[llvm] r267006 - [instcombine][unordered] Implement *-load forwarding for unordered atomics

[Philip Reames via llvm-commits]

Author: reames
Date: Thu Apr 21 12:03:33 2016
New Revision: 267006

URL: http://llvm.org/viewvc/llvm-project?rev=267006&view=rev
Log:
[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.


Modified:
    llvm/trunk/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
    llvm/trunk/test/Transforms/InstCombine/atomic.ll

Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp?rev=267006&r1=267005&r2=267006&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp (original)
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp Thu Apr 21 12:03:33 2016
@@ -809,10 +809,6 @@ Instruction *InstCombiner::visitLoadInst
       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
                                             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);

Modified: llvm/trunk/test/Transforms/InstCombine/atomic.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/atomic.ll?rev=267006&r1=267005&r2=267006&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstCombine/atomic.ll (original)
+++ llvm/trunk/test/Transforms/InstCombine/atomic.ll Thu Apr 21 12:03:33 2016
@@ -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
@@ -59,4 +59,37 @@ define i32 @test5(i32* %p) {
   %z = add i32 %x, %y
   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
+}