[PATCH] D95102: AMDGPU: Remove v_rsq_f64 patterns

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added reviewers: rampitec, b-sumner.
Herald added subscribers: kerbowa, hiraditya, t-tye, tpr, dstuttard, yaxunl, nhaehnle, jvesely, kzhuravl.
arsenm requested review of this revision.
Herald added a subscriber: wdng.
Herald added a project: LLVM.

This isn't accurate enough without correction


https://reviews.llvm.org/D95102

Files:
  llvm/lib/Target/AMDGPU/SIInstructions.td
  llvm/test/CodeGen/AMDGPU/llvm.amdgcn.rcp.ll
  llvm/test/CodeGen/AMDGPU/rsq.ll


Index: llvm/test/CodeGen/AMDGPU/rsq.ll
===================================================================
--- llvm/test/CodeGen/AMDGPU/rsq.ll
+++ llvm/test/CodeGen/AMDGPU/rsq.ll
@@ -17,8 +17,8 @@
 }
 
 ; SI-LABEL: {{^}}rsq_f64:
-; SI-UNSAFE: v_rsq_f64_e32
-; SI-SAFE: v_sqrt_f64_e32
+; SI: v_sqrt_f64
+; SI: v_rcp_f64
 ; SI: s_endpgm
 define amdgpu_kernel void @rsq_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #0 {
   %val = load double, double addrspace(1)* %in, align 4
@@ -97,8 +97,8 @@
 
 ; SI-UNSAFE: buffer_load_dwordx2 [[VAL:v\[[0-9]+:[0-9]+\]]]
 ; SI-UNSAFE: v_sqrt_f64_e32 [[SQRT:v\[[0-9]+:[0-9]+\]]], [[VAL]]
-; SI-UNSAFE: v_rsq_f64_e32 [[RSQ:v\[[0-9]+:[0-9]+\]]], [[VAL]]
-; SI-UNSAFE: v_fma_f64 {{v\[[0-9]+:[0-9]+\]}}, -{{v\[[0-9]+:[0-9]+\]}}, [[RSQ]], 1.0
+; SI-UNSAFE: v_rcp_f64_e32 [[RCP:v\[[0-9]+:[0-9]+\]]], [[VAL]]
+; SI-UNSAFE: v_fma_f64 {{v\[[0-9]+:[0-9]+\]}}, -{{v\[[0-9]+:[0-9]+\]}}, [[RCP]], 1.0
 ; SI-UNSAFE: v_fma_f64
 ; SI-UNSAFE: v_fma_f64
 ; SI-UNSAFE: v_fma_f64
@@ -135,8 +135,7 @@
 
 ; SI-UNSAFE: buffer_load_dwordx2 [[VAL:v\[[0-9]+:[0-9]+\]]]
 ; SI-UNSAFE-DAG: v_sqrt_f64_e64 [[SQRT:v\[[0-9]+:[0-9]+\]]], -[[VAL]]
-; SI-UNSAFE-DAG: v_xor_b32_e32 v[[HI:[0-9]+]], 0x80000000, v{{[0-9]+}}
-; SI-UNSAFE: v_rsq_f64_e32 [[RSQ:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+}}:[[HI]]{{\]}}
+; SI-UNSAFE: v_rcp_f64_e32 [[RSQ:v\[[0-9]+:[0-9]+\]]], [[SQRT]]
 ; SI-UNSAFE: v_fma_f64 {{v\[[0-9]+:[0-9]+\]}}, -{{v\[[0-9]+:[0-9]+\]}}, [[RSQ]], 1.0
 ; SI-UNSAFE: v_fma_f64
 ; SI-UNSAFE: v_fma_f64
Index: llvm/test/CodeGen/AMDGPU/llvm.amdgcn.rcp.ll
===================================================================
--- llvm/test/CodeGen/AMDGPU/llvm.amdgcn.rcp.ll
+++ llvm/test/CodeGen/AMDGPU/llvm.amdgcn.rcp.ll
@@ -1,4 +1,4 @@
-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=SI -check-prefix=FUNC %s
 
 declare float @llvm.amdgcn.rcp.f32(float) #0
 declare double @llvm.amdgcn.rcp.f64(double) #0
@@ -132,8 +132,8 @@
 }
 
 ; FUNC-LABEL: {{^}}unsafe_rsq_rcp_pat_f64:
-; SI: v_rsq_f64_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}}
-; SI-NOT: [[RESULT]]
+; SI: v_sqrt_f64_e32 [[SQRT:v\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}}
+; SI: v_rcp_f64_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[SQRT]]
 ; SI: buffer_store_dwordx2 [[RESULT]]
 define amdgpu_kernel void @unsafe_rsq_rcp_pat_f64(double addrspace(1)* %out, double %src) #2 {
   %sqrt = call double @llvm.sqrt.f64(double %src)
Index: llvm/lib/Target/AMDGPU/SIInstructions.td
===================================================================
--- llvm/lib/Target/AMDGPU/SIInstructions.td
+++ llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -786,11 +786,9 @@
 
 let OtherPredicates = [UnsafeFPMath] in {
 
-//defm : RsqPat<V_RSQ_F64_e32, f64>;
 //defm : RsqPat<V_RSQ_F32_e32, f32>;
 
 def : RsqPat<V_RSQ_F32_e32, f32>;
-def : RsqPat<V_RSQ_F64_e32, f64>;
 
 // Convert (x - floor(x)) to fract(x)
 def : GCNPat <


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D95102.318036.patch
Type: text/x-patch
Size: 3048 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210120/7e56c22c/attachment.bin>