`llvm.canonicalize` in common code #142105

dominik-steenken · 2025-05-30T08:16:00Z

This PR takes the work previously done by @pawan-nirpal-031 on X86 in #106370, and makes it available in common code. This should enable all targets to use __builtin_canonicalize at least for f32 and f64 data types.

Canonicalization is implemented here as multiplication by 1.0, as suggested in the docs.

Moving the X86 implementation to common code was suggested in the reviews of #106370 by @arsenm .

@pawan-nirpal-031 I hope you don't mind that i went ahead with this to get it working on SystemZ.

I don't have a lot of experience with this part of the code base yet, so reviews are very welcome.

dominik-steenken · 2025-05-30T08:16:13Z

@uweigand FYI

llvmbot · 2025-05-30T08:16:35Z

@llvm/pr-subscribers-backend-loongarch
@llvm/pr-subscribers-backend-powerpc
@llvm/pr-subscribers-backend-amdgpu
@llvm/pr-subscribers-backend-aarch64

@llvm/pr-subscribers-llvm-selectiondag

Author: Dominik Steenken (dominik-steenken)

Changes

This PR takes the work previously done by @pawan-nirpal-031 on X86 in #106370, and makes it available in common code. This should enable all targets to use __builtin_canonicalize at least for f32 and f64 data types.

Canonicalization is implemented here as multiplication by 1.0, as suggested in the docs.

Moving the X86 implementation to common code was suggested in the reviews of #106370 by @arsenm .

@pawan-nirpal-031 I hope you don't mind that i went ahead with this to getit working on SystemZ.

I don't have a lot of experience with this part of the code base yet, so reviews are very welcome.

Full diff: https://github.com/llvm/llvm-project/pull/142105.diff

3 Files Affected:

(modified) llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp (+17)
(modified) llvm/lib/CodeGen/TargetLoweringBase.cpp (+3)
(added) llvm/test/CodeGen/SystemZ/canonicalize-vars.ll (+141)

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index 528c07cc5549d..6599704aaf290 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -3356,6 +3356,23 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
     Results.push_back(Op);
     break;
   }
+  case ISD::FCANONICALIZE: {
+    // This implements llvm.canonicalize.f* by multiplication with 1.0,
+    // as suggested in https://llvm.org/docs/LangRef.html#id2335.
+    // Get operand x.
+    SDValue Operand = Node->getOperand(0);
+    // Get fp value type used.
+    EVT VT = Operand.getValueType();
+    // Produce appropriately-typed 1.0 constant.
+    SDValue One = DAG.getConstantFP(1.0, dl, VT);
+    // Produce multiplication node x * 1.0.
+    SDValue Chain = DAG.getEntryNode();
+    SDValue Mul = DAG.getNode(ISD::STRICT_FMUL, dl, {VT, MVT::Other},
+                              {Chain, Operand, One});
+
+    Results.push_back(Mul);
+    break;
+  }
   case ISD::SIGN_EXTEND_INREG: {
     EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
     EVT VT = Node->getValueType(0);
diff --git a/llvm/lib/CodeGen/TargetLoweringBase.cpp b/llvm/lib/CodeGen/TargetLoweringBase.cpp
index 935afaf9dd550..ba765afd80bf1 100644
--- a/llvm/lib/CodeGen/TargetLoweringBase.cpp
+++ b/llvm/lib/CodeGen/TargetLoweringBase.cpp
@@ -870,6 +870,9 @@ void TargetLoweringBase::initActions() {
                       ISD::FATAN2},
                      {MVT::f32, MVT::f64, MVT::f128}, Expand);
 
