[ARM] Enable shrink-wrapping by default.

[oota-llvm.git] / test / CodeGen / ARM / sub-cmp-peephole.ll

diff --git a/test/CodeGen/ARM/sub-cmp-peephole.ll b/test/CodeGen/ARM/sub-cmp-peephole.ll
index 1b411e3af1cd7b09e043fe204048099f097af3c9..32038883469b73e360af66992bd4ebd47188c87f 100644 (file)
--- a/test/CodeGen/ARM/sub-cmp-peephole.ll
+++ b/test/CodeGen/ARM/sub-cmp-peephole.ll
@@ -1,4 +1,7 @@
 ; RUN: llc < %s -mtriple=arm-apple-darwin | FileCheck %s
+; RUN: llc < %s -mtriple=arm-apple-darwin | FileCheck %s --check-prefix=V7
+; RUN: llc < %s -mtriple=armv8-none-linux-gnueabi | FileCheck %s -check-prefix=V8
+
 
 define i32 @f(i32 %a, i32 %b) nounwind ssp {
 entry:
@@ -84,3 +87,120 @@ land.lhs.true:                                    ; preds = %num2long.exit
 if.end11:                                         ; preds = %num2long.exit
   ret i32 23
 }
+
+; When considering the producer of cmp's src as the subsuming instruction,
+; only consider that when the comparison is to 0.
+define i32 @cmp_src_nonzero(i32 %a, i32 %b, i32 %x, i32 %y) {
+entry:
+; CHECK-LABEL: cmp_src_nonzero:
+; CHECK: sub
+; CHECK: cmp
+  %sub = sub i32 %a, %b
+  %cmp = icmp eq i32 %sub, 17
+  %ret = select i1 %cmp, i32 %x, i32 %y
+  ret i32 %ret
+}
+
+define float @float_sel(i32 %a, i32 %b, float %x, float %y) {
+entry:
+; CHECK-LABEL: float_sel:
+; CHECK-NOT: cmp
+; V8-LABEL: float_sel:
+; V8-NOT: cmp
+; V8: vseleq.f32
+  %sub = sub i32 %a, %b
+  %cmp = icmp eq i32 %sub, 0
+  %ret = select i1 %cmp, float %x, float %y
+  ret float %ret
+}
+
+define double @double_sel(i32 %a, i32 %b, double %x, double %y) {
+entry:
+; CHECK-LABEL: double_sel:
+; CHECK-NOT: cmp
+; V8-LABEL: double_sel:
+; V8-NOT: cmp
+; V8: vseleq.f64
+  %sub = sub i32 %a, %b
+  %cmp = icmp eq i32 %sub, 0
+  %ret = select i1 %cmp, double %x, double %y
+  ret double %ret
+}
+
+@t = common global i32 0
+define double @double_sub(i32 %a, i32 %b, double %x, double %y) {
+entry:
+; CHECK-LABEL: double_sub:
+; CHECK: subs
+; CHECK-NOT: cmp
+; V8-LABEL: double_sub:
+; V8: vsel
+  %cmp = icmp sgt i32 %a, %b
+  %sub = sub i32 %a, %b
+  store i32 %sub, i32* @t
+  %ret = select i1 %cmp, double %x, double %y
+  ret double %ret
+}
+
+define double @double_sub_swap(i32 %a, i32 %b, double %x, double %y) {
+entry:
+; V7-LABEL: double_sub_swap:
+; V7-NOT: cmp
+; V7: subs
+; V8-LABEL: double_sub_swap:
+; V8-NOT: subs
+; V8: cmp
+; V8: vsel
+  %cmp = icmp sgt i32 %a, %b
+  %sub = sub i32 %b, %a
+  %ret = select i1 %cmp, double %x, double %y
+  store i32 %sub, i32* @t
+  ret double %ret
+}
+
+declare void @abort()
+declare void @exit(i32)
+
+; If the comparison uses the V bit (signed overflow/underflow), we can't
+; omit the comparison.
+define i32 @cmp_slt0(i32 %a, i32 %b, i32 %x, i32 %y) {
+entry:
+; CHECK-LABEL: cmp_slt0
+; CHECK: sub
+; CHECK: cmp
+; CHECK: bge
+  %load = load i32, i32* @t, align 4
+  %sub = sub i32 %load, 17
+  %cmp = icmp slt i32 %sub, 0
+  br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+  call void @abort()
+  unreachable
+
+if.else:
+  call void @exit(i32 0)
+  unreachable
+}
+
+; Same for the C bit. (Note the ult X, 0 is trivially
+; false, so the DAG combiner may or may not optimize it).
+define i32 @cmp_ult0(i32 %a, i32 %b, i32 %x, i32 %y) {
+entry:
+; CHECK-LABEL: cmp_ult0
+; CHECK: sub
+; CHECK: cmp
+; CHECK: bhs
+  %load = load i32, i32* @t, align 4
+  %sub = sub i32 %load, 17
+  %cmp = icmp ult i32 %sub, 0
+  br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+  call void @abort()
+  unreachable
+
+if.else:
+  call void @exit(i32 0)
+  unreachable
+}