llvm-project/llvm/test/Transforms/InstCombine/fdiv.ll

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

558 lines
19 KiB
LLVM
Raw Normal View History

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -S -instcombine < %s | FileCheck %s
define float @exact_inverse(float %x) {
; CHECK-LABEL: @exact_inverse(
; CHECK-NEXT: [[DIV:%.*]] = fmul float [[X:%.*]], 1.250000e-01
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv float %x, 8.0
ret float %div
}
; Min normal float = 1.17549435E-38
define float @exact_inverse2(float %x) {
; CHECK-LABEL: @exact_inverse2(
; CHECK-NEXT: [[DIV:%.*]] = fmul float [[X:%.*]], 0x47D0000000000000
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv float %x, 0x3810000000000000
ret float %div
}
; Max exponent = 1.70141183E+38; don't transform to multiply with denormal.
define float @exact_inverse_but_denorm(float %x) {
; CHECK-LABEL: @exact_inverse_but_denorm(
; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[X:%.*]], 0x47E0000000000000
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv float %x, 0x47E0000000000000
ret float %div
}
; Denormal = float 1.40129846E-45; inverse can't be represented.
define float @not_exact_inverse2(float %x) {
; CHECK-LABEL: @not_exact_inverse2(
; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[X:%.*]], 0x36A0000000000000
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv float %x, 0x36A0000000000000
ret float %div
}
; Fast math allows us to replace this fdiv.
define float @not_exact_but_allow_recip(float %x) {
; CHECK-LABEL: @not_exact_but_allow_recip(
; CHECK-NEXT: [[DIV:%.*]] = fmul arcp float [[X:%.*]], 0x3FD5555560000000
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv arcp float %x, 3.0
ret float %div
}
; Fast math allows us to replace this fdiv, but we don't to avoid a denormal.
; TODO: What if the function attributes tell us that denormals are flushed?
define float @not_exact_but_allow_recip_but_denorm(float %x) {
; CHECK-LABEL: @not_exact_but_allow_recip_but_denorm(
; CHECK-NEXT: [[DIV:%.*]] = fdiv arcp float [[X:%.*]], 0x47E0000100000000
; CHECK-NEXT: ret float [[DIV]]
;
%div = fdiv arcp float %x, 0x47E0000100000000
ret float %div
}
define <2 x float> @exact_inverse_splat(<2 x float> %x) {
; CHECK-LABEL: @exact_inverse_splat(
; CHECK-NEXT: [[DIV:%.*]] = fmul <2 x float> [[X:%.*]], <float 2.500000e-01, float 2.500000e-01>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv <2 x float> %x, <float 4.0, float 4.0>
ret <2 x float> %div
}
; Fast math allows us to replace this fdiv.
define <2 x float> @not_exact_but_allow_recip_splat(<2 x float> %x) {
; CHECK-LABEL: @not_exact_but_allow_recip_splat(
; CHECK-NEXT: [[DIV:%.*]] = fmul arcp <2 x float> [[X:%.*]], <float 0x3FD5555560000000, float 0x3FD5555560000000>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv arcp <2 x float> %x, <float 3.0, float 3.0>
ret <2 x float> %div
}
define <2 x float> @exact_inverse_vec(<2 x float> %x) {
; CHECK-LABEL: @exact_inverse_vec(
; CHECK-NEXT: [[DIV:%.*]] = fmul <2 x float> [[X:%.*]], <float 2.500000e-01, float 1.250000e-01>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv <2 x float> %x, <float 4.0, float 8.0>
ret <2 x float> %div
}
define <2 x float> @not_exact_inverse_splat(<2 x float> %x) {
; CHECK-LABEL: @not_exact_inverse_splat(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], <float 3.000000e+00, float 3.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv <2 x float> %x, <float 3.0, float 3.0>
ret <2 x float> %div
}
define <2 x float> @not_exact_inverse_vec(<2 x float> %x) {
; CHECK-LABEL: @not_exact_inverse_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], <float 4.000000e+00, float 3.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv <2 x float> %x, <float 4.0, float 3.0>
