[llvm-branch-commits] [llvm] release/19.x: [AMDGPU] Fix folding clamp into pseudo scalar instructions (#100568) (PR #102446)

[via llvm-branch-commits]

https://github.com/llvmbot created https://github.com/llvm/llvm-project/pull/102446

Backport 817cd726454f01e990cd84e5e1d339b120b5ebaa

Requested by: @mbrkusanin

>From ea8163f942bad0a251c2c7eca57f79c00e13b909 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Mirko=20Brku=C5=A1anin?= <Mirko.Brkusanin at amd.com>
Date: Thu, 25 Jul 2024 18:19:26 +0200
Subject: [PATCH] [AMDGPU] Fix folding clamp into pseudo scalar instructions
 (#100568)

Clamp is canonically a v_max* instruction with a VGPR dst. Folding clamp
into a pseudo scalar instruction can cause issues due to a change in
regbank. We fix this with a copy.

(cherry picked from commit 817cd726454f01e990cd84e5e1d339b120b5ebaa)
---
 llvm/lib/Target/AMDGPU/SIFoldOperands.cpp     | 13 +++++++++-
 .../CodeGen/AMDGPU/si-fold-scalar-clamp.mir   | 26 +++++++++++++++++++
 2 files changed, 38 insertions(+), 1 deletion(-)
 create mode 100644 llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir

diff --git a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
index 9b2cab2eb73a39..32ecf350db59cf 100644
--- a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -1581,7 +1581,18 @@ bool SIFoldOperands::tryFoldClamp(MachineInstr &MI) {
 
   // Clamp is applied after omod, so it is OK if omod is set.
   DefClamp->setImm(1);
-  MRI->replaceRegWith(MI.getOperand(0).getReg(), Def->getOperand(0).getReg());
+
+  Register DefReg = Def->getOperand(0).getReg();
+  Register MIDstReg = MI.getOperand(0).getReg();
+  if (TRI->isSGPRReg(*MRI, DefReg)) {
+    // Pseudo scalar instructions have a SGPR for dst and clamp is a v_max*
+    // instruction with a VGPR dst.
+    BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), TII->get(AMDGPU::COPY),
+            MIDstReg)
+        .addReg(DefReg);
+  } else {
+    MRI->replaceRegWith(MIDstReg, DefReg);
+  }
   MI.eraseFromParent();
 
   // Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac
diff --git a/llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir b/llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir
new file mode 100644
index 00000000000000..1f4d046a8739fa
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir
@@ -0,0 +1,26 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -run-pass=si-fold-operands -verify-machineinstrs -o - %s | FileCheck %s
+---
+name:            test
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0
+
+    ; CHECK-LABEL: name: test
+    ; CHECK: liveins: $sgpr0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+    ; CHECK-NEXT: [[V_S_RSQ_F32_e64_:%[0-9]+]]:sgpr_32 = nofpexcept V_S_RSQ_F32_e64 0, [[COPY]], 1, 0, implicit $mode, implicit $exec
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[V_S_RSQ_F32_e64_]]
+    ; CHECK-NEXT: [[V_ADD_F32_e32_:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e32 [[COPY1]], [[COPY1]], implicit $mode, implicit $exec
+    ; CHECK-NEXT: $vgpr0 = COPY [[V_ADD_F32_e32_]]
+    ; CHECK-NEXT: S_ENDPGM 0
+    %0:sgpr_32 = COPY $sgpr0
+    %1:sgpr_32 = nofpexcept V_S_RSQ_F32_e64 0, %0, 0, 0, implicit $mode, implicit $exec
+    %2:vgpr_32 = nofpexcept V_MAX_F32_e64 0, %1, 0, %1, -1, 0, implicit $mode, implicit $exec
+    %3:vgpr_32 = nofpexcept V_ADD_F32_e32 %2:vgpr_32, %2:vgpr_32, implicit $mode, implicit $exec
+    $vgpr0 = COPY %3
+    S_ENDPGM 0
+
+...