Reapply the new LoopStrengthReduction code, with compile time and

bug fixes, and with improved heuristics for analyzing foreign-loop
addrecs.

This change also flattens IVUsers, eliminating the stride-oriented
groupings, which makes it easier to work with.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@95975 91177308-0d34-0410-b5e6-96231b3b80d8

test/Transforms/LoopStrengthReduce/2008-08-06-CmpStride.ll

@@ -1,5 +1,4 @@
-; RUN: opt < %s -loop-reduce -S | grep ugt
-; PR2535
+; RUN: llc -march=x86-64 < %s -o - | grep {cmpl	\\$\[1\], %}
 
 @.str = internal constant [4 x i8] c"%d\0A\00"
 
@@ -16,7 +15,7 @@ forbody:
         %add166 = or i32 %mul15, 1              ; <i32> [#uses=1] *
         call i32 (i8*, ...)* @printf( i8* noalias  getelementptr ([4 x i8]* @.str, i32 0, i32 0), i32 %add166 ) nounwind
         %inc = add i32 %i.0, 1          ; <i32> [#uses=3]
-        %cmp = icmp ult i32 %inc, 1027          ; <i1> [#uses=1]
+        %cmp = icmp ne i32 %inc, 1027          ; <i1> [#uses=1]
         br i1 %cmp, label %forbody, label %afterfor
 
 afterfor:               ; preds = %forcond

test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-0.ll

@@ -1,10 +1,15 @@
-; RUN: llc %s -o - --x86-asm-syntax=att | grep {cmpl	\$4}
+; RUN: llc < %s -o - | FileCheck %s
 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-apple-darwin9"
 
-; This is like change-compare-stride-trickiness-1.ll except the comparison
-; happens before the relevant use, so the comparison stride can't be
-; easily changed.
+; The comparison happens before the relevant use, but it can still be rewritten
+; to compare with zero.
+
+; CHECK: foo:
+; CHECK: align
+; CHECK: incl  %eax
+; CHECK-NEXT: decl  %ecx
+; CHECK-NEXT: jne
 
 define void @foo() nounwind {
 entry:

test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-1.ll

@@ -1,10 +1,12 @@
-; RUN: llc %s -o - --x86-asm-syntax=att | grep {cmp.	\$8}
+; RUN: llc %s -o - --x86-asm-syntax=att | grep {cmp.	\$10}
 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-apple-darwin9"
 
 ; The comparison happens after the relevant use, so the stride can easily
 ; be changed. The comparison can be done in a narrower mode than the
 ; induction variable.
+; TODO: By making the first store post-increment as well, the loop setup
+; could be made simpler.
 
 define void @foo() nounwind {
 entry:

test/Transforms/LoopStrengthReduce/count-to-zero.ll

@@ -19,7 +19,7 @@ bb3:                                              ; preds = %bb1
   %tmp4 = add i32 %c_addr.1, -1                   ; <i32> [#uses=1]
   %c_addr.1.be = select i1 %tmp2, i32 %tmp3, i32 %tmp4 ; <i32> [#uses=1]
   %indvar.next = add i32 %indvar, 1               ; <i32> [#uses=1]
-; CHECK: sub i32 %lsr.iv, 1
+; CHECK: add i32 %lsr.iv, -1
   br label %bb6
 
 bb6:                                              ; preds = %bb3, %entry

test/Transforms/LoopStrengthReduce/invariant_value_first.ll

@@ -1,5 +1,5 @@
 ; Check that the index of 'P[outer]' is pulled out of the loop.
-; RUN: opt < %s -loop-reduce -S | \
+; RUN: opt < %s -loop-reduce -S -default-data-layout="e-p:32:32:32" | \
 ; RUN:   not grep {getelementptr.*%outer.*%INDVAR}
 
 declare i1 @pred()

test/Transforms/LoopStrengthReduce/invariant_value_first_arg.ll

@@ -1,5 +1,5 @@
 ; Check that the index of 'P[outer]' is pulled out of the loop.
-; RUN: opt < %s -loop-reduce -S | \
+; RUN: opt < %s -loop-reduce -S -default-data-layout="e-p:32:32:32" | \
 ; RUN:   not grep {getelementptr.*%outer.*%INDVAR}
 
 declare i1 @pred()

test/Transforms/LoopStrengthReduce/ops_after_indvar.ll

@@ -1,7 +1,7 @@
 ; Check that this test makes INDVAR and related stuff dead, because P[indvar]
 ; gets reduced, making INDVAR dead.
 
-; RUN: opt < %s -loop-reduce -S | not grep INDVAR
+; RUN: opt < %s -loop-reduce -S -default-data-layout="e-p:32:32:32" | not grep INDVAR
 
 declare i1 @pred()
 

test/Transforms/LoopStrengthReduce/quadradic-exit-value.ll

@@ -1,4 +1,4 @@
-; RUN: opt < %s -analyze -iv-users | grep {Stride i64 {3,+,2}<%loop>:}
+; RUN: opt < %s -analyze -iv-users | grep {\{1,+,3,+,2\}<%loop> (post-inc)}
 
 ; The value of %r is dependent on a polynomial iteration expression.
 

test/Transforms/LoopStrengthReduce/remove_indvar.ll

@@ -7,10 +7,12 @@ define void @test(i32* %P) {
 ; <label>:0
 	br label %Loop
 Loop:		; preds = %Loop, %0
+        %i = phi i32 [ 0, %0 ], [ %i.next, %Loop ]
 	%INDVAR = phi i32 [ 0, %0 ], [ %INDVAR2, %Loop ]		; <i32> [#uses=2]
 	%STRRED = getelementptr i32* %P, i32 %INDVAR		; <i32*> [#uses=1]
 	store i32 0, i32* %STRRED
 	%INDVAR2 = add i32 %INDVAR, 1		; <i32> [#uses=1]
+        %i.next = add i32 %i, 1
 	%cond = call i1 @pred( )		; <i1> [#uses=1]
 	br i1 %cond, label %Loop, label %Out
 Out:		; preds = %Loop

test/Transforms/LoopStrengthReduce/use_postinc_value_outside_loop.ll

@@ -1,5 +1,5 @@
 ; RUN: opt < %s -loop-reduce -S | \
-; RUN:   grep {add i32 %lsr.iv.next, 1}
+; RUN:   grep {add i32 %indvar630.ui, 1}
 ;
 ; Make sure that the use of the IV outside of the loop (the store) uses the 
 ; post incremented value of the IV, not the preincremented value.  This