ret <2 x float> %div
}
define <2 x float> @not_exact_inverse_vec_arcp(<2 x float> %x) {
; CHECK-LABEL: @not_exact_inverse_vec_arcp(
; CHECK-NEXT: [[DIV:%.*]] = fmul arcp <2 x float> [[X:%.*]], <float 2.500000e-01, float 0x3FD5555560000000>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv arcp <2 x float> %x, <float 4.0, float 3.0>
ret <2 x float> %div
}
define <2 x float> @not_exact_inverse_vec_arcp_with_undef_elt(<2 x float> %x) {
; CHECK-LABEL: @not_exact_inverse_vec_arcp_with_undef_elt(
; CHECK-NEXT: [[DIV:%.*]] = fdiv arcp <2 x float> [[X:%.*]], <float undef, float 3.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%div = fdiv arcp <2 x float> %x, <float undef, float 3.0>
ret <2 x float> %div
}
; (X / Y) / Z --> X / (Y * Z)
define float @div_with_div_numerator(float %x, float %y, float %z) {
; CHECK-LABEL: @div_with_div_numerator(
; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc arcp float [[Y:%.*]], [[Z:%.*]]
; CHECK-NEXT: [[DIV2:%.*]] = fdiv reassoc arcp float [[X:%.*]], [[TMP1]]
; CHECK-NEXT: ret float [[DIV2]]
;
%div1 = fdiv ninf float %x, %y
%div2 = fdiv arcp reassoc float %div1, %z
ret float %div2
}
; Z / (X / Y) --> (Z * Y) / X
define <2 x float> @div_with_div_denominator(<2 x float> %x, <2 x float> %y, <2 x float> %z) {
; CHECK-LABEL: @div_with_div_denominator(
; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc arcp <2 x float> [[Y:%.*]], [[Z:%.*]]
; CHECK-NEXT: [[DIV2:%.*]] = fdiv reassoc arcp <2 x float> [[TMP1]], [[X:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV2]]
;
%div1 = fdiv nnan <2 x float> %x, %y
%div2 = fdiv arcp reassoc <2 x float> %z, %div1
ret <2 x float> %div2
}
; Don't create an extra multiply if we can't eliminate the first div.
declare void @use_f32(float)
define float @div_with_div_numerator_extra_use(float %x, float %y, float %z) {
; CHECK-LABEL: @div_with_div_numerator_extra_use(
; CHECK-NEXT: [[DIV1:%.*]] = fdiv float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[DIV2:%.*]] = fdiv fast float [[DIV1]], [[Z:%.*]]
; CHECK-NEXT: call void @use_f32(float [[DIV1]])
; CHECK-NEXT: ret float [[DIV2]]
;
%div1 = fdiv float %x, %y
%div2 = fdiv fast float %div1, %z
call void @use_f32(float %div1)
ret float %div2
}
define float @div_with_div_denominator_extra_use(float %x, float %y, float %z) {
; CHECK-LABEL: @div_with_div_denominator_extra_use(
; CHECK-NEXT: [[DIV1:%.*]] = fdiv float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[DIV2:%.*]] = fdiv fast float [[Z:%.*]], [[DIV1]]
; CHECK-NEXT: call void @use_f32(float [[DIV1]])
; CHECK-NEXT: ret float [[DIV2]]
;
%div1 = fdiv float %x, %y
%div2 = fdiv fast float %z, %div1
call void @use_f32(float %div1)
ret float %div2
}
define float @fneg_fneg(float %x, float %y) {
; CHECK-LABEL: @fneg_fneg(
; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%x.fneg = fsub float -0.0, %x
%y.fneg = fsub float -0.0, %y
%div = fdiv float %x.fneg, %y.fneg
ret float %div
}
define float @unary_fneg_unary_fneg(float %x, float %y) {
; CHECK-LABEL: @unary_fneg_unary_fneg(
; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%x.fneg = fneg float %x
%y.fneg = fneg float %y
%div = fdiv float %x.fneg, %y.fneg
ret float %div
}
define float @unary_fneg_fneg(float %x, float %y) {
; CHECK-LABEL: @unary_fneg_fneg(
; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%x.fneg = fneg float %x
%y.fneg = fsub float -0.0, %y
%div = fdiv float %x.fneg, %y.fneg
ret float %div
}
define float @fneg_unary_fneg(float %x, float %y) {
; CHECK-LABEL: @fneg_unary_fneg(
; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%x.fneg = fsub float -0.0, %x
%y.fneg = fneg float %y
%div = fdiv float %x.fneg, %y.fneg
ret float %div
}
; The test above shows that no FMF are needed, but show that we are not dropping FMF.