+  // Insert custom handling default for llvm.canonicalize.*.
+  setOperationAction(ISD::FCANONICALIZE, {MVT::f32, MVT::f64}, Expand);
+
   // FIXME: Query RuntimeLibCalls to make the decision.
   setOperationAction({ISD::LRINT, ISD::LLRINT, ISD::LROUND, ISD::LLROUND},
                      {MVT::f32, MVT::f64, MVT::f128}, LibCall);
diff --git a/llvm/test/CodeGen/SystemZ/canonicalize-vars.ll b/llvm/test/CodeGen/SystemZ/canonicalize-vars.ll
new file mode 100644
index 0000000000000..477f5e1547567
--- /dev/null
+++ b/llvm/test/CodeGen/SystemZ/canonicalize-vars.ll
@@ -0,0 +1,141 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --default-march s390x-unknown-linux-gnu --version 5
+; RUN: llc -mtriple=s390x-linux-gnu -mcpu=z16 < %s | FileCheck %s -check-prefixes=Z16
+
+define float @canonicalize_fp32(float %a) {
+; Z16-LABEL: canonicalize_fp32:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vgmf %v1, 2, 8
+; Z16-NEXT:    meebr %f0, %f1
+; Z16-NEXT:    br %r14
+  %canonicalized = call float @llvm.canonicalize.f32(float %a)
+  ret float %canonicalized
+}
+
+define double @canonicalize_fp64(double %a) {
+; Z16-LABEL: canonicalize_fp64:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vgmg %v1, 2, 11
+; Z16-NEXT:    mdbr %f0, %f1
+; Z16-NEXT:    br %r14
+  %canonicalized = call double @llvm.canonicalize.f64(double %a)
+  ret double %canonicalized
+}
+
+define void @canonicalize_ptr_f32(float * %out) {
+; Z16-LABEL: canonicalize_ptr_f32:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vgmf %v0, 2, 8
+; Z16-NEXT:    meeb %f0, 0(%r2)
+; Z16-NEXT:    ste %f0, 0(%r2)
+; Z16-NEXT:    br %r14
+  %val = load float, float * %out
+  %canonicalized = call float @llvm.canonicalize.f32(float %val)
+  store float %canonicalized, float * %out
+  ret void
+}
+
+define void @canonicalize_ptr_f64(double * %out) {
+; Z16-LABEL: canonicalize_ptr_f64:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vgmg %v0, 2, 11
+; Z16-NEXT:    mdb %f0, 0(%r2)
+; Z16-NEXT:    std %f0, 0(%r2)
+; Z16-NEXT:    br %r14
+  %val = load double, double * %out
+  %canonicalized = call double @llvm.canonicalize.f64(double %val)
+  store double %canonicalized, double * %out
+  ret void
+}
+
+define <4 x float> @canonicalize_v4f32(<4 x float> %a) {
+; Z16-LABEL: canonicalize_v4f32:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vrepf %v0, %v24, 3
+; Z16-NEXT:    vgmf %v1, 2, 8
+; Z16-NEXT:    vrepf %v2, %v24, 2
+; Z16-NEXT:    meebr %f0, %f1
+; Z16-NEXT:    meebr %f2, %f1
+; Z16-NEXT:    vrepf %v3, %v24, 1
+; Z16-NEXT:    vmrhf %v0, %v2, %v0
+; Z16-NEXT:    wfmsb %f2, %v24, %f1
+; Z16-NEXT:    wfmsb %f1, %f3, %f1
+; Z16-NEXT:    vmrhf %v1, %v2, %v1
+; Z16-NEXT:    vmrhg %v24, %v1, %v0
+; Z16-NEXT:    br %r14
+  %canonicalized = call <4 x float> @llvm.canonicalize.v4f32(<4 x float> %a)
+  ret <4 x float> %canonicalized
+}
+
+define <4 x double> @canonicalize_v4f64(<4 x double> %a) {
+; Z16-LABEL: canonicalize_v4f64:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vgmg %v0, 2, 11
+; Z16-NEXT:    vrepg %v2, %v24, 1
+; Z16-NEXT:    wfmdb %f1, %v24, %f0
+; Z16-NEXT:    mdbr %f2, %f0
+; Z16-NEXT:    vmrhg %v24, %v1, %v2
+; Z16-NEXT:    vrepg %v2, %v26, 1
+; Z16-NEXT:    wfmdb %f1, %v26, %f0
+; Z16-NEXT:    wfmdb %f0, %f2, %f0
+; Z16-NEXT:    vmrhg %v26, %v1, %v0
+; Z16-NEXT:    br %r14
+  %canonicalized = call <4 x double> @llvm.canonicalize.v4f64(<4 x double> %a)
+  ret <4 x double> %canonicalized
+}
+
+define void @canonicalize_ptr_v4f32(<4 x float> * %out) {
+; Z16-LABEL: canonicalize_ptr_v4f32:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vl %v0, 0(%r2), 3
+; Z16-NEXT:    vrepf %v1, %v0, 3
+; Z16-NEXT:    vgmf %v2, 2, 8
+; Z16-NEXT:    vrepf %v3, %v0, 2
+; Z16-NEXT:    meebr %f1, %f2
+; Z16-NEXT:    meebr %f3, %f2
+; Z16-NEXT:    vmrhf %v1, %v3, %v1
+; Z16-NEXT:    wfmsb %f3, %f0, %f2
+; Z16-NEXT:    vrepf %v0, %v0, 1
+; Z16-NEXT:    meebr %f0, %f2
+; Z16-NEXT:    vmrhf %v0, %v3, %v0
+; Z16-NEXT:    vmrhg %v0, %v0, %v1
+; Z16-NEXT:    vst %v0, 0(%r2), 3
+; Z16-NEXT:    br %r14
+  %val = load <4 x float>, <4 x float> * %out
+  %canonicalized = call <4 x float> @llvm.canonicalize.v4f32(<4 x float> %val)
+  store <4 x float> %canonicalized, <4 x float> * %out
+  ret void
+}
+
+define void @canonicalize_ptr_v4f64(<4 x double> * %out) {
+; Z16-LABEL: canonicalize_ptr_v4f64:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vl %v1, 16(%r2), 4
+; Z16-NEXT:    vgmg %v2, 2, 11
+; Z16-NEXT:    wfmdb %f3, %f1, %f2
+; Z16-NEXT:    vrepg %v1, %v1, 1
+; Z16-NEXT:    mdbr %f1, %f2
+; Z16-NEXT:    vl %v0, 0(%r2), 4
+; Z16-NEXT:    vmrhg %v1, %v3, %v1
+; Z16-NEXT:    wfmdb %f3, %f0, %f2
+; Z16-NEXT:    vrepg %v0, %v0, 1
+; Z16-NEXT:    mdbr %f0, %f2
+; Z16-NEXT:    vmrhg %v0, %v3, %v0
+; Z16-NEXT:    vst %v0, 0(%r2), 4
+; Z16-NEXT:    vst %v1, 16(%r2), 4
+; Z16-NEXT:    br %r14
+  %val = load <4 x double>, <4 x double> * %out
+  %canonicalized = call <4 x double> @llvm.canonicalize.v4f64(<4 x double> %val)
+  store <4 x double> %canonicalized, <4 x double> * %out
+  ret void
+}
+
+define void @canonicalize_undef(double * %out) {
+; Z16-LABEL: canonicalize_undef:
+; Z16:       # %bb.0:
+; Z16-NEXT:    llihh %r0, 32760
+; Z16-NEXT:    stg %r0, 0(%r2)
+; Z16-NEXT:    br %r14
+  %canonicalized = call double @llvm.canonicalize.f64(double undef)
+  store double %canonicalized, double * %out
+  ret void
+}

