[llvm] [DAG] Use copysign in frem power-2 fold. (PR #91751)
David Green via llvm-commits
llvm-commits at lists.llvm.org
Fri May 10 08:44:36 PDT 2024
https://github.com/davemgreen updated https://github.com/llvm/llvm-project/pull/91751
>From 862abfa5282dff991d0b57e24e69bc63addb1cae Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Fri, 10 May 2024 16:44:25 +0100
Subject: [PATCH] [DAG] Use copysign in frem power-2 fold.
As a small addition to #91148, this uses copysign to produce the correct sign
for zero when converting frem to div/trunc/mul when we do not know that the
input is positive (and we care about sign bits). The copysign lets us get the
sign of zero correct.
In testing, the only case this produced different results thant fmod was:
frem -inf, 4.0 -> nan vs -nan
---
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 25 +++--
llvm/test/CodeGen/AArch64/frem-power2.ll | 92 +++++++++++++++++--
llvm/test/CodeGen/ARM/frem-power2.ll | 24 ++++-
3 files changed, 121 insertions(+), 20 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index be919b7a8922f..22fecc2b0d0a6 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -17383,15 +17383,22 @@ SDValue DAGCombiner::visitFREM(SDNode *N) {
TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
TLI.isOperationLegalOrCustom(ISD::FDIV, VT) &&
TLI.isOperationLegalOrCustom(ISD::FTRUNC, VT) &&
- DAG.isKnownToBeAPowerOfTwoFP(N1) &&
- (Flags.hasNoSignedZeros() || DAG.cannotBeOrderedNegativeFP(N0))) {
- SDValue Div = DAG.getNode(ISD::FDIV, DL, VT, N0, N1);
- SDValue Rnd = DAG.getNode(ISD::FTRUNC, DL, VT, Div);
- if (TLI.isFMAFasterThanFMulAndFAdd(DAG.getMachineFunction(), VT))
- return DAG.getNode(ISD::FMA, DL, VT, DAG.getNode(ISD::FNEG, DL, VT, Rnd),
- N1, N0);
- SDValue Mul = DAG.getNode(ISD::FMUL, DL, VT, Rnd, N1);
- return DAG.getNode(ISD::FSUB, DL, VT, N0, Mul);
+ DAG.isKnownToBeAPowerOfTwoFP(N1)) {
+ bool NeedsCopySign =
+ !Flags.hasNoSignedZeros() && !DAG.cannotBeOrderedNegativeFP(N0);
+ if (!NeedsCopySign || TLI.isOperationLegalOrCustom(ISD::FCOPYSIGN, VT)) {
+ SDValue Div = DAG.getNode(ISD::FDIV, DL, VT, N0, N1);
+ SDValue Rnd = DAG.getNode(ISD::FTRUNC, DL, VT, Div);
+ SDValue MLA;
+ if (TLI.isFMAFasterThanFMulAndFAdd(DAG.getMachineFunction(), VT)) {
+ MLA = DAG.getNode(ISD::FMA, DL, VT, DAG.getNode(ISD::FNEG, DL, VT, Rnd),
+ N1, N0);
+ } else {
+ SDValue Mul = DAG.getNode(ISD::FMUL, DL, VT, Rnd, N1);
+ MLA = DAG.getNode(ISD::FSUB, DL, VT, N0, Mul);
+ }
+ return NeedsCopySign ? DAG.getNode(ISD::FCOPYSIGN, DL, VT, MLA, N0) : MLA;
+ }
}
return SDValue();
diff --git a/llvm/test/CodeGen/AArch64/frem-power2.ll b/llvm/test/CodeGen/AArch64/frem-power2.ll
index 402e03c5e2652..4192745abd347 100644
--- a/llvm/test/CodeGen/AArch64/frem-power2.ll