define float @fneg_fneg_fast(float %x, float %y) {
; CHECK-LABEL: @fneg_fneg_fast(
; CHECK-NEXT: [[DIV:%.*]] = fdiv fast float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%x.fneg = fsub float -0.0, %x
%y.fneg = fsub float -0.0, %y
%div = fdiv fast float %x.fneg, %y.fneg
ret float %div
}
define float @unary_fneg_unary_fneg_fast(float %x, float %y) {
; CHECK-LABEL: @unary_fneg_unary_fneg_fast(
; CHECK-NEXT: [[DIV:%.*]] = fdiv fast float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%x.fneg = fneg float %x
%y.fneg = fneg float %y
%div = fdiv fast float %x.fneg, %y.fneg
ret float %div
}
define <2 x float> @fneg_fneg_vec(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @fneg_fneg_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%xneg = fsub <2 x float> <float -0.0, float -0.0>, %x
%yneg = fsub <2 x float> <float -0.0, float -0.0>, %y
%div = fdiv <2 x float> %xneg, %yneg
ret <2 x float> %div
}
define <2 x float> @unary_fneg_unary_fneg_vec(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @unary_fneg_unary_fneg_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%xneg = fneg <2 x float> %x
%yneg = fneg <2 x float> %y
%div = fdiv <2 x float> %xneg, %yneg
ret <2 x float> %div
}
define <2 x float> @fneg_unary_fneg_vec(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @fneg_unary_fneg_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%xneg = fsub <2 x float> <float -0.0, float -0.0>, %x
%yneg = fneg <2 x float> %y
%div = fdiv <2 x float> %xneg, %yneg
ret <2 x float> %div
}
define <2 x float> @unary_fneg_fneg_vec(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @unary_fneg_fneg_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%xneg = fneg <2 x float> %x
%yneg = fsub <2 x float> <float -0.0, float -0.0>, %y
%div = fdiv <2 x float> %xneg, %yneg
ret <2 x float> %div
}
define <2 x float> @fneg_fneg_vec_undef_elts(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @fneg_fneg_vec_undef_elts(
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%xneg = fsub <2 x float> <float undef, float -0.0>, %x
%yneg = fsub <2 x float> <float -0.0, float undef>, %y
%div = fdiv <2 x float> %xneg, %yneg
ret <2 x float> %div
}
define float @fneg_dividend_constant_divisor(float %x) {
; CHECK-LABEL: @fneg_dividend_constant_divisor(
; CHECK-NEXT: [[DIV:%.*]] = fdiv nsz float [[X:%.*]], -3.000000e+00
; CHECK-NEXT: ret float [[DIV]]
;
%neg = fsub float -0.0, %x
%div = fdiv nsz float %neg, 3.0
ret float %div
}
define float @unary_fneg_dividend_constant_divisor(float %x) {
; CHECK-LABEL: @unary_fneg_dividend_constant_divisor(
; CHECK-NEXT: [[DIV:%.*]] = fdiv nsz float [[X:%.*]], -3.000000e+00
; CHECK-NEXT: ret float [[DIV]]
;
%neg = fneg float %x
%div = fdiv nsz float %neg, 3.0
ret float %div
}
define float @fneg_divisor_constant_dividend(float %x) {
; CHECK-LABEL: @fneg_divisor_constant_dividend(
; CHECK-NEXT: [[DIV:%.*]] = fdiv nnan float 3.000000e+00, [[X:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%neg = fsub float -0.0, %x
%div = fdiv nnan float -3.0, %neg
ret float %div
}
define float @unary_fneg_divisor_constant_dividend(float %x) {
