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

13 files changed, 455 insertions, 36 deletions

diff --git a/test/CodeGen/X86/2006-05-11-InstrSched.ll b/test/CodeGen/X86/2006-05-11-InstrSched.ll
index bdbe713a29..56d6aa960e 100644
--- a/test/CodeGen/X86/2006-05-11-InstrSched.ll
+++ b/test/CodeGen/X86/2006-05-11-InstrSched.ll
@@ -1,5 +1,5 @@
 ; RUN: llc < %s -march=x86 -mattr=+sse2 -stats -realign-stack=0 |&\
-; RUN:     grep {asm-printer} | grep 31
+; RUN:     grep {asm-printer} | grep 34
 
 target datalayout = "e-p:32:32"
 define void @foo(i32* %mc, i32* %bp, i32* %ms, i32* %xmb, i32* %mpp, i32* %tpmm, i32* %ip, i32* %tpim, i32* %dpp, i32* %tpdm, i32* %bpi, i32 %M) nounwind {
@@ -40,7 +40,7 @@ cond_true:		; preds = %cond_true, %entry
 	%tmp137.upgrd.7 = bitcast i32* %tmp137 to <2 x i64>*		; <<2 x i64>*> [#uses=1]
 	store <2 x i64> %tmp131, <2 x i64>* %tmp137.upgrd.7
 	%tmp147 = add nsw i32 %tmp.10, 8		; <i32> [#uses=1]
-	%tmp.upgrd.8 = icmp slt i32 %tmp147, %M		; <i1> [#uses=1]
+	%tmp.upgrd.8 = icmp ne i32 %tmp147, %M		; <i1> [#uses=1]
 	%indvar.next = add i32 %indvar, 1		; <i32> [#uses=1]
 	br i1 %tmp.upgrd.8, label %cond_true, label %return
 
diff --git a/test/CodeGen/X86/2007-03-15-GEP-Idx-Sink.ll b/test/CodeGen/X86/2007-03-15-GEP-Idx-Sink.ll
index 4cac9b4c4a..e1f890192d 100644
--- a/test/CodeGen/X86/2007-03-15-GEP-Idx-Sink.ll
+++ b/test/CodeGen/X86/2007-03-15-GEP-Idx-Sink.ll
@@ -1,7 +1,7 @@
 ; RUN: llc < %s -march=x86 -mtriple=i686-darwin | \
 ; RUN:   grep push | count 3
 
-define void @foo(i8** %buf, i32 %size, i32 %col, i8* %p) {
+define void @foo(i8** %buf, i32 %size, i32 %col, i8* %p) nounwind {
 entry:
 	icmp sgt i32 %size, 0		; <i1>:0 [#uses=1]
 	br i1 %0, label %bb.preheader, label %return
diff --git a/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll b/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll
index 721d4c945b..8e315f4d80 100644
--- a/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll
+++ b/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll
@@ -35,7 +35,7 @@ cond_next36.i:		; preds = %cond_next.i
 bb.i28.i:		; preds = %bb.i28.i, %cond_next36.i
 ; CHECK: %bb.i28.i
 ; CHECK: addl $2
-; CHECK: addl $2
+; CHECK: addl $-2
 	%j.0.reg2mem.0.i16.i = phi i32 [ 0, %cond_next36.i ], [ %indvar.next39.i, %bb.i28.i ]		; <i32> [#uses=2]
 	%din_addr.1.reg2mem.0.i17.i = phi double [ 0.000000e+00, %cond_next36.i ], [ %tmp16.i25.i, %bb.i28.i ]		; <double> [#uses=1]
 	%tmp1.i18.i = fptosi double %din_addr.1.reg2mem.0.i17.i to i32		; <i32> [#uses=2]
diff --git a/test/CodeGen/X86/full-lsr.ll b/test/CodeGen/X86/full-lsr.ll
index 3bd58b65be..ff9b1b0b6a 100644
--- a/test/CodeGen/X86/full-lsr.ll
+++ b/test/CodeGen/X86/full-lsr.ll
@@ -1,12 +1,7 @@
 ; RUN: llc < %s -march=x86 >%t
 