+++ b/llvm/test/CodeGen/AArch64/frem-power2.ll
@@ -3,10 +3,22 @@
; RUN: llc -mtriple=aarch64 -mattr=+fullfp16 -global-isel -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-GI
define float @frem2(float %x) {
-; CHECK-LABEL: frem2:
-; CHECK: // %bb.0: // %entry
-; CHECK-NEXT: fmov s1, #2.00000000
-; CHECK-NEXT: b fmodf
+; CHECK-SD-LABEL: frem2:
+; CHECK-SD: // %bb.0: // %entry
+; CHECK-SD-NEXT: fmov s1, #2.00000000
+; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
+; CHECK-SD-NEXT: fdiv s2, s0, s1
+; CHECK-SD-NEXT: frintz s2, s2
+; CHECK-SD-NEXT: fmsub s1, s2, s1, s0
+; CHECK-SD-NEXT: mvni v2.4s, #128, lsl #24
+; CHECK-SD-NEXT: bit v0.16b, v1.16b, v2.16b
+; CHECK-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-SD-NEXT: ret
+;
+; CHECK-GI-LABEL: frem2:
+; CHECK-GI: // %bb.0: // %entry
+; CHECK-GI-NEXT: fmov s1, #2.00000000
+; CHECK-GI-NEXT: b fmodf
entry:
%fmod = frem float %x, 2.0
ret float %fmod
@@ -311,6 +323,67 @@ entry:
ret float %fmod
}
+define <4 x float> @frem2_vec(<4 x float> %x) {
+; CHECK-SD-LABEL: frem2_vec:
+; CHECK-SD: // %bb.0: // %entry
+; CHECK-SD-NEXT: movi v1.4s, #64, lsl #24
+; CHECK-SD-NEXT: mov v3.16b, v0.16b
+; CHECK-SD-NEXT: fdiv v2.4s, v0.4s, v1.4s
+; CHECK-SD-NEXT: frintz v2.4s, v2.4s
+; CHECK-SD-NEXT: fmls v3.4s, v1.4s, v2.4s
+; CHECK-SD-NEXT: mvni v1.4s, #128, lsl #24
+; CHECK-SD-NEXT: bit v0.16b, v3.16b, v1.16b
+; CHECK-SD-NEXT: ret
+;
+; CHECK-GI-LABEL: frem2_vec:
+; CHECK-GI: // %bb.0: // %entry
+; CHECK-GI-NEXT: sub sp, sp, #80
+; CHECK-GI-NEXT: str d10, [sp, #48] // 8-byte Folded Spill
+; CHECK-GI-NEXT: stp d9, d8, [sp, #56] // 16-byte Folded Spill
+; CHECK-GI-NEXT: str x30, [sp, #72] // 8-byte Folded Spill
+; CHECK-GI-NEXT: .cfi_def_cfa_offset 80
+; CHECK-GI-NEXT: .cfi_offset w30, -8
+; CHECK-GI-NEXT: .cfi_offset b8, -16
+; CHECK-GI-NEXT: .cfi_offset b9, -24
+; CHECK-GI-NEXT: .cfi_offset b10, -32
+; CHECK-GI-NEXT: fmov s1, #2.00000000
+; CHECK-GI-NEXT: mov s8, v0.s[1]
+; CHECK-GI-NEXT: mov s9, v0.s[2]
+; CHECK-GI-NEXT: mov s10, v0.s[3]
+; CHECK-GI-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-GI-NEXT: bl fmodf
+; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
+; CHECK-GI-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
+; CHECK-GI-NEXT: fmov s1, #2.00000000
+; CHECK-GI-NEXT: fmov s0, s8
+; CHECK-GI-NEXT: bl fmodf
+; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
+; CHECK-GI-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
+; CHECK-GI-NEXT: fmov s1, #2.00000000
+; CHECK-GI-NEXT: fmov s0, s9
+; CHECK-GI-NEXT: bl fmodf
+; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
+; CHECK-GI-NEXT: str q0, [sp] // 16-byte Folded Spill
+; CHECK-GI-NEXT: fmov s1, #2.00000000
+; CHECK-GI-NEXT: fmov s0, s10
+; CHECK-GI-NEXT: bl fmodf