; CHECK-LABEL: @unary_fneg_divisor_constant_dividend(
; CHECK-NEXT: [[DIV:%.*]] = fdiv nnan float 3.000000e+00, [[X:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%neg = fneg float %x
%div = fdiv nnan float -3.0, %neg
ret float %div
}
define <2 x float> @fneg_dividend_constant_divisor_vec(<2 x float> %x) {
; CHECK-LABEL: @fneg_dividend_constant_divisor_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv ninf <2 x float> [[X:%.*]], <float -3.000000e+00, float 8.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%neg = fsub <2 x float> <float -0.0, float -0.0>, %x
%div = fdiv ninf <2 x float> %neg, <float 3.0, float -8.0>
ret <2 x float> %div
}
define <2 x float> @unary_fneg_dividend_constant_divisor_vec(<2 x float> %x) {
; CHECK-LABEL: @unary_fneg_dividend_constant_divisor_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv ninf <2 x float> [[X:%.*]], <float -3.000000e+00, float 8.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%neg = fneg <2 x float> %x
%div = fdiv ninf <2 x float> %neg, <float 3.0, float -8.0>
ret <2 x float> %div
}
define <2 x float> @fneg_dividend_constant_divisor_vec_undef_elt(<2 x float> %x) {
; CHECK-LABEL: @fneg_dividend_constant_divisor_vec_undef_elt(
; CHECK-NEXT: [[DIV:%.*]] = fdiv ninf <2 x float> [[X:%.*]], <float -3.000000e+00, float 8.000000e+00>
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%neg = fsub <2 x float> <float undef, float -0.0>, %x
%div = fdiv ninf <2 x float> %neg, <float 3.0, float -8.0>
ret <2 x float> %div
}
define <2 x float> @fneg_divisor_constant_dividend_vec(<2 x float> %x) {
; CHECK-LABEL: @fneg_divisor_constant_dividend_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv afn <2 x float> <float 3.000000e+00, float -5.000000e+00>, [[X:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%neg = fsub <2 x float> <float -0.0, float -0.0>, %x
%div = fdiv afn <2 x float> <float -3.0, float 5.0>, %neg
ret <2 x float> %div
}
define <2 x float> @unary_fneg_divisor_constant_dividend_vec(<2 x float> %x) {
; CHECK-LABEL: @unary_fneg_divisor_constant_dividend_vec(
; CHECK-NEXT: [[DIV:%.*]] = fdiv afn <2 x float> <float 3.000000e+00, float -5.000000e+00>, [[X:%.*]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%neg = fneg <2 x float> %x
%div = fdiv afn <2 x float> <float -3.0, float 5.0>, %neg
ret <2 x float> %div
}
; X / (X * Y) --> 1.0 / Y
define float @div_factor(float %x, float %y) {
; CHECK-LABEL: @div_factor(
; CHECK-NEXT: [[D:%.*]] = fdiv reassoc nnan float 1.000000e+00, [[Y:%.*]]
; CHECK-NEXT: ret float [[D]]
;
%m = fmul float %x, %y
%d = fdiv nnan reassoc float %x, %m
ret float %d;
}
; We can't do the transform without 'nnan' because if x is NAN and y is a number, this should return NAN.
define float @div_factor_too_strict(float %x, float %y) {
; CHECK-LABEL: @div_factor_too_strict(
; CHECK-NEXT: [[M:%.*]] = fmul float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[D:%.*]] = fdiv reassoc float [[X]], [[M]]
; CHECK-NEXT: ret float [[D]]
;
%m = fmul float %x, %y
%d = fdiv reassoc float %x, %m
ret float %d
}
; Commute, verify vector types, and show that we are not dropping extra FMF.