-; TODO: Enhance full lsr mode to get this:
-; RUNX: grep {addl	\\\$4,} %t | count 3
-; RUNX: not grep {,%} %t
-
-; For now, it should find this, which is still pretty good:
-; RUN: not grep {addl	\\\$4,} %t
-; RUN: grep {,%} %t | count 6
+; RUN: grep {addl	\\\$4,} %t | count 3
+; RUN: not grep {,%} %t
 
 define void @foo(float* nocapture %A, float* nocapture %B, float* nocapture %C, i32 %N) nounwind {
 entry:
diff --git a/test/CodeGen/X86/iv-users-in-other-loops.ll b/test/CodeGen/X86/iv-users-in-other-loops.ll
index c695c29e06..408fb20b8d 100644
--- a/test/CodeGen/X86/iv-users-in-other-loops.ll
+++ b/test/CodeGen/X86/iv-users-in-other-loops.ll
@@ -1,11 +1,11 @@
 ; RUN: llc < %s -march=x86-64 -o %t
-; RUN: grep inc %t | count 1
+; RUN: not grep inc %t
 ; RUN: grep dec %t | count 2
 ; RUN: grep addq %t | count 13
 ; RUN: not grep addb %t
-; RUN: grep leaq %t | count 9
-; RUN: grep leal %t | count 3
-; RUN: grep movq %t | count 5
+; RUN: not grep leaq %t
+; RUN: not grep leal %t
+; RUN: not grep movq %t
 
 ; IV users in each of the loops from other loops shouldn't cause LSR
 ; to insert new induction variables. Previously it would create a
diff --git a/test/CodeGen/X86/loop-strength-reduce-2.ll b/test/CodeGen/X86/loop-strength-reduce-2.ll
index 30b5114349..b546462b68 100644
--- a/test/CodeGen/X86/loop-strength-reduce-2.ll
+++ b/test/CodeGen/X86/loop-strength-reduce-2.ll
@@ -1,11 +1,24 @@
-; RUN: llc < %s -march=x86 -relocation-model=pic | \
-; RUN:   grep {, 4} | count 1
-; RUN: llc < %s -march=x86 | not grep lea
+; RUN: llc < %s -march=x86 -relocation-model=pic | FileCheck %s -check-prefix=PIC
+; RUN: llc < %s -march=x86 -relocation-model=static | FileCheck %s -check-prefix=STATIC
 ;
 ; Make sure the common loop invariant A is hoisted up to preheader,
 ; since too many registers are needed to subsume it into the addressing modes.
 ; It's safe to sink A in when it's not pic.
 
+; PIC:  align
+; PIC:  movl  $4, -4([[REG:%e[a-z]+]])
+; PIC:  movl  $5, ([[REG]])
+; PIC:  addl  $4, [[REG]]
+; PIC:  decl  {{%e[[a-z]+}}
+; PIC:  jne
+
+; STATIC: align
+; STATIC: movl  $4, -4(%ecx)
+; STATIC: movl  $5, (%ecx)
+; STATIC: addl  $4, %ecx
+; STATIC: decl  %eax
+; STATIC: jne
+
 @A = global [16 x [16 x i32]] zeroinitializer, align 32		; <[16 x [16 x i32]]*> [#uses=2]
 
 define void @test(i32 %row, i32 %N.in) nounwind {
diff --git a/test/CodeGen/X86/loop-strength-reduce-3.ll b/test/CodeGen/X86/loop-strength-reduce-3.ll
index 70c91340c9..b1c9fb9c07 100644
--- a/test/CodeGen/X86/loop-strength-reduce-3.ll
+++ b/test/CodeGen/X86/loop-strength-reduce-3.ll
@@ -1,8 +1,11 @@
-; RUN: llc < %s -mtriple=i386-apple-darwin -relocation-model=dynamic-no-pic | \
-; RUN:   grep {A+} | count 2
-;
-; Make sure the common loop invariant A is not hoisted up to preheader,
-; since it can be subsumed it into the addressing modes.
+; RUN: llc < %s -mtriple=i386-apple-darwin -relocation-model=dynamic-no-pic | FileCheck %s
+
+; CHECK: align
+; CHECK: movl  $4, -4(%ecx)
+; CHECK: movl  $5, (%ecx)
+; CHECK: addl  $4, %ecx
+; CHECK: decl  %eax
+; CHECK: jne
 
