; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; RUN: opt < %s -passes=aggressive-instcombine -S | FileCheck %s

; The LIT tests rely on i32, i16 and i8 being valid machine types.
; The bounds checking tests require also i64 and i128.
target datalayout = "n8:16:32:64:128"

; This LIT test checks if TruncInstCombine pass correctly recognizes the
; constraints from a signed min-max clamp. The clamp is a sequence of smin and
; smax instructions limiting a variable into a range, smin <= x <= smax.
;
; Each LIT test (except the last ones) has two versions depending on the order
; of smin and smax:
; a) y = smax(smin(x, upper_limit), lower_limit)
; b) y = smin(smax(x, lower_limit), upper_limit)
;
; The clamp is used in TruncInstCombine.cpp pass (as part of aggressive-instcombine)
; to optimize extensions and truncations of lshr. This is what is tested here.
; The pass also optimizes extensions and truncations of other binary operators,
; but in such cases the smin-smax clamp may not be used.

define i8 @test_0a(i16 %x) {
; CHECK-LABEL: define i8 @test_0a(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 31)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 0)
; CHECK-NEXT:    [[A:%.*]] = trunc i16 [[TMP2]] to i8
; CHECK-NEXT:    [[B:%.*]] = lshr i8 [[A]], 2
; CHECK-NEXT:    ret i8 [[B]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 31)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 0)
  %a = sext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

define i8 @test_0b(i16 %x) {
; CHECK-LABEL: define i8 @test_0b(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smax.i16(i16 [[X]], i16 0)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smin.i16(i16 [[TMP1]], i16 31)
; CHECK-NEXT:    [[A:%.*]] = trunc i16 [[TMP2]] to i8
; CHECK-NEXT:    [[B:%.*]] = lshr i8 [[A]], 2
; CHECK-NEXT:    ret i8 [[B]]
;
  %1 = tail call i16 @llvm.smax.i16(i16 %x, i16 0)
  %2 = tail call i16 @llvm.smin.i16(i16 %1, i16 31)
  %a = sext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

; The following two tests contain add instead of lshr.
; The optimization works here as well.
define i8 @test_1a(i16 %x) {
; CHECK-LABEL: define i8 @test_1a(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 31)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 0)
; CHECK-NEXT:    [[A:%.*]] = trunc i16 [[TMP2]] to i8
; CHECK-NEXT:    [[B:%.*]] = add i8 [[A]], 2
; CHECK-NEXT:    ret i8 [[B]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 31)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 0)
  %a = sext i16 %2 to i32
  %b = add i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

define i8 @test_1b(i16 %x) {
; CHECK-LABEL: define i8 @test_1b(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smax.i16(i16 [[X]], i16 0)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smin.i16(i16 [[TMP1]], i16 31)
; CHECK-NEXT:    [[A:%.*]] = trunc i16 [[TMP2]] to i8
; CHECK-NEXT:    [[B:%.*]] = add i8 [[A]], 2
; CHECK-NEXT:    ret i8 [[B]]
;
  %1 = tail call i16 @llvm.smax.i16(i16 %x, i16 0)
  %2 = tail call i16 @llvm.smin.i16(i16 %1, i16 31)
  %a = sext i16 %2 to i32
  %b = add i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

; Tests for clamping with negative min and max.

; With sext no optimization occurs.
define i8 @test_2a(i16 %x) {
; CHECK-LABEL: define i8 @test_2a(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 -1)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 -31)
; CHECK-NEXT:    [[A:%.*]] = sext i16 [[TMP2]] to i32
; CHECK-NEXT:    [[B:%.*]] = lshr i32 [[A]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i32 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 -1)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 -31)
  %a = sext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

define i8 @test_2b(i16 %x) {
; CHECK-LABEL: define i8 @test_2b(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smax.i16(i16 [[X]], i16 -31)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smin.i16(i16 [[TMP1]], i16 -1)
; CHECK-NEXT:    [[A:%.*]] = sext i16 [[TMP2]] to i32
; CHECK-NEXT:    [[B:%.*]] = lshr i32 [[A]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i32 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i16 @llvm.smax.i16(i16 %x, i16 -31)
  %2 = tail call i16 @llvm.smin.i16(i16 %1, i16 -1)
  %a = sext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

; With zext the optimization occurs.
define i8 @test_2c(i16 %x) {
; CHECK-LABEL: define i8 @test_2c(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 -1)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 -31)
; CHECK-NEXT:    [[B:%.*]] = lshr i16 [[TMP2]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 -1)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 -31)
  %a = zext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

define i8 @test_2d(i16 %x) {
; CHECK-LABEL: define i8 @test_2d(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smax.i16(i16 [[X]], i16 -31)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smin.i16(i16 [[TMP1]], i16 -1)
; CHECK-NEXT:    [[B:%.*]] = lshr i16 [[TMP2]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i16 @llvm.smax.i16(i16 %x, i16 -31)
  %2 = tail call i16 @llvm.smin.i16(i16 %1, i16 -1)
  %a = zext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

