32>* %a
+  %t2 = mul <8 x i32> %t1, %t1
+  store <8 x i32> %t2, <8 x i32>* %a
+  ret void
+}