Index: llvm/test/Transforms/InstCombine/abs-intrinsic.ll =================================================================== --- llvm/test/Transforms/InstCombine/abs-intrinsic.ll +++ llvm/test/Transforms/InstCombine/abs-intrinsic.ll @@ -2,6 +2,7 @@ ; RUN: opt < %s -instcombine -S | FileCheck %s declare i32 @llvm.abs.i32(i32, i1) +declare <4 x i32> @llvm.abs.v4i32(<4 x i32>, i1) define i1 @abs_nsw_must_be_positive(i32 %x) { ; CHECK-LABEL: @abs_nsw_must_be_positive( @@ -12,6 +13,15 @@ ret i1 %c2 } +define <4 x i1> @abs_nsw_must_be_positive_vec(<4 x i32> %x) { +; CHECK-LABEL: @abs_nsw_must_be_positive_vec( +; CHECK-NEXT: ret <4 x i1> +; + %abs = call <4 x i32> @llvm.abs.v4i32(<4 x i32> %x, i1 true) + %c2 = icmp sge <4 x i32> %abs, zeroinitializer + ret <4 x i1> %c2 +} + ; Negative test, no nsw provides no information about the sign bit of the result. define i1 @abs_nonsw(i32 %x) { ; CHECK-LABEL: @abs_nonsw( @@ -24,6 +34,17 @@ ret i1 %c2 } +define <4 x i1> @abs_nonsw_vec(<4 x i32> %x) { +; CHECK-LABEL: @abs_nonsw_vec( +; CHECK-NEXT: [[ABS:%.*]] = call <4 x i32> @llvm.abs.v4i32(<4 x i32> [[X:%.*]], i1 false) +; CHECK-NEXT: [[C2:%.*]] = icmp sgt <4 x i32> [[ABS]], +; CHECK-NEXT: ret <4 x i1> [[C2]] +; + %abs = call <4 x i32> @llvm.abs.v4i32(<4 x i32> %x, i1 false) + %c2 = icmp sge <4 x i32> %abs, zeroinitializer + ret <4 x i1> %c2 +} + ; abs preserves trailing zeros so the second and is unneeded define i32 @abs_trailing_zeros(i32 %x) { ; CHECK-LABEL: @abs_trailing_zeros( @@ -37,6 +58,18 @@ ret i32 %and2 } +define <4 x i32> @abs_trailing_zeros_vec(<4 x i32> %x) { +; CHECK-LABEL: @abs_trailing_zeros_vec( +; CHECK-NEXT: [[AND:%.*]] = and <4 x i32> [[X:%.*]], +; CHECK-NEXT: [[ABS:%.*]] = call <4 x i32> @llvm.abs.v4i32(<4 x i32> [[AND]], i1 false) +; CHECK-NEXT: ret <4 x i32> [[ABS]] +; + %and = and <4 x i32> %x, + %abs = call <4 x i32> @llvm.abs.v4i32(<4 x i32> %and, i1 false) + %and2 = and <4 x i32> %abs, + ret <4 x i32> %and2 +} + ; negative test, can't remove the second and based on trailing zeroes. ; FIXME: Could remove the first and using demanded bits. define i32 @abs_trailing_zeros_negative(i32 %x) { @@ -52,6 +85,19 @@ ret i32 %and2 } +define <4 x i32> @abs_trailing_zeros_negative_vec(<4 x i32> %x) { +; CHECK-LABEL: @abs_trailing_zeros_negative_vec( +; CHECK-NEXT: [[AND:%.*]] = and <4 x i32> [[X:%.*]], +; CHECK-NEXT: [[ABS:%.*]] = call <4 x i32> @llvm.abs.v4i32(<4 x i32> [[AND]], i1 false) +; CHECK-NEXT: [[AND2:%.*]] = and <4 x i32> [[ABS]], +; CHECK-NEXT: ret <4 x i32> [[AND2]] +; + %and = and <4 x i32> %x, + %abs = call <4 x i32> @llvm.abs.v4i32(<4 x i32> %and, i1 false) + %and2 = and <4 x i32> %abs, + ret <4 x i32> %and2 +} + ; Make sure we infer this add doesn't overflow. The input to the abs has 3 ; sign bits, the abs reduces this to 2 sign bits. define i32 @abs_signbits(i30 %x) { @@ -67,6 +113,19 @@ ret i32 %add } +define <4 x i32> @abs_signbits_vec(<4 x i30> %x) { +; CHECK-LABEL: @abs_signbits_vec( +; CHECK-NEXT: [[EXT:%.*]] = sext <4 x i30> [[X:%.*]] to <4 x i32> +; CHECK-NEXT: [[ABS:%.*]] = call <4 x i32> @llvm.abs.v4i32(<4 x i32> [[EXT]], i1 false) +; CHECK-NEXT: [[ADD:%.*]] = add nsw <4 x i32> [[ABS]], +; CHECK-NEXT: ret <4 x i32> [[ADD]] +; + %ext = sext <4 x i30> %x to <4 x i32> + %abs = call <4 x i32> @llvm.abs.v4i32(<4 x i32> %ext, i1 false) + %add = add <4 x i32> %abs, + ret <4 x i32> %add +} + define i1 @abs_known_positive_input_compare(i31 %x) { ; CHECK-LABEL: @abs_known_positive_input_compare( ; CHECK-NEXT: ret i1 true @@ -77,6 +136,16 @@ ret i1 %c2 } +define <4 x i1> @abs_known_positive_input_compare_vec(<4 x i31> %x) { +; CHECK-LABEL: @abs_known_positive_input_compare_vec( +; CHECK-NEXT: ret <4 x i1> +; + %zext = zext <4 x i31> %x to <4 x i32> + %abs = call <4 x i32> @llvm.abs.v4i32(<4 x i32> %zext, i1 false) + %c2 = icmp sge <4 x i32> %abs, zeroinitializer + ret <4 x i1> %c2 +} + define i1 @abs_known_not_int_min(i32 %x) { ; CHECK-LABEL: @abs_known_not_int_min( ; CHECK-NEXT: ret i1 true @@ -87,14 +156,12 @@ ret i1 %c2 } -; FIXME: We should remove this abs -define i32 @abs_known_positive_input(i31 %x) { -; CHECK-LABEL: @abs_known_positive_input( -; CHECK-NEXT: [[ZEXT:%.*]] = zext i31 [[X:%.*]] to i32 -; CHECK-NEXT: [[ABS:%.*]] = call i32 @llvm.abs.i32(i32 [[ZEXT]], i1 false) -; CHECK-NEXT: ret i32 [[ABS]] +define <4 x i1> @abs_known_not_int_min_vec(<4 x i32> %x) { +; CHECK-LABEL: @abs_known_not_int_min_vec( +; CHECK-NEXT: ret <4 x i1> ; - %zext = zext i31 %x to i32 - %abs = call i32 @llvm.abs.i32(i32 %zext, i1 false) - ret i32 %abs + %or = or <4 x i32> %x, + %abs = call <4 x i32> @llvm.abs.v4i32(<4 x i32> %or, i1 false) + %c2 = icmp sge <4 x i32> %abs, zeroinitializer + ret <4 x i1> %c2 }