[llvm] [DAG] Lower frem of power-2 using div/trunc/mul+sub (PR #91148)

David Green via llvm-commits llvm-commits at lists.llvm.org
Fri May 10 03:18:02 PDT 2024


https://github.com/davemgreen updated https://github.com/llvm/llvm-project/pull/91148

>From 532116aa3e1584930b299faa545d019d26a4391d Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Tue, 7 May 2024 13:50:50 +0100
Subject: [PATCH 1/5] [DAG] Lower frem of power-2 using div/trunk/mul+sub.

If we are lowering a frem and the divisor is known to me an integer power-2, we
can use the formula 'frem = x - trunc(x / d) * d'. This avoids the more
expensive call to fmod. The results are identical as fmod so long as d is a
power-2 (so the mul does not round incorrectly), and the sign of the return is
either always positive or signed-zero not important (nsz).

Unfortunately Alive2 does not handle this well at the moment. I was using
exhaustive checking to test this, hopefully I didn't make a mistake in it
(https://gist.github.com/davemgreen/6078015f30d3bacd1e9572f8db5d4b64).

I found this in cpythons implementation of float_pow. I currently added it as a
DAG combine for frem with power-2 fp constants, with some extra utility
functions for checking if a floating-point value is known non-negative or a
integer power-2.
---
 llvm/include/llvm/CodeGen/SelectionDAG.h      |   8 +
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp |  17 +-
 .../lib/CodeGen/SelectionDAG/SelectionDAG.cpp |  17 +
 llvm/test/CodeGen/AArch64/frem-power2.ll      | 388 +++++++++++-------
 llvm/test/CodeGen/ARM/frem-power2.ll          |  18 +-
 5 files changed, 285 insertions(+), 163 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/SelectionDAG.h b/llvm/include/llvm/CodeGen/SelectionDAG.h
index 4b1b58d4af0bb..bc4200a68f2d8 100644
--- a/llvm/include/llvm/CodeGen/SelectionDAG.h
+++ b/llvm/include/llvm/CodeGen/SelectionDAG.h
@@ -1996,6 +1996,10 @@ class SelectionDAG {
   /// is set.
   bool isKnownToBeAPowerOfTwo(SDValue Val, unsigned Depth = 0) const;
 
+  /// Test if the given _fp_ value is known to be an integer power-of-2, either
+  /// positive or negative.
+  bool isKnownToBeAPowerOfTwoFP(SDValue Val, unsigned Depth = 0) const;
+
   /// Return the number of times the sign bit of the register is replicated into
   /// the other bits. We know that at least 1 bit is always equal to the sign
   /// bit (itself), but other cases can give us information. For example,
@@ -2111,6 +2115,10 @@ class SelectionDAG {
   /// Test whether the given SDValue is known to contain non-zero value(s).
   bool isKnownNeverZero(SDValue Op, unsigned Depth = 0) const;
 
+  /// Test whether the given float value is known to not be negative. 0.0 is
+  /// considered non-negative, -0.0 is considered negative.
+  bool isKnownNonNegativeFP(SDValue Op) const;
+
   /// Test whether two SDValues are known to compare equal. This
   /// is true if they are the same value, or if one is negative zero and the
   /// other positive zero.
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index fddc97d8901a2..5fe7ab0188254 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -17365,17 +17365,32 @@ SDValue DAGCombiner::visitFREM(SDNode *N) {
   EVT VT = N->getValueType(0);
   SDNodeFlags Flags = N->getFlags();
   SelectionDAG::FlagInserter FlagsInserter(DAG, N);
+  SDLoc DL(N);
 
   if (SDValue R = DAG.simplifyFPBinop(N->getOpcode(), N0, N1, Flags))
     return R;
 
   // fold (frem c1, c2) -> fmod(c1,c2)
-  if (SDValue C = DAG.FoldConstantArithmetic(ISD::FREM, SDLoc(N), VT, {N0, N1}))
+  if (SDValue C = DAG.FoldConstantArithmetic(ISD::FREM, DL, VT, {N0, N1}))
     return C;
 
   if (SDValue NewSel = foldBinOpIntoSelect(N))
     return NewSel;
 
+  // Lower frem N0, N1 => x - trunc(N0 / N1) * N1, providing N1 is an integer
+  // power of 2.
+  if (DAG.isKnownToBeAPowerOfTwoFP(N1) &&
+      (Flags.hasNoSignedZeros() || DAG.isKnownNonNegativeFP(N0)) &&
+      !TLI.isOperationLegal(ISD::FREM, VT) &&
+      TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
+      TLI.isOperationLegalOrCustom(ISD::FDIV, VT) &&
+      TLI.isOperationLegalOrCustom(ISD::FTRUNC, VT)) {
+    SDValue Div = DAG.getNode(ISD::FDIV, DL, VT, N0, N1);
+    SDValue Rnd = DAG.getNode(ISD::FTRUNC, DL, VT, Div);
+    SDValue Mul = DAG.getNode(ISD::FMUL, DL, VT, Rnd, N1);
+    return DAG.getNode(ISD::FSUB, DL, VT, N0, Mul);
+  }
+
   return SDValue();
 }
 
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index eef5acd032345..af5a3c8fde3a7 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -4373,6 +4373,16 @@ bool SelectionDAG::isKnownToBeAPowerOfTwo(SDValue Val, unsigned Depth) const {
   return false;
 }
 
