[llvm] GlobalISel needs fdiv 1 / sqrt(x) to rsq combine (PR #78673)

Nick Anderson via llvm-commits llvm-commits at lists.llvm.org
Thu Feb 8 23:53:03 PST 2024


================
@@ -0,0 +1,419 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 4
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -run-pass=amdgpu-postlegalizer-combiner -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s
+
+---
+name:            rsq_f16
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0
+
+    ; GCN-LABEL: name: rsq_f16
+    ; GCN: liveins: $vgpr0
+    ; GCN-NEXT: {{  $}}
+    ; GCN-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; GCN-NEXT: %x:_(s16) = G_TRUNC [[COPY]](s32)
+    ; GCN-NEXT: [[INT:%[0-9]+]]:_(s16) = contract G_INTRINSIC intrinsic(@llvm.amdgcn.rsq), %x(s16)
+    ; GCN-NEXT: %ext:_(s32) = G_ANYEXT [[INT]](s16)
+    ; GCN-NEXT: $vgpr0 = COPY %ext(s32)
+    %0:_(s32) = COPY $vgpr0
+    %x:_(s16) = G_TRUNC %0:_(s32)
+    %sqrt:_(s16) = contract G_FSQRT %x
+    %one:_(s16) = G_FCONSTANT half 1.0
+    %rsq:_(s16) = contract G_FDIV %one, %sqrt
+    %ext:_(s32) = G_ANYEXT %rsq:_(s16)
+    $vgpr0 = COPY %ext
+
+...
+
----------------
nickleus27 wrote:

> Probably doesn't really matter, eventually the real testing should be the end to end IR tests shared with the DAG

What is the end to end IR tests shared with the DAG?

https://github.com/llvm/llvm-project/pull/78673


More information about the llvm-commits mailing list