[llvm] [NVPTX] Lower -1/x to neg.f64(rcp.rn.f64) instead of fdiv (PR #98343)

[Rajat Bajpai via llvm-commits]

https://github.com/rajatbajpai updated https://github.com/llvm/llvm-project/pull/98343

>From cc31d5590fefad319142a54ae24c968c3aee8749 Mon Sep 17 00:00:00 2001
From: rbajpai <rbajpai at nvidia.com>
Date: Wed, 10 Jul 2024 12:35:55 +0530
Subject: [PATCH 1/3] [NVPTX] Lower -1/x to neg.f64(recp.rn.f64) instead of
 fdiv

The NVPTX backend lowers 1/x to rcp.rn.f64 instruction instead
of slower fdiv instruction. However, in the case of -1/x, it uses the
slower fdiv instruction. After this change, -1/x will be lowered
into neg.f64 (rcp.rn.f64).
---
 llvm/lib/Target/NVPTX/NVPTXInstrInfo.td | 14 ++++++++++++++
 llvm/test/CodeGen/NVPTX/rcp-opt.ll      | 13 +++++++++++++
 2 files changed, 27 insertions(+)
 create mode 100644 llvm/test/CodeGen/NVPTX/rcp-opt.ll

diff --git a/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td b/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
index 827febe845a4c..4183d6ac2537c 100644
--- a/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
+++ b/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
@@ -1150,6 +1150,16 @@ def DoubleConst1 : PatLeaf<(fpimm), [{
   return &N->getValueAPF().getSemantics() == &llvm::APFloat::IEEEdouble() &&
          N->getValueAPF().convertToDouble() == 1.0;
 }]>;
+// Constant -1.0 (double)
+def DoubleConstNeg1 : PatLeaf<(fpimm), [{
+  return &N->getValueAPF().getSemantics() == &llvm::APFloat::IEEEdouble() &&
+         N->getValueAPF().convertToDouble() == -1.0;
+}]>;
+// Constant -1.0 -> 1.0 (double)
+def NegDoubleConst : SDNodeXForm<fpimm, [{
+  return CurDAG->getTargetConstantFP(-(N->getValueAPF()),
+                                     SDLoc(N), MVT::f64);
+}]>;
 
 // Loads FP16 constant into a register.
 //
@@ -1225,6 +1235,10 @@ def FDIV64ri :
             "div.rn.f64 \t$dst, $a, $b;",
             [(set Float64Regs:$dst, (fdiv Float64Regs:$a, fpimm:$b))]>;
 
+// fdiv -1.0, X => fneg (rcp.rn X)
+def : Pat<(fdiv DoubleConstNeg1:$a, Float64Regs:$b),
+          (FNEGf64 (FDIV641r (NegDoubleConst node:$a), Float64Regs:$b))>;
+
 //
 // F32 Approximate reciprocal
 //
diff --git a/llvm/test/CodeGen/NVPTX/rcp-opt.ll b/llvm/test/CodeGen/NVPTX/rcp-opt.ll
new file mode 100644
index 0000000000000..889ba63610881
--- /dev/null
+++ b/llvm/test/CodeGen/NVPTX/rcp-opt.ll
@@ -0,0 +1,13 @@
+; RUN: llc < %s -march=nvptx64 | FileCheck %s
+; RUN: %if ptxas %{ llc < %s -march=nvptx64 | %ptxas-verify %}
+
+;; Check if fdiv -1, X lowers to fneg (rcp.rn X).
+
+; CHECK-LABEL: .func{{.*}}test1
+define double @test1(double %in) {
+; CHECK: rcp.rn.f64 [[RCP:%.*]], [[X:%.*]];
+; CHECK-NEXT: neg.f64 [[FNEG:%.*]], [[RCP]];
+  %div = fdiv double 1.000000e+00, %in
+  %neg = fsub double -0.000000e+00, %div
+  ret double %neg
+}

>From 88c5d5d433423f535c2dd5dc44fa73440425bd23 Mon Sep 17 00:00:00 2001
From: rbajpai <rbajpai at nvidia.com>
Date: Fri, 12 Jul 2024 12:28:26 +0530
Subject: [PATCH 2/3] [NFC] Cosmetic changes and added two test scenarios.

