[AMDGPU] User SGPR count increased to 32 on gfx1250

[rampitec]

No description provided.

[rampitec]

• [AMDGPU] User SGPR count increased to 32 on gfx1250 #154205 👈 (View in Graphite)
• main

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

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Stanislav Mekhanoshin (rampitec)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/154205.diff

3 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp (+5-1)
• (modified) llvm/test/CodeGen/AMDGPU/preload-implicit-kernargs.ll (+1-3)
• (modified) llvm/test/CodeGen/AMDGPU/preload-kernargs.ll (+5-7)

diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index c41d62748c4be..42a64f0601cb8 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -2417,7 +2417,11 @@ unsigned getNSAMaxSize(const MCSubtargetInfo &STI, bool HasSampler) {
   return 0;
 }
 
-unsigned getMaxNumUserSGPRs(const MCSubtargetInfo &STI) { return 16; }
+unsigned getMaxNumUserSGPRs(const MCSubtargetInfo &STI) {
+  if (isGFX1250(STI))
+    return 32;
+  return 16;
+}
 
 bool isSI(const MCSubtargetInfo &STI) {
   return STI.hasFeature(AMDGPU::FeatureSouthernIslands);
diff --git a/llvm/test/CodeGen/AMDGPU/preload-implicit-kernargs.ll b/llvm/test/CodeGen/AMDGPU/preload-implicit-kernargs.ll
index c87f723086a41..546054cba4700 100644
--- a/llvm/test/CodeGen/AMDGPU/preload-implicit-kernargs.ll
+++ b/llvm/test/CodeGen/AMDGPU/preload-implicit-kernargs.ll
@@ -117,9 +117,7 @@ define amdgpu_kernel void @no_free_sgprs_block_count_x(ptr addrspace(1) inreg %o
 ;
 ; GFX1250-LABEL: no_free_sgprs_block_count_x:
 ; GFX1250:       ; %bb.0:
-; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x28
-; GFX1250-NEXT:    s_wait_kmcnt 0x0
-; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s0
+; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s18
 ; GFX1250-NEXT:    global_store_b32 v0, v1, s[8:9]
 ; GFX1250-NEXT:    s_endpgm
   %imp_arg_ptr = call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()
diff --git a/llvm/test/CodeGen/AMDGPU/preload-kernargs.ll b/llvm/test/CodeGen/AMDGPU/preload-kernargs.ll
index d5edfb42fa6d1..be86fd18cd737 100644
--- a/llvm/test/CodeGen/AMDGPU/preload-kernargs.ll
+++ b/llvm/test/CodeGen/AMDGPU/preload-kernargs.ll
@@ -471,13 +471,11 @@ define amdgpu_kernel void @v8i32_arg(ptr addrspace(1) nocapture inreg %out, <8 x
 ;
 ; GFX1250-LABEL: v8i32_arg:
 ; GFX1250:       ; %bb.0:
-; GFX1250-NEXT:    s_load_b256 s[4:11], s[0:1], 0x20
-; GFX1250-NEXT:    s_wait_kmcnt 0x0
-; GFX1250-NEXT:    v_dual_mov_b32 v8, 0 :: v_dual_mov_b32 v0, s8
-; GFX1250-NEXT:    v_dual_mov_b32 v1, s9 :: v_dual_mov_b32 v2, s10
-; GFX1250-NEXT:    v_dual_mov_b32 v3, s11 :: v_dual_mov_b32 v4, s4
-; GFX1250-NEXT:    v_dual_mov_b32 v5, s5 :: v_dual_mov_b32 v6, s6
-; GFX1250-NEXT:    v_mov_b32_e32 v7, s7
+; GFX1250-NEXT:    v_dual_mov_b32 v8, 0 :: v_dual_mov_b32 v0, s14
+; GFX1250-NEXT:    v_dual_mov_b32 v1, s15 :: v_dual_mov_b32 v2, s16
+; GFX1250-NEXT:    v_dual_mov_b32 v3, s17 :: v_dual_mov_b32 v4, s10
+; GFX1250-NEXT:    v_dual_mov_b32 v5, s11 :: v_dual_mov_b32 v6, s12
+; GFX1250-NEXT:    v_mov_b32_e32 v7, s13
 ; GFX1250-NEXT:    s_clause 0x1
 ; GFX1250-NEXT:    global_store_b128 v8, v[0:3], s[2:3] offset:16
 ; GFX1250-NEXT:    global_store_b128 v8, v[4:7], s[2:3]