Work around MemorySanitizer movmsk false positives

MemorySanitizer does not recognize the x86 movmskps and pmovmskb instructions. In such cases it falls back to checking all input operands for uninitialized bits. This causes false positives specifically with Reactor's Byte8 type which is often loaded from memory, leaving the upper 64-bit of an SSE register undefined. This change masks out the unused parts of the input vector, leaving just the sign bits. Bug: b/172238865 Change-Id: I50c921a7ff8a4ebdba89136bb82c9f46ccdc3769 Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/54268 Kokoro-Result: kokoro <noreply+kokoro@google.com> Tested-by: Nicolas Capens <nicolascapens@google.com> Reviewed-by: Alexis Hétu <sugoi@google.com>

Work around MemorySanitizer movmsk false positives
b162fcf9 · Nicolas Capens · Nicolas Capens · be169ef3 · b162fcf9
Commit b162fcf9 authored May 12, 2021 by Nicolas Capens Committed by Nicolas Capens May 12, 2021
Hide whitespace changes
Inline Side-by-side

Showing with 22 additions and 2 deletions

LLVMReactor.cpp src/Reactor/LLVMReactor.cpp +22 -2

No files found.
--- a/src/Reactor/LLVMReactor.cpp
+++ b/src/Reactor/LLVMReactor.cpp
@@ -3960,12 +3960,32 @@ RValue<Int4> pmaddwd(RValue<Short8> x, RValue<Short8> y)

 RValue<Int> movmskps(RValue<Float4> x)
 {
-	return RValue<Int>(createInstruction(llvm::Intrinsic::x86_sse_movmsk_ps, x.value()));
+	Value *v = x.value();
+
+	// TODO(b/172238865): MemorySanitizer does not support movmsk instructions,
+	// which makes it look at the entire 128-bit input for undefined bits. Mask off
+	// just the sign bits to avoid false positives.
+	if(__has_feature(memory_sanitizer))
+	{
+		v = As<Float4>(As<Int4>(v) & Int4(0x80000000u)).value();
+	}
+
+	return RValue<Int>(createInstruction(llvm::Intrinsic::x86_sse_movmsk_ps, v));
 }

 RValue<Int> pmovmskb(RValue<Byte8> x)
 {
-	return RValue<Int>(createInstruction(llvm::Intrinsic::x86_sse2_pmovmskb_128, x.value())) & 0xFF;
+	Value *v = x.value();
+
+	// TODO(b/172238865): MemorySanitizer does not support movmsk instructions,
+	// which makes it look at the entire 128-bit input for undefined bits. Mask off
+	// just the sign bits in the lower 64-bit vector to avoid false positives.
+	if(__has_feature(memory_sanitizer))
+	{
+		v = As<Byte16>(As<Int4>(v) & Int4(0x80808080u, 0x80808080u, 0, 0)).value();
+	}
+
+	return RValue<Int>(createInstruction(llvm::Intrinsic::x86_sse2_pmovmskb_128, v)) & 0xFF;
 }

 RValue<Int4> pmovzxbd(RValue<Byte16> x)