[llvm-commits] [llvm] r102157 - in /llvm/trunk: lib/Analysis/README.txt lib/Transforms/Scalar/LoopStrengthReduce.cpp test/CodeGen/X86/lsr-delayed-fold.ll

[Dan Gohman]

Author: djg
Date: Thu Apr 22 20:55:05 2010
New Revision: 102157

URL: http://llvm.org/viewvc/llvm-project?rev=102157&view=rev
Log:
Fix LSR to tolerate cases where ScalarEvolution initially
misses an opportunity to fold add operands, but folds them
after LSR has separated them out. This fixes rdar://7886751.

Added:
    llvm/trunk/test/CodeGen/X86/lsr-delayed-fold.ll
Modified:
    llvm/trunk/lib/Analysis/README.txt
    llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp

Modified: llvm/trunk/lib/Analysis/README.txt
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/README.txt?rev=102157&r1=102156&r2=102157&view=diff
==============================================================================
--- llvm/trunk/lib/Analysis/README.txt (original)
+++ llvm/trunk/lib/Analysis/README.txt Thu Apr 22 20:55:05 2010
@@ -16,3 +16,15 @@
 which is very inefficient when expanded into code.
 
 //===---------------------------------------------------------------------===//
+
+In test/CodeGen/X86/lsr-delayed-fold.ll,
+
+ScalarEvolution is forming this expression:
+
+((trunc i64 (-1 * %arg5) to i32) + (trunc i64 %arg5 to i32) + (-1 * (trunc i64 undef to i32)))
+
+This could be folded to
+
+(-1 * (trunc i64 undef to i32))
+
+//===---------------------------------------------------------------------===//

Modified: llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp?rev=102157&r1=102156&r2=102157&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp Thu Apr 22 20:55:05 2010
@@ -2060,8 +2060,11 @@
                            LU.Kind, LU.AccessTy, TLI, SE))
         continue;
 
+      const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
+      if (InnerSum->isZero())
+        continue;
       Formula F = Base;
-      F.BaseRegs[i] = SE.getAddExpr(InnerAddOps);
+      F.BaseRegs[i] = InnerSum;
       F.BaseRegs.push_back(*J);
       if (InsertFormula(LU, LUIdx, F))
         // If that formula hadn't been seen before, recurse to find more like

Added: llvm/trunk/test/CodeGen/X86/lsr-delayed-fold.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/lsr-delayed-fold.ll?rev=102157&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/lsr-delayed-fold.ll (added)
+++ llvm/trunk/test/CodeGen/X86/lsr-delayed-fold.ll Thu Apr 22 20:55:05 2010
@@ -0,0 +1,28 @@
+; RUN: llc -march=x86-64 < %s > /dev/null
+; rdar://7886751
+
+; ScalarEvolution misses an opportunity to fold ((trunc x) + (trunc -x) + y),
+; but LSR should tolerate this.
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-apple-darwin11.0"
+
+define fastcc void @formatValue(i64 %arg5) nounwind {
+bb12:                                             ; preds = %bb11
+  %t = trunc i64 %arg5 to i32                   ; <i32> [#uses=1]
+  %t13 = sub i64 0, %arg5                       ; <i64> [#uses=1]
+  %t14 = and i64 %t13, 4294967295             ; <i64> [#uses=1]
+  br label %bb15
+
+bb15:                                             ; preds = %bb21, %bb12
+  %t16 = phi i64 [ 0, %bb12 ], [ %t23, %bb15 ] ; <i64> [#uses=2]
+  %t17 = mul i64 %t14, %t16                 ; <i64> [#uses=1]
+  %t18 = add i64 undef, %t17                  ; <i64> [#uses=1]
+  %t19 = trunc i64 %t18 to i32                ; <i32> [#uses=1]
+  %t22 = icmp eq i32 %t19, %t               ; <i1> [#uses=1]
+  %t23 = add i64 %t16, 1                      ; <i64> [#uses=1]
+  br i1 %t22, label %bb24, label %bb15
+
+bb24:                                             ; preds = %bb21, %bb11
+  unreachable
+}