llvmbot · 2025-05-30T08:16:36Z

@llvm/pr-subscribers-backend-systemz

Author: Dominik Steenken (dominik-steenken)

Changes

This PR takes the work previously done by @pawan-nirpal-031 on X86 in #106370, and makes it available in common code. This should enable all targets to use __builtin_canonicalize at least for f32 and f64 data types.

Canonicalization is implemented here as multiplication by 1.0, as suggested in the docs.

Moving the X86 implementation to common code was suggested in the reviews of #106370 by @arsenm .

@pawan-nirpal-031 I hope you don't mind that i went ahead with this to getit working on SystemZ.

I don't have a lot of experience with this part of the code base yet, so reviews are very welcome.

Full diff: https://github.com/llvm/llvm-project/pull/142105.diff

3 Files Affected:

(modified) llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp (+17)
(modified) llvm/lib/CodeGen/TargetLoweringBase.cpp (+3)
(added) llvm/test/CodeGen/SystemZ/canonicalize-vars.ll (+141)

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index 528c07cc5549d..6599704aaf290 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -3356,6 +3356,23 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
     Results.push_back(Op);
     break;
   }
+  case ISD::FCANONICALIZE: {
+    // This implements llvm.canonicalize.f* by multiplication with 1.0,
+    // as suggested in https://llvm.org/docs/LangRef.html#id2335.
+    // Get operand x.
+    SDValue Operand = Node->getOperand(0);
+    // Get fp value type used.
+    EVT VT = Operand.getValueType();
+    // Produce appropriately-typed 1.0 constant.
+    SDValue One = DAG.getConstantFP(1.0, dl, VT);
+    // Produce multiplication node x * 1.0.
+    SDValue Chain = DAG.getEntryNode();
+    SDValue Mul = DAG.getNode(ISD::STRICT_FMUL, dl, {VT, MVT::Other},
+                              {Chain, Operand, One});
+
+    Results.push_back(Mul);
+    break;
+  }
   case ISD::SIGN_EXTEND_INREG: {
     EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
     EVT VT = Node->getValueType(0);
diff --git a/llvm/lib/CodeGen/TargetLoweringBase.cpp b/llvm/lib/CodeGen/TargetLoweringBase.cpp
index 935afaf9dd550..ba765afd80bf1 100644
--- a/llvm/lib/CodeGen/TargetLoweringBase.cpp
+++ b/llvm/lib/CodeGen/TargetLoweringBase.cpp
@@ -870,6 +870,9 @@ void TargetLoweringBase::initActions() {
                       ISD::FATAN2},
                      {MVT::f32, MVT::f64, MVT::f128}, Expand);
 
