[PATCH] D124871: Handle instrumentation of scalar single-precision (_ss) intrinsics

[Nicolas Capens via Phabricator via llvm-commits]

capn created this revision.
capn added reviewers: eugenis, vitalybuka.
Herald added subscribers: pengfei, hiraditya.
Herald added a project: All.
capn requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Instrumentation of scalar double-precision intrinsics such as
x86_sse41_round_sd was already handled by https://reviews.llvm.org/D82398,
but not their single-precision counterparts. Also rcpss and rsqrtss are
useful to support.

https://issuetracker.google.com/172238865


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D124871

Files:
  llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp


Index: llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp
===================================================================
--- llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp
+++ llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp
@@ -3202,27 +3202,39 @@
     SOC.Done(&I);
   }
 
-  // Instrument _mm_*_sd intrinsics
-  void handleUnarySdIntrinsic(IntrinsicInst &I) {
+  // Instrument _mm_*_sd|ss intrinsics
+  void handleUnarySdSsIntrinsic(IntrinsicInst &I) {
     IRBuilder<> IRB(&I);
+    unsigned Width =
+        cast<FixedVectorType>(I.getArgOperand(0)->getType())->getNumElements();
     Value *First = getShadow(&I, 0);
     Value *Second = getShadow(&I, 1);
-    // High word of first operand, low word of second
-    Value *Shadow =
-        IRB.CreateShuffleVector(First, Second, llvm::makeArrayRef<int>({2, 1}));
+    // First element of second operand, remaining elements of first operand
+    SmallVector<int, 16> Mask;
+    Mask.push_back(Width);
+    for (int i = 1; i < Width; i++) {
+      Mask.push_back(i);
+    }
+    Value *Shadow = IRB.CreateShuffleVector(First, Second, Mask);
 
     setShadow(&I, Shadow);
     setOriginForNaryOp(I);
   }
 
-  void handleBinarySdIntrinsic(IntrinsicInst &I) {
+  void handleBinarySdSsIntrinsic(IntrinsicInst &I) {
     IRBuilder<> IRB(&I);
+    unsigned Width =
+        cast<FixedVectorType>(I.getArgOperand(0)->getType())->getNumElements();
     Value *First = getShadow(&I, 0);
     Value *Second = getShadow(&I, 1);
     Value *OrShadow = IRB.CreateOr(First, Second);
-    // High word of first operand, low word of both OR'd together
-    Value *Shadow = IRB.CreateShuffleVector(First, OrShadow,
-                                            llvm::makeArrayRef<int>({2, 1}));
+    // First element of both OR'd together, remaining elements of first operand
+    SmallVector<int, 16> Mask;
+    Mask.push_back(Width);
+    for (int i = 1; i < Width; i++) {
+      Mask.push_back(i);
+    }
+    Value *Shadow = IRB.CreateShuffleVector(First, OrShadow, Mask);
 
     setShadow(&I, Shadow);
     setOriginForNaryOp(I);
@@ -3497,11 +3509,16 @@
       break;
 
     case Intrinsic::x86_sse41_round_sd:
-      handleUnarySdIntrinsic(I);
+    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:
+    case Intrinsic::x86_sse_max_ss:
     case Intrinsic::x86_sse2_min_sd:
-      handleBinarySdIntrinsic(I);
+    case Intrinsic::x86_sse_min_ss:
+      handleBinarySdSsIntrinsic(I);
       break;
 
     case Intrinsic::fshl:


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D124871.426812.patch
Type: text/x-patch
Size: 2646 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220503/e505b79d/attachment.bin>