[PATCH] D82398: [MSAN] Handle x86 {round,min,max}sd intrinsics

[Evgenii Stepanov via Phabricator via cfe-commits]

eugenis added inline comments.


================
Comment at: llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp:3077
+    Value *LowShadow = IRB.CreateOr(LowA, LowB);
+    Value *Shadow = IRB.CreateInsertElement(Second, LowShadow, IRB.getInt32(0));
+
----------------
guiand wrote:
> eugenis wrote:
> > You probably want to insert in First, not Second.
> > 
> > Is the generated code any better if you OR the vectors, and then shuffle to put the top element of First into the top element of the output? That's what LLVM generates if I express this logic in C.
> > 
> > 
> The codegen is basically identical either way, but if you'd like I can still upload a patch to change these into shufflevector instructions.
This is much better.
Use makeArrayRef({2, 1}).


================
Comment at: llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp:3331
+      // case Intrinsic::x86_avx512_mask_sub_sd_round:
+      // case Intrinsic::x86_avx512_mask_mul_sd_round:
+      // case Intrinsic::x86_avx512_mask_div_sd_round:
----------------
Unrelated change.


================
Comment at: llvm/test/Instrumentation/MemorySanitizer/vector_sd.ll:17
+  ; CHECK: [[OP2VEC_SHADOW:%.+]] = insertelement <2 x i64> <i64 -1, i64 -1>, i64 [[LOWVAL_SHADOW]], i32 0
+  %op2vec = insertelement <2 x double> undef, double %lowval, i32 0
+  ; CHECK: [[OP1VEC_SHADOW:%.+]] = insertelement <2 x i64> <i64 -1, i64 -1>, i64 [[HIVAL_SHADOW]], i32 1
----------------
pass the vectors as arguments, it will make the test case a lot simpler


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D82398/new/

https://reviews.llvm.org/D82398