+  // Insert custom handling default for llvm.canonicalize.*.
+  setOperationAction(ISD::FCANONICALIZE, {MVT::f32, MVT::f64}, Expand);
+
   // FIXME: Query RuntimeLibCalls to make the decision.
   setOperationAction({ISD::LRINT, ISD::LLRINT, ISD::LROUND, ISD::LLROUND},
                      {MVT::f32, MVT::f64, MVT::f128}, LibCall);
diff --git a/llvm/test/CodeGen/SystemZ/canonicalize-vars.ll b/llvm/test/CodeGen/SystemZ/canonicalize-vars.ll
new file mode 100644
index 0000000000000..477f5e1547567
--- /dev/null
+++ b/llvm/test/CodeGen/SystemZ/canonicalize-vars.ll
@@ -0,0 +1,141 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --default-march s390x-unknown-linux-gnu --version 5
+; RUN: llc -mtriple=s390x-linux-gnu -mcpu=z16 < %s | FileCheck %s -check-prefixes=Z16
+
+define float @canonicalize_fp32(float %a) {
+; Z16-LABEL: canonicalize_fp32:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vgmf %v1, 2, 8
+; Z16-NEXT:    meebr %f0, %f1
+; Z16-NEXT:    br %r14
+  %canonicalized = call float @llvm.canonicalize.f32(float %a)
+  ret float %canonicalized
+}
+
+define double @canonicalize_fp64(double %a) {
+; Z16-LABEL: canonicalize_fp64:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vgmg %v1, 2, 11
+; Z16-NEXT:    mdbr %f0, %f1
+; Z16-NEXT:    br %r14
+  %canonicalized = call double @llvm.canonicalize.f64(double %a)
+  ret double %canonicalized
+}
+
+define void @canonicalize_ptr_f32(float * %out) {
+; Z16-LABEL: canonicalize_ptr_f32:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vgmf %v0, 2, 8
+; Z16-NEXT:    meeb %f0, 0(%r2)
+; Z16-NEXT:    ste %f0, 0(%r2)
+; Z16-NEXT:    br %r14
+  %val = load float, float * %out
+  %canonicalized = call float @llvm.canonicalize.f32(float %val)
+  store float %canonicalized, float * %out
+  ret void
+}
+
+define void @canonicalize_ptr_f64(double * %out) {
+; Z16-LABEL: canonicalize_ptr_f64:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vgmg %v0, 2, 11
+; Z16-NEXT:    mdb %f0, 0(%r2)
+; Z16-NEXT:    std %f0, 0(%r2)
+; Z16-NEXT:    br %r14
+  %val = load double, double * %out
+  %canonicalized = call double @llvm.canonicalize.f64(double %val)
+  store double %canonicalized, double * %out
+  ret void
+}
+
+define <4 x float> @canonicalize_v4f32(<4 x float> %a) {
+; Z16-LABEL: canonicalize_v4f32:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vrepf %v0, %v24, 3
+; Z16-NEXT:    vgmf %v1, 2, 8
+; Z16-NEXT:    vrepf %v2, %v24, 2
+; Z16-NEXT:    meebr %f0, %f1
+; Z16-NEXT:    meebr %f2, %f1
+; Z16-NEXT:    vrepf %v3, %v24, 1
+; Z16-NEXT:    vmrhf %v0, %v2, %v0
+; Z16-NEXT:    wfmsb %f2, %v24, %f1
+; Z16-NEXT:    wfmsb %f1, %f3, %f1
+; Z16-NEXT:    vmrhf %v1, %v2, %v1