+bool SelectionDAG::isKnownToBeAPowerOfTwoFP(SDValue Val, unsigned Depth) const {
+  if (ConstantFPSDNode *C1 = isConstOrConstSplatFP(Val, true))
+    return C1->getValueAPF().getExactLog2Abs() >= 0;
+
+  if (Val.getOpcode() == ISD::UINT_TO_FP || Val.getOpcode() == ISD::SINT_TO_FP)
+    return isKnownToBeAPowerOfTwo(Val.getOperand(0), Depth + 1);
+
+  return false;
+}
+
 unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const {
   EVT VT = Op.getValueType();
 
@@ -5555,6 +5565,13 @@ bool SelectionDAG::isKnownNeverZero(SDValue Op, unsigned Depth) const {
   return computeKnownBits(Op, Depth).isNonZero();
 }
 
+bool SelectionDAG::isKnownNonNegativeFP(SDValue Op) const {
+  if (ConstantFPSDNode *C1 = isConstOrConstSplatFP(Op, true))
+    return !C1->isNegative();
+
+  return Op.getOpcode() == ISD::FABS;
+}
+
 bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const {
   // Check the obvious case.
   if (A == B) return true;
diff --git a/llvm/test/CodeGen/AArch64/frem-power2.ll b/llvm/test/CodeGen/AArch64/frem-power2.ll
index 5d627fcd6b656..94a9f701e1f57 100644
--- a/llvm/test/CodeGen/AArch64/frem-power2.ll
+++ b/llvm/test/CodeGen/AArch64/frem-power2.ll
@@ -13,31 +13,59 @@ entry:
 }
 
 define float @frem2_nsz(float %x) {
-; CHECK-LABEL: frem2_nsz:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fmov s1, #2.00000000
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem2_nsz:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fmov s1, #2.00000000
+; CHECK-SD-NEXT:    fdiv s1, s0, s1
+; CHECK-SD-NEXT:    frintz s1, s1
+; CHECK-SD-NEXT:    fadd s1, s1, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem2_nsz:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fmov s1, #2.00000000
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %fmod = frem nsz float %x, 2.0
   ret float %fmod
 }
 
 define float @frem2_fast(float %x) {
-; CHECK-LABEL: frem2_fast:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fmov s1, #2.00000000
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem2_fast:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fmov s1, #0.50000000
+; CHECK-SD-NEXT:    fmov s2, #-2.00000000
+; CHECK-SD-NEXT:    fmul s1, s0, s1
+; CHECK-SD-NEXT:    frintz s1, s1
+; CHECK-SD-NEXT:    fmadd s0, s1, s2, s0
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem2_fast:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fmov s1, #2.00000000
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %fmod = frem fast float %x, 2.0
   ret float %fmod
 }
 
 define float @frem2_abs(float %x) {
-; CHECK-LABEL: frem2_abs:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fabs s0, s0
-; CHECK-NEXT:    fmov s1, #2.00000000
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem2_abs:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fabs s0, s0
+; CHECK-SD-NEXT:    fmov s1, #2.00000000
+; CHECK-SD-NEXT:    fdiv s1, s0, s1
+; CHECK-SD-NEXT:    frintz s1, s1
+; CHECK-SD-NEXT:    fadd s1, s1, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem2_abs:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fabs s0, s0
+; CHECK-GI-NEXT:    fmov s1, #2.00000000
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %a = tail call float @llvm.fabs.f32(float %x)
   %fmod = frem float %a, 2.0
@@ -47,14 +75,11 @@ entry:
 define half @hrem2_nsz(half %x) {
 ; CHECK-SD-LABEL: hrem2_nsz:
 ; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
-; CHECK-SD-NEXT:    .cfi_def_cfa_offset 16
-; CHECK-SD-NEXT:    .cfi_offset w30, -16
-; CHECK-SD-NEXT:    fcvt s0, h0
-; CHECK-SD-NEXT:    fmov s1, #2.00000000
-; CHECK-SD-NEXT:    bl fmodf
-; CHECK-SD-NEXT:    fcvt h0, s0
-; CHECK-SD-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-SD-NEXT:    fmov h1, #2.00000000
+; CHECK-SD-NEXT:    fdiv h1, h0, h1
+; CHECK-SD-NEXT:    frintz h1, h1
+; CHECK-SD-NEXT:    fadd h1, h1, h1
+; CHECK-SD-NEXT:    fsub h0, h0, h1
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: hrem2_nsz:
@@ -75,10 +100,19 @@ entry:
 }
 
 define double @drem2_nsz(double %x) {
-; CHECK-LABEL: drem2_nsz:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fmov d1, #2.00000000
-; CHECK-NEXT:    b fmod
+; CHECK-SD-LABEL: drem2_nsz:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fmov d1, #2.00000000
+; CHECK-SD-NEXT:    fdiv d1, d0, d1
+; CHECK-SD-NEXT:    frintz d1, d1
+; CHECK-SD-NEXT:    fadd d1, d1, d1
+; CHECK-SD-NEXT:    fsub d0, d0, d1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: drem2_nsz:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fmov d1, #2.00000000
+; CHECK-GI-NEXT:    b fmod
 entry:
   %fmod = frem nsz double %x, 2.0
   ret double %fmod
@@ -105,10 +139,16 @@ entry:
 }
 
 define float @frem1_nsz(float %x) {
-; CHECK-LABEL: frem1_nsz:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fmov s1, #1.00000000
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem1_nsz:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    frintz s1, s0
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem1_nsz:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fmov s1, #1.00000000
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %fmod = frem nsz float %x, 1.0
   ret float %fmod
@@ -125,21 +165,40 @@ entry:
 }
 
