[SPARC] Use FMA instructions when we have UA2007

[koachan]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-sparc

Author: Koakuma (koachan)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/148434.diff

4 Files Affected:

• (modified) llvm/lib/Target/Sparc/SparcISelLowering.cpp (+13-2)
• (modified) llvm/lib/Target/Sparc/SparcISelLowering.h (+5)
• (modified) llvm/lib/Target/Sparc/SparcInstrUAOSA.td (+12)
• (added) llvm/test/CodeGen/SPARC/float-ua2007.ll (+138)

diff --git a/llvm/lib/Target/Sparc/SparcISelLowering.cpp b/llvm/lib/Target/Sparc/SparcISelLowering.cpp
index 21dbe8f585b3e..8dd8071fc6fab 100644
--- a/llvm/lib/Target/Sparc/SparcISelLowering.cpp
+++ b/llvm/lib/Target/Sparc/SparcISelLowering.cpp
@@ -1799,12 +1799,14 @@ SparcTargetLowering::SparcTargetLowering(const TargetMachine &TM,
   setOperationAction(ISD::FCOS , MVT::f64, Expand);
   setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
   setOperationAction(ISD::FREM , MVT::f64, Expand);
-  setOperationAction(ISD::FMA  , MVT::f64, Expand);
+  setOperationAction(ISD::FMA, MVT::f64,
+                     Subtarget->isUA2007() ? Legal : Expand);
   setOperationAction(ISD::FSIN , MVT::f32, Expand);
   setOperationAction(ISD::FCOS , MVT::f32, Expand);
   setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
   setOperationAction(ISD::FREM , MVT::f32, Expand);
-  setOperationAction(ISD::FMA, MVT::f32, Expand);
+  setOperationAction(ISD::FMA, MVT::f32,
+                     Subtarget->isUA2007() ? Legal : Expand);
   setOperationAction(ISD::ROTL , MVT::i32, Expand);
   setOperationAction(ISD::ROTR , MVT::i32, Expand);
   setOperationAction(ISD::BSWAP, MVT::i32, Expand);
@@ -3570,6 +3572,15 @@ bool SparcTargetLowering::isCheapToSpeculateCttz(Type *Ty) const {
   return isCheapToSpeculateCtlz(Ty);
 }
 
+bool SparcTargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
+                                                     EVT VT) const {
+  return Subtarget->isUA2007();
+}
+
+bool SparcTargetLowering::enableAggressiveFMAFusion(EVT VT) const {
+  return Subtarget->isUA2007();
+}
+
 // Override to disable global variable loading on Linux.
 void SparcTargetLowering::insertSSPDeclarations(Module &M) const {
   if (!Subtarget->isTargetLinux())
diff --git a/llvm/lib/Target/Sparc/SparcISelLowering.h b/llvm/lib/Target/Sparc/SparcISelLowering.h
index 0d220f8c3d32e..58dcd4c9bfa2f 100644
--- a/llvm/lib/Target/Sparc/SparcISelLowering.h
+++ b/llvm/lib/Target/Sparc/SparcISelLowering.h
@@ -177,6 +177,11 @@ namespace llvm {
 
     bool isCheapToSpeculateCttz(Type *Ty) const override;
 