+; Z16-NEXT:    vmrhg %v24, %v1, %v0
+; Z16-NEXT:    br %r14
+  %canonicalized = call <4 x float> @llvm.canonicalize.v4f32(<4 x float> %a)
+  ret <4 x float> %canonicalized
+}
+
+define <4 x double> @canonicalize_v4f64(<4 x double> %a) {
+; Z16-LABEL: canonicalize_v4f64:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vgmg %v0, 2, 11
+; Z16-NEXT:    vrepg %v2, %v24, 1
+; Z16-NEXT:    wfmdb %f1, %v24, %f0
+; Z16-NEXT:    mdbr %f2, %f0
+; Z16-NEXT:    vmrhg %v24, %v1, %v2
+; Z16-NEXT:    vrepg %v2, %v26, 1
+; Z16-NEXT:    wfmdb %f1, %v26, %f0
+; Z16-NEXT:    wfmdb %f0, %f2, %f0
+; Z16-NEXT:    vmrhg %v26, %v1, %v0
+; Z16-NEXT:    br %r14
+  %canonicalized = call <4 x double> @llvm.canonicalize.v4f64(<4 x double> %a)
+  ret <4 x double> %canonicalized
+}
+
+define void @canonicalize_ptr_v4f32(<4 x float> * %out) {
+; Z16-LABEL: canonicalize_ptr_v4f32:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vl %v0, 0(%r2), 3
+; Z16-NEXT:    vrepf %v1, %v0, 3
+; Z16-NEXT:    vgmf %v2, 2, 8
+; Z16-NEXT:    vrepf %v3, %v0, 2
+; Z16-NEXT:    meebr %f1, %f2
+; Z16-NEXT:    meebr %f3, %f2
+; Z16-NEXT:    vmrhf %v1, %v3, %v1
+; Z16-NEXT:    wfmsb %f3, %f0, %f2
+; Z16-NEXT:    vrepf %v0, %v0, 1
+; Z16-NEXT:    meebr %f0, %f2
+; Z16-NEXT:    vmrhf %v0, %v3, %v0
+; Z16-NEXT:    vmrhg %v0, %v0, %v1
+; Z16-NEXT:    vst %v0, 0(%r2), 3
+; Z16-NEXT:    br %r14
+  %val = load <4 x float>, <4 x float> * %out
+  %canonicalized = call <4 x float> @llvm.canonicalize.v4f32(<4 x float> %val)
+  store <4 x float> %canonicalized, <4 x float> * %out
+  ret void
+}
+
+define void @canonicalize_ptr_v4f64(<4 x double> * %out) {
+; Z16-LABEL: canonicalize_ptr_v4f64:
+; Z16:       # %bb.0:
+; Z16-NEXT:    vl %v1, 16(%r2), 4
+; Z16-NEXT:    vgmg %v2, 2, 11
+; Z16-NEXT:    wfmdb %f3, %f1, %f2
+; Z16-NEXT:    vrepg %v1, %v1, 1
+; Z16-NEXT:    mdbr %f1, %f2
+; Z16-NEXT:    vl %v0, 0(%r2), 4
+; Z16-NEXT:    vmrhg %v1, %v3, %v1
+; Z16-NEXT:    wfmdb %f3, %f0, %f2
+; Z16-NEXT:    vrepg %v0, %v0, 1
+; Z16-NEXT:    mdbr %f0, %f2
+; Z16-NEXT:    vmrhg %v0, %v3, %v0
+; Z16-NEXT:    vst %v0, 0(%r2), 4
+; Z16-NEXT:    vst %v1, 16(%r2), 4
+; Z16-NEXT:    br %r14
+  %val = load <4 x double>, <4 x double> * %out
+  %canonicalized = call <4 x double> @llvm.canonicalize.v4f64(<4 x double> %val)
+  store <4 x double> %canonicalized, <4 x double> * %out
+  ret void
+}
+
+define void @canonicalize_undef(double * %out) {
+; Z16-LABEL: canonicalize_undef:
+; Z16:       # %bb.0:
+; Z16-NEXT:    llihh %r0, 32760
+; Z16-NEXT:    stg %r0, 0(%r2)
+; Z16-NEXT:    br %r14
+  %canonicalized = call double @llvm.canonicalize.f64(double undef)
+  store double %canonicalized, double * %out
+  ret void
+}