 define float @fremm2_nsz(float %x) {
-; CHECK-LABEL: fremm2_nsz:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fmov s1, #-2.00000000
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: fremm2_nsz:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fmov s1, #-2.00000000
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmul s1, s2, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fremm2_nsz:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fmov s1, #-2.00000000
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %fmod = frem nsz float %x, -2.0
   ret float %fmod
 }
 
 define float @frem4_abs(float %x) {
-; CHECK-LABEL: frem4_abs:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fabs s0, s0
-; CHECK-NEXT:    fmov s1, #4.00000000
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem4_abs:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fabs s0, s0
+; CHECK-SD-NEXT:    fmov s1, #4.00000000
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmul s1, s2, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem4_abs:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fabs s0, s0
+; CHECK-GI-NEXT:    fmov s1, #4.00000000
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %a = tail call float @llvm.fabs.f32(float %x)
   %fmod = frem float %a, 4.0
@@ -147,11 +206,21 @@ entry:
 }
 
 define float @frem16_abs(float %x) {
-; CHECK-LABEL: frem16_abs:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fabs s0, s0
-; CHECK-NEXT:    fmov s1, #16.00000000
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem16_abs:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fabs s0, s0
+; CHECK-SD-NEXT:    fmov s1, #16.00000000
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmul s1, s2, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem16_abs:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fabs s0, s0
+; CHECK-GI-NEXT:    fmov s1, #16.00000000
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %a = tail call float @llvm.fabs.f32(float %x)
   %fmod = frem float %a, 16.0
@@ -159,12 +228,23 @@ entry:
 }
 
 define float @frem4294967296_abs(float %x) {
-; CHECK-LABEL: frem4294967296_abs:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fabs s0, s0
-; CHECK-NEXT:    mov w8, #1333788672 // =0x4f800000
-; CHECK-NEXT:    fmov s1, w8
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem4294967296_abs:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fabs s0, s0
+; CHECK-SD-NEXT:    mov w8, #1333788672 // =0x4f800000
+; CHECK-SD-NEXT:    fmov s1, w8
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmul s1, s2, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem4294967296_abs:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fabs s0, s0
+; CHECK-GI-NEXT:    mov w8, #1333788672 // =0x4f800000
+; CHECK-GI-NEXT:    fmov s1, w8
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %a = tail call float @llvm.fabs.f32(float %x)
   %fmod = frem float %a, 4294967296.0
@@ -172,12 +252,23 @@ entry:
 }
 
 define float @frem1152921504606846976_abs(float %x) {
-; CHECK-LABEL: frem1152921504606846976_abs:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fabs s0, s0
-; CHECK-NEXT:    mov w8, #1568669696 // =0x5d800000
-; CHECK-NEXT:    fmov s1, w8
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem1152921504606846976_abs:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fabs s0, s0
+; CHECK-SD-NEXT:    mov w8, #1568669696 // =0x5d800000
+; CHECK-SD-NEXT:    fmov s1, w8
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmul s1, s2, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem1152921504606846976_abs:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fabs s0, s0
+; CHECK-GI-NEXT:    mov w8, #1568669696 // =0x5d800000
+; CHECK-GI-NEXT:    fmov s1, w8
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %a = tail call float @llvm.fabs.f32(float %x)
   %fmod = frem float %a, 1152921504606846976.0
@@ -185,12 +276,23 @@ entry:
 }
 
 define float @frem4611686018427387904_abs(float %x) {
-; CHECK-LABEL: frem4611686018427387904_abs:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fabs s0, s0
-; CHECK-NEXT:    mov w8, #1585446912 // =0x5e800000
-; CHECK-NEXT:    fmov s1, w8
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem4611686018427387904_abs:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fabs s0, s0
+; CHECK-SD-NEXT:    mov w8, #1585446912 // =0x5e800000
+; CHECK-SD-NEXT:    fmov s1, w8
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmul s1, s2, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem4611686018427387904_abs:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fabs s0, s0
+; CHECK-GI-NEXT:    mov w8, #1585446912 // =0x5e800000
+; CHECK-GI-NEXT:    fmov s1, w8
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %a = tail call float @llvm.fabs.f32(float %x)
   %fmod = frem float %a, 4611686018427387904.0
@@ -198,11 +300,21 @@ entry:
 }
 