+; CHECK-GI-NEXT: ldp q2, q1, [sp, #16] // 32-byte Folded Reload
+; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
+; CHECK-GI-NEXT: ldr x30, [sp, #72] // 8-byte Folded Reload
+; CHECK-GI-NEXT: ldp d9, d8, [sp, #56] // 16-byte Folded Reload
+; CHECK-GI-NEXT: ldr d10, [sp, #48] // 8-byte Folded Reload
+; CHECK-GI-NEXT: mov v1.s[1], v2.s[0]
+; CHECK-GI-NEXT: ldr q2, [sp] // 16-byte Folded Reload
+; CHECK-GI-NEXT: mov v1.s[2], v2.s[0]
+; CHECK-GI-NEXT: mov v1.s[3], v0.s[0]
+; CHECK-GI-NEXT: mov v0.16b, v1.16b
+; CHECK-GI-NEXT: add sp, sp, #80
+; CHECK-GI-NEXT: ret
+entry:
+ %fmod = frem <4 x float> %x, <float 2.0, float 2.0, float 2.0, float 2.0>
+ ret <4 x float> %fmod
+}
+
define <4 x float> @frem2_nsz_vec(<4 x float> %x) {
; CHECK-SD-LABEL: frem2_nsz_vec:
; CHECK-SD: // %bb.0: // %entry
@@ -514,10 +587,15 @@ define float @frem2_constneg_sitofp(float %x, i32 %sa) {
; CHECK-SD-LABEL: frem2_constneg_sitofp:
; CHECK-SD: // %bb.0: // %entry
; CHECK-SD-NEXT: mov w8, #1 // =0x1
-; CHECK-SD-NEXT: fmov s0, #-12.50000000
+; CHECK-SD-NEXT: fmov s1, #-12.50000000
; CHECK-SD-NEXT: lsl w8, w8, w0
-; CHECK-SD-NEXT: scvtf s1, w8
-; CHECK-SD-NEXT: b fmodf
+; CHECK-SD-NEXT: scvtf s0, w8
+; CHECK-SD-NEXT: fdiv s2, s1, s0
+; CHECK-SD-NEXT: frintz s2, s2
+; CHECK-SD-NEXT: fmsub s0, s2, s0, s1
+; CHECK-SD-NEXT: fabs s0, s0
+; CHECK-SD-NEXT: fneg s0, s0
+; CHECK-SD-NEXT: ret
;
; CHECK-GI-LABEL: frem2_constneg_sitofp:
; CHECK-GI: // %bb.0: // %entry
diff --git a/llvm/test/CodeGen/ARM/frem-power2.ll b/llvm/test/CodeGen/ARM/frem-power2.ll
index 7f52943175ac9..71c2c09c0105c 100644
--- a/llvm/test/CodeGen/ARM/frem-power2.ll
+++ b/llvm/test/CodeGen/ARM/frem-power2.ll
@@ -14,13 +14,29 @@ define float @frem4(float %x) {
;
; CHECK-FP-LABEL: frem4:
; CHECK-FP: @ %bb.0: @ %entry
-; CHECK-FP-NEXT: mov.w r1, #1082130432
-; CHECK-FP-NEXT: b fmodf
+; CHECK-FP-NEXT: vmov.f32 s0, #4.000000e+00
+; CHECK-FP-NEXT: vmov s2, r0
+; CHECK-FP-NEXT: lsrs r0, r0, #31
+; CHECK-FP-NEXT: vdiv.f32 s4, s2, s0
+; CHECK-FP-NEXT: vrintz.f32 s4, s4
+; CHECK-FP-NEXT: vfms.f32 s2, s4, s0
+; CHECK-FP-NEXT: vmov r1, s2
+; CHECK-FP-NEXT: bfi r1, r0, #31, #1
+; CHECK-FP-NEXT: mov r0, r1
+; CHECK-FP-NEXT: bx lr
;
; CHECK-M33-LABEL: frem4:
; CHECK-M33: @ %bb.0: @ %entry
-; CHECK-M33-NEXT: mov.w r1, #1082130432
-; CHECK-M33-NEXT: b fmodf
+; CHECK-M33-NEXT: vmov.f32 s0, #4.000000e+00
+; CHECK-M33-NEXT: vmov s2, r0
+; CHECK-M33-NEXT: lsrs r0, r0, #31
+; CHECK-M33-NEXT: vdiv.f32 s4, s2, s0
+; CHECK-M33-NEXT: vrintz.f32 s4, s4
+; CHECK-M33-NEXT: vmls.f32 s2, s4, s0
+; CHECK-M33-NEXT: vmov r1, s2
+; CHECK-M33-NEXT: bfi r1, r0, #31, #1
+; CHECK-M33-NEXT: mov r0, r1
+; CHECK-M33-NEXT: bx lr
entry:
%fmod = frem float %x, 4.0
ret float %fmod
More information about the llvm-commits
mailing list