 @A = global [16 x [16 x i32]] zeroinitializer, align 32		; <[16 x [16 x i32]]*> [#uses=2]
 
diff --git a/test/CodeGen/X86/loop-strength-reduce.ll b/test/CodeGen/X86/loop-strength-reduce.ll
index 4cb56ca9ed..42c6ac4983 100644
--- a/test/CodeGen/X86/loop-strength-reduce.ll
+++ b/test/CodeGen/X86/loop-strength-reduce.ll
@@ -1,8 +1,11 @@
-; RUN: llc < %s -march=x86 -relocation-model=static | \
-; RUN:   grep {A+} | count 2
-;
-; Make sure the common loop invariant A is not hoisted up to preheader,
-; since it can be subsumed into the addressing mode in all uses.
+; RUN: llc < %s -march=x86 -relocation-model=static | FileCheck %s
+
+; CHECK: align
+; CHECK: movl  $4, -4(%ecx)
+; CHECK: movl  $5, (%ecx)
+; CHECK: addl  $4, %ecx
+; CHECK: decl  %eax
+; CHECK: jne
 
 @A = internal global [16 x [16 x i32]] zeroinitializer, align 32		; <[16 x [16 x i32]]*> [#uses=2]
 
diff --git a/test/CodeGen/X86/loop-strength-reduce4.ll b/test/CodeGen/X86/loop-strength-reduce4.ll
index 07e46eca75..6c0eb8c0df 100644
--- a/test/CodeGen/X86/loop-strength-reduce4.ll
+++ b/test/CodeGen/X86/loop-strength-reduce4.ll
@@ -1,5 +1,19 @@
-; RUN: llc < %s -march=x86 | grep cmp | grep 64
-; RUN: llc < %s -march=x86 | not grep inc
+; RUN: llc < %s -march=x86 -relocation-model=static -mtriple=i686-apple-darwin | FileCheck %s -check-prefix=STATIC
+; RUN: llc < %s -march=x86 -relocation-model=pic | FileCheck %s -check-prefix=PIC
+
+; By starting the IV at -64 instead of 0, a cmp is eliminated,
+; as the flags from the add can be used directly.
+
+; STATIC: movl    $-64, %ecx
+
+; STATIC: movl    %eax, _state+76(%ecx)
+; STATIC: addl    $16, %ecx
+; STATIC: jne
+
+; In PIC mode the symbol can't be folded, so the change-compare-stride
+; trick applies.
+
+; PIC: cmpl $64
 
 @state = external global [0 x i32]		; <[0 x i32]*> [#uses=4]
 @S = external global [0 x i32]		; <[0 x i32]*> [#uses=4]
diff --git a/test/CodeGen/X86/loop-strength-reduce8.ll b/test/CodeGen/X86/loop-strength-reduce8.ll
index e14cd8a99e..6b2247d1d6 100644
--- a/test/CodeGen/X86/loop-strength-reduce8.ll
+++ b/test/CodeGen/X86/loop-strength-reduce8.ll
@@ -1,4 +1,10 @@
-; RUN: llc < %s -mtriple=i386-apple-darwin | grep leal | not grep 16
+; RUN: llc < %s -mtriple=i386-apple-darwin | FileCheck %s
+
+; CHECK: leal 16(%eax), %edx
+; CHECK: align
+; CHECK: addl    $4, %edx
+; CHECK: decl    %ecx
+; CHECK: jne     LBB1_2
 
