[Local] Consider atomic loads from constant global as dead

Per the guidance in
https://llvm.org/docs/Atomics.html#atomics-and-ir-optimization,
an atomic load from a constant global can be dropped, as there can
be no stores to synchronize with. Any write to the constant global
would be UB.

IPSCCP will already drop such loads, but the main helper in Local
doesn't recognize this currently. This is motivated by D118387.

Differential Revision: https://reviews.llvm.org/D124241

llvm/lib/Transforms/Utils/Local.cpp

@@ -500,6 +500,13 @@ bool llvm::wouldInstructionBeTriviallyDead(Instruction *I,
     if (isMathLibCallNoop(Call, TLI))
       return true;
 
+  // Non-volatile atomic loads from constants can be removed.
+  if (auto *LI = dyn_cast<LoadInst>(I))
+    if (auto *GV = dyn_cast<GlobalVariable>(
+            LI->getPointerOperand()->stripPointerCasts()))
+      if (!LI->isVolatile() && GV->isConstant())
+        return true;
+
   return false;
 }
 

llvm/test/CodeGen/PowerPC/atomics-constant.ll

@@ -3,16 +3,14 @@
 
 target triple = "powerpc64le-unknown-linux-gnu"
 
-@a = dso_local constant i64 zeroinitializer
+@a = dso_local global i64 zeroinitializer
 
 define i64 @foo() {
 ; CHECK-LABEL: foo:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    li 4, 0
 ; CHECK-NEXT:    addis 3, 2, a@toc@ha
 ; CHECK-NEXT:    ld 3, a@toc@l(3)
-; CHECK-NEXT:    cmpd 7, 4, 4
-; CHECK-NEXT:    li 3, 0
+; CHECK-NEXT:    cmpd 7, 3, 3
 ; CHECK-NEXT:    bne- 7, .+4
 ; CHECK-NEXT:    isync
 ; CHECK-NEXT:    blr

llvm/test/Transforms/InstCombine/atomic.ll

@@ -426,7 +426,6 @@ define void @no_atomic_vector_store(<2 x float> %p, i8* %p2) {
 
 define i32 @atomic_load_from_constant_global() {
 ; CHECK-LABEL: @atomic_load_from_constant_global(
-; CHECK-NEXT:    [[V:%.*]] = load atomic i32, i32* @c seq_cst, align 4
 ; CHECK-NEXT:    ret i32 42
 ;
   %v = load atomic i32, i32* @c seq_cst, align 4
@@ -435,7 +434,6 @@ define i32 @atomic_load_from_constant_global() {
 
 define i8 @atomic_load_from_constant_global_bitcast() {
 ; CHECK-LABEL: @atomic_load_from_constant_global_bitcast(
-; CHECK-NEXT:    [[V:%.*]] = load atomic i8, i8* bitcast (i32* @c to i8*) seq_cst, align 1
 ; CHECK-NEXT:    ret i8 42
 ;
   %v = load atomic i8, i8* bitcast (i32* @c to i8*) seq_cst, align 1