---
 llvm/lib/Target/NVPTX/NVPTXInstrInfo.td |  3 ++-
 llvm/test/CodeGen/NVPTX/rcp-opt.ll      | 23 ++++++++++++++++++++++-
 2 files changed, 24 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td b/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
index 4183d6ac2537c..5394a88d0724e 100644
--- a/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
+++ b/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
@@ -1235,7 +1235,8 @@ def FDIV64ri :
             "div.rn.f64 \t$dst, $a, $b;",
             [(set Float64Regs:$dst, (fdiv Float64Regs:$a, fpimm:$b))]>;
 
-// fdiv -1.0, X => fneg (rcp.rn X)
+// fdiv will be converted to rcp
+// fneg (fdiv 1.0, X) => fneg (rcp.rn X)
 def : Pat<(fdiv DoubleConstNeg1:$a, Float64Regs:$b),
           (FNEGf64 (FDIV641r (NegDoubleConst node:$a), Float64Regs:$b))>;
 
diff --git a/llvm/test/CodeGen/NVPTX/rcp-opt.ll b/llvm/test/CodeGen/NVPTX/rcp-opt.ll
index 889ba63610881..c320d3551943f 100644
--- a/llvm/test/CodeGen/NVPTX/rcp-opt.ll
+++ b/llvm/test/CodeGen/NVPTX/rcp-opt.ll
@@ -1,7 +1,7 @@
 ; RUN: llc < %s -march=nvptx64 | FileCheck %s
 ; RUN: %if ptxas %{ llc < %s -march=nvptx64 | %ptxas-verify %}
 
-;; Check if fdiv -1, X lowers to fneg (rcp.rn X).
+;; Check if fneg (fdiv 1, X) lowers to fneg (rcp.rn X).
 
 ; CHECK-LABEL: .func{{.*}}test1
 define double @test1(double %in) {
@@ -11,3 +11,24 @@ define double @test1(double %in) {
   %neg = fsub double -0.000000e+00, %div
   ret double %neg
 }
+
+;; Check if fdiv -1, X lowers to fneg (rcp.rn X).
+
+; CHECK-LABEL: .func{{.*}}test2
+define double @test2(double %in) {
+; CHECK: rcp.rn.f64 [[RCP:%.*]], [[X:%.*]];
+; CHECK-NEXT: neg.f64 [[FNEG:%.*]], [[RCP]];
+  %div = fdiv double -1.000000e+00, %in
+  ret double %div
+}
+
+;; Check if fdiv 1, (fneg X) lowers to fneg (rcp.rn X).
+
+; CHECK-LABEL: .func{{.*}}test3
+define double @test3(double %in) {
+; CHECK: rcp.rn.f64 [[RCP:%.*]], [[X:%.*]];
+; CHECK-NEXT: neg.f64 [[FNEG:%.*]], [[RCP]];
+  %neg = fsub double -0.000000e+00, %in
+  %div = fdiv double 1.000000e+00, %neg
+  ret double %div
+}

>From 43b539b300c3a8cb590975ad0cb3c3268c50d716 Mon Sep 17 00:00:00 2001
From: rbajpai <rbajpai at nvidia.com>
Date: Sat, 13 Jul 2024 12:37:25 +0530
Subject: [PATCH 3/3] [NFC] Fixed the comment before the Node transform

---
 llvm/lib/Target/NVPTX/NVPTXInstrInfo.td | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td b/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
index 5394a88d0724e..cd17a9de541ad 100644
--- a/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
+++ b/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
@@ -1155,7 +1155,9 @@ def DoubleConstNeg1 : PatLeaf<(fpimm), [{
   return &N->getValueAPF().getSemantics() == &llvm::APFloat::IEEEdouble() &&
          N->getValueAPF().convertToDouble() == -1.0;
 }]>;
-// Constant -1.0 -> 1.0 (double)
+
+
+// Constant -X -> X (double)
 def NegDoubleConst : SDNodeXForm<fpimm, [{
   return CurDAG->getTargetConstantFP(-(N->getValueAPF()),
                                      SDLoc(N), MVT::f64);