 	%struct.CUMULATIVE_ARGS = type { i32, i32, i32, i32, i32, i32, i32 }
 	%struct.bitmap_element = type { %struct.bitmap_element*, %struct.bitmap_element*, i32, [2 x i64] }
diff --git a/test/CodeGen/X86/lsr-reuse.ll b/test/CodeGen/X86/lsr-reuse.ll
new file mode 100644
index 0000000000..7f2b8cc8f8
--- /dev/null
+++ b/test/CodeGen/X86/lsr-reuse.ll
@@ -0,0 +1,386 @@
+; RUN: llc < %s -march=x86-64 -O3 | FileCheck %s
+target datalayout = "e-p:64:64:64"
+target triple = "x86_64-unknown-unknown"
+
+; Full strength reduction reduces register pressure from 5 to 4 here.
+; Instruction selection should use the FLAGS value from the dec for
+; the branch. Scheduling should push the adds upwards.
+
+; CHECK: full_me_0:
+; CHECK: movsd   (%rsi), %xmm0
+; CHECK: addq    $8, %rsi
+; CHECK: mulsd   (%rdx), %xmm0
+; CHECK: addq    $8, %rdx
+; CHECK: movsd   %xmm0, (%rdi)
+; CHECK: addq    $8, %rdi
+; CHECK: decq    %rcx
+; CHECK: jne
+
+define void @full_me_0(double* nocapture %A, double* nocapture %B, double* nocapture %C, i64 %n) nounwind {
+entry:
+  %t0 = icmp sgt i64 %n, 0
+  br i1 %t0, label %loop, label %return
+
+loop:
+  %i = phi i64 [ %i.next, %loop ], [ 0, %entry ]
+  %Ai = getelementptr inbounds double* %A, i64 %i
+  %Bi = getelementptr inbounds double* %B, i64 %i
+  %Ci = getelementptr inbounds double* %C, i64 %i
+  %t1 = load double* %Bi
+  %t2 = load double* %Ci
+  %m = fmul double %t1, %t2
+  store double %m, double* %Ai
+  %i.next = add nsw i64 %i, 1
+  %exitcond = icmp eq i64 %i.next, %n
+  br i1 %exitcond, label %return, label %loop
+
+return:
+  ret void
+}
+
+; Mostly-full strength reduction means we do full strength reduction on all
+; except for the offsets.
+;
+; Given a choice between constant offsets -2048 and 2048, choose the negative
+; value, because at boundary conditions it has a smaller encoding.
+; TODO: That's an over-general heuristic. It would be better for the target
+; to indicate what the encoding cost would be. Then using a 2048 offset
+; would be better on x86-64, since the start value would be 0 instead of
+; 2048.
+
+; CHECK: mostly_full_me_0:
+; CHECK: movsd   -2048(%rsi), %xmm0
+; CHECK: mulsd   -2048(%rdx), %xmm0
+; CHECK: movsd   %xmm0, -2048(%rdi)
+; CHECK: movsd   (%rsi), %xmm0
+; CHECK: addq    $8, %rsi
+; CHECK: divsd   (%rdx), %xmm0
+; CHECK: addq    $8, %rdx
+; CHECK: movsd   %xmm0, (%rdi)
+; CHECK: addq    $8, %rdi
+; CHECK: decq    %rcx
+; CHECK: jne
+
+define void @mostly_full_me_0(double* nocapture %A, double* nocapture %B, double* nocapture %C, i64 %n) nounwind {
+entry:
+  %t0 = icmp sgt i64 %n, 0
+  br i1 %t0, label %loop, label %return
+
+loop:
+  %i = phi i64 [ %i.next, %loop ], [ 0, %entry ]
+  %Ai = getelementptr inbounds double* %A, i64 %i
+  %Bi = getelementptr inbounds double* %B, i64 %i
