llvm/test/Transforms/InstCombine/fsh.ll

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt < %s -instcombine -S | FileCheck %s

declare i32 @llvm.fshl.i32(i32, i32, i32)
declare i33 @llvm.fshr.i33(i33, i33, i33)
declare <2 x i32> @llvm.fshr.v2i32(<2 x i32>, <2 x i32>, <2 x i32>)
declare <2 x i31> @llvm.fshl.v2i31(<2 x i31>, <2 x i31>, <2 x i31>)

; If the shift mask doesn't include any demanded bits, the funnel shift can be eliminated.

define i32 @fshl_mask_simplify1(i32 %x, i32 %y, i32 %sh) {
; CHECK-LABEL: @fshl_mask_simplify1(
; CHECK-NEXT:    ret i32 [[X:%.*]]
;
  %maskedsh = and i32 %sh, 32
  %r = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %maskedsh)
  ret i32 %r
}

define <2 x i32> @fshr_mask_simplify2(<2 x i32> %x, <2 x i32> %y, <2 x i32> %sh) {
; CHECK-LABEL: @fshr_mask_simplify2(
; CHECK-NEXT:    ret <2 x i32> [[Y:%.*]]
;
  %maskedsh = and <2 x i32> %sh, <i32 64, i32 64>
  %r = call <2 x i32> @llvm.fshr.v2i32(<2 x i32> %x, <2 x i32> %y, <2 x i32> %maskedsh)
  ret <2 x i32> %r
}

; Negative test.

define i32 @fshl_mask_simplify3(i32 %x, i32 %y, i32 %sh) {
; CHECK-LABEL: @fshl_mask_simplify3(
; CHECK-NEXT:    [[MASKEDSH:%.*]] = and i32 [[SH:%.*]], 16
; CHECK-NEXT:    [[R:%.*]] = call i32 @llvm.fshl.i32(i32 [[X:%.*]], i32 [[Y:%.*]], i32 [[MASKEDSH]])
; CHECK-NEXT:    ret i32 [[R]]
;
  %maskedsh = and i32 %sh, 16
  %r = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %maskedsh)
  ret i32 %r
}

; Check again with weird bitwidths - the analysis is invalid with non-power-of-2.

define i33 @fshr_mask_simplify1(i33 %x, i33 %y, i33 %sh) {
; CHECK-LABEL: @fshr_mask_simplify1(
; CHECK-NEXT:    [[MASKEDSH:%.*]] = and i33 [[SH:%.*]], 64
; CHECK-NEXT:    [[R:%.*]] = call i33 @llvm.fshr.i33(i33 [[X:%.*]], i33 [[Y:%.*]], i33 [[MASKEDSH]])
; CHECK-NEXT:    ret i33 [[R]]
;
  %maskedsh = and i33 %sh, 64
  %r = call i33 @llvm.fshr.i33(i33 %x, i33 %y, i33 %maskedsh)
  ret i33 %r
}

; Check again with weird bitwidths - the analysis is invalid with non-power-of-2.

define <2 x i31> @fshl_mask_simplify2(<2 x i31> %x, <2 x i31> %y, <2 x i31> %sh) {
; CHECK-LABEL: @fshl_mask_simplify2(
; CHECK-NEXT:    [[MASKEDSH:%.*]] = and <2 x i31> [[SH:%.*]], <i31 32, i31 32>
; CHECK-NEXT:    [[R:%.*]] = call <2 x i31> @llvm.fshl.v2i31(<2 x i31> [[X:%.*]], <2 x i31> [[Y:%.*]], <2 x i31> [[MASKEDSH]])
; CHECK-NEXT:    ret <2 x i31> [[R]]
;
  %maskedsh = and <2 x i31> %sh, <i31 32, i31 32>
  %r = call <2 x i31> @llvm.fshl.v2i31(<2 x i31> %x, <2 x i31> %y, <2 x i31> %maskedsh)
  ret <2 x i31> %r
}

; Check again with weird bitwidths - the analysis is invalid with non-power-of-2.

define i33 @fshr_mask_simplify3(i33 %x, i33 %y, i33 %sh) {
; CHECK-LABEL: @fshr_mask_simplify3(
; CHECK-NEXT:    [[MASKEDSH:%.*]] = and i33 [[SH:%.*]], 32
; CHECK-NEXT:    [[R:%.*]] = call i33 @llvm.fshr.i33(i33 [[X:%.*]], i33 [[Y:%.*]], i33 [[MASKEDSH]])
; CHECK-NEXT:    ret i33 [[R]]
;
  %maskedsh = and i33 %sh, 32
  %r = call i33 @llvm.fshr.i33(i33 %x, i33 %y, i33 %maskedsh)
  ret i33 %r
}

