summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorTom Stellard <thomas.stellard@amd.com>2014-04-02 19:53:29 +0000
committerTom Stellard <thomas.stellard@amd.com>2014-04-02 19:53:29 +0000
commit36a031870bb98f220e4a2e467b3f10de10080509 (patch)
tree3b0b4038478c69ffd2ed4dfa7d73aee008500ab0
parent0f58c7cc3a2081f1a21549f71dcfac07ca8d1713 (diff)
downloadbcm5719-llvm-36a031870bb98f220e4a2e467b3f10de10080509.tar.gz
bcm5719-llvm-36a031870bb98f220e4a2e467b3f10de10080509.zip
TargetLibraryInfo: Disable memcpy and memset on R600
There are no implementations of these for R600. llvm-svn: 205455
-rw-r--r--llvm/lib/Target/TargetLibraryInfo.cpp11
-rw-r--r--llvm/test/CodeGen/R600/loop-idiom.ll54
2 files changed, 64 insertions, 1 deletions
diff --git a/llvm/lib/Target/TargetLibraryInfo.cpp b/llvm/lib/Target/TargetLibraryInfo.cpp
index 7f8919772f7..6ec0b1f929c 100644
--- a/llvm/lib/Target/TargetLibraryInfo.cpp
+++ b/llvm/lib/Target/TargetLibraryInfo.cpp
@@ -378,7 +378,16 @@ static void initialize(TargetLibraryInfo &TLI, const Triple &T,
llvm_unreachable("TargetLibraryInfo function names must be sorted");
}
#endif // !NDEBUG
-
+
+ // There are no library implementations of mempcy and memset for r600 and
+ // these can be difficult to lower in the backend.
+ if (T.getArch() == Triple::r600) {
+ TLI.setUnavailable(LibFunc::memcpy);
+ TLI.setUnavailable(LibFunc::memset);
+ TLI.setUnavailable(LibFunc::memset_pattern16);
+ return;
+ }
+
// memset_pattern16 is only available on iOS 3.0 and Mac OS X 10.5 and later.
if (T.isMacOSX()) {
if (T.isMacOSXVersionLT(10, 5))
diff --git a/llvm/test/CodeGen/R600/loop-idiom.ll b/llvm/test/CodeGen/R600/loop-idiom.ll
new file mode 100644
index 00000000000..8a9cba2796c
--- /dev/null
+++ b/llvm/test/CodeGen/R600/loop-idiom.ll
@@ -0,0 +1,54 @@
+; RUN: opt -basicaa -loop-idiom -S < %s -march=r600 -mcpu=redwood | FileCheck --check-prefix=R600 --check-prefix=FUNC %s
+; RUN: opt -basicaa -loop-idiom -S < %s -march=r600 -mcpu=SI | FileCheck --check-prefix=SI --check-prefix=FUNC %s
+
+target datalayout = "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:32:32-p5:64:64-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64"
+target triple = "r600--"
+
+
+
+; Make sure loop-idiom doesn't create memcpy or memset. There are no library
+; implementations of these for R600.
+
+; FUNC: @no_memcpy
+; R600-NOT: @llvm.memcpy
+; SI-NOT: @llvm.memcpy
+define void @no_memcpy(i8 addrspace(3)* %in, i32 %size) {
+entry:
+ %dest = alloca i8, i32 32
+ br label %for.body
+
+for.body:
+ %0 = phi i32 [0, %entry], [%4, %for.body]
+ %1 = getelementptr i8 addrspace(3)* %in, i32 %0
+ %2 = getelementptr i8* %dest, i32 %0
+ %3 = load i8 addrspace(3)* %1
+ store i8 %3, i8* %2
+ %4 = add i32 %0, 1
+ %5 = icmp eq i32 %4, %size
+ br i1 %5, label %for.end, label %for.body
+
+for.end:
+ ret void
+}
+
+; FUNC: @no_memset
+; R600-NOT: @llvm.memset
+; R600-NOT: @memset_pattern16
+; SI-NOT: @llvm.memset
+; SI-NOT: @memset_pattern16
+define void @no_memset(i32 %size) {
+entry:
+ %dest = alloca i8, i32 32
+ br label %for.body
+
+for.body:
+ %0 = phi i32 [0, %entry], [%2, %for.body]
+ %1 = getelementptr i8* %dest, i32 %0
+ store i8 0, i8* %1
+ %2 = add i32 %0, 1
+ %3 = icmp eq i32 %2, %size
+ br i1 %3, label %for.end, label %for.body
+
+for.end:
+ ret void
+}
OpenPOWER on IntegriCloud