+  %Ci = getelementptr inbounds double* %C, i64 %i
+  %t1 = load double* %Bi
+  %t2 = load double* %Ci
+  %m = fmul double %t1, %t2
+  store double %m, double* %Ai
+  %j = add i64 %i, 256
+  %Aj = getelementptr inbounds double* %A, i64 %j
+  %Bj = getelementptr inbounds double* %B, i64 %j
+  %Cj = getelementptr inbounds double* %C, i64 %j
+  %t3 = load double* %Bj
+  %t4 = load double* %Cj
+  %o = fdiv double %t3, %t4
+  store double %o, double* %Aj
+  %i.next = add nsw i64 %i, 1
+  %exitcond = icmp eq i64 %i.next, %n
+  br i1 %exitcond, label %return, label %loop
+
+return:
+  ret void
+}
+
+; A minor variation on mostly_full_me_0.
+; Prefer to start the indvar at 0.
+
+; CHECK: mostly_full_me_1:
+; CHECK: movsd   (%rsi), %xmm0
+; CHECK: mulsd   (%rdx), %xmm0
+; CHECK: movsd   %xmm0, (%rdi)
+; CHECK: movsd   -2048(%rsi), %xmm0
+; CHECK: addq    $8, %rsi
+; CHECK: divsd   -2048(%rdx), %xmm0
+; CHECK: addq    $8, %rdx
+; CHECK: movsd   %xmm0, -2048(%rdi)
+; CHECK: addq    $8, %rdi
+; CHECK: decq    %rcx
+; CHECK: jne
+
+define void @mostly_full_me_1(double* nocapture %A, double* nocapture %B, double* nocapture %C, i64 %n) nounwind {
+entry:
+  %t0 = icmp sgt i64 %n, 0
+  br i1 %t0, label %loop, label %return
+
+loop:
+  %i = phi i64 [ %i.next, %loop ], [ 0, %entry ]
+  %Ai = getelementptr inbounds double* %A, i64 %i
+  %Bi = getelementptr inbounds double* %B, i64 %i
+  %Ci = getelementptr inbounds double* %C, i64 %i
+  %t1 = load double* %Bi
+  %t2 = load double* %Ci
+  %m = fmul double %t1, %t2
+  store double %m, double* %Ai
+  %j = sub i64 %i, 256
+  %Aj = getelementptr inbounds double* %A, i64 %j
+  %Bj = getelementptr inbounds double* %B, i64 %j
+  %Cj = getelementptr inbounds double* %C, i64 %j
+  %t3 = load double* %Bj
+  %t4 = load double* %Cj
+  %o = fdiv double %t3, %t4
+  store double %o, double* %Aj
+  %i.next = add nsw i64 %i, 1
+  %exitcond = icmp eq i64 %i.next, %n
+  br i1 %exitcond, label %return, label %loop
+
+return:
+  ret void
+}
+
+; A slightly less minor variation on mostly_full_me_0.
+
+; CHECK: mostly_full_me_2:
+; CHECK: movsd   (%rsi), %xmm0
+; CHECK: mulsd   (%rdx), %xmm0
+; CHECK: movsd   %xmm0, (%rdi)
+; CHECK: movsd   -4096(%rsi), %xmm0
+; CHECK: addq    $8, %rsi
+; CHECK: divsd   -4096(%rdx), %xmm0
+; CHECK: addq    $8, %rdx
+; CHECK: movsd   %xmm0, -4096(%rdi)
+; CHECK: addq    $8, %rdi
+; CHECK: decq    %rcx
+; CHECK: jne
+
+define void @mostly_full_me_2(double* nocapture %A, double* nocapture %B, double* nocapture %C, i64 %n) nounwind {
+entry:
+  %t0 = icmp sgt i64 %n, 0
+  br i1 %t0, label %loop, label %return
+
+loop:
+  %i = phi i64 [ %i.next, %loop ], [ 0, %entry ]
+  %k = add i64 %i, 256
+  %Ak = getelementptr inbounds double* %A, i64 %k
+  %Bk = getelementptr inbounds double* %B, i64 %k
