[InstCombine][KnownBits] Use KnownBits better to detect nsw adds
Change checkRippleForAdd from a heuristic to a full check -
if it is provable that the add does not overflow return true, otherwise false.
Patch by Yoav Ben-Shalom
Differential Revision: https://reviews.llvm.org/D32686
llvm-svn: 302093
diff --git a/llvm/test/Transforms/InstCombine/AddOverFlow.ll b/llvm/test/Transforms/InstCombine/AddOverFlow.ll
index a341cb0..91fa86e 100644
--- a/llvm/test/Transforms/InstCombine/AddOverFlow.ll
+++ b/llvm/test/Transforms/InstCombine/AddOverFlow.ll
@@ -95,6 +95,44 @@
ret i16 %c
}
+; CHECK-LABEL: @ripple_nsw3
+; CHECK: add nsw i16 %a, %b
+define i16 @ripple_nsw3(i16 %x, i16 %y) {
+ %a = and i16 %y, 43691
+ %b = and i16 %x, 21843
+ %c = add i16 %a, %b
+ ret i16 %c
+}
+
+; Like the previous test, but flip %a and %b
+; CHECK-LABEL: @ripple_nsw4
+; CHECK: add nsw i16 %b, %a
+define i16 @ripple_nsw4(i16 %x, i16 %y) {
+ %a = and i16 %y, 43691
+ %b = and i16 %x, 21843
+ %c = add i16 %b, %a
+ ret i16 %c
+}
+
+; CHECK-LABEL: @ripple_nsw5
+; CHECK: add nsw i16 %a, %b
+define i16 @ripple_nsw5(i16 %x, i16 %y) {
+ %a = or i16 %y, 43691
+ %b = or i16 %x, 54613
+ %c = add i16 %a, %b
+ ret i16 %c
+}
+
+; Like the previous test, but flip %a and %b
+; CHECK-LABEL: @ripple_nsw6
+; CHECK: add nsw i16 %b, %a
+define i16 @ripple_nsw6(i16 %x, i16 %y) {
+ %a = or i16 %y, 43691
+ %b = or i16 %x, 54613
+ %c = add i16 %b, %a
+ ret i16 %c
+}
+
; CHECK-LABEL: @ripple_no_nsw1
; CHECK: add i32 %a, %x
define i32 @ripple_no_nsw1(i32 %x, i32 %y) {
@@ -116,3 +154,41 @@
%c = add i16 %a, %b
ret i16 %c
}
+
+; CHECK-LABEL: @ripple_no_nsw3
+; CHECK: add i16 %a, %b
+define i16 @ripple_no_nsw3(i16 %x, i16 %y) {
+ %a = and i16 %y, 43691
+ %b = and i16 %x, 21845
+ %c = add i16 %a, %b
+ ret i16 %c
+}
+
+; Like the previous test, but flip %a and %b
+; CHECK-LABEL: @ripple_no_nsw4
+; CHECK: add i16 %b, %a
+define i16 @ripple_no_nsw4(i16 %x, i16 %y) {
+ %a = and i16 %y, 43691
+ %b = and i16 %x, 21845
+ %c = add i16 %b, %a
+ ret i16 %c
+}
+
+; CHECK-LABEL: @ripple_no_nsw5
+; CHECK: add i16 %a, %b
+define i16 @ripple_no_nsw5(i16 %x, i16 %y) {
+ %a = or i16 %y, 43689
+ %b = or i16 %x, 54613
+ %c = add i16 %a, %b
+ ret i16 %c
+}
+
+; Like the previous test, but flip %a and %b
+; CHECK-LABEL: @ripple_no_nsw6
+; CHECK: add i16 %b, %a
+define i16 @ripple_no_nsw6(i16 %x, i16 %y) {
+ %a = or i16 %y, 43689
+ %b = or i16 %x, 54613
+ %c = add i16 %b, %a
+ ret i16 %c
+}
diff --git a/llvm/test/Transforms/InstCombine/demand_shrink_nsw.ll b/llvm/test/Transforms/InstCombine/demand_shrink_nsw.ll
index f491742..4f7d00e 100644
--- a/llvm/test/Transforms/InstCombine/demand_shrink_nsw.ll
+++ b/llvm/test/Transforms/InstCombine/demand_shrink_nsw.ll
@@ -3,7 +3,7 @@
; The constant at %v35 should be shrunk, but this must lead to the nsw flag of
; %v43 getting removed so that %v44 is not illegally optimized away.
; CHECK-LABEL: @foo
-; CHECK: %v35 = add nuw i32 %v34, 1362915575
+; CHECK: %v35 = add nuw nsw i32 %v34, 1362915575
; ...
; CHECK: add nuw i32 %v42, 1533579450
; CHECK-NEXT: %v44 = or i32 %v43, -2147483648