; Tests for clamping with mixed-signed min and max.
; With zext the optimization occurs.
define i8 @test_3a(i16 %x) {
; CHECK-LABEL: define i8 @test_3a(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 31)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 -31)
; CHECK-NEXT:    [[B:%.*]] = lshr i16 [[TMP2]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 31)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 -31)
  %a = zext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

define i8 @test_3b(i16 %x) {
; CHECK-LABEL: define i8 @test_3b(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smax.i16(i16 [[X]], i16 -31)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smin.i16(i16 [[TMP1]], i16 31)
; CHECK-NEXT:    [[B:%.*]] = lshr i16 [[TMP2]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i16 @llvm.smax.i16(i16 %x, i16 -31)
  %2 = tail call i16 @llvm.smin.i16(i16 %1, i16 31)
  %a = zext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

; Optimizations with vector types.
define <16 x i8> @test_vec_1a(<16 x i16> %x) {
; CHECK-LABEL: define <16 x i8> @test_vec_1a(
; CHECK-SAME: <16 x i16> [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call <16 x i16> @llvm.smin.v16i16(<16 x i16> [[X]], <16 x i16> splat (i16 127))
; CHECK-NEXT:    [[TMP2:%.*]] = tail call <16 x i16> @llvm.smax.v16i16(<16 x i16> [[TMP1]], <16 x i16> zeroinitializer)
; CHECK-NEXT:    [[A:%.*]] = trunc <16 x i16> [[TMP2]] to <16 x i8>
; CHECK-NEXT:    [[B:%.*]] = lshr <16 x i8> [[A]], splat (i8 2)
; CHECK-NEXT:    ret <16 x i8> [[B]]
;
  %1 = tail call <16 x i16> @llvm.smin.v16i16(<16 x i16> %x, <16 x i16> splat (i16 127))
  %2 = tail call <16 x i16> @llvm.smax.v16i16(<16 x i16> %1, <16 x i16> zeroinitializer)
  %a = sext <16 x i16> %2 to <16 x i32>
  %b = lshr <16 x i32> %a, splat (i32 2)
  %b.trunc = trunc <16 x i32> %b to <16 x i8>
  ret <16 x i8> %b.trunc
}

define <16 x i8> @test_vec_1b(<16 x i16> %x) {
; CHECK-LABEL: define <16 x i8> @test_vec_1b(
; CHECK-SAME: <16 x i16> [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call <16 x i16> @llvm.smax.v16i16(<16 x i16> [[X]], <16 x i16> zeroinitializer)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call <16 x i16> @llvm.smin.v16i16(<16 x i16> [[TMP1]], <16 x i16> splat (i16 127))
; CHECK-NEXT:    [[A:%.*]] = trunc <16 x i16> [[TMP2]] to <16 x i8>
; CHECK-NEXT:    [[B:%.*]] = lshr <16 x i8> [[A]], splat (i8 2)
; CHECK-NEXT:    ret <16 x i8> [[B]]
;
  %1 = tail call <16 x i16> @llvm.smax.v16i16(<16 x i16> %x, <16 x i16> zeroinitializer)
  %2 = tail call <16 x i16> @llvm.smin.v16i16(<16 x i16> %1, <16 x i16> splat (i16 127))
  %a = sext <16 x i16> %2 to <16 x i32>
  %b = lshr <16 x i32> %a, splat (i32 2)
  %b.trunc = trunc <16 x i32> %b to <16 x i8>
  ret <16 x i8> %b.trunc
}

; A longer test that was the original motivation for the smin-smax clamping.
define i8 @test_final(i16 %x, i16 %y) {
; CHECK-LABEL: define i8 @test_final(
; CHECK-SAME: i16 [[X:%.*]], i16 [[Y:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 127)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 0)
; CHECK-NEXT:    [[TMP3:%.*]] = tail call i16 @llvm.smax.i16(i16 [[Y]], i16 0)
; CHECK-NEXT:    [[TMP4:%.*]] = tail call i16 @llvm.smin.i16(i16 [[TMP3]], i16 127)
; CHECK-NEXT:    [[MUL:%.*]] = mul i16 [[TMP2]], [[TMP4]]
; CHECK-NEXT:    [[SHR:%.*]] = lshr i16 [[MUL]], 7
; CHECK-NEXT:    [[TRUNC:%.*]] = trunc i16 [[SHR]] to i8
; CHECK-NEXT:    ret i8 [[TRUNC]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 127)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 0)
  %x.clamp = zext nneg i16 %2 to i32
  %3 = tail call i16 @llvm.smax.i16(i16 %y, i16 0)
  %4 = tail call i16 @llvm.smin.i16(i16 %3, i16 127)
  %y.clamp = zext nneg i16 %4 to i32
  %mul = mul nuw nsw i32 %x.clamp, %y.clamp
  %shr = lshr i32 %mul, 7
  %trunc= trunc nuw nsw i32 %shr to i8
  ret i8 %trunc
}

; Range tests below check if the bounds are dealt with correctly.