; This mask op is unnecessary.

define i32 @fshl_mask_not_required(i32 %x, i32 %y, i32 %sh) {
; CHECK-LABEL: @fshl_mask_not_required(
; CHECK-NEXT:    [[R:%.*]] = call i32 @llvm.fshl.i32(i32 [[X:%.*]], i32 [[Y:%.*]], i32 [[SH:%.*]])
; CHECK-NEXT:    ret i32 [[R]]
;
  %maskedsh = and i32 %sh, 31
  %r = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %maskedsh)
  ret i32 %r
}

; This mask op can be reduced.

define i32 @fshl_mask_reduce_constant(i32 %x, i32 %y, i32 %sh) {
; CHECK-LABEL: @fshl_mask_reduce_constant(
; CHECK-NEXT:    [[MASKEDSH:%.*]] = and i32 [[SH:%.*]], 1
; CHECK-NEXT:    [[R:%.*]] = call i32 @llvm.fshl.i32(i32 [[X:%.*]], i32 [[Y:%.*]], i32 [[MASKEDSH]])
; CHECK-NEXT:    ret i32 [[R]]
;
  %maskedsh = and i32 %sh, 33
  %r = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %maskedsh)
  ret i32 %r
}

; But this mask op is required.

define i32 @fshl_mask_negative(i32 %x, i32 %y, i32 %sh) {
; CHECK-LABEL: @fshl_mask_negative(
; CHECK-NEXT:    [[MASKEDSH:%.*]] = and i32 [[SH:%.*]], 15
; CHECK-NEXT:    [[R:%.*]] = call i32 @llvm.fshl.i32(i32 [[X:%.*]], i32 [[Y:%.*]], i32 [[MASKEDSH]])
; CHECK-NEXT:    ret i32 [[R]]
;
  %maskedsh = and i32 %sh, 15
  %r = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %maskedsh)
  ret i32 %r
}

; The transform is not limited to mask ops.

define <2 x i32> @fshr_set_but_not_demanded_vec(<2 x i32> %x, <2 x i32> %y, <2 x i32> %sh) {
; CHECK-LABEL: @fshr_set_but_not_demanded_vec(
; CHECK-NEXT:    [[R:%.*]] = call <2 x i32> @llvm.fshr.v2i32(<2 x i32> [[X:%.*]], <2 x i32> [[Y:%.*]], <2 x i32> [[SH:%.*]])
; CHECK-NEXT:    ret <2 x i32> [[R]]
;
  %bogusbits = or <2 x i32> %sh, <i32 32, i32 32>
  %r = call <2 x i32> @llvm.fshr.v2i32(<2 x i32> %x, <2 x i32> %y, <2 x i32> %bogusbits)
  ret <2 x i32> %r
}

; Check again with weird bitwidths - the analysis is invalid with non-power-of-2.

define <2 x i31> @fshl_set_but_not_demanded_vec(<2 x i31> %x, <2 x i31> %y, <2 x i31> %sh) {
; CHECK-LABEL: @fshl_set_but_not_demanded_vec(
; CHECK-NEXT:    [[BOGUSBITS:%.*]] = or <2 x i31> [[SH:%.*]], <i31 32, i31 32>
; CHECK-NEXT:    [[R:%.*]] = call <2 x i31> @llvm.fshl.v2i31(<2 x i31> [[X:%.*]], <2 x i31> [[Y:%.*]], <2 x i31> [[BOGUSBITS]])
; CHECK-NEXT:    ret <2 x i31> [[R]]
;
  %bogusbits = or <2 x i31> %sh, <i31 32, i31 32>
  %r = call <2 x i31> @llvm.fshl.v2i31(<2 x i31> %x, <2 x i31> %y, <2 x i31> %bogusbits)
  ret <2 x i31> %r
}

; Simplify one undef or zero operand and constant shift amount.

