| Message ID | 20231114100320.47373-1-xry111@xry111.site |
|---|---|
| State | New |
| Headers | show |
| Series | Only allow (copysign x, NEG_CONST) -> (fneg (fabs x)) simplification for constant folding [PR112483] | expand |
On Tue, Nov 14, 2023 at 11:04 AM Xi Ruoyao <xry111@xry111.site> wrote: > > From: Andrew Pinski <pinskia@gmail.com> > > On targets with native copysign instructions, (copysign x, -1) is > usually more efficient than (fneg (fabs x)). Since r14-5284, in the > middle end we always optimize (fneg (fabs x)) to (copysign x, -1), not > vice versa. If the target does not support native fcopysign, > expand_COPYSIGN will expand it as (fneg (fabs x)) anyway. > > gcc/ChangeLog: > > PR rtl-optimization/112483 > * simplify-rtx.cc (simplify_binary_operation_1) <case COPYSIGN>: > Call simplify_unary_operation for NEG instead of > simplify_gen_unary. > --- > > [xry111]: > > Following Andrew's suggestion, I bootstrapped and regtested this > patch on loongarch64-linux-gnu. Now with > > float t(float x) > { > return __builtin_copysignf(x, -0.1234); > } > > It correctly generates > > pcalau12i $r12,%pc_hi20(.LC0) > fld.s $f1,$r12,%pc_lo12(.LC0) > fcopysign.s $f0,$f0,$f1 > jr $r1 > > instead of the de-optimized fabs.s and fneg.s > > Ok for trunk? > > gcc/simplify-rtx.cc | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/gcc/simplify-rtx.cc b/gcc/simplify-rtx.cc > index 2d2e5a3c1ca..f3745d86aea 100644 > --- a/gcc/simplify-rtx.cc > +++ b/gcc/simplify-rtx.cc > @@ -4392,7 +4392,7 @@ simplify_ashift: > real_convert (&f1, mode, CONST_DOUBLE_REAL_VALUE (trueop1)); > rtx tmp = simplify_gen_unary (ABS, mode, op0, mode); > if (REAL_VALUE_NEGATIVE (f1)) > - tmp = simplify_gen_unary (NEG, mode, tmp, mode); > + tmp = simplify_unary_operation (NEG, mode, tmp, mode); shouldn't that be when either the ABS or the NEG simplify? And I wonder when that happens - I suppose when op0 is CONST_DOUBLE only? > return tmp; > } > if (GET_CODE (op0) == NEG || GET_CODE (op0) == ABS) > -- > 2.42.1 >
On Tue, 2023-11-14 at 11:44 +0100, Richard Biener wrote: > > diff --git a/gcc/simplify-rtx.cc b/gcc/simplify-rtx.cc > > index 2d2e5a3c1ca..f3745d86aea 100644 > > --- a/gcc/simplify-rtx.cc > > +++ b/gcc/simplify-rtx.cc > > @@ -4392,7 +4392,7 @@ simplify_ashift: > > real_convert (&f1, mode, CONST_DOUBLE_REAL_VALUE (trueop1)); > > rtx tmp = simplify_gen_unary (ABS, mode, op0, mode); > > if (REAL_VALUE_NEGATIVE (f1)) > > - tmp = simplify_gen_unary (NEG, mode, tmp, mode); > > + tmp = simplify_unary_operation (NEG, mode, tmp, mode); > > return tmp; > > } > > shouldn't that be when either the ABS or the NEG simplify? Simplify (copysign x, POSTIVE_CONST) to (abs x) is an optimization. So for a positive f1, tmp will just be (abs x) and we return it. > And I wonder when that happens - I suppose when op0 is CONST_DOUBLE only? Yes, it's Andrew's intention.
On Tue, Nov 14, 2023 at 10:14 PM Xi Ruoyao <xry111@xry111.site> wrote: > > On Tue, 2023-11-14 at 11:44 +0100, Richard Biener wrote: > > > diff --git a/gcc/simplify-rtx.cc b/gcc/simplify-rtx.cc > > > index 2d2e5a3c1ca..f3745d86aea 100644 > > > --- a/gcc/simplify-rtx.cc > > > +++ b/gcc/simplify-rtx.cc > > > @@ -4392,7 +4392,7 @@ simplify_ashift: > > > real_convert (&f1, mode, CONST_DOUBLE_REAL_VALUE (trueop1)); > > > rtx tmp = simplify_gen_unary (ABS, mode, op0, mode); > > > if (REAL_VALUE_NEGATIVE (f1)) > > > - tmp = simplify_gen_unary (NEG, mode, tmp, mode); > > > + tmp = simplify_unary_operation (NEG, mode, tmp, mode); > > > return tmp; > > > } > > > > shouldn't that be when either the ABS or the NEG simplify? > > Simplify (copysign x, POSTIVE_CONST) to (abs x) is an optimization. So > for a positive f1, tmp will just be (abs x) and we return it. Ah, OK. > > And I wonder when that happens - I suppose when op0 is CONST_DOUBLE only? > > Yes, it's Andrew's intention. The patch is fine then. Richard. > -- > Xi Ruoyao <xry111@xry111.site> > School of Aerospace Science and Technology, Xidian University
diff --git a/gcc/simplify-rtx.cc b/gcc/simplify-rtx.cc index 2d2e5a3c1ca..f3745d86aea 100644 --- a/gcc/simplify-rtx.cc +++ b/gcc/simplify-rtx.cc @@ -4392,7 +4392,7 @@ simplify_ashift: real_convert (&f1, mode, CONST_DOUBLE_REAL_VALUE (trueop1)); rtx tmp = simplify_gen_unary (ABS, mode, op0, mode); if (REAL_VALUE_NEGATIVE (f1)) - tmp = simplify_gen_unary (NEG, mode, tmp, mode); + tmp = simplify_unary_operation (NEG, mode, tmp, mode); return tmp; } if (GET_CODE (op0) == NEG || GET_CODE (op0) == ABS)
From: Andrew Pinski <pinskia@gmail.com> On targets with native copysign instructions, (copysign x, -1) is usually more efficient than (fneg (fabs x)). Since r14-5284, in the middle end we always optimize (fneg (fabs x)) to (copysign x, -1), not vice versa. If the target does not support native fcopysign, expand_COPYSIGN will expand it as (fneg (fabs x)) anyway. gcc/ChangeLog: PR rtl-optimization/112483 * simplify-rtx.cc (simplify_binary_operation_1) <case COPYSIGN>: Call simplify_unary_operation for NEG instead of simplify_gen_unary. --- [xry111]: Following Andrew's suggestion, I bootstrapped and regtested this patch on loongarch64-linux-gnu. Now with float t(float x) { return __builtin_copysignf(x, -0.1234); } It correctly generates pcalau12i $r12,%pc_hi20(.LC0) fld.s $f1,$r12,%pc_lo12(.LC0) fcopysign.s $f0,$f0,$f1 jr $r1 instead of the de-optimized fabs.s and fneg.s Ok for trunk? gcc/simplify-rtx.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)