; This gets optimized.
define i8 @test_bounds_1(i16 %x) {
; CHECK-LABEL: define i8 @test_bounds_1(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 127)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 0)
; CHECK-NEXT:    [[A:%.*]] = trunc i16 [[TMP2]] to i8
; CHECK-NEXT:    [[B:%.*]] = lshr i8 [[A]], 7
; CHECK-NEXT:    ret i8 [[B]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 127)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 0)
  %a = sext i16 %2 to i32
  %b = lshr i32 %a, 7
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

; While this does not.
define i8 @test_bounds_2(i16 %x) {
; CHECK-LABEL: define i8 @test_bounds_2(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 128)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 0)
; CHECK-NEXT:    [[A:%.*]] = trunc i16 [[TMP2]] to i8
; CHECK-NEXT:    [[B:%.*]] = lshr i8 [[A]], 7
; CHECK-NEXT:    ret i8 [[B]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 128)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 0)
  %a = sext i16 %2 to i32
  %b = lshr i32 %a, 7
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

; This should get optimized. We test here if the optimization works correctly
; if the upper limit is signed max int.
define i8 @test_bounds_3(i16 %x) {
; CHECK-LABEL: define i8 @test_bounds_3(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 32767)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 32752)
; CHECK-NEXT:    [[B:%.*]] = lshr i16 [[TMP2]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 32767)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 32752)
  %a = sext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

; Here min = 128 is greater than max = 0.
define i8 @test_bounds_4(i16 %x) {
; CHECK-LABEL: define i8 @test_bounds_4(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 0)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 128)
; CHECK-NEXT:    [[B:%.*]] = lshr i16 [[TMP2]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 0)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 128)
  %a = sext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

; The following 3 tests check the situation where min and max are minimal and
; maximal signed values. No transformations should occur here.
define i8 @test_bounds_5(i16 %x) {
; CHECK-LABEL: define i8 @test_bounds_5(
; CHECK-SAME: i16 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 32767)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 -32768)
; CHECK-NEXT:    [[B:%.*]] = lshr i16 [[TMP2]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i16 @llvm.smin.i16(i16 %x, i16 32767)
  %2 = tail call i16 @llvm.smax.i16(i16 %1, i16 -32768)
  %a = zext i16 %2 to i32
  %b = lshr i32 %a, 2
  %b.trunc = trunc i32 %b to i8
  ret i8 %b.trunc
}

define i8 @test_bounds_6(i32 %x) {
; CHECK-LABEL: define i8 @test_bounds_6(
; CHECK-SAME: i32 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i32 @llvm.smin.i32(i32 [[X]], i32 2147483647)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i32 @llvm.smax.i32(i32 [[TMP1]], i32 -2147483648)
; CHECK-NEXT:    [[B:%.*]] = lshr i32 [[TMP2]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i32 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i32 @llvm.smin.i32(i32 %x, i32 2147483647)
  %2 = tail call i32 @llvm.smax.i32(i32 %1, i32 -2147483648)
  %a = zext i32 %2 to i64
  %b = lshr i64 %a, 2
  %b.trunc = trunc i64 %b to i8
  ret i8 %b.trunc
}

define i8 @test_bounds_7(i64 %x) {
; CHECK-LABEL: define i8 @test_bounds_7(
; CHECK-SAME: i64 [[X:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = tail call i64 @llvm.smin.i64(i64 [[X]], i64 9223372036854775807)
; CHECK-NEXT:    [[TMP2:%.*]] = tail call i64 @llvm.smax.i64(i64 [[TMP1]], i64 -9223372036854775808)
; CHECK-NEXT:    [[B:%.*]] = lshr i64 [[TMP2]], 2
; CHECK-NEXT:    [[B_TRUNC:%.*]] = trunc i64 [[B]] to i8
; CHECK-NEXT:    ret i8 [[B_TRUNC]]
;
  %1 = tail call i64 @llvm.smin.i64(i64 %x, i64 9223372036854775807)
  %2 = tail call i64 @llvm.smax.i64(i64 %1, i64 -9223372036854775808)
  %a = zext i64 %2 to i128
  %b = lshr i128 %a, 2
  %b.trunc = trunc i128 %b to i8
  ret i8 %b.trunc
}