 define float @frem9223372036854775808_abs(float %x) {
-; CHECK-LABEL: frem9223372036854775808_abs:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    fabs s0, s0
-; CHECK-NEXT:    movi v1.2s, #95, lsl #24
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem9223372036854775808_abs:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    movi v1.2s, #95, lsl #24
+; CHECK-SD-NEXT:    fabs s0, s0
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmul s1, s2, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem9223372036854775808_abs:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fabs s0, s0
+; CHECK-GI-NEXT:    movi v1.2s, #95, lsl #24
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %a = tail call float @llvm.fabs.f32(float %x)
   %fmod = frem float %a, 9223372036854775808.0
@@ -212,42 +324,11 @@ entry:
 define <4 x float> @frem2_nsz_vec(<4 x float> %x) {
 ; CHECK-SD-LABEL: frem2_nsz_vec:
 ; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    sub sp, sp, #48
-; CHECK-SD-NEXT:    str x30, [sp, #32] // 8-byte Folded Spill
-; CHECK-SD-NEXT:    .cfi_def_cfa_offset 48
-; CHECK-SD-NEXT:    .cfi_offset w30, -16
-; CHECK-SD-NEXT:    str q0, [sp, #16] // 16-byte Folded Spill
-; CHECK-SD-NEXT:    mov s0, v0.s[1]
-; CHECK-SD-NEXT:    fmov s1, #2.00000000
-; CHECK-SD-NEXT:    bl fmodf
-; CHECK-SD-NEXT:    fmov s1, #2.00000000
-; CHECK-SD-NEXT:    // kill: def $s0 killed $s0 def $q0
-; CHECK-SD-NEXT:    str q0, [sp] // 16-byte Folded Spill
-; CHECK-SD-NEXT:    ldr q0, [sp, #16] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    // kill: def $s0 killed $s0 killed $q0
-; CHECK-SD-NEXT:    bl fmodf
-; CHECK-SD-NEXT:    ldr q1, [sp] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    // kill: def $s0 killed $s0 def $q0
-; CHECK-SD-NEXT:    mov v0.s[1], v1.s[0]
-; CHECK-SD-NEXT:    fmov s1, #2.00000000
-; CHECK-SD-NEXT:    str q0, [sp] // 16-byte Folded Spill
-; CHECK-SD-NEXT:    ldr q0, [sp, #16] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    mov s0, v0.s[2]
-; CHECK-SD-NEXT:    bl fmodf
-; CHECK-SD-NEXT:    ldr q1, [sp] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    // kill: def $s0 killed $s0 def $q0
-; CHECK-SD-NEXT:    mov v1.s[2], v0.s[0]
-; CHECK-SD-NEXT:    ldr q0, [sp, #16] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    mov s0, v0.s[3]
-; CHECK-SD-NEXT:    str q1, [sp] // 16-byte Folded Spill
-; CHECK-SD-NEXT:    fmov s1, #2.00000000
-; CHECK-SD-NEXT:    bl fmodf
-; CHECK-SD-NEXT:    ldr q1, [sp] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    // kill: def $s0 killed $s0 def $q0
-; CHECK-SD-NEXT:    ldr x30, [sp, #32] // 8-byte Folded Reload
-; CHECK-SD-NEXT:    mov v1.s[3], v0.s[0]
-; CHECK-SD-NEXT:    mov v0.16b, v1.16b
-; CHECK-SD-NEXT:    add sp, sp, #48
+; CHECK-SD-NEXT:    movi v1.4s, #64, lsl #24
+; CHECK-SD-NEXT:    fdiv v1.4s, v0.4s, v1.4s
+; CHECK-SD-NEXT:    frintz v1.4s, v1.4s
+; CHECK-SD-NEXT:    fadd v1.4s, v1.4s, v1.4s
+; CHECK-SD-NEXT:    fsub v0.4s, v0.4s, v1.4s
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem2_nsz_vec:
@@ -302,48 +383,13 @@ entry:
 define <4 x float> @frem1152921504606846976_absv(<4 x float> %x) {
 ; CHECK-SD-LABEL: frem1152921504606846976_absv:
 ; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    sub sp, sp, #48
-; CHECK-SD-NEXT:    str d8, [sp, #32] // 8-byte Folded Spill
-; CHECK-SD-NEXT:    str x30, [sp, #40] // 8-byte Folded Spill
-; CHECK-SD-NEXT:    .cfi_def_cfa_offset 48
-; CHECK-SD-NEXT:    .cfi_offset w30, -8
-; CHECK-SD-NEXT:    .cfi_offset b8, -16
-; CHECK-SD-NEXT:    fabs v0.4s, v0.4s
 ; CHECK-SD-NEXT:    mov w8, #1568669696 // =0x5d800000
-; CHECK-SD-NEXT:    fmov s8, w8
-; CHECK-SD-NEXT:    str q0, [sp, #16] // 16-byte Folded Spill
-; CHECK-SD-NEXT:    mov s0, v0.s[1]
-; CHECK-SD-NEXT:    fmov s1, s8
-; CHECK-SD-NEXT:    bl fmodf
-; CHECK-SD-NEXT:    fmov s1, s8
-; CHECK-SD-NEXT:    // kill: def $s0 killed $s0 def $q0
-; CHECK-SD-NEXT:    str q0, [sp] // 16-byte Folded Spill
-; CHECK-SD-NEXT:    ldr q0, [sp, #16] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    // kill: def $s0 killed $s0 killed $q0
-; CHECK-SD-NEXT:    bl fmodf
-; CHECK-SD-NEXT:    ldr q1, [sp] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    // kill: def $s0 killed $s0 def $q0
-; CHECK-SD-NEXT:    mov v0.s[1], v1.s[0]
-; CHECK-SD-NEXT:    fmov s1, s8
-; CHECK-SD-NEXT:    str q0, [sp] // 16-byte Folded Spill
-; CHECK-SD-NEXT:    ldr q0, [sp, #16] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    mov s0, v0.s[2]
-; CHECK-SD-NEXT:    bl fmodf
-; CHECK-SD-NEXT:    ldr q1, [sp] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    // kill: def $s0 killed $s0 def $q0
-; CHECK-SD-NEXT:    mov v1.s[2], v0.s[0]
-; CHECK-SD-NEXT:    ldr q0, [sp, #16] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    mov s0, v0.s[3]
-; CHECK-SD-NEXT:    str q1, [sp] // 16-byte Folded Spill
-; CHECK-SD-NEXT:    fmov s1, s8
-; CHECK-SD-NEXT:    bl fmodf
-; CHECK-SD-NEXT:    ldr q1, [sp] // 16-byte Folded Reload
-; CHECK-SD-NEXT:    // kill: def $s0 killed $s0 def $q0
-; CHECK-SD-NEXT:    ldr x30, [sp, #40] // 8-byte Folded Reload
-; CHECK-SD-NEXT:    ldr d8, [sp, #32] // 8-byte Folded Reload
-; CHECK-SD-NEXT:    mov v1.s[3], v0.s[0]
-; CHECK-SD-NEXT:    mov v0.16b, v1.16b
-; CHECK-SD-NEXT:    add sp, sp, #48
+; CHECK-SD-NEXT:    fabs v0.4s, v0.4s
+; CHECK-SD-NEXT:    dup v1.4s, w8
+; CHECK-SD-NEXT:    fdiv v2.4s, v0.4s, v1.4s
+; CHECK-SD-NEXT:    frintz v2.4s, v2.4s
+; CHECK-SD-NEXT:    fmul v1.4s, v2.4s, v1.4s
+; CHECK-SD-NEXT:    fsub v0.4s, v0.4s, v1.4s
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem1152921504606846976_absv:
@@ -401,12 +447,23 @@ entry:
 }
 