+  %Ck = getelementptr inbounds double* %C, i64 %k
+  %t1 = load double* %Bk
+  %t2 = load double* %Ck
+  %m = fmul double %t1, %t2
+  store double %m, double* %Ak
+  %j = sub i64 %i, 256
+  %Aj = getelementptr inbounds double* %A, i64 %j
+  %Bj = getelementptr inbounds double* %B, i64 %j
+  %Cj = getelementptr inbounds double* %C, i64 %j
+  %t3 = load double* %Bj
+  %t4 = load double* %Cj
+  %o = fdiv double %t3, %t4
+  store double %o, double* %Aj
+  %i.next = add nsw i64 %i, 1
+  %exitcond = icmp eq i64 %i.next, %n
+  br i1 %exitcond, label %return, label %loop
+
+return:
+  ret void
+}
+
+; In this test, the counting IV exit value is used, so full strength reduction
+; would not reduce register pressure. IndVarSimplify ought to simplify such
+; cases away, but it's useful here to verify that LSR's register pressure
+; heuristics are working as expected.
+
+; CHECK: count_me_0:
+; CHECK: movsd   (%rsi,%rax,8), %xmm0
+; CHECK: mulsd   (%rdx,%rax,8), %xmm0
+; CHECK: movsd   %xmm0, (%rdi,%rax,8)
+; CHECK: incq    %rax
+; CHECK: cmpq    %rax, %rcx
+; CHECK: jne
+
+define i64 @count_me_0(double* nocapture %A, double* nocapture %B, double* nocapture %C, i64 %n) nounwind {
+entry:
+  %t0 = icmp sgt i64 %n, 0
+  br i1 %t0, label %loop, label %return
+
+loop:
+  %i = phi i64 [ %i.next, %loop ], [ 0, %entry ]
+  %Ai = getelementptr inbounds double* %A, i64 %i
+  %Bi = getelementptr inbounds double* %B, i64 %i
+  %Ci = getelementptr inbounds double* %C, i64 %i
+  %t1 = load double* %Bi
+  %t2 = load double* %Ci
+  %m = fmul double %t1, %t2
+  store double %m, double* %Ai
+  %i.next = add nsw i64 %i, 1
+  %exitcond = icmp eq i64 %i.next, %n
+  br i1 %exitcond, label %return, label %loop
+
+return:
+  %q = phi i64 [ 0, %entry ], [ %i.next, %loop ]
+  ret i64 %q
+}
+
+; In this test, the trip count value is used, so full strength reduction
+; would not reduce register pressure.
+; (though it would reduce register pressure inside the loop...)
+
+; CHECK: count_me_1:
+; CHECK: movsd   (%rsi,%rax,8), %xmm0
+; CHECK: mulsd   (%rdx,%rax,8), %xmm0
+; CHECK: movsd   %xmm0, (%rdi,%rax,8)
+; CHECK: incq    %rax
+; CHECK: cmpq    %rax, %rcx
+; CHECK: jne
+
+define i64 @count_me_1(double* nocapture %A, double* nocapture %B, double* nocapture %C, i64 %n) nounwind {
+entry:
+  %t0 = icmp sgt i64 %n, 0
+  br i1 %t0, label %loop, label %return
+
+loop:
+  %i = phi i64 [ %i.next, %loop ], [ 0, %entry ]
+  %Ai = getelementptr inbounds double* %A, i64 %i
+  %Bi = getelementptr inbounds double* %B, i64 %i
+  %Ci = getelementptr inbounds double* %C, i64 %i
+  %t1 = load double* %Bi
+  %t2 = load double* %Ci
+  %m = fmul double %t1, %t2
+  store double %m, double* %Ai
+  %i.next = add nsw i64 %i, 1
+  %exitcond = icmp eq i64 %i.next, %n
