[PATCH] D80813: [AMDGPU] Propagate fast-math flags when lowering FSIN and FCOS

[Jay Foad via Phabricator via llvm-commits]

This revision was automatically updated to reflect the committed changes.
Closed by commit rG2768edfff19a: [AMDGPU] Propagate fast-math flags when lowering FSIN and FCOS (authored by foad).

Changed prior to commit:
  https://reviews.llvm.org/D80813?vs=267265&id=267488#toc

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D80813/new/

https://reviews.llvm.org/D80813

Files:
  llvm/lib/Target/AMDGPU/SIISelLowering.cpp
  llvm/test/CodeGen/AMDGPU/llvm.sin.ll


Index: llvm/test/CodeGen/AMDGPU/llvm.sin.ll
===================================================================
--- llvm/test/CodeGen/AMDGPU/llvm.sin.ll
+++ llvm/test/CodeGen/AMDGPU/llvm.sin.ll
@@ -52,7 +52,8 @@
 }
 
 ; FUNC-LABEL: {{^}}fmf_sin_3x_f32:
-; GCN: v_mul_f32
+; GCN-NOT: v_add_f32
+; GCN: 0x3ef47644
 ; GCN: v_mul_f32
 ; SICIVI: v_fract_f32
 ; GFX9-NOT: v_fract_f32
@@ -95,7 +96,8 @@
 }
 
 ; FUNC-LABEL: {{^}}fmf_sin_2x_f32:
-; GCN: v_add_f32
+; GCN-NOT: v_add_f32
+; GCN: 0x3ea2f983
 ; GCN: v_mul_f32
 ; SICIVI: v_fract_f32
 ; GFX9-NOT: v_fract_f32
@@ -137,8 +139,8 @@
 }
 
 ; FUNC-LABEL: {{^}}fmf_sin_cancel_f32:
-; GCN: v_mul_f32
-; GCN: v_mul_f32
+; GCN-NOT: v_add_f32
+; GCN-NOT: v_mul_f32
 ; SICIVI: v_fract_f32
 ; GFX9-NOT: v_fract_f32
 ; GCN: v_sin_f32
Index: llvm/lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -8288,22 +8288,24 @@
   SDValue Arg = Op.getOperand(0);
   SDValue TrigVal;
 
-  // TODO: Should this propagate fast-math-flags?
+  // Propagate fast-math flags so that the multiply we introduce can be folded
+  // if Arg is already the result of a multiply by constant.
+  auto Flags = Op->getFlags();
 
   SDValue OneOver2Pi = DAG.getConstantFP(0.5 * numbers::inv_pi, DL, VT);
 
   if (Subtarget->hasTrigReducedRange()) {
-    SDValue MulVal = DAG.getNode(ISD::FMUL, DL, VT, Arg, OneOver2Pi);
-    TrigVal = DAG.getNode(AMDGPUISD::FRACT, DL, VT, MulVal);
+    SDValue MulVal = DAG.getNode(ISD::FMUL, DL, VT, Arg, OneOver2Pi, Flags);
+    TrigVal = DAG.getNode(AMDGPUISD::FRACT, DL, VT, MulVal, Flags);
   } else {
-    TrigVal = DAG.getNode(ISD::FMUL, DL, VT, Arg, OneOver2Pi);
+    TrigVal = DAG.getNode(ISD::FMUL, DL, VT, Arg, OneOver2Pi, Flags);
   }
 
   switch (Op.getOpcode()) {
   case ISD::FCOS:
-    return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, TrigVal);
+    return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, TrigVal, Flags);
   case ISD::FSIN:
-    return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, TrigVal);
+    return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, TrigVal, Flags);
   default:
     llvm_unreachable("Wrong trig opcode");
   }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D80813.267488.patch
Type: text/x-patch
Size: 2257 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200531/7b7d8358/attachment.bin>