 define float @frem2_nsz_sitofp(float %x, i32 %sa) {
-; CHECK-LABEL: frem2_nsz_sitofp:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    mov w8, #1 // =0x1
-; CHECK-NEXT:    lsl w8, w8, w0
-; CHECK-NEXT:    scvtf s1, w8
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem2_nsz_sitofp:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    mov w8, #1 // =0x1
+; CHECK-SD-NEXT:    lsl w8, w8, w0
+; CHECK-SD-NEXT:    scvtf s1, w8
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmul s1, s2, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem2_nsz_sitofp:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    mov w8, #1 // =0x1
+; CHECK-GI-NEXT:    lsl w8, w8, w0
+; CHECK-GI-NEXT:    scvtf s1, w8
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %s = shl i32 1, %sa
   %y = sitofp i32 %s to float
@@ -415,12 +472,23 @@ entry:
 }
 
 define float @frem2_nsz_uitofp(float %x, i32 %sa) {
-; CHECK-LABEL: frem2_nsz_uitofp:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    mov w8, #1 // =0x1
-; CHECK-NEXT:    lsl w8, w8, w0
-; CHECK-NEXT:    ucvtf s1, w8
-; CHECK-NEXT:    b fmodf
+; CHECK-SD-LABEL: frem2_nsz_uitofp:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    mov w8, #1 // =0x1
+; CHECK-SD-NEXT:    lsl w8, w8, w0
+; CHECK-SD-NEXT:    ucvtf s1, w8
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmul s1, s2, s1
+; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: frem2_nsz_uitofp:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    mov w8, #1 // =0x1
+; CHECK-GI-NEXT:    lsl w8, w8, w0
+; CHECK-GI-NEXT:    ucvtf s1, w8
+; CHECK-GI-NEXT:    b fmodf
 entry:
   %s = shl i32 1, %sa
   %y = uitofp i32 %s to float
@@ -432,10 +500,14 @@ define float @frem2_const_sitofp(float %x, i32 %sa) {
 ; CHECK-SD-LABEL: frem2_const_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:    fmul s0, s2, s0
+; CHECK-SD-NEXT:    fsub s0, s1, s0
+; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem2_const_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 8052c8c35bcfb..efce515af439c 100644
--- a/llvm/test/CodeGen/ARM/frem-power2.ll
+++ b/llvm/test/CodeGen/ARM/frem-power2.ll
@@ -37,13 +37,23 @@ define float @frem4_nsz(float %x) {
 ;
 ; CHECK-FP-LABEL: frem4_nsz:
 ; 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:    vdiv.f32 s4, s2, s0
+; CHECK-FP-NEXT:    vrintz.f32 s4, s4
+; CHECK-FP-NEXT:    vmls.f32 s2, s4, s0
+; CHECK-FP-NEXT:    vmov r0, s2
+; CHECK-FP-NEXT:    bx lr
 ;
 ; CHECK-M33-LABEL: frem4_nsz:
 ; 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:    vdiv.f32 s4, s2, s0
+; CHECK-M33-NEXT:    vrintz.f32 s4, s4
+; CHECK-M33-NEXT:    vmls.f32 s2, s4, s0
+; CHECK-M33-NEXT:    vmov r0, s2
+; CHECK-M33-NEXT:    bx lr
 entry:
   %fmod = frem nsz float %x, 4.0
   ret float %fmod

>From f3a99cc901732fe4a824d54375b93b0d2a04054a Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Tue, 7 May 2024 16:12:59 +0100
Subject: [PATCH 2/5] Do the legality checks first

---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 5fe7ab0188254..beee8e8b949c7 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -17379,12 +17379,12 @@ SDValue DAGCombiner::visitFREM(SDNode *N) {
 
   // Lower frem N0, N1 => x - trunc(N0 / N1) * N1, providing N1 is an integer
   // power of 2.
-  if (DAG.isKnownToBeAPowerOfTwoFP(N1) &&
-      (Flags.hasNoSignedZeros() || DAG.isKnownNonNegativeFP(N0)) &&
-      !TLI.isOperationLegal(ISD::FREM, VT) &&
+  if (!TLI.isOperationLegal(ISD::FREM, VT) &&
       TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
       TLI.isOperationLegalOrCustom(ISD::FDIV, VT) &&
-      TLI.isOperationLegalOrCustom(ISD::FTRUNC, VT)) {
+      TLI.isOperationLegalOrCustom(ISD::FTRUNC, VT) &&
+      DAG.isKnownToBeAPowerOfTwoFP(N1) &&
+      (Flags.hasNoSignedZeros() || DAG.isKnownNonNegativeFP(N0))) {
     SDValue Div = DAG.getNode(ISD::FDIV, DL, VT, N0, N1);
     SDValue Rnd = DAG.getNode(ISD::FTRUNC, DL, VT, Div);
     SDValue Mul = DAG.getNode(ISD::FMUL, DL, VT, Rnd, N1);

