[llvm] [AMDGPU] Fix folding clamp into pseudo scalar instructions (PR #100568)

[Mirko Brkušanin via llvm-commits]

https://github.com/mbrkusanin updated https://github.com/llvm/llvm-project/pull/100568

>From 578adb36870da0e980990f64db760250445d4eb2 Mon Sep 17 00:00:00 2001
From: Mirko Brkusanin <Mirko.Brkusanin at amd.com>
Date: Thu, 25 Jul 2024 15:47:55 +0200
Subject: [PATCH 1/3] [AMDGPU] Fix SIFoldOperands when folding clamp into
 pseudo scalar instructions

Clamp is canonicaly a v_max* instruction with a VGPR dst. Folding clamp into
a pseudo scalar instruction can cause issue due to a change in regbank. We
fix this with a copy.
---
 llvm/lib/Target/AMDGPU/SIFoldOperands.cpp     | 13 +++++++++-
 .../CodeGen/AMDGPU/si-fold-scalar-clamp.mir   | 24 +++++++++++++++++++
 2 files changed, 36 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 9b2cab2eb73a3..f11c4dc46e27c 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)) {
+    // Psuedo 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 0000000000000..928479534dcdd
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir
@@ -0,0 +1,24 @@
+# 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: EXP_DONE 0, killed [[COPY1]], [[COPY1]], [[COPY1]], [[COPY1]], -1, 0, 15, implicit $exec
+    ; 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
+    EXP_DONE 0, killed %2, %2, %2, %2, -1, 0, 15, implicit $exec
+    S_ENDPGM 0
+
+...

>From ed083f0bbfe5756b38a09d2e64133ce9a1cf62a2 Mon Sep 17 00:00:00 2001
From: Mirko Brkusanin <Mirko.Brkusanin at amd.com>
Date: Thu, 25 Jul 2024 16:06:34 +0200
Subject: [PATCH 2/3] fix typo

---
 llvm/lib/Target/AMDGPU/SIFoldOperands.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
index f11c4dc46e27c..32ecf350db59c 100644
--- a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -1585,7 +1585,7 @@ bool SIFoldOperands::tryFoldClamp(MachineInstr &MI) {
   Register DefReg = Def->getOperand(0).getReg();
   Register MIDstReg = MI.getOperand(0).getReg();
   if (TRI->isSGPRReg(*MRI, DefReg)) {
-    // Psuedo scalar instructions have a SGPR for dst and clamp is a v_max*
+    // 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)

>From e206cce7893026b33e4a959368390b2a039a2ac6 Mon Sep 17 00:00:00 2001
From: Mirko Brkusanin <Mirko.Brkusanin at amd.com>
Date: Thu, 25 Jul 2024 17:34:19 +0200
Subject: [PATCH 3/3] test update

---
 llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir b/llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir
index 928479534dcdd..1f4d046a8739f 100644
--- a/llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir
+++ b/llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir
@@ -13,12 +13,14 @@ body:             |
     ; 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: EXP_DONE 0, killed [[COPY1]], [[COPY1]], [[COPY1]], [[COPY1]], -1, 0, 15, implicit $exec
+    ; 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
-    EXP_DONE 0, killed %2, %2, %2, %2, -1, 0, 15, 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
 
 ...