diff options
| author | Evan Cheng <evan.cheng@apple.com> | 2011-03-29 23:06:19 +0000 |
|---|---|---|
| committer | Evan Cheng <evan.cheng@apple.com> | 2011-03-29 23:06:19 +0000 |
| commit | 92e3916c3b750f7eb4f41e14e401434b713e558b (patch) | |
| tree | 0c10c6ff7bc874bb3f979faab5d232c3aae3ed71 /lib/VMCore | |
| parent | 75c7563f834b06cfc71ef53bd4c37e58d2d96ff6 (diff) | |
| download | external_llvm-92e3916c3b750f7eb4f41e14e401434b713e558b.zip external_llvm-92e3916c3b750f7eb4f41e14e401434b713e558b.tar.gz external_llvm-92e3916c3b750f7eb4f41e14e401434b713e558b.tar.bz2 | |
Add intrinsics @llvm.arm.neon.vmulls and @llvm.arm.neon.vmullu.* back. Frontends
was lowering them to sext / uxt + mul instructions. Unfortunately the
optimization passes may hoist the extensions out of the loop and separate them.
When that happens, the long multiplication instructions can be broken into
several scalar instructions, causing significant performance issue.
Note the vmla and vmls intrinsics are not added back. Frontend will codegen them
as intrinsics vmull* + add / sub. Also note the isel optimizations for catching
mul + sext / zext are not changed either.
First part of rdar://8832507, rdar://9203134
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@128502 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/VMCore')
| -rw-r--r-- | lib/VMCore/AutoUpgrade.cpp | 1 |
1 files changed, 0 insertions, 1 deletions
diff --git a/lib/VMCore/AutoUpgrade.cpp b/lib/VMCore/AutoUpgrade.cpp index b323540..4e578ed 100644 --- a/lib/VMCore/AutoUpgrade.cpp +++ b/lib/VMCore/AutoUpgrade.cpp @@ -84,7 +84,6 @@ static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) { Name.compare(14, 5, "vsubl", 5) == 0 || Name.compare(14, 5, "vaddw", 5) == 0 || Name.compare(14, 5, "vsubw", 5) == 0 || - Name.compare(14, 5, "vmull", 5) == 0 || Name.compare(14, 5, "vmlal", 5) == 0 || Name.compare(14, 5, "vmlsl", 5) == 0 || Name.compare(14, 5, "vabdl", 5) == 0 || |