AMDGPU: Add some leaf intrinsics to isAlwaysUniform

[arsenm]

These would always be uniform anyway, but it shouldn't hurt to
mark them as always uniform. This will help use TTI::isAlwaysUniform
in place of proper uniformity analysis in trivial situations.

[arsenm]

• AMDGPU: Add some leaf intrinsics to isAlwaysUniform #101925 👈
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

Join @arsenm and the rest of your teammates on Graphite

[llvmbot]

@llvm/pr-subscribers-llvm-analysis

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

Changes

These would always be uniform anyway, but it shouldn't hurt to
mark them as always uniform. This will help use TTI::isAlwaysUniform
in place of proper uniformity analysis in trivial situations.

---

Full diff: https://github.com/llvm/llvm-project/pull/101925.diff

2 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td (+4)
• (modified) llvm/test/Analysis/UniformityAnalysis/AMDGPU/always_uniform.ll (+33)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td b/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td
index a323f63767737..98c93b1c6801f 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td
@@ -378,3 +378,7 @@ def : AlwaysUniform<int_amdgcn_icmp>;
 def : AlwaysUniform<int_amdgcn_fcmp>;
 def : AlwaysUniform<int_amdgcn_ballot>;
 def : AlwaysUniform<int_amdgcn_if_break>;
+def : AlwaysUniform<int_amdgcn_s_getpc>;
+def : AlwaysUniform<int_amdgcn_s_getreg>;
+def : AlwaysUniform<int_amdgcn_s_memrealtime>;
+def : AlwaysUniform<int_amdgcn_s_memtime>;
diff --git a/llvm/test/Analysis/UniformityAnalysis/AMDGPU/always_uniform.ll b/llvm/test/Analysis/UniformityAnalysis/AMDGPU/always_uniform.ll
index d7d8d29fbfd42..285bd04a93bbc 100644
--- a/llvm/test/Analysis/UniformityAnalysis/AMDGPU/always_uniform.ll
+++ b/llvm/test/Analysis/UniformityAnalysis/AMDGPU/always_uniform.ll
@@ -79,6 +79,39 @@ define void @no_divergent_args_if_inreg(i32 inreg %i32, i1 inreg %i1) {
  ret void
 }
 
+; CHECK-LABEL: for function 's_getpc':
+; CHECK: ALL VALUES UNIFORM
+define void @s_getpc(ptr addrspace(1) inreg %out) {
+  %result = call i64 @llvm.amdgcn.s.getpc()
+  store i64 %result, ptr addrspace(1) %out, align 8
+  ret void
+}
+
+; CHECK-LABEL: for function 's_getreg':
+; CHECK: ALL VALUES UNIFORM
+define void @s_getreg(ptr addrspace(1) inreg %out) {
+  %result = call i32 @llvm.amdgcn.s.getreg(i32 123)
+  store i32 %result, ptr addrspace(1) %out, align 4
+  ret void
+}
+
+; CHECK-LABEL: for function 's_memtime':
+; CHECK: ALL VALUES UNIFORM
+define void @s_memtime(ptr addrspace(1) inreg %out) {
+  %result = call i64 @llvm.amdgcn.s.memtime()
+  store i64 %result, ptr addrspace(1) %out, align 8
+  ret void
+}
+
+; CHECK-LABEL: for function 's_memrealtime':
+; CHECK: ALL VALUES UNIFORM
+define void @s_memrealtime(ptr addrspace(1) inreg %out) {
+  %result = call i64 @llvm.amdgcn.s.memrealtime()
+  store i64 %result, ptr addrspace(1) %out, align 8
+  ret void
+}
+
+
 declare i32 @llvm.amdgcn.workitem.id.x() #0
 declare i32 @llvm.amdgcn.readfirstlane(i32) #0
 declare i64 @llvm.amdgcn.icmp.i32(i32, i32, i32) #1

[ssahasra]

These would always be uniform anyway, but it shouldn't hurt to mark them as always uniform. This will help use TTI::isAlwaysUniform in place of proper uniformity analysis in trivial situations.

I don't think that so much justification is required. Things that are known to be uniform should be marked as such, liberally.

[rampitec]

I think uniformity analysis will always tell it is uniform because it has no divergent arguments. Do you think searching this table is faster?

[arsenm]

I think uniformity analysis will always tell it is uniform because it has no divergent arguments. Do you think searching this table is faster?

The table is only used once to initialize special overriding map entries. The point of this is we can use it for trivial checks without using full uniformity analysis