github-actions · 2025-05-30T08:18:22Z

✅ With the latest revision this PR passed the undef deprecator.

arsenm · 2025-05-30T09:04:34Z

llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

+    // This implements llvm.canonicalize.f* by multiplication with 1.0,
+    // as suggested in https://llvm.org/docs/LangRef.html#id2335.


This should probably elaborate on how this is hacking in strictfp operations in non-strict functions

From what i understand, the point of the strictfp operation here is to prevent any optimization that would, for instance, recognize that 1.0 * x can be converted to just x. So would this suffice as an addition to the comment?

// We use `STRICT_FMUL` here to prevent this operation from being optimized out.

pawan-nirpal-031 · 2025-05-30T13:54:46Z

Hi @dominik-steenken absolutely, Please go ahead with this, I have been wanting to get back and finish on the commitments as following PRs, this is much welcome. I will make some time to support canonicalization of denormals ( I have a PR just haven't been able to attend to it, if you need it sooner please let me know, I'll give you the patch and you may go ahead with it ) Thanks for your PR.

Previously, this intrinsic has been implemented in the X86 backend. As suggested in the comments for the PR that merges that implementation (llvm#106370), this commit replicates that implementation (at least for `f32` and `f64`) in common code. Canonicalization is implemented as multiplication with 1.0.

These tests include only the floating point types supported in common code. Once a SystemZ-specific implementation for `f16` and `f128` is available, these tests will be expanded to include those types as well. The tests are heavily based on the equivalent tests for X86.