+  br i1 %exitcond, label %return, label %loop
+
+return:
+  %q = phi i64 [ 0, %entry ], [ %n, %loop ]
+  ret i64 %q
+}
+
+; Full strength reduction doesn't save any registers here because the
+; loop tripcount is a constant.
+
+; CHECK: count_me_2:
+; CHECK: movl    $10, %eax
+; CHECK: align
+; CHECK: BB7_1:
+; CHECK: movsd   -40(%rdi,%rax,8), %xmm0
+; CHECK: addsd   -40(%rsi,%rax,8), %xmm0
+; CHECK: movsd   %xmm0, -40(%rdx,%rax,8)
+; CHECK: movsd   (%rdi,%rax,8), %xmm0
+; CHECK: subsd   (%rsi,%rax,8), %xmm0
+; CHECK: movsd   %xmm0, (%rdx,%rax,8)
+; CHECK: incq    %rax
+; CHECK: cmpq    $5010, %rax
+; CHECK: jne
+
+define void @count_me_2(double* nocapture %A, double* nocapture %B, double* nocapture %C) nounwind {
+entry:
+  br label %loop
+
+loop:
+  %i = phi i64 [ 0, %entry ], [ %i.next, %loop ]
+  %i5 = add i64 %i, 5
+  %Ai = getelementptr double* %A, i64 %i5
+  %t2 = load double* %Ai
+  %Bi = getelementptr double* %B, i64 %i5
+  %t4 = load double* %Bi
+  %t5 = fadd double %t2, %t4
+  %Ci = getelementptr double* %C, i64 %i5
+  store double %t5, double* %Ci
+  %i10 = add i64 %i, 10
+  %Ai10 = getelementptr double* %A, i64 %i10
+  %t9 = load double* %Ai10
+  %Bi10 = getelementptr double* %B, i64 %i10
+  %t11 = load double* %Bi10
+  %t12 = fsub double %t9, %t11
+  %Ci10 = getelementptr double* %C, i64 %i10
+  store double %t12, double* %Ci10
+  %i.next = add i64 %i, 1
+  %exitcond = icmp eq i64 %i.next, 5000
+  br i1 %exitcond, label %return, label %loop
+
+return:
+  ret void
+}
+
+; This should be fully strength-reduced to reduce register pressure.
+
+; CHECK: full_me_1:
+; CHECK: align
+; CHECK: BB8_1:
+; CHECK: movsd   (%rdi), %xmm0
+; CHECK: addsd   (%rsi), %xmm0
+; CHECK: movsd   %xmm0, (%rdx)
+; CHECK: movsd   40(%rdi), %xmm0
+; CHECK: addq    $8, %rdi
+; CHECK: subsd   40(%rsi), %xmm0
+; CHECK: addq    $8, %rsi
+; CHECK: movsd   %xmm0, 40(%rdx)
+; CHECK: addq    $8, %rdx
+; CHECK: decq    %rcx
+; CHECK: jne
+
+define void @full_me_1(double* nocapture %A, double* nocapture %B, double* nocapture %C, i64 %n) nounwind {
+entry:
+  br label %loop
+
+loop:
+  %i = phi i64 [ 0, %entry ], [ %i.next, %loop ]
+  %i5 = add i64 %i, 5
+  %Ai = getelementptr double* %A, i64 %i5
+  %t2 = load double* %Ai
+  %Bi = getelementptr double* %B, i64 %i5
+  %t4 = load double* %Bi
+  %t5 = fadd double %t2, %t4
+  %Ci = getelementptr double* %C, i64 %i5
+  store double %t5, double* %Ci
+  %i10 = add i64 %i, 10
+  %Ai10 = getelementptr double* %A, i64 %i10
+  %t9 = load double* %Ai10
+  %Bi10 = getelementptr double* %B, i64 %i10
+  %t11 = load double* %Bi10
+  %t12 = fsub double %t9, %t11
+  %Ci10 = getelementptr double* %C, i64 %i10
+  store double %t12, double* %Ci10
+  %i.next = add i64 %i, 1
+  %exitcond = icmp eq i64 %i.next, %n
+  br i1 %exitcond, label %return, label %loop