; X / (Y * X) --> 1.0 / Y
define <2 x float> @div_factor_commute(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @div_factor_commute(
; CHECK-NEXT: [[D:%.*]] = fdiv reassoc nnan ninf nsz <2 x float> <float 1.000000e+00, float 1.000000e+00>, [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[D]]
;
%m = fmul <2 x float> %y, %x
%d = fdiv nnan ninf nsz reassoc <2 x float> %x, %m
ret <2 x float> %d
}
; C1/(X*C2) => (C1/C2) / X
define <2 x float> @div_constant_dividend1(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend1(
; CHECK-NEXT: [[T2:%.*]] = fdiv reassoc arcp <2 x float> <float 5.000000e+00, float 1.000000e+00>, [[X:%.*]]
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fmul <2 x float> %x, <float 3.0e0, float 7.0e0>
%t2 = fdiv arcp reassoc <2 x float> <float 15.0e0, float 7.0e0>, %t1
ret <2 x float> %t2
}
define <2 x float> @div_constant_dividend1_arcp_only(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend1_arcp_only(
; CHECK-NEXT: [[T1:%.*]] = fmul <2 x float> [[X:%.*]], <float 3.000000e+00, float 7.000000e+00>
; CHECK-NEXT: [[T2:%.*]] = fdiv arcp <2 x float> <float 1.500000e+01, float 7.000000e+00>, [[T1]]
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fmul <2 x float> %x, <float 3.0e0, float 7.0e0>
%t2 = fdiv arcp <2 x float> <float 15.0e0, float 7.0e0>, %t1
ret <2 x float> %t2
}
; C1/(X/C2) => (C1*C2) / X
define <2 x float> @div_constant_dividend2(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend2(
; CHECK-NEXT: [[T2:%.*]] = fdiv reassoc arcp <2 x float> <float 4.500000e+01, float 4.900000e+01>, [[X:%.*]]
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fdiv <2 x float> %x, <float 3.0e0, float -7.0e0>
%t2 = fdiv arcp reassoc <2 x float> <float 15.0e0, float -7.0e0>, %t1
ret <2 x float> %t2
}
define <2 x float> @div_constant_dividend2_reassoc_only(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend2_reassoc_only(
; CHECK-NEXT: [[T1:%.*]] = fdiv <2 x float> [[X:%.*]], <float 3.000000e+00, float -7.000000e+00>
; CHECK-NEXT: [[T2:%.*]] = fdiv reassoc <2 x float> <float 1.500000e+01, float -7.000000e+00>, [[T1]]
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fdiv <2 x float> %x, <float 3.0e0, float -7.0e0>
%t2 = fdiv reassoc <2 x float> <float 15.0e0, float -7.0e0>, %t1
ret <2 x float> %t2
}
; C1/(C2/X) => (C1/C2) * X
; This tests the combination of 2 folds: (C1 * X) / C2 --> (C1 / C2) * X
define <2 x float> @div_constant_dividend3(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend3(
; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc arcp <2 x float> [[X:%.*]], <float 1.500000e+01, float -7.000000e+00>
; CHECK-NEXT: [[T2:%.*]] = fmul reassoc arcp <2 x float> [[TMP1]], <float 0x3FD5555560000000, float 0x3FC24924A0000000>
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fdiv <2 x float> <float 3.0e0, float 7.0e0>, %x
%t2 = fdiv arcp reassoc <2 x float> <float 15.0e0, float -7.0e0>, %t1
ret <2 x float> %t2
}
define double @fdiv_fneg1(double %x, double %y) {
; CHECK-LABEL: @fdiv_fneg1(
[InstCombine] canonicalize fneg before fmul/fdiv Reverse the canonicalization of fneg relative to fmul/fdiv. That makes it easier to implement the transforms (and possibly other fneg transforms) in 1 place because we can always start the pattern match from fneg (either the legacy binop or the new unop). There's a secondary practical benefit seen in PR21914 and PR42681: https://bugs.llvm.org/show_bug.cgi?id=21914 https://bugs.llvm.org/show_bug.cgi?id=42681 ...hoisting fneg rather than sinking seems to play nicer with LICM in IR (although this change may expose analysis holes in the other direction). 1. The instcombine test changes show the expected neutral IR diffs from reversing the order. 