[PATCH] D110076: [AMDGPU][GlobalISel] Code quality: Combine V_RSQ

[Mateja Marjanovic via Phabricator via llvm-commits]

matejam created this revision.
matejam added reviewers: foad, arsenm, Petar.Avramovic, mbrkusanin.
matejam added a project: LLVM.
Herald added subscribers: kerbowa, hiraditya, t-tye, tpr, dstuttard, rovka, yaxunl, nhaehnle, jvesely, kzhuravl.
matejam requested review of this revision.
Herald added a subscriber: wdng.

Combine V_RCP and V_SQRT into V_RSQ on GlobalISel.
This combiner already existed but was used only when unsafe-fp-math is on, unlike SDag.


https://reviews.llvm.org/D110076

Files:
  llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
  llvm/lib/Target/AMDGPU/SIInstructions.td
  llvm/test/CodeGen/AMDGPU/GlobalISel/combine-rsq.mir


Index: llvm/test/CodeGen/AMDGPU/GlobalISel/combine-rsq.mir
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AMDGPU/GlobalISel/combine-rsq.mir
@@ -0,0 +1,29 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -global-isel -march=amdgcn -mcpu=gfx1010 -run-pass=instruction-select -verify-machineinstrs %s -o - | FileCheck %s
+
+---
+name:            test
+alignment:       1
+legalized:       true
+regBankSelected: true
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0
+
+    ; CHECK-LABEL: name: test
+    ; CHECK: liveins: $sgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[COPY]]
+    ; CHECK: %3:vgpr_32 = afn nofpexcept V_RSQ_F32_e32 [[COPY1]], implicit $mode, implicit $exec
+    ; CHECK: $vgpr0 = COPY %3
+    ; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
+    %0:sgpr(s32) = COPY $sgpr0
+    %4:vgpr(s32) = COPY %0(s32)
+    %2:vgpr(s32) = afn G_FSQRT %4
+    %3:vgpr(s32) = afn G_INTRINSIC intrinsic(@llvm.amdgcn.rcp), %2(s32)
+    $vgpr0 = COPY %3(s32)
+    SI_RETURN_TO_EPILOG implicit $vgpr0
+
+...
Index: llvm/lib/Target/AMDGPU/SIInstructions.td
===================================================================
--- llvm/lib/Target/AMDGPU/SIInstructions.td
+++ llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -825,12 +825,10 @@
 // VOP1 Patterns
 //===----------------------------------------------------------------------===//
 
-let OtherPredicates = [UnsafeFPMath] in {
-
-//defm : RsqPat<V_RSQ_F32_e32, f32>;
-
 def : RsqPat<V_RSQ_F32_e32, f32>;
 
+let OtherPredicates = [UnsafeFPMath] in {
+
 // Convert (x - floor(x)) to fract(x)
 def : GCNPat <
   (f32 (fsub (f32 (VOP3Mods f32:$x, i32:$mods)),
Index: llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -658,7 +658,7 @@
 
 class RsqPat<Instruction RsqInst, ValueType vt> : AMDGPUPat <
   (AMDGPUrcp (fsqrt vt:$src)),
-  (RsqInst $src)
+  (RsqInst vt:$src)
 >;
 
 // Instructions which select to the same v_min_f*


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D110076.373594.patch
Type: text/x-patch
Size: 2248 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210920/eecaf992/attachment.bin>