…r f32,f64 Since the default OperationAction for `ISD::FCANONICALIZE` is now `Expand`, these targets need to explicitly set it to `Legal` in their backends in order to preserve their functionality.

dominik-steenken · 2025-06-05T11:35:02Z

Ok, so i updated the comment as requested, got rid of the undef check in order to remove the error from the undef deprecator and added a bunch of setOperationAction(..., Legal) to various backends that had independently handled fcanonicalize via their Tablegen files in order to get the tests clean again. Would appreciate another review, especially from folks of the backends that i had to touch in order to fix the tests :).

jayfoad · 2025-06-05T12:19:09Z

__builtin.canonicalize in common code

This patch is about the LLVM intrinsic, not the Clang builtin, so the title should say llvm.canonicalize

llvmbot added backend:SystemZ llvm:SelectionDAG SelectionDAGISel as well labels May 30, 2025

arsenm reviewed May 30, 2025

View reviewed changes

llvmbot added backend:AArch64 backend:AMDGPU labels Jun 3, 2025

dominik-steenken force-pushed the generic-canonicalize branch from f08a196 to ce1fea8 Compare June 5, 2025 08:41

llvmbot added backend:PowerPC llvm:codegen labels Jun 5, 2025

dominik-steenken force-pushed the generic-canonicalize branch from ce1fea8 to 524580c Compare June 5, 2025 09:11

llvmbot added the backend:loongarch label Jun 5, 2025

dominik-steenken added 2 commits June 5, 2025 11:12

[AArch64][AMDGPU][Mips][PowerPC][LoongArch] Legalize fcanonicalize fo…

07f5741

…r f32,f64 Since the default OperationAction for `ISD::FCANONICALIZE` is now `Expand`, these targets need to explicitly set it to `Legal` in their backends in order to preserve their functionality.

dominik-steenken force-pushed the generic-canonicalize branch from 524580c to 07f5741 Compare June 5, 2025 09:13

dominik-steenken added 2 commits June 5, 2025 11:17

Update comment based on review

cd13388

Straighten out code formatting

510da45

dominik-steenken changed the title ~~__builtin.canonicalize in common code~~ llvm.canonicalize in common code Jun 6, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

`llvm.canonicalize` in common code #142105

`llvm.canonicalize` in common code #142105

dominik-steenken commented May 30, 2025 •

edited

Loading

Uh oh!

dominik-steenken commented May 30, 2025

Uh oh!

llvmbot commented May 30, 2025 •

edited

Loading

Uh oh!

llvmbot commented May 30, 2025

Uh oh!

github-actions bot commented May 30, 2025 •

edited

Loading

Uh oh!

arsenm May 30, 2025

Uh oh!

dominik-steenken May 30, 2025

Uh oh!

pawan-nirpal-031 commented May 30, 2025 •

edited

Loading

Uh oh!

dominik-steenken commented Jun 5, 2025

Uh oh!

jayfoad commented Jun 5, 2025

Uh oh!

Uh oh!

		// This implements llvm.canonicalize.f* by multiplication with 1.0,
		// as suggested in https://llvm.org/docs/LangRef.html#id2335.

llvm.canonicalize in common code #142105

Are you sure you want to change the base?

llvm.canonicalize in common code #142105

Conversation

dominik-steenken commented May 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dominik-steenken commented May 30, 2025

Uh oh!

llvmbot commented May 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented May 30, 2025

Uh oh!

github-actions bot commented May 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

arsenm May 30, 2025

Choose a reason for hiding this comment

Uh oh!

dominik-steenken May 30, 2025

Choose a reason for hiding this comment

Uh oh!

pawan-nirpal-031 commented May 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dominik-steenken commented Jun 5, 2025

Uh oh!

jayfoad commented Jun 5, 2025

Uh oh!

Uh oh!

`llvm.canonicalize` in common code #142105

`llvm.canonicalize` in common code #142105

dominik-steenken commented May 30, 2025 •

edited

Loading

llvmbot commented May 30, 2025 •

edited

Loading

github-actions bot commented May 30, 2025 •

edited

Loading

pawan-nirpal-031 commented May 30, 2025 •

edited

Loading