define i32 @fshl_op0_undef(i32 %x) {
; CHECK-LABEL: @fshl_op0_undef(
; CHECK-NEXT:    [[R:%.*]] = lshr i32 [[X:%.*]], 25
; CHECK-NEXT:    ret i32 [[R]]
;
  %r = call i32 @llvm.fshl.i32(i32 undef, i32 %x, i32 7)
  ret i32 %r
}

define i32 @fshl_op0_zero(i32 %x) {
; CHECK-LABEL: @fshl_op0_zero(
; CHECK-NEXT:    [[R:%.*]] = lshr i32 [[X:%.*]], 25
; CHECK-NEXT:    ret i32 [[R]]
;
  %r = call i32 @llvm.fshl.i32(i32 0, i32 %x, i32 7)
  ret i32 %r
}

define i33 @fshr_op0_undef(i33 %x) {
; CHECK-LABEL: @fshr_op0_undef(
; CHECK-NEXT:    [[R:%.*]] = lshr i33 [[X:%.*]], 7
; CHECK-NEXT:    ret i33 [[R]]
;
  %r = call i33 @llvm.fshr.i33(i33 undef, i33 %x, i33 7)
  ret i33 %r
}

define i33 @fshr_op0_zero(i33 %x) {
; CHECK-LABEL: @fshr_op0_zero(
; CHECK-NEXT:    [[R:%.*]] = lshr i33 [[X:%.*]], 7
; CHECK-NEXT:    ret i33 [[R]]
;
  %r = call i33 @llvm.fshr.i33(i33 0, i33 %x, i33 7)
  ret i33 %r
}

define i32 @fshl_op1_undef(i32 %x) {
; CHECK-LABEL: @fshl_op1_undef(
; CHECK-NEXT:    [[R:%.*]] = shl i32 [[X:%.*]], 7
; CHECK-NEXT:    ret i32 [[R]]
;
  %r = call i32 @llvm.fshl.i32(i32 %x, i32 undef, i32 7)
  ret i32 %r
}

define i32 @fshl_op1_zero(i32 %x) {
; CHECK-LABEL: @fshl_op1_zero(
; CHECK-NEXT:    [[R:%.*]] = shl i32 [[X:%.*]], 7
; CHECK-NEXT:    ret i32 [[R]]
;
  %r = call i32 @llvm.fshl.i32(i32 %x, i32 0, i32 7)
  ret i32 %r
}

define i33 @fshr_op1_undef(i33 %x) {
; CHECK-LABEL: @fshr_op1_undef(
; CHECK-NEXT:    [[R:%.*]] = shl i33 [[X:%.*]], 26
; CHECK-NEXT:    ret i33 [[R]]
;
  %r = call i33 @llvm.fshr.i33(i33 %x, i33 undef, i33 7)
  ret i33 %r
}

define i33 @fshr_op1_zero(i33 %x) {
; CHECK-LABEL: @fshr_op1_zero(
; CHECK-NEXT:    [[R:%.*]] = shl i33 [[X:%.*]], 26
; CHECK-NEXT:    ret i33 [[R]]
;
  %r = call i33 @llvm.fshr.i33(i33 %x, i33 0, i33 7)
  ret i33 %r
}

define <2 x i31> @fshl_op0_zero_vec(<2 x i31> %x) {
; CHECK-LABEL: @fshl_op0_zero_vec(
; CHECK-NEXT:    [[R:%.*]] = lshr <2 x i31> [[X:%.*]], <i31 24, i31 24>
; CHECK-NEXT:    ret <2 x i31> [[R]]
;
  %r = call <2 x i31> @llvm.fshl.v2i31(<2 x i31> zeroinitializer, <2 x i31> %x, <2 x i31> <i31 7, i31 7>)
  ret <2 x i31> %r
}

define <2 x i31> @fshl_op1_undef_vec(<2 x i31> %x) {
; CHECK-LABEL: @fshl_op1_undef_vec(
; CHECK-NEXT:    [[R:%.*]] = shl <2 x i31> [[X:%.*]], <i31 7, i31 7>
; CHECK-NEXT:    ret <2 x i31> [[R]]
;
  %r = call <2 x i31> @llvm.fshl.v2i31(<2 x i31> %x, <2 x i31> undef, <2 x i31> <i31 7, i31 7>)
  ret <2 x i31> %r
}

define <2 x i32> @fshr_op0_undef_vec(<2 x i32> %x) {
; CHECK-LABEL: @fshr_op0_undef_vec(
; CHECK-NEXT:    [[R:%.*]] = lshr <2 x i32> [[X:%.*]], <i32 7, i32 7>
; CHECK-NEXT:    ret <2 x i32> [[R]]
;
  %r = call <2 x i32> @llvm.fshr.v2i32(<2 x i32> undef, <2 x i32> %x, <2 x i32> <i32 7, i32 7>)
  ret <2 x i32> %r
}