2. The reassociation tests show that we were missing an optimization opportunity to fold away fneg-of-fneg. My reading of IEEE-754 says that all of these transforms are allowed (regardless of binop/unop fneg version) because: "For all other operations [besides copy/abs/negate/copysign], this standard does not specify the sign bit of a NaN result." In all of these transforms, we always have some other binop (fadd/fsub/fmul/fdiv), so we are free to flip the sign bit of a potential intermediate NaN operand. (If that interpretation is wrong, then we must already have a bug in the existing transforms?) 3. The clang tests shouldn't exist as-is, but that's effectively a revert of rL367149 (the test broke with an extension of the pre-existing fneg canonicalization in rL367146). Differential Revision: https://reviews.llvm.org/D65399 llvm-svn: 367447
2019-08-01 00:53:22 +08:00
; CHECK-NEXT: [[NEG:%.*]] = fsub double -0.000000e+00, [[X:%.*]]
; CHECK-NEXT: [[DIV:%.*]] = fdiv double [[NEG]], [[Y:%.*]]
; CHECK-NEXT: ret double [[DIV]]
;
%neg = fsub double -0.0, %x
%div = fdiv double %neg, %y
ret double %div
}
define double @fdiv_unary_fneg1(double %x, double %y) {
; CHECK-LABEL: @fdiv_unary_fneg1(
[InstCombine] canonicalize fneg before fmul/fdiv Reverse the canonicalization of fneg relative to fmul/fdiv. That makes it easier to implement the transforms (and possibly other fneg transforms) in 1 place because we can always start the pattern match from fneg (either the legacy binop or the new unop). There's a secondary practical benefit seen in PR21914 and PR42681: https://bugs.llvm.org/show_bug.cgi?id=21914 https://bugs.llvm.org/show_bug.cgi?id=42681 ...hoisting fneg rather than sinking seems to play nicer with LICM in IR (although this change may expose analysis holes in the other direction). 1. The instcombine test changes show the expected neutral IR diffs from reversing the order. 2. The reassociation tests show that we were missing an optimization opportunity to fold away fneg-of-fneg. My reading of IEEE-754 says that all of these transforms are allowed (regardless of binop/unop fneg version) because: "For all other operations [besides copy/abs/negate/copysign], this standard does not specify the sign bit of a NaN result." In all of these transforms, we always have some other binop (fadd/fsub/fmul/fdiv), so we are free to flip the sign bit of a potential intermediate NaN operand. (If that interpretation is wrong, then we must already have a bug in the existing transforms?) 3. The clang tests shouldn't exist as-is, but that's effectively a revert of rL367149 (the test broke with an extension of the pre-existing fneg canonicalization in rL367146). Differential Revision: https://reviews.llvm.org/D65399 llvm-svn: 367447
2019-08-01 00:53:22 +08:00
; CHECK-NEXT: [[NEG:%.*]] = fneg double [[X:%.*]]
; CHECK-NEXT: [[DIV:%.*]] = fdiv double [[NEG]], [[Y:%.*]]
; CHECK-NEXT: ret double [[DIV]]
;
%neg = fneg double %x
%div = fdiv double %neg, %y
ret double %div
}
define <2 x float> @fdiv_fneg2(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @fdiv_fneg2(
[InstCombine] canonicalize fneg before fmul/fdiv Reverse the canonicalization of fneg relative to fmul/fdiv. That makes it easier to implement the transforms (and possibly other fneg transforms) in 1 place because we can always start the pattern match from fneg (either the legacy binop or the new unop). There's a secondary practical benefit seen in PR21914 and PR42681: https://bugs.llvm.org/show_bug.cgi?id=21914 https://bugs.llvm.org/show_bug.cgi?id=42681 ...hoisting fneg rather than sinking seems to play nicer with LICM in IR (although this change may expose analysis holes in the other direction). 