[llvm-branch-commits] [llvm] [AMDGPU] Test precommit for subreg reload (PR #175001)

[via llvm-branch-commits]

https://github.com/easyonaadit updated https://github.com/llvm/llvm-project/pull/175001

>From 788b0d3f495ced424b64966dcebf982051ac1fc1 Mon Sep 17 00:00:00 2001
From: Christudasan Devadasan <Christudasan.Devadasan at amd.com>
Date: Thu, 8 Jan 2026 06:23:19 +0000
Subject: [PATCH 1/2] [AMDGPU] Test precommit for subreg reload

This test currently fails due to insufficient
registers during allocation. Once the subreg
reload is implemented, it will begin to pass
as the partial reload help mitigate register
pressure.
---
 ...pressure-mitigation-with-subreg-reload.mir | 37 +++++++++++++++++++
 1 file changed, 37 insertions(+)
 create mode 100644 llvm/test/CodeGen/AMDGPU/regpressure-mitigation-with-subreg-reload.mir

diff --git a/llvm/test/CodeGen/AMDGPU/regpressure-mitigation-with-subreg-reload.mir b/llvm/test/CodeGen/AMDGPU/regpressure-mitigation-with-subreg-reload.mir
new file mode 100644
index 0000000000000..43289230c587d
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/regpressure-mitigation-with-subreg-reload.mir
@@ -0,0 +1,37 @@
+# RUN: not llc -mcpu=gfx1100 -mtriple=amdgcn-amd-amdhsa -stress-regalloc=4 -run-pass=greedy -filetype=null %s 2>&1 | FileCheck %s
+
+# This test demonstrates register pressure mitigation through subreg-aware
+# reloads during register allocation. Currently, the test would fail during
+# RA due to insufficient registers. Subreg-aware reload, when implemented,
+# mitigates register pressure by reloading only the used portions of tuple
+# registers from their spilled locations instead of reloading entire tuples,
+# thereby enhancing register allocation.
+
+# CHECK: error: <unknown>:0:0: ran out of registers during register allocation
+
+---
+name:            subreg-reload
+tracksRegLiveness: true
+machineFunctionInfo:
+  frameOffsetReg:  '$sgpr33'
+  stackPtrOffsetReg: '$sgpr32'
+body:             |
+  bb.0.entry:
+    liveins: $vgpr1, $sgpr0_sgpr1, $sgpr2_sgpr3
+
+    %8:vgpr_32 = COPY $vgpr1
+    %12:vreg_64 = COPY killed renamable $sgpr0_sgpr1
+    %11:vreg_128 = FLAT_LOAD_DWORDX4 %12, 0, 0, implicit $exec, implicit $flat_scr
+    undef early-clobber %49.sub0_sub1:vreg_128, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %11.sub0, 42, 0, 0, implicit $exec
+    undef %50.sub0:vreg_64 = COPY %49.sub1
+    early-clobber %49.sub1_sub2:vreg_128, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %11.sub1, 42, %50, 0, implicit $exec
+    early-clobber %15:vreg_64, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %11.sub2, 42, 0, 0, implicit $exec
+    undef %52.sub0:vreg_64 = COPY %15.sub1
+    early-clobber %27:vreg_64, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %11.sub3, 42, %52, 0, implicit $exec
+    %49.sub2:vreg_128 = COPY %15.sub0
+    %49.sub3:vreg_128 = COPY %27.sub0
+    $vgpr31 = COPY %8
+    INLINEASM &"; use v1", 1, 327690, $vgpr1
+    FLAT_STORE_DWORDX4 %12, %49, 0, 0, implicit $exec, implicit $flat_scr
+    SI_RETURN
+...

>From 841b269094897d289c9ef2e335c489489c969036 Mon Sep 17 00:00:00 2001
From: Christudasan Devadasan <Christudasan.Devadasan at amd.com>
Date: Thu, 8 Jan 2026 17:30:02 +0000
Subject: [PATCH 2/2] compacted the virt-reg numbers

---
 ...pressure-mitigation-with-subreg-reload.mir | 28 +++++++++----------
 1 file changed, 14 insertions(+), 14 deletions(-)

diff --git a/llvm/test/CodeGen/AMDGPU/regpressure-mitigation-with-subreg-reload.mir b/llvm/test/CodeGen/AMDGPU/regpressure-mitigation-with-subreg-reload.mir
index 43289230c587d..a0993d7e35176 100644
--- a/llvm/test/CodeGen/AMDGPU/regpressure-mitigation-with-subreg-reload.mir
+++ b/llvm/test/CodeGen/AMDGPU/regpressure-mitigation-with-subreg-reload.mir
@@ -16,22 +16,22 @@ machineFunctionInfo:
   frameOffsetReg:  '$sgpr33'
   stackPtrOffsetReg: '$sgpr32'
 body:             |
-  bb.0.entry:
+  bb.0:
     liveins: $vgpr1, $sgpr0_sgpr1, $sgpr2_sgpr3
 
-    %8:vgpr_32 = COPY $vgpr1
-    %12:vreg_64 = COPY killed renamable $sgpr0_sgpr1
-    %11:vreg_128 = FLAT_LOAD_DWORDX4 %12, 0, 0, implicit $exec, implicit $flat_scr
-    undef early-clobber %49.sub0_sub1:vreg_128, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %11.sub0, 42, 0, 0, implicit $exec
-    undef %50.sub0:vreg_64 = COPY %49.sub1
-    early-clobber %49.sub1_sub2:vreg_128, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %11.sub1, 42, %50, 0, implicit $exec
-    early-clobber %15:vreg_64, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %11.sub2, 42, 0, 0, implicit $exec
-    undef %52.sub0:vreg_64 = COPY %15.sub1
-    early-clobber %27:vreg_64, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %11.sub3, 42, %52, 0, implicit $exec
-    %49.sub2:vreg_128 = COPY %15.sub0
-    %49.sub3:vreg_128 = COPY %27.sub0
-    $vgpr31 = COPY %8
+    %1:vgpr_32 = COPY $vgpr1
+    %2:vreg_64 = COPY killed renamable $sgpr0_sgpr1
+    %3:vreg_128 = FLAT_LOAD_DWORDX4 %2, 0, 0, implicit $exec, implicit $flat_scr
+    undef early-clobber %4.sub0_sub1:vreg_128, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %3.sub0, 42, 0, 0, implicit $exec
+    undef %5.sub0:vreg_64 = COPY %4.sub1
+    early-clobber %4.sub1_sub2:vreg_128, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %3.sub1, 42, %5, 0, implicit $exec
+    early-clobber %6:vreg_64, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %3.sub2, 42, 0, 0, implicit $exec
+    undef %7.sub0:vreg_64 = COPY %6.sub1
+    early-clobber %8:vreg_64, $sgpr_null = V_MAD_U64_U32_gfx11_e64 %3.sub3, 42, %7, 0, implicit $exec
+    %4.sub2:vreg_128 = COPY %6.sub0
+    %4.sub3:vreg_128 = COPY %8.sub0
+    $vgpr31 = COPY %1
     INLINEASM &"; use v1", 1, 327690, $vgpr1
-    FLAT_STORE_DWORDX4 %12, %49, 0, 0, implicit $exec, implicit $flat_scr
+    FLAT_STORE_DWORDX4 %2, %4, 0, 0, implicit $exec, implicit $flat_scr
     SI_RETURN
 ...