From c641e9d6ffd76cccfc597c18191308a31a8daff7 Mon Sep 17 00:00:00 2001 From: Sanjay Patel Date: Sun, 4 Sep 2016 20:58:27 +0000 Subject: [InstCombine] allow icmp (and X, C2), C1 folds for splat constant vectors The code to calculate 'UsesRemoved' could be simplified. As-is, that code is a victim of PR30273: https://llvm.org/bugs/show_bug.cgi?id=30273 llvm-svn: 280637 --- llvm/test/Transforms/InstCombine/icmp.ll | 34 +++++++++++++++++++++----------- 1 file changed, 22 insertions(+), 12 deletions(-) (limited to 'llvm/test/Transforms') diff --git a/llvm/test/Transforms/InstCombine/icmp.ll b/llvm/test/Transforms/InstCombine/icmp.ll index 9ad8250f66c..73059321a6b 100644 --- a/llvm/test/Transforms/InstCombine/icmp.ll +++ b/llvm/test/Transforms/InstCombine/icmp.ll @@ -1013,15 +1013,29 @@ define i1 @test67(i32 %x) { ret i1 %cmp } -; FIXME: Vectors should fold the same way. +; The test above relies on 3 different folds. +; This test only checks the last of those (icmp ugt -> icmp ne). + define <2 x i1> @test67vec(<2 x i32> %x) { ; CHECK-LABEL: @test67vec( +; CHECK-NEXT: [[AND:%.*]] = and <2 x i32> %x, +; CHECK-NEXT: [[CMP:%.*]] = icmp ne <2 x i32> [[AND]], zeroinitializer +; CHECK-NEXT: ret <2 x i1> [[CMP]] +; + %and = and <2 x i32> %x, + %cmp = icmp ugt <2 x i32> %and, + ret <2 x i1> %cmp +} + +; FIXME: Vector constant for the 'and' should use less bits. +define <2 x i1> @test67vec2(<2 x i32> %x) { +; CHECK-LABEL: @test67vec2( ; CHECK-NEXT: [[AND:%.*]] = and <2 x i32> %x, ; CHECK-NEXT: [[CMP:%.*]] = icmp ugt <2 x i32> [[AND]], ; CHECK-NEXT: ret <2 x i1> [[CMP]] ; %and = and <2 x i32> %x, - %cmp = icmp sgt <2 x i32> %and, + %cmp = icmp ugt <2 x i32> %and, ret <2 x i1> %cmp } @@ -2059,13 +2073,12 @@ define i1 @icmp_and_or_lshr(i32 %x, i32 %y) { ret i1 %ret } -; FIXME: Vectors should fold the same way. define <2 x i1> @icmp_and_or_lshr_vec(<2 x i32> %x, <2 x i32> %y) { ; CHECK-LABEL: @icmp_and_or_lshr_vec( -; CHECK-NEXT: [[SHF:%.*]] = lshr <2 x i32> %x, %y -; CHECK-NEXT: [[OR:%.*]] = or <2 x i32> [[SHF]], %x -; CHECK-NEXT: [[AND:%.*]] = and <2 x i32> [[OR]], -; CHECK-NEXT: [[RET:%.*]] = icmp ne <2 x i32> [[AND]], zeroinitializer +; CHECK-NEXT: [[SHF1:%.*]] = shl nuw <2 x i32> , %y +; CHECK-NEXT: [[OR2:%.*]] = or <2 x i32> [[SHF1]], +; CHECK-NEXT: [[AND3:%.*]] = and <2 x i32> [[OR2]], %x +; CHECK-NEXT: [[RET:%.*]] = icmp ne <2 x i32> [[AND3]], zeroinitializer ; CHECK-NEXT: ret <2 x i1> [[RET]] ; %shf = lshr <2 x i32> %x, %y @@ -2088,13 +2101,10 @@ define i1 @icmp_and_or_lshr_cst(i32 %x) { ret i1 %ret } -; FIXME: Vectors should fold the same way. define <2 x i1> @icmp_and_or_lshr_cst_vec(<2 x i32> %x) { ; CHECK-LABEL: @icmp_and_or_lshr_cst_vec( -; CHECK-NEXT: [[SHF:%.*]] = lshr <2 x i32> %x, -; CHECK-NEXT: [[OR:%.*]] = or <2 x i32> [[SHF]], %x -; CHECK-NEXT: [[AND:%.*]] = and <2 x i32> [[OR]], -; CHECK-NEXT: [[RET:%.*]] = icmp ne <2 x i32> [[AND]], zeroinitializer +; CHECK-NEXT: [[AND1:%.*]] = and <2 x i32> %x, +; CHECK-NEXT: [[RET:%.*]] = icmp ne <2 x i32> [[AND1]], zeroinitializer ; CHECK-NEXT: ret <2 x i1> [[RET]] ; %shf = lshr <2 x i32> %x, -- cgit v1.2.3