>From 52ac9860b8f3a8c73ea6401e9fcb09fbac94e864 Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Thu, 9 May 2024 11:53:42 +0100
Subject: [PATCH 3/5] Use FMA

---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp |  9 ++-
 .../lib/CodeGen/SelectionDAG/SelectionDAG.cpp |  2 +-
 llvm/test/CodeGen/AArch64/frem-power2.ll      | 65 +++++++------------
 llvm/test/CodeGen/ARM/frem-power2.ll          |  2 +-
 4 files changed, 34 insertions(+), 44 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index beee8e8b949c7..9f5dc24e2376c 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -17387,8 +17387,13 @@ SDValue DAGCombiner::visitFREM(SDNode *N) {
       (Flags.hasNoSignedZeros() || DAG.isKnownNonNegativeFP(N0))) {
     SDValue Div = DAG.getNode(ISD::FDIV, DL, VT, N0, N1);
     SDValue Rnd = DAG.getNode(ISD::FTRUNC, DL, VT, Div);
-    SDValue Mul = DAG.getNode(ISD::FMUL, DL, VT, Rnd, N1);
-    return DAG.getNode(ISD::FSUB, DL, VT, N0, Mul);
+    if (TLI.isFMAFasterThanFMulAndFAdd(DAG.getMachineFunction(), VT)) {
+      return 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);
+      return DAG.getNode(ISD::FSUB, DL, VT, N0, Mul);
+    }
   }
 
   return SDValue();
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index af5a3c8fde3a7..3b950f8ace5da 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -5567,7 +5567,7 @@ bool SelectionDAG::isKnownNeverZero(SDValue Op, unsigned Depth) const {
 
 bool SelectionDAG::isKnownNonNegativeFP(SDValue Op) const {
   if (ConstantFPSDNode *C1 = isConstOrConstSplatFP(Op, true))
-    return !C1->isNegative();
+    return !C1->isNegative() && !C1->isNaN();
 