+    bool enableAggressiveFMAFusion(EVT VT) const override;
+
+    bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
+                                    EVT VT) const override;
+
     bool shouldInsertFencesForAtomic(const Instruction *I) const override {
       // FIXME: We insert fences for each atomics and generate
       // sub-optimal code for PSO/TSO. (Approximately nobody uses any
diff --git a/llvm/lib/Target/Sparc/SparcInstrUAOSA.td b/llvm/lib/Target/Sparc/SparcInstrUAOSA.td
index 3a30e552e6db1..ffd4423137e3f 100644
--- a/llvm/lib/Target/Sparc/SparcInstrUAOSA.td
+++ b/llvm/lib/Target/Sparc/SparcInstrUAOSA.td
@@ -66,3 +66,15 @@ defm CXBCOND : F2_56<"cxb", 1>;
 def FPMADDX   : FourOp<"fpmaddx", 0b110111, 0b0000, DFPRegs>;
 def FPMADDXHI : FourOp<"fpmaddxhi", 0b110111, 0b0100, DFPRegs>;
 } // Predicates = [HasOSA2011]
+
+// UA2007 instruction patterns.
+let Predicates = [HasUA2007] in {
+def : Pat<(f32 (any_fma f32:$rs1, f32:$rs2, f32:$add)), (FMADDS $rs1, $rs2, $add)>;
+def : Pat<(f64 (any_fma f64:$rs1, f64:$rs2, f64:$add)), (FMADDD $rs1, $rs2, $add)>;
+def : Pat<(f32 (any_fma f32:$rs1, f32:$rs2, (fneg f32:$sub))), (FMSUBS $rs1, $rs2, $sub)>;
+def : Pat<(f64 (any_fma f64:$rs1, f64:$rs2, (fneg f64:$sub))), (FMSUBD $rs1, $rs2, $sub)>;
+def : Pat<(f32 (fneg (any_fma f32:$rs1, f32:$rs2, f32:$add))), (FNMADDS $rs1, $rs2, $add)>;
+def : Pat<(f64 (fneg (any_fma f64:$rs1, f64:$rs2, f64:$add))), (FNMADDD $rs1, $rs2, $add)>;
+def : Pat<(f32 (fneg (any_fma f32:$rs1, f32:$rs2, (fneg f32:$sub)))), (FNMSUBS $rs1, $rs2, $sub)>;
+def : Pat<(f64 (fneg (any_fma f64:$rs1, f64:$rs2, (fneg f64:$sub)))), (FNMSUBD $rs1, $rs2, $sub)>;
+} // Predicates = [HasUA2007]
diff --git a/llvm/test/CodeGen/SPARC/float-ua2007.ll b/llvm/test/CodeGen/SPARC/float-ua2007.ll
new file mode 100644
index 0000000000000..9534333927639
--- /dev/null
+++ b/llvm/test/CodeGen/SPARC/float-ua2007.ll
@@ -0,0 +1,138 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=sparc64 -mattr=-ua2007 < %s | FileCheck %s -check-prefix=NO-UA2007
+; RUN: llc -mtriple=sparc64 -mattr=+ua2007 < %s | FileCheck %s -check-prefix=UA2007
+
+define float @fmadds(float %a, float %b, float %c) nounwind {
+; NO-UA2007-LABEL: fmadds:
+; NO-UA2007:       ! %bb.0:
+; NO-UA2007-NEXT:    fmuls %f1, %f3, %f0
+; NO-UA2007-NEXT:    retl
+; NO-UA2007-NEXT:    fadds %f0, %f5, %f0
+;
+; UA2007-LABEL: fmadds:
+; UA2007:       ! %bb.0:
+; UA2007-NEXT:    retl
+; UA2007-NEXT:    fmadds %f1, %f3, %f5, %f0
+  %ret = call float @llvm.fmuladd.f32(float %a, float %b, float %c)
+  ret float %ret
+}
+
+define double @fmaddd(double %a, double %b, double %c) nounwind {
+; NO-UA2007-LABEL: fmaddd:
+; NO-UA2007:       ! %bb.0:
+; NO-UA2007-NEXT:    fmuld %f0, %f2, %f0
+; NO-UA2007-NEXT:    retl
+; NO-UA2007-NEXT:    faddd %f0, %f4, %f0
+;
+; UA2007-LABEL: fmaddd:
+; UA2007:       ! %bb.0:
+; UA2007-NEXT:    retl
+; UA2007-NEXT:    fmaddd %f0, %f2, %f4, %f0
+  %ret = call double @llvm.fmuladd.f64(double %a, double %b, double %c)
+  ret double %ret
+}
+
+define float @fmsubs(float %a, float %b, float %c) nounwind {
+; NO-UA2007-LABEL: fmsubs:
+; NO-UA2007:       ! %bb.0:
+; NO-UA2007-NEXT:    fmuls %f1, %f3, %f0