+
+return:
+  ret void
+}
+
+; This is a variation on full_me_0 in which the 0,+,1 induction variable
+; has a non-address use, pinning that value in a register.
+
+; CHECK: count_me_3:
+; CHECK: call
+; CHECK: movsd   (%r15,%r13,8), %xmm0
+; CHECK: mulsd   (%r14,%r13,8), %xmm0
+; CHECK: movsd   %xmm0, (%r12,%r13,8)
+; CHECK: incq    %r13
+; CHECK: cmpq    %r13, %rbx
+; CHECK: jne
+
+declare void @use(i64)
+
+define void @count_me_3(double* nocapture %A, double* nocapture %B, double* nocapture %C, i64 %n) nounwind {
+entry:
+  %t0 = icmp sgt i64 %n, 0
+  br i1 %t0, label %loop, label %return
+
+loop:
+  %i = phi i64 [ %i.next, %loop ], [ 0, %entry ]
+  call void @use(i64 %i)
+  %Ai = getelementptr inbounds double* %A, i64 %i
+  %Bi = getelementptr inbounds double* %B, i64 %i
+  %Ci = getelementptr inbounds double* %C, i64 %i
+  %t1 = load double* %Bi
+  %t2 = load double* %Ci
+  %m = fmul double %t1, %t2
+  store double %m, double* %Ai
+  %i.next = add nsw i64 %i, 1
+  %exitcond = icmp eq i64 %i.next, %n
+  br i1 %exitcond, label %return, label %loop
+
+return:
+  ret void
+}
diff --git a/test/CodeGen/X86/masked-iv-safe.ll b/test/CodeGen/X86/masked-iv-safe.ll
index bc493bd8f7..0b4d73a683 100644
--- a/test/CodeGen/X86/masked-iv-safe.ll
+++ b/test/CodeGen/X86/masked-iv-safe.ll
@@ -169,7 +169,7 @@ loop:
 	%indvar.i24 = and i64 %indvar, 16777215
 	%t3 = getelementptr double* %d, i64 %indvar.i24
 	%t4 = load double* %t3
-	%t5 = fmul double %t4, 2.3
+	%t5 = fdiv double %t4, 2.3
 	store double %t5, double* %t3
 	%t6 = getelementptr double* %d, i64 %indvar
 	%t7 = load double* %t6
@@ -199,7 +199,7 @@ loop:
 	%indvar.i24 = ashr i64 %s1, 24
 	%t3 = getelementptr double* %d, i64 %indvar.i24
 	%t4 = load double* %t3
-	%t5 = fmul double %t4, 2.3
+	%t5 = fdiv double %t4, 2.3
 	store double %t5, double* %t3
 	%t6 = getelementptr double* %d, i64 %indvar
 	%t7 = load double* %t6
@@ -229,7 +229,7 @@ loop:
 	%indvar.i24 = ashr i64 %s1, 24
 	%t3 = getelementptr double* %d, i64 %indvar.i24
 	%t4 = load double* %t3
-	%t5 = fmul double %t4, 2.3
+	%t5 = fdiv double %t4, 2.3
 	store double %t5, double* %t3
 	%t6 = getelementptr double* %d, i64 %indvar
 	%t7 = load double* %t6
diff --git a/test/CodeGen/X86/pr3495.ll b/test/CodeGen/X86/pr3495.ll
index 14f2a54486..e84a84f59b 100644
--- a/test/CodeGen/X86/pr3495.ll
+++ b/test/CodeGen/X86/pr3495.ll
@@ -1,8 +1,7 @@
 ; RUN: llc < %s -march=x86 -stats |& grep {Number of loads added} | grep 2
 ; RUN: llc < %s -march=x86 -stats |& grep {Number of register spills} | grep 1
-; RUN: llc < %s -march=x86 -stats |& grep {Number of machine instrs printed} | grep 37
+; RUN: llc < %s -march=x86 -stats |& grep {Number of machine instrs printed} | grep 34
 ; 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]