[llvm] r335411 - Fix invariant fdiv hoisting in LICM

[Stanislav Mekhanoshin via llvm-commits]

Author: rampitec
Date: Fri Jun 22 21:01:28 2018
New Revision: 335411

URL: http://llvm.org/viewvc/llvm-project?rev=335411&view=rev
Log:
Fix invariant fdiv hoisting in LICM

FDiv is replaced with multiplication by reciprocal and invariant
reciprocal is hoisted out of the loop, while multiplication remains
even if invariant.

Switch checks for all invariant operands and only invariant
denominator to fix the issue.

Differential Revision: https://reviews.llvm.org/D48447

Modified:
    llvm/trunk/lib/Transforms/Scalar/LICM.cpp
    llvm/trunk/test/Transforms/LICM/hoist-fast-fdiv.ll

Modified: llvm/trunk/lib/Transforms/Scalar/LICM.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/LICM.cpp?rev=335411&r1=335410&r2=335411&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Scalar/LICM.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/LICM.cpp Fri Jun 22 21:01:28 2018
@@ -477,6 +477,20 @@ bool llvm::hoistRegion(DomTreeNode *N, A
         continue;
       }
 
+      // Try hoisting the instruction out to the preheader.  We can only do
+      // this if all of the operands of the instruction are loop invariant and
+      // if it is safe to hoist the instruction.
+      //
+      if (CurLoop->hasLoopInvariantOperands(&I) &&
+          canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, SafetyInfo, ORE) &&
+          (IsMustExecute ||
+           isSafeToExecuteUnconditionally(
+               I, DT, CurLoop, SafetyInfo, ORE,
+               CurLoop->getLoopPreheader()->getTerminator()))) {
+        Changed |= hoist(I, DT, CurLoop, SafetyInfo, ORE);
+        continue;
+      }
+
       // Attempt to remove floating point division out of the loop by
       // converting it to a reciprocal multiplication.
       if (I.getOpcode() == Instruction::FDiv &&
@@ -500,20 +514,6 @@ bool llvm::hoistRegion(DomTreeNode *N, A
         continue;
       }
 
-      // Try hoisting the instruction out to the preheader.  We can only do
-      // this if all of the operands of the instruction are loop invariant and
-      // if it is safe to hoist the instruction.
-      //
-      if (CurLoop->hasLoopInvariantOperands(&I) &&
-          canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, SafetyInfo, ORE) &&
-          (IsMustExecute ||
-           isSafeToExecuteUnconditionally(
-               I, DT, CurLoop, SafetyInfo, ORE,
-               CurLoop->getLoopPreheader()->getTerminator()))) {
-        Changed |= hoist(I, DT, CurLoop, SafetyInfo, ORE);
-        continue;
-      }
-
       if (IsMustExecute)
         IsMustExecute = isGuaranteedToTransferExecutionToSuccessor(&I);
     }

Modified: llvm/trunk/test/Transforms/LICM/hoist-fast-fdiv.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/LICM/hoist-fast-fdiv.ll?rev=335411&r1=335410&r2=335411&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/LICM/hoist-fast-fdiv.ll (original)
+++ llvm/trunk/test/Transforms/LICM/hoist-fast-fdiv.ll Fri Jun 22 21:01:28 2018
@@ -1,9 +1,9 @@
 ; RUN: opt -licm -S < %s | FileCheck %s
 
 ; Function Attrs: noinline norecurse nounwind readnone ssp uwtable
-define zeroext i1 @f(double %v) #0 {
+define zeroext i1 @invariant_denom(double %v) #0 {
 entry:
-; CHECK-LABEL: @f(
+; CHECK-LABEL: @invariant_denom(
 ; CHECK-NEXT: entry:
 ; CHECK-NEXT: fdiv fast double 1.000000e+00, %v
   br label %loop
@@ -32,3 +32,29 @@ end:
   ret i1 %v16
 }
 
+define void @invariant_fdiv(float* %out, float %arg) {
+; CHECK-LABEL: @invariant_fdiv(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: %div = fdiv fast float 4.000000e+00, %arg
+; CHECK-NEXT: fmul fast float %div, 0x41F0000000000000
+entry:
+  br label %loop
+
+loop:                                              ; preds = %loop, %entry
+  %ind = phi i32 [ 0, %entry ], [ %inc, %loop ]
+
+; CHECK-LABEL: loop:
+; CHECK: getelementptr
+; CHECK-NOT: fdiv
+; CHECK-NOT: fmul
+  %div = fdiv fast float 4.000000e+00, %arg
+  %mul = fmul fast float %div, 0x41F0000000000000
+  %gep = getelementptr inbounds float, float* %out, i32 %ind
+  store float %mul, float* %gep, align 4
+  %inc = add nuw nsw i32 %ind, 1
+  %cond = icmp eq i32 %inc, 1024
+  br i1 %cond, label %exit, label %loop
+
+exit:                                              ; preds = %loop
+  ret void
+}