   return Op.getOpcode() == ISD::FABS;
 }
diff --git a/llvm/test/CodeGen/AArch64/frem-power2.ll b/llvm/test/CodeGen/AArch64/frem-power2.ll
index 94a9f701e1f57..402e03c5e2652 100644
--- a/llvm/test/CodeGen/AArch64/frem-power2.ll
+++ b/llvm/test/CodeGen/AArch64/frem-power2.ll
@@ -16,10 +16,9 @@ define float @frem2_nsz(float %x) {
 ; CHECK-SD-LABEL: frem2_nsz:
 ; CHECK-SD:       // %bb.0: // %entry
 ; CHECK-SD-NEXT:    fmov s1, #2.00000000
-; CHECK-SD-NEXT:    fdiv s1, s0, s1
-; CHECK-SD-NEXT:    frintz s1, s1
-; CHECK-SD-NEXT:    fadd s1, s1, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem2_nsz:
@@ -55,10 +54,9 @@ define float @frem2_abs(float %x) {
 ; CHECK-SD:       // %bb.0: // %entry
 ; CHECK-SD-NEXT:    fabs s0, s0
 ; CHECK-SD-NEXT:    fmov s1, #2.00000000
-; CHECK-SD-NEXT:    fdiv s1, s0, s1
-; CHECK-SD-NEXT:    frintz s1, s1
-; CHECK-SD-NEXT:    fadd s1, s1, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fdiv s2, s0, s1
+; CHECK-SD-NEXT:    frintz s2, s2
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem2_abs:
@@ -76,10 +74,10 @@ define half @hrem2_nsz(half %x) {
 ; CHECK-SD-LABEL: hrem2_nsz:
 ; CHECK-SD:       // %bb.0: // %entry
 ; CHECK-SD-NEXT:    fmov h1, #2.00000000
+; CHECK-SD-NEXT:    fmov h2, #-2.00000000
 ; CHECK-SD-NEXT:    fdiv h1, h0, h1
 ; CHECK-SD-NEXT:    frintz h1, h1
-; CHECK-SD-NEXT:    fadd h1, h1, h1
-; CHECK-SD-NEXT:    fsub h0, h0, h1
+; CHECK-SD-NEXT:    fmadd h0, h1, h2, h0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: hrem2_nsz:
@@ -103,10 +101,9 @@ define double @drem2_nsz(double %x) {
 ; CHECK-SD-LABEL: drem2_nsz:
 ; CHECK-SD:       // %bb.0: // %entry
 ; CHECK-SD-NEXT:    fmov d1, #2.00000000
-; CHECK-SD-NEXT:    fdiv d1, d0, d1
-; CHECK-SD-NEXT:    frintz d1, d1
-; CHECK-SD-NEXT:    fadd d1, d1, d1
-; CHECK-SD-NEXT:    fsub d0, d0, d1
+; CHECK-SD-NEXT:    fdiv d2, d0, d1
+; CHECK-SD-NEXT:    frintz d2, d2
+; CHECK-SD-NEXT:    fmsub d0, d2, d1, d0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: drem2_nsz:
@@ -170,8 +167,7 @@ define float @fremm2_nsz(float %x) {
 ; CHECK-SD-NEXT:    fmov s1, #-2.00000000
 ; CHECK-SD-NEXT:    fdiv s2, s0, s1
 ; CHECK-SD-NEXT:    frintz s2, s2
-; CHECK-SD-NEXT:    fmul s1, s2, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: fremm2_nsz:
@@ -190,8 +186,7 @@ define float @frem4_abs(float %x) {
 ; CHECK-SD-NEXT:    fmov s1, #4.00000000
 ; CHECK-SD-NEXT:    fdiv s2, s0, s1
 ; CHECK-SD-NEXT:    frintz s2, s2
-; CHECK-SD-NEXT:    fmul s1, s2, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem4_abs:
@@ -212,8 +207,7 @@ define float @frem16_abs(float %x) {
 ; CHECK-SD-NEXT:    fmov s1, #16.00000000
 ; CHECK-SD-NEXT:    fdiv s2, s0, s1
 ; CHECK-SD-NEXT:    frintz s2, s2
-; CHECK-SD-NEXT:    fmul s1, s2, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem16_abs:
@@ -235,8 +229,7 @@ define float @frem4294967296_abs(float %x) {
 ; CHECK-SD-NEXT:    fmov s1, w8
 ; CHECK-SD-NEXT:    fdiv s2, s0, s1
 ; CHECK-SD-NEXT:    frintz s2, s2
-; CHECK-SD-NEXT:    fmul s1, s2, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem4294967296_abs:
@@ -259,8 +252,7 @@ define float @frem1152921504606846976_abs(float %x) {
 ; CHECK-SD-NEXT:    fmov s1, w8
 ; CHECK-SD-NEXT:    fdiv s2, s0, s1
 ; CHECK-SD-NEXT:    frintz s2, s2
-; CHECK-SD-NEXT:    fmul s1, s2, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem1152921504606846976_abs:
@@ -283,8 +275,7 @@ define float @frem4611686018427387904_abs(float %x) {
 ; CHECK-SD-NEXT:    fmov s1, w8
 ; CHECK-SD-NEXT:    fdiv s2, s0, s1
 ; CHECK-SD-NEXT:    frintz s2, s2
-; CHECK-SD-NEXT:    fmul s1, s2, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem4611686018427387904_abs:
@@ -306,8 +297,7 @@ define float @frem9223372036854775808_abs(float %x) {
 ; CHECK-SD-NEXT:    fabs s0, s0
 ; CHECK-SD-NEXT:    fdiv s2, s0, s1
 ; CHECK-SD-NEXT:    frintz s2, s2
-; CHECK-SD-NEXT:    fmul s1, s2, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem9223372036854775808_abs:
@@ -325,10 +315,9 @@ define <4 x float> @frem2_nsz_vec(<4 x float> %x) {
 ; CHECK-SD-LABEL: frem2_nsz_vec:
 ; CHECK-SD:       // %bb.0: // %entry
 ; CHECK-SD-NEXT:    movi v1.4s, #64, lsl #24
-; CHECK-SD-NEXT:    fdiv v1.4s, v0.4s, v1.4s
-; CHECK-SD-NEXT:    frintz v1.4s, v1.4s
-; CHECK-SD-NEXT:    fadd v1.4s, v1.4s, v1.4s
-; CHECK-SD-NEXT:    fsub v0.4s, v0.4s, v1.4s
+; CHECK-SD-NEXT:    fdiv v2.4s, v0.4s, v1.4s
+; CHECK-SD-NEXT:    frintz v2.4s, v2.4s
+; CHECK-SD-NEXT:    fmls v0.4s, v1.4s, v2.4s
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem2_nsz_vec:
@@ -388,8 +377,7 @@ define <4 x float> @frem1152921504606846976_absv(<4 x float> %x) {
 ; CHECK-SD-NEXT:    dup v1.4s, w8
 ; CHECK-SD-NEXT:    fdiv v2.4s, v0.4s, v1.4s
 ; CHECK-SD-NEXT:    frintz v2.4s, v2.4s
-; CHECK-SD-NEXT:    fmul v1.4s, v2.4s, v1.4s
-; CHECK-SD-NEXT:    fsub v0.4s, v0.4s, v1.4s
+; CHECK-SD-NEXT:    fmls v0.4s, v1.4s, v2.4s
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem1152921504606846976_absv:
@@ -454,8 +442,7 @@ define float @frem2_nsz_sitofp(float %x, i32 %sa) {
 ; CHECK-SD-NEXT:    scvtf s1, w8
 ; CHECK-SD-NEXT:    fdiv s2, s0, s1
 ; CHECK-SD-NEXT:    frintz s2, s2
-; CHECK-SD-NEXT:    fmul s1, s2, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem2_nsz_sitofp:
@@ -479,8 +466,7 @@ define float @frem2_nsz_uitofp(float %x, i32 %sa) {
 ; CHECK-SD-NEXT:    ucvtf s1, w8
 ; CHECK-SD-NEXT:    fdiv s2, s0, s1
 ; CHECK-SD-NEXT:    frintz s2, s2
-; CHECK-SD-NEXT:    fmul s1, s2, s1
-; CHECK-SD-NEXT:    fsub s0, s0, s1
+; CHECK-SD-NEXT:    fmsub s0, s2, s1, s0
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem2_nsz_uitofp:
@@ -505,8 +491,7 @@ define float @frem2_const_sitofp(float %x, i32 %sa) {
 ; CHECK-SD-NEXT:    scvtf s0, w8
 ; CHECK-SD-NEXT:    fdiv s2, s1, s0
 ; CHECK-SD-NEXT:    frintz s2, s2
-; CHECK-SD-NEXT:    fmul s0, s2, s0
-; CHECK-SD-NEXT:    fsub s0, s1, s0
+; CHECK-SD-NEXT:    fmsub s0, s2, s0, s1
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: frem2_const_sitofp:
diff --git a/llvm/test/CodeGen/ARM/frem-power2.ll b/llvm/test/CodeGen/ARM/frem-power2.ll
index efce515af439c..7f52943175ac9 100644
--- a/llvm/test/CodeGen/ARM/frem-power2.ll
+++ b/llvm/test/CodeGen/ARM/frem-power2.ll
@@ -41,7 +41,7 @@ define float @frem4_nsz(float %x) {
 ; CHECK-FP-NEXT:    vmov s2, r0
 ; CHECK-FP-NEXT:    vdiv.f32 s4, s2, s0
 ; CHECK-FP-NEXT:    vrintz.f32 s4, s4
-; CHECK-FP-NEXT:    vmls.f32 s2, s4, s0
+; CHECK-FP-NEXT:    vfms.f32 s2, s4, s0
 ; CHECK-FP-NEXT:    vmov r0, s2
 ; CHECK-FP-NEXT:    bx lr
 ;