define <2 x i32> @fshr_op1_zero_vec(<2 x i32> %x) {
; CHECK-LABEL: @fshr_op1_zero_vec(
; CHECK-NEXT:    [[R:%.*]] = shl <2 x i32> [[X:%.*]], <i32 25, i32 25>
; CHECK-NEXT:    ret <2 x i32> [[R]]
;
  %r = call <2 x i32> @llvm.fshr.v2i32(<2 x i32> %x, <2 x i32> zeroinitializer, <2 x i32> <i32 7, i32 7>)
  ret <2 x i32> %r
}

; Only demand bits from one of the operands.

define i32 @fshl_only_op0_demanded(i32 %x, i32 %y) {
; CHECK-LABEL: @fshl_only_op0_demanded(
; CHECK-NEXT:    [[Z:%.*]] = shl i32 [[X:%.*]], 7
; CHECK-NEXT:    [[R:%.*]] = and i32 [[Z]], 128
; CHECK-NEXT:    ret i32 [[R]]
;
  %z = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 7)
  %r = and i32 %z, 128
  ret i32 %r
}

define i32 @fshl_only_op1_demanded(i32 %x, i32 %y) {
; CHECK-LABEL: @fshl_only_op1_demanded(
; CHECK-NEXT:    [[Z:%.*]] = lshr i32 [[Y:%.*]], 25
; CHECK-NEXT:    [[R:%.*]] = and i32 [[Z]], 63
; CHECK-NEXT:    ret i32 [[R]]
;
  %z = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 7)
  %r = and i32 %z, 63
  ret i32 %r
}

define i33 @fshr_only_op1_demanded(i33 %x, i33 %y) {
; CHECK-LABEL: @fshr_only_op1_demanded(
; CHECK-NEXT:    [[Z:%.*]] = lshr i33 [[Y:%.*]], 7
; CHECK-NEXT:    [[R:%.*]] = and i33 [[Z]], 12392
; CHECK-NEXT:    ret i33 [[R]]
;
  %z = call i33 @llvm.fshr.i33(i33 %x, i33 %y, i33 7)
  %r = and i33 %z, 12392
  ret i33 %r
}

define i33 @fshr_only_op0_demanded(i33 %x, i33 %y) {
; CHECK-LABEL: @fshr_only_op0_demanded(
; CHECK-NEXT:    [[TMP1:%.*]] = lshr i33 [[X:%.*]], 4
; CHECK-NEXT:    [[R:%.*]] = and i33 [[TMP1]], 7
; CHECK-NEXT:    ret i33 [[R]]
;
  %z = call i33 @llvm.fshr.i33(i33 %x, i33 %y, i33 7)
  %r = lshr i33 %z, 30
  ret i33 %r
}

define <2 x i31> @fshl_only_op1_demanded_vec_splat(<2 x i31> %x, <2 x i31> %y) {
; CHECK-LABEL: @fshl_only_op1_demanded_vec_splat(
; CHECK-NEXT:    [[Z:%.*]] = lshr <2 x i31> [[Y:%.*]], <i31 24, i31 24>
; CHECK-NEXT:    [[R:%.*]] = and <2 x i31> [[Z]], <i31 63, i31 31>
; CHECK-NEXT:    ret <2 x i31> [[R]]
;
  %z = call <2 x i31> @llvm.fshl.v2i31(<2 x i31> %x, <2 x i31> %y, <2 x i31> <i31 7, i31 7>)
  %r = and <2 x i31> %z, <i31 63, i31 31>
  ret <2 x i31> %r
}

; The shift modulo bitwidth is the same for all vector elements, but this is not simplified yet.
define <2 x i31> @fshl_only_op1_demanded_vec_nonsplat(<2 x i31> %x, <2 x i31> %y) {
; CHECK-LABEL: @fshl_only_op1_demanded_vec_nonsplat(
; CHECK-NEXT:    [[Z:%.*]] = call <2 x i31> @llvm.fshl.v2i31(<2 x i31> [[X:%.*]], <2 x i31> [[Y:%.*]], <2 x i31> <i31 7, i31 38>)
; CHECK-NEXT:    [[R:%.*]] = and <2 x i31> [[Z]], <i31 63, i31 31>
; CHECK-NEXT:    ret <2 x i31> [[R]]
;
  %z = call <2 x i31> @llvm.fshl.v2i31(<2 x i31> %x, <2 x i31> %y, <2 x i31> <i31 7, i31 38>)
  %r = and <2 x i31> %z, <i31 63, i31 31>
  ret <2 x i31> %r
}