+; NO-UA2007-NEXT:    retl
+; NO-UA2007-NEXT:    fsubs %f0, %f5, %f0
+;
+; UA2007-LABEL: fmsubs:
+; UA2007:       ! %bb.0:
+; UA2007-NEXT:    retl
+; UA2007-NEXT:    fmsubs %f1, %f3, %f5, %f0
+  %neg = fneg float %c
+  %ret = call float @llvm.fmuladd.f32(float %a, float %b, float %neg)
+  ret float %ret
+}
+
+define double @fmsubd(double %a, double %b, double %c) nounwind {
+; NO-UA2007-LABEL: fmsubd:
+; NO-UA2007:       ! %bb.0:
+; NO-UA2007-NEXT:    fmuld %f0, %f2, %f0
+; NO-UA2007-NEXT:    retl
+; NO-UA2007-NEXT:    fsubd %f0, %f4, %f0
+;
+; UA2007-LABEL: fmsubd:
+; UA2007:       ! %bb.0:
+; UA2007-NEXT:    retl
+; UA2007-NEXT:    fmsubd %f0, %f2, %f4, %f0
+  %neg = fneg double %c
+  %ret = call double @llvm.fmuladd.f64(double %a, double %b, double %neg)
+  ret double %ret
+}
+
+define float @fnmadds(float %a, float %b, float %c) nounwind {
+; NO-UA2007-LABEL: fnmadds:
+; NO-UA2007:       ! %bb.0:
+; NO-UA2007-NEXT:    fmuls %f1, %f3, %f0
+; NO-UA2007-NEXT:    fadds %f0, %f5, %f0
+; NO-UA2007-NEXT:    retl
+; NO-UA2007-NEXT:    fnegs %f0, %f0
+;
+; UA2007-LABEL: fnmadds:
+; UA2007:       ! %bb.0:
+; UA2007-NEXT:    retl
+; UA2007-NEXT:    fnmadds %f1, %f3, %f5, %f0
+  %fma = call float @llvm.fmuladd.f32(float %a, float %b, float %c)
+  %ret = fneg float %fma
+  ret float %ret
+}
+
+define double @fnmaddd(double %a, double %b, double %c) nounwind {
+; NO-UA2007-LABEL: fnmaddd:
+; NO-UA2007:       ! %bb.0:
+; NO-UA2007-NEXT:    fmuld %f0, %f2, %f0
+; NO-UA2007-NEXT:    faddd %f0, %f4, %f0
+; NO-UA2007-NEXT:    retl
+; NO-UA2007-NEXT:    fnegd %f0, %f0
+;
+; UA2007-LABEL: fnmaddd:
+; UA2007:       ! %bb.0:
+; UA2007-NEXT:    retl
+; UA2007-NEXT:    fnmaddd %f0, %f2, %f4, %f0
+  %fma = call double @llvm.fmuladd.f64(double %a, double %b, double %c)
+  %ret = fneg double %fma
+  ret double %ret
+}
+
+define float @fnmsubs(float %a, float %b, float %c) nounwind {
+; NO-UA2007-LABEL: fnmsubs:
+; NO-UA2007:       ! %bb.0:
+; NO-UA2007-NEXT:    fmuls %f1, %f3, %f0
+; NO-UA2007-NEXT:    fsubs %f0, %f5, %f0
+; NO-UA2007-NEXT:    retl
+; NO-UA2007-NEXT:    fnegs %f0, %f0
+;
+; UA2007-LABEL: fnmsubs:
+; UA2007:       ! %bb.0:
+; UA2007-NEXT:    retl
+; UA2007-NEXT:    fnmsubs %f1, %f3, %f5, %f0
+  %neg = fneg float %c
+  %fma = call float @llvm.fmuladd.f32(float %a, float %b, float %neg)
+  %ret = fneg float %fma
+  ret float %ret
+}
+
+define double @fnmsubd(double %a, double %b, double %c) nounwind {
+; NO-UA2007-LABEL: fnmsubd:
+; NO-UA2007:       ! %bb.0:
+; NO-UA2007-NEXT:    fmuld %f0, %f2, %f0
+; NO-UA2007-NEXT:    fsubd %f0, %f4, %f0
+; NO-UA2007-NEXT:    retl
+; NO-UA2007-NEXT:    fnegd %f0, %f0
+;
+; UA2007-LABEL: fnmsubd:
+; UA2007:       ! %bb.0:
+; UA2007-NEXT:    retl
+; UA2007-NEXT:    fnmsubd %f0, %f2, %f4, %f0
+  %neg = fneg double %c
+  %fma = call double @llvm.fmuladd.f64(double %a, double %b, double %neg)
+  %ret = fneg double %fma
+  ret double %ret
+}
+
+declare float @llvm.fmuladd.f32(float, float, float)
+declare double @llvm.fmuladd.f64(double, double, double)

[koachan]

Would it be safe if I change those to return a constant true?
Given that on targets without UA2007 FMAs are going to be expanded anyway...