>From b6bbf9cd086cc36376f85cfd953fdee515de30ab Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Thu, 9 May 2024 12:37:41 +0100
Subject: [PATCH 4/5] No else after return

---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp  | 8 +++-----
 llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 2 +-
 2 files changed, 4 insertions(+), 6 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 9f5dc24e2376c..002576e43e3be 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -17387,13 +17387,11 @@ SDValue DAGCombiner::visitFREM(SDNode *N) {
       (Flags.hasNoSignedZeros() || DAG.isKnownNonNegativeFP(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)) {
+    if (TLI.isFMAFasterThanFMulAndFAdd(DAG.getMachineFunction(), VT))
       return 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);
-      return DAG.getNode(ISD::FSUB, DL, VT, N0, Mul);
-    }
+    SDValue Mul = DAG.getNode(ISD::FMUL, DL, VT, Rnd, N1);
+    return DAG.getNode(ISD::FSUB, DL, VT, N0, Mul);
   }
 
   return SDValue();
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 3b950f8ace5da..af5a3c8fde3a7 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -5567,7 +5567,7 @@ bool SelectionDAG::isKnownNeverZero(SDValue Op, unsigned Depth) const {
 
 bool SelectionDAG::isKnownNonNegativeFP(SDValue Op) const {
   if (ConstantFPSDNode *C1 = isConstOrConstSplatFP(Op, true))
-    return !C1->isNegative() && !C1->isNaN();
+    return !C1->isNegative();
 
   return Op.getOpcode() == ISD::FABS;
 }

>From d5013f18bffdf7f1ba3e85e839b887cbb051e339 Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Fri, 10 May 2024 11:17:50 +0100
Subject: [PATCH 5/5] isKnownPositiveOrNaNFP

---
 llvm/include/llvm/CodeGen/SelectionDAG.h       | 6 +++---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp  | 2 +-
 llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 2 +-
 3 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/SelectionDAG.h b/llvm/include/llvm/CodeGen/SelectionDAG.h
index bc4200a68f2d8..3e4f0e1f6d3a0 100644
--- a/llvm/include/llvm/CodeGen/SelectionDAG.h
+++ b/llvm/include/llvm/CodeGen/SelectionDAG.h
@@ -2115,9 +2115,9 @@ class SelectionDAG {
   /// Test whether the given SDValue is known to contain non-zero value(s).
   bool isKnownNeverZero(SDValue Op, unsigned Depth = 0) const;
 
-  /// Test whether the given float value is known to not be negative. 0.0 is
-  /// considered non-negative, -0.0 is considered negative.
-  bool isKnownNonNegativeFP(SDValue Op) const;
+  /// Test whether the given float value is known to be positive. +0.0, +inf and
+  /// +nan are considered positive, -0.0, -inf and -nan are not.
+  bool isKnownPositiveOrNaNFP(SDValue Op) const;
 
   /// Test whether two SDValues are known to compare equal. This
   /// is true if they are the same value, or if one is negative zero and the
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 002576e43e3be..394de9203e5ab 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -17384,7 +17384,7 @@ SDValue DAGCombiner::visitFREM(SDNode *N) {
       TLI.isOperationLegalOrCustom(ISD::FDIV, VT) &&
       TLI.isOperationLegalOrCustom(ISD::FTRUNC, VT) &&
       DAG.isKnownToBeAPowerOfTwoFP(N1) &&
-      (Flags.hasNoSignedZeros() || DAG.isKnownNonNegativeFP(N0))) {
+      (Flags.hasNoSignedZeros() || DAG.isKnownPositiveOrNaNFP(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))
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index af5a3c8fde3a7..04584f40547db 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -5565,7 +5565,7 @@ bool SelectionDAG::isKnownNeverZero(SDValue Op, unsigned Depth) const {
   return computeKnownBits(Op, Depth).isNonZero();
 }
 
-bool SelectionDAG::isKnownNonNegativeFP(SDValue Op) const {
+bool SelectionDAG::isKnownPositiveOrNaNFP(SDValue Op) const {
   if (ConstantFPSDNode *C1 = isConstOrConstSplatFP(Op, true))
     return !C1->isNegative();
 



More information about the llvm-commits mailing list