[instcombine][unordered] Implement *-load forwarding for unordered atomics

This builds on 266999 which made FindAvailableValue do the right thing.  Tests included show the newly enabled transforms and those which disabled either due to conservatism or correctness requirements.

llvm-svn: 267006

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp b/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
index c855663..fc3de6b 100644 (file)
--- a/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
@@ -809,10 +809,6 @@ Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
       return &LI;
   }
 
-  // None of the following transforms are legal for volatile/atomic loads.
-  // FIXME: Some of it is okay for atomic loads; needs refactoring.
-  if (!LI.isSimple()) return nullptr;
-
   if (Instruction *Res = unpackLoadToAggregate(*this, LI))
     return Res;
 
@@ -840,6 +836,10 @@ Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
                                             LI.getName() + ".cast"));
   }
 
+  // None of the following transforms are legal for volatile/atomic loads.
+  // FIXME: Some of it is okay for atomic loads; needs refactoring.
+  if (!LI.isSimple()) return nullptr;
+
   // load(gep null, ...) -> unreachable
   if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
     const Value *GEPI0 = GEPI->getOperand(0);

diff --git a/llvm/test/Transforms/InstCombine/atomic.ll b/llvm/test/Transforms/InstCombine/atomic.ll
index bf576ba..75b63ff 100644 (file)
--- a/llvm/test/Transforms/InstCombine/atomic.ll
+++ b/llvm/test/Transforms/InstCombine/atomic.ll
@@ -37,11 +37,11 @@ define i32 @test3(i32* %p) {
   ret i32 %z
 }
 
-; FIXME: Forwarding from a stronger atomic is fine
+; Forwarding from a stronger ordered atomic is fine
 define i32 @test4(i32* %p) {
 ; CHECK-LABEL: define i32 @test4(
 ; CHECK: %x = load atomic i32, i32* %p seq_cst, align 4
-; CHECK: %y = load atomic i32, i32* %p unordered, align 4
+; CHECK: shl i32 %x, 1
   %x = load atomic i32, i32* %p seq_cst, align 4
   %y = load atomic i32, i32* %p unordered, align 4
   %z = add i32 %x, %y
@@ -60,3 +60,36 @@ define i32 @test5(i32* %p) {
   ret i32 %z
 }
 
+; Forwarding atomic to atomic is fine
+define i32 @test6(i32* %p) {
+; CHECK-LABEL: define i32 @test6(
+; CHECK: %x = load atomic i32, i32* %p unordered, align 4
+; CHECK: shl i32 %x, 1
+  %x = load atomic i32, i32* %p unordered, align 4
+  %y = load atomic i32, i32* %p unordered, align 4
+  %z = add i32 %x, %y
+  ret i32 %z
+}
+
+; FIXME: we currently don't do anything for monotonic
+define i32 @test7(i32* %p) {
+; CHECK-LABEL: define i32 @test7(
+; CHECK: %x = load atomic i32, i32* %p seq_cst, align 4
+; CHECK: %y = load atomic i32, i32* %p monotonic, align 4
+  %x = load atomic i32, i32* %p seq_cst, align 4
+  %y = load atomic i32, i32* %p monotonic, align 4
+  %z = add i32 %x, %y
+  ret i32 %z
+}
+
+; FIXME: We could forward in racy code
+define i32 @test8(i32* %p) {
+; CHECK-LABEL: define i32 @test8(
+; CHECK: %x = load atomic i32, i32* %p seq_cst, align 4
+; CHECK: %y = load atomic i32, i32* %p acquire, align 4
+  %x = load atomic i32, i32* %p seq_cst, align 4
+  %y = load atomic i32, i32* %p acquire, align 4
+  %z = add i32 %x, %y
+  ret i32 %z
+}
+