1. The instcombine test changes show the expected neutral IR diffs from reversing the order. 2. The reassociation tests show that we were missing an optimization opportunity to fold away fneg-of-fneg. My reading of IEEE-754 says that all of these transforms are allowed (regardless of binop/unop fneg version) because: "For all other operations [besides copy/abs/negate/copysign], this standard does not specify the sign bit of a NaN result." In all of these transforms, we always have some other binop (fadd/fsub/fmul/fdiv), so we are free to flip the sign bit of a potential intermediate NaN operand. (If that interpretation is wrong, then we must already have a bug in the existing transforms?) 3. The clang tests shouldn't exist as-is, but that's effectively a revert of rL367149 (the test broke with an extension of the pre-existing fneg canonicalization in rL367146). Differential Revision: https://reviews.llvm.org/D65399 llvm-svn: 367447
2019-08-01 00:53:22 +08:00
; CHECK-NEXT: [[NEG:%.*]] = fsub <2 x float> <float -0.000000e+00, float -0.000000e+00>, [[X:%.*]]
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[Y:%.*]], [[NEG]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%neg = fsub <2 x float> <float -0.0, float -0.0>, %x
%div = fdiv <2 x float> %y, %neg
ret <2 x float> %div
}
define <2 x float> @fdiv_unary_fneg2(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @fdiv_unary_fneg2(
[InstCombine] canonicalize fneg before fmul/fdiv Reverse the canonicalization of fneg relative to fmul/fdiv. That makes it easier to implement the transforms (and possibly other fneg transforms) in 1 place because we can always start the pattern match from fneg (either the legacy binop or the new unop). There's a secondary practical benefit seen in PR21914 and PR42681: https://bugs.llvm.org/show_bug.cgi?id=21914 https://bugs.llvm.org/show_bug.cgi?id=42681 ...hoisting fneg rather than sinking seems to play nicer with LICM in IR (although this change may expose analysis holes in the other direction). 1. The instcombine test changes show the expected neutral IR diffs from reversing the order. 2. The reassociation tests show that we were missing an optimization opportunity to fold away fneg-of-fneg. My reading of IEEE-754 says that all of these transforms are allowed (regardless of binop/unop fneg version) because: "For all other operations [besides copy/abs/negate/copysign], this standard does not specify the sign bit of a NaN result." In all of these transforms, we always have some other binop (fadd/fsub/fmul/fdiv), so we are free to flip the sign bit of a potential intermediate NaN operand. (If that interpretation is wrong, then we must already have a bug in the existing transforms?) 3. The clang tests shouldn't exist as-is, but that's effectively a revert of rL367149 (the test broke with an extension of the pre-existing fneg canonicalization in rL367146). Differential Revision: https://reviews.llvm.org/D65399 llvm-svn: 367447
2019-08-01 00:53:22 +08:00
; CHECK-NEXT: [[NEG:%.*]] = fneg <2 x float> [[X:%.*]]
; CHECK-NEXT: [[DIV:%.*]] = fdiv <2 x float> [[Y:%.*]], [[NEG]]
; CHECK-NEXT: ret <2 x float> [[DIV]]
;
%neg = fneg <2 x float> %x
%div = fdiv <2 x float> %y, %neg
ret <2 x float> %div
}
define float @fdiv_fneg1_extra_use(float %x, float %y) {
; CHECK-LABEL: @fdiv_fneg1_extra_use(
; CHECK-NEXT: [[NEG:%.*]] = fsub float -0.000000e+00, [[X:%.*]]
; CHECK-NEXT: call void @use_f32(float [[NEG]])
; CHECK-NEXT: [[DIV:%.*]] = fdiv float [[NEG]], [[Y:%.*]]
; CHECK-NEXT: ret float [[DIV]]
;
%neg = fsub float -0.0, %x
call void @use_f32(float %neg)
%div = fdiv float %neg, %y
ret float %div
}