Revert rcpss/rsqrtss instrumentation to default propagation
_mm_round_ss() takes two vector arguments, while _mm_rcp_ss() takes only
one, despite both underlying instructions taking one source operand.
LLVM uses the C++ intrinsics behavior, so x86_sse_rcp_ss and
x86_sse_rsqrt_ss don't support the shuffle operation between the first
and second argument that is performed by handleUnarySdSsIntrinsic().
Instead the behavior of the fallback path, which calls
maybeHandleSimpleNomemIntrinsic() and propagates the shadow, was fine.
Bug: b/172238865
Change-Id: I1bfbcaf518dd90f5a113cdef2b75d1a602e43d90
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/65808
Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Sean Risser <srisser@google.com>
Kokoro-Result: kokoro <noreply+kokoro@google.com>
diff --git a/third_party/llvm-10.0/llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp b/third_party/llvm-10.0/llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp
index 347af58..dddd15a 100644
--- a/third_party/llvm-10.0/llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp
+++ b/third_party/llvm-10.0/llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp
@@ -3321,8 +3321,6 @@
case Intrinsic::x86_sse41_round_sd:
case Intrinsic::x86_sse41_round_ss:
- case Intrinsic::x86_sse_rcp_ss:
- case Intrinsic::x86_sse_rsqrt_ss:
handleUnarySdSsIntrinsic(I);
break;
case Intrinsic::x86_sse2_max_sd:
