Reverse a loop that is counting up to a maximum to

count down to 0 instead, under very restricted
circumstances.  Adjust 4 testcases in which this
optimization fires.

llvm-svn: 71439

4 files changed, 11 insertions, 5 deletions

diff --git a/llvm/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll b/llvm/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll
index df6d76a0930..1b36fcec67a 100644
--- a/llvm/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll
+++ b/llvm/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll
@@ -1,5 +1,7 @@
 ; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 -stats |& \
 ; RUN:   grep {1 .*folded into instructions}
+; Increment in loop bb.128.i adjusted to 2, to prevent loop reversal from
+; kicking in.
 
 declare fastcc void @rdft(i32, i32, double*, i32*, double*)
 
@@ -41,7 +43,7 @@ bb.i28.i:		; preds = %bb.i28.i, %cond_next36.i
 	%tmp1213.i23.i = sitofp i32 %x.0.i21.i to double		; <double> [#uses=1]
 	%tmp15.i24.i = sub double 0.000000e+00, %tmp1213.i23.i		; <double> [#uses=1]
 	%tmp16.i25.i = mul double 0.000000e+00, %tmp15.i24.i		; <double> [#uses=1]
-	%indvar.next39.i = add i32 %j.0.reg2mem.0.i16.i, 1		; <i32> [#uses=2]
+	%indvar.next39.i = add i32 %j.0.reg2mem.0.i16.i, 2		; <i32> [#uses=2]
 	%exitcond40.i = icmp eq i32 %indvar.next39.i, %tmp8.i14.i		; <i1> [#uses=1]
 	br i1 %exitcond40.i, label %mp_unexp_d2mp.exit29.i, label %bb.i28.i
 
diff --git a/llvm/test/CodeGen/X86/iv-users-in-other-loops.ll b/llvm/test/CodeGen/X86/iv-users-in-other-loops.ll
index 67d9d49a44f..d11b025d589 100644
--- a/llvm/test/CodeGen/X86/iv-users-in-other-loops.ll
+++ b/llvm/test/CodeGen/X86/iv-users-in-other-loops.ll
@@ -1,5 +1,6 @@
 ; RUN: llvm-as < %s | llc -march=x86-64 -f -o %t
-; RUN: grep inc %t | count 2
+; RUN: grep inc %t | count 1
+; RUN: grep dec %t | count 2
 ; RUN: grep addq %t | count 13
 ; RUN: grep leaq %t | count 8
 ; RUN: grep leal %t | count 4
@@ -8,6 +9,7 @@
 ; IV users in each of the loops from other loops shouldn't cause LSR
 ; to insert new induction variables. Previously it would create a
 ; flood of new induction variables.
+; Also, the loop reversal should kick in once.
 
 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"
 target triple = "x86_64-unknown-linux-gnu"
diff --git a/llvm/test/CodeGen/X86/masked-iv-safe.ll b/llvm/test/CodeGen/X86/masked-iv-safe.ll
index 2ba3f830fdb..e9b80a4c42c 100644
--- a/llvm/test/CodeGen/X86/masked-iv-safe.ll
+++ b/llvm/test/CodeGen/X86/masked-iv-safe.ll
@@ -4,12 +4,13 @@
 ; RUN: not grep sar %t
 ; RUN: not grep shl %t
 ; RUN: grep add %t | count 6
-; RUN: grep inc %t | count 4
-; RUN: grep dec %t | count 2
+; RUN: grep inc %t | count 2
+; RUN: grep dec %t | count 4
 ; RUN: grep lea %t | count 2
 
 ; Optimize away zext-inreg and sext-inreg on the loop induction
 ; variable using trip-count information.
+; Also, the loop-reversal algorithm kicks in twice.
 
 define void @count_up(double* %d, i64 %n) nounwind {
 entry:
diff --git a/llvm/test/CodeGen/X86/pr3495.ll b/llvm/test/CodeGen/X86/pr3495.ll
index 726ad7413b0..62382c6d78c 100644
--- a/llvm/test/CodeGen/X86/pr3495.ll
+++ b/llvm/test/CodeGen/X86/pr3495.ll
@@ -1,7 +1,8 @@
 ; RUN: llvm-as < %s | llc -march=x86 -stats |& grep {Number of reloads omited} | grep 2
 ; RUN: llvm-as < %s | llc -march=x86 -stats |& not grep {Number of available reloads turned into copies}
-; RUN: llvm-as < %s | llc -march=x86 -stats |& grep {Number of machine instrs printed} | grep 39
+; RUN: llvm-as < %s | llc -march=x86 -stats |& grep {Number of machine instrs printed} | grep 38
 ; PR3495
+; The loop reversal kicks in once here, resulting in one fewer instruction.
 
 target triple = "i386-pc-linux-gnu"
 @x = external global [8 x i32], align 32		; <[8 x i32]*> [#uses=1]