; Demand bits from both operands -- cannot simplify.

define i32 @fshl_both_ops_demanded(i32 %x, i32 %y) {
; CHECK-LABEL: @fshl_both_ops_demanded(
; CHECK-NEXT:    [[Z:%.*]] = call i32 @llvm.fshl.i32(i32 [[X:%.*]], i32 [[Y:%.*]], i32 7)
; CHECK-NEXT:    [[R:%.*]] = and i32 [[Z]], 192
; CHECK-NEXT:    ret i32 [[R]]
;
  %z = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 7)
  %r = and i32 %z, 192
  ret i32 %r
}

define i33 @fshr_both_ops_demanded(i33 %x, i33 %y) {
; CHECK-LABEL: @fshr_both_ops_demanded(
; CHECK-NEXT:    [[Z:%.*]] = call i33 @llvm.fshr.i33(i33 [[X:%.*]], i33 [[Y:%.*]], i33 26)
; CHECK-NEXT:    [[R:%.*]] = and i33 [[Z]], 192
; CHECK-NEXT:    ret i33 [[R]]
;
  %z = call i33 @llvm.fshr.i33(i33 %x, i33 %y, i33 26)
  %r = and i33 %z, 192
  ret i33 %r
}

; Both operands are demanded, but there are known bits.

define i32 @fshl_known_bits(i32 %x, i32 %y) {
; CHECK-LABEL: @fshl_known_bits(
; CHECK-NEXT:    ret i32 128
;
  %x2 = or i32 %x, 1   ; lo bit set
  %y2 = lshr i32 %y, 1 ; hi bit clear
  %z = call i32 @llvm.fshl.i32(i32 %x2, i32 %y2, i32 7)
  %r = and i32 %z, 192
  ret i32 %r
}

define i33 @fshr_known_bits(i33 %x, i33 %y) {
; CHECK-LABEL: @fshr_known_bits(
; CHECK-NEXT:    ret i33 128
;
  %x2 = or i33 %x, 1 ; lo bit set
  %y2 = lshr i33 %y, 1 ; hi bit set
  %z = call i33 @llvm.fshr.i33(i33 %x2, i33 %y2, i33 26)
  %r = and i33 %z, 192
  ret i33 %r
}

; This case fails to simplify due to multiple uses.

define i33 @fshr_multi_use(i33 %a) {
; CHECK-LABEL: @fshr_multi_use(
; CHECK-NEXT:    [[B:%.*]] = tail call i33 @llvm.fshr.i33(i33 [[A:%.*]], i33 [[A]], i33 1)
; CHECK-NEXT:    [[C:%.*]] = lshr i33 [[B]], 23
; CHECK-NEXT:    [[D:%.*]] = xor i33 [[C]], [[B]]
; CHECK-NEXT:    [[E:%.*]] = and i33 [[D]], 31
; CHECK-NEXT:    ret i33 [[E]]
;
  %b = tail call i33 @llvm.fshr.i33(i33 %a, i33 %a, i33 1)
  %c = lshr i33 %b, 23
  %d = xor i33 %c, %b
  %e = and i33 %d, 31
  ret i33 %e
}

; This demonstrates the same simplification working if the fshr intrinsic
; is expanded into shifts and or.

define i33 @expanded_fshr_multi_use(i33 %a) {
; CHECK-LABEL: @expanded_fshr_multi_use(
; CHECK-NEXT:    [[TMP:%.*]] = lshr i33 [[A:%.*]], 1
; CHECK-NEXT:    [[C:%.*]] = lshr i33 [[A]], 24
; CHECK-NEXT:    [[D:%.*]] = xor i33 [[C]], [[TMP]]
; CHECK-NEXT:    [[E:%.*]] = and i33 [[D]], 31
; CHECK-NEXT:    ret i33 [[E]]
;
  %tmp = lshr i33 %a, 1
  %tmp2 = shl i33 %a, 32
  %b = or i33 %tmp, %tmp2
  %c = lshr i33 %b, 23
  %d = xor i33 %c, %b
  %e = and i33 %d, 31
  ret i33 %e
}