From: Dan Gohman <gohman@apple.com>
Date: Tue, 22 Jun 2010 23:07:13 +0000 (+0000)
Subject: Fix OptimizeMax to handle an odd case where one of the max operands
X-Git-Url: http://plrg.eecs.uci.edu/git/?a=commitdiff_plain;h=caf71ab4735d6b80300cf123a8c3c0759d2e2873;p=oota-llvm.git

Fix OptimizeMax to handle an odd case where one of the max operands
is another max which folds. This fixes PR7454.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106594 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Transforms/Scalar/LoopStrengthReduce.cpp b/lib/Transforms/Scalar/LoopStrengthReduce.cpp
index d14d7d99f27..e8da89a12cd 100644
--- a/lib/Transforms/Scalar/LoopStrengthReduce.cpp
+++ b/lib/Transforms/Scalar/LoopStrengthReduce.cpp
@@ -1643,8 +1643,11 @@ ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
     NewRHS = Sel->getOperand(1);
   else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
     NewRHS = Sel->getOperand(2);
+  else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
+    NewRHS = SU->getValue();
   else
-    llvm_unreachable("Max doesn't match expected pattern!");
+    // Max doesn't match expected pattern.
+    return Cond;
 
   // Determine the new comparison opcode. It may be signed or unsigned,
   // and the original comparison may be either equality or inequality.
diff --git a/test/CodeGen/X86/optimize-max-3.ll b/test/CodeGen/X86/optimize-max-3.ll
index bf8bfa28daf..11c3507f359 100644
--- a/test/CodeGen/X86/optimize-max-3.ll
+++ b/test/CodeGen/X86/optimize-max-3.ll
@@ -1,4 +1,4 @@
-; RUN: llc < %s -march=x86-64 | FileCheck %s
+; RUN: llc < %s -march=x86-64 -asm-verbose=false | FileCheck %s
 
 ; LSR's OptimizeMax should eliminate the select (max).
 
@@ -30,3 +30,46 @@ for.body:                                         ; preds = %for.body.preheader,
 for.end:                                          ; preds = %for.body, %entry
   ret void
 }
+
+; In this case, one of the max operands is another max, which folds,
+; leaving a two-operand max which doesn't fit the usual pattern.
+; OptimizeMax should handle this case.
+; PR7454
+
+;      CHECK: _Z18GenerateStatusPagei:
+
+;      CHECK:         jle
+; CHECK-NEXT:         xorl    %edi, %edi
+; CHECK-NEXT:         align
+; CHECK-NEXT: BB1_2:
+; CHECK-NEXT:         callq
+; CHECK-NEXT:         incl    %ebx
+; CHECK-NEXT:         cmpl    %r14d, %ebx
+; CHECK-NEXT:         movq    %rax, %rdi
+; CHECK-NEXT:         jl
+
+define void @_Z18GenerateStatusPagei(i32 %jobs_to_display) nounwind {
+entry:
+  %cmp.i = icmp sgt i32 %jobs_to_display, 0       ; <i1> [#uses=1]
+  %tmp = select i1 %cmp.i, i32 %jobs_to_display, i32 0 ; <i32> [#uses=3]
+  %cmp8 = icmp sgt i32 %tmp, 0                    ; <i1> [#uses=1]
+  br i1 %cmp8, label %bb.nph, label %for.end
+
+bb.nph:                                           ; preds = %entry
+  %tmp11 = icmp sgt i32 %tmp, 1                   ; <i1> [#uses=1]
+  %smax = select i1 %tmp11, i32 %tmp, i32 1       ; <i32> [#uses=1]
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %bb.nph
+  %i.010 = phi i32 [ 0, %bb.nph ], [ %inc, %for.body ] ; <i32> [#uses=1]
+  %it.0.09 = phi float* [ null, %bb.nph ], [ %call.i, %for.body ] ; <float*> [#uses=1]
+  %call.i = call float* @_ZSt18_Rb_tree_decrementPKSt18_Rb_tree_node_base(float* %it.0.09) ; <float*> [#uses=1]
+  %inc = add nsw i32 %i.010, 1                    ; <i32> [#uses=2]
+  %exitcond = icmp eq i32 %inc, %smax             ; <i1> [#uses=1]
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+declare float* @_ZSt18_Rb_tree_decrementPKSt18_Rb_tree_node_base(float*)