[DAG] Use copysign in frem power-2 fold. #91751

davemgreen · 2024-05-10T14:39:22Z

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 than fmod was:
frem -inf, 4.0 -> nan vs -nan

llvmbot · 2024-05-10T14:39:52Z

@llvm/pr-subscribers-backend-arm

@llvm/pr-subscribers-llvm-selectiondag

Author: David Green (davemgreen)

Changes

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 (or we care about sign bits). The copysign lets us get the sign of zero correct.

In testing, the only case this produced different results than fmod was:
frem -inf, 4.0 -> nan vs -nan

Full diff: https://github.com/llvm/llvm-project/pull/91751.diff

3 Files Affected:

(modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+16-9)
(modified) llvm/test/CodeGen/AArch64/frem-power2.ll (+85-7)
(modified) llvm/test/CodeGen/ARM/frem-power2.ll (+20-4)

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

llvmbot · 2024-05-10T14:39:52Z

@llvm/pr-subscribers-backend-aarch64

Author: David Green (davemgreen)

Changes

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 (or we care about sign bits). The copysign lets us get the sign of zero correct.

In testing, the only case this produced different results than fmod was:
frem -inf, 4.0 -> nan vs -nan

Full diff: https://github.com/llvm/llvm-project/pull/91751.diff

3 Files Affected:

(modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+16-9)
(modified) llvm/test/CodeGen/AArch64/frem-power2.ll (+85-7)
(modified) llvm/test/CodeGen/ARM/frem-power2.ll (+20-4)

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

jayfoad · 2024-05-10T15:31:40Z

when we do not know that the input is positive (or we care about sign bits)

"and we care about sign bits", surely?

davemgreen · 2024-05-10T15:45:48Z

when we do not know that the input is positive (or we care about sign bits)

"and we care about sign bits", surely?

Sure, sounds good. The original was even more wrong, and I messed up the reference commit too.

arsenm · 2024-05-11T07:32:53Z

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

+      DAG.isKnownToBeAPowerOfTwoFP(N1)) {
+    bool NeedsCopySign =
+        !Flags.hasNoSignedZeros() && !DAG.cannotBeOrderedNegativeFP(N0);
+    if (!NeedsCopySign || TLI.isOperationLegalOrCustom(ISD::FCOPYSIGN, VT)) {


I don't see why the legality of copysign matters here. You either need it or not

I was aiming for either knowing that the COPYSIGN isn't necessary, or checking that the lowering was going to be at least somewhat sensible. I think you might be right though, if we've already checked legality the conditions above we can assume the copysign is fine too.

As a small addition to llvm#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

…7a657b78d Local branch amd-gfx bfb7a65 Merged main:ba8a2ade84f4c1bfc531fe3673470377c038f31d into amd-gfx:de5d17eeaae8 Remote branch main 4c98f5b [DAG] Use copysign in frem power-2 fold. (llvm#91751)

davemgreen requested review from jayfoad, arsenm and RKSimon May 10, 2024 14:39

llvmbot added backend:ARM backend:AArch64 llvm:SelectionDAG SelectionDAGISel as well labels May 10, 2024

davemgreen force-pushed the gh-frem-power2-copysign branch from 1f2945e to 862abfa Compare May 10, 2024 15:44

arsenm reviewed May 11, 2024

View reviewed changes

davemgreen added 2 commits May 17, 2024 16:42

Remove COPYSIGN legality check

f43c25b

davemgreen force-pushed the gh-frem-power2-copysign branch from 862abfa to f43c25b Compare May 17, 2024 15:53

arsenm approved these changes May 17, 2024

View reviewed changes

davemgreen merged commit 4c98f5b into llvm:main May 18, 2024
4 checks passed

davemgreen deleted the gh-frem-power2-copysign branch May 18, 2024 21:50

PeddleSpam mentioned this pull request May 20, 2024

[AMDGPU] Fix error in #88512. #92770

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[DAG] Use copysign in frem power-2 fold. #91751

[DAG] Use copysign in frem power-2 fold. #91751

Uh oh!

davemgreen commented May 10, 2024 •

edited

Loading

Uh oh!

llvmbot commented May 10, 2024 •

edited

Loading

Uh oh!

llvmbot commented May 10, 2024

Uh oh!

jayfoad commented May 10, 2024

Uh oh!

davemgreen commented May 10, 2024

Uh oh!

arsenm May 11, 2024

Uh oh!

davemgreen May 17, 2024

Uh oh!

Uh oh!

Uh oh!

[DAG] Use copysign in frem power-2 fold. #91751

[DAG] Use copysign in frem power-2 fold. #91751

Uh oh!

Conversation

davemgreen commented May 10, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented May 10, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented May 10, 2024

Uh oh!

jayfoad commented May 10, 2024

Uh oh!

davemgreen commented May 10, 2024

Uh oh!

arsenm May 11, 2024

Choose a reason for hiding this comment

Uh oh!

davemgreen May 17, 2024

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

davemgreen commented May 10, 2024 •

edited

Loading

llvmbot commented May 10, 2024 •

edited

Loading