[AMDGPU] Disable inline constants for pseudo scalar transcendentals (l…

…lvm#104395)

Prevent operand folding from inlining constants into pseudo scalar
transcendental f16 instructions.
However still allow literal constants.

(cherry picked from commit fc6300a)

Change-Id: I5cd412741939cc812150dbb24bd2735a64573b70

llvm/lib/Target/AMDGPU/GCNSubtarget.h

@@ -1218,6 +1218,12 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   /// and STOREcnt rather than VMcnt, LGKMcnt and VScnt respectively.
   bool hasExtendedWaitCounts() const { return getGeneration() >= GFX12; }
 
+  /// \returns true if inline constants are not supported for F16 pseudo
+  /// scalar transcendentals.
+  bool hasNoF16PseudoScalarTransInlineConstants() const {
+    return getGeneration() == GFX12;
+  }
+
   /// \returns The maximum number of instructions that can be enclosed in an
   /// S_CLAUSE on the given subtarget, or 0 for targets that do not support that
   /// instruction.

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

@@ -5804,6 +5804,10 @@ bool SIInstrInfo::isOperandLegal(const MachineInstr &MI, unsigned OpIdx,
           return false;
       }
     }
+  } else if (ST.hasNoF16PseudoScalarTransInlineConstants() && !MO->isReg() &&
+             isF16PseudoScalarTrans(MI.getOpcode()) &&
+             isInlineConstant(*MO, OpInfo)) {
+    return false;
   }
 
   if (MO->isReg()) {

llvm/lib/Target/AMDGPU/SIInstrInfo.h

@@ -918,6 +918,14 @@ class SIInstrInfo final : public AMDGPUGenInstrInfo {
     return MI.getDesc().TSFlags & SIInstrFlags::IsNeverUniform;
   }
 
+  static bool isF16PseudoScalarTrans(unsigned Opcode) {
+    return Opcode == AMDGPU::V_S_EXP_F16_e64 ||
+           Opcode == AMDGPU::V_S_LOG_F16_e64 ||
+           Opcode == AMDGPU::V_S_RCP_F16_e64 ||
+           Opcode == AMDGPU::V_S_RSQ_F16_e64 ||
+           Opcode == AMDGPU::V_S_SQRT_F16_e64;
+  }
+
   static bool doesNotReadTiedSource(const MachineInstr &MI) {
     return MI.getDesc().TSFlags & SIInstrFlags::TiedSourceNotRead;
   }

llvm/test/CodeGen/AMDGPU/pseudo-scalar-transcendental.mir

@@ -0,0 +1,120 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 -run-pass=si-fold-operands -verify-machineinstrs -o - %s | FileCheck --check-prefix=GCN %s
+
+# Do not use inline constants for f16 pseudo scalar transcendentals.
+# But allow literal constants.
+
+---
+name: exp_f16_imm
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: exp_f16_imm
+    ; GCN: [[S_MOV_B32_:%[0-9]+]]:sgpr_32 = S_MOV_B32 15360
+    ; GCN-NEXT: [[V_S_EXP_F16_e64_:%[0-9]+]]:sgpr_32 = V_S_EXP_F16_e64 1, [[S_MOV_B32_]], 0, 0, implicit $mode, implicit $exec
+    %0:sgpr_32 = S_MOV_B32 15360
+    %1:sgpr_32 = V_S_EXP_F16_e64 1, %0:sgpr_32, 0, 0, implicit $mode, implicit $exec
+...
+
+---
+name: exp_f16_literal
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: exp_f16_literal
+    ; GCN: [[V_S_EXP_F16_e64_:%[0-9]+]]:sgpr_32 = V_S_EXP_F16_e64 1, 16960, 0, 0, implicit $mode, implicit $exec
+    %0:sgpr_32 = S_MOV_B32 16960
+    %1:sgpr_32 = V_S_EXP_F16_e64 1, %0:sgpr_32, 0, 0, implicit $mode, implicit $exec
+...
+
+---
+name: log_f16_imm
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: log_f16_imm
+    ; GCN: [[S_MOV_B32_:%[0-9]+]]:sgpr_32 = S_MOV_B32 15360
+    ; GCN-NEXT: [[V_S_LOG_F16_e64_:%[0-9]+]]:sgpr_32 = V_S_LOG_F16_e64 1, [[S_MOV_B32_]], 0, 0, implicit $mode, implicit $exec
+    %0:sgpr_32 = S_MOV_B32 15360
+    %1:sgpr_32 = V_S_LOG_F16_e64 1, %0:sgpr_32, 0, 0, implicit $mode, implicit $exec
+...
+
+---
+name: log_f16_literal
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: log_f16_literal
+    ; GCN: [[V_S_LOG_F16_e64_:%[0-9]+]]:sgpr_32 = V_S_LOG_F16_e64 1, 16960, 0, 0, implicit $mode, implicit $exec
+    %0:sgpr_32 = S_MOV_B32 16960
+    %1:sgpr_32 = V_S_LOG_F16_e64 1, %0:sgpr_32, 0, 0, implicit $mode, implicit $exec
+...
+
+---
+name: rcp_f16_imm
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: rcp_f16_imm
+    ; GCN: [[S_MOV_B32_:%[0-9]+]]:sgpr_32 = S_MOV_B32 15360
+    ; GCN-NEXT: [[V_S_RCP_F16_e64_:%[0-9]+]]:sgpr_32 = V_S_RCP_F16_e64 1, [[S_MOV_B32_]], 0, 0, implicit $mode, implicit $exec
+    %0:sgpr_32 = S_MOV_B32 15360
+    %1:sgpr_32 = V_S_RCP_F16_e64 1, %0:sgpr_32, 0, 0, implicit $mode, implicit $exec
+...
+
+---
+name: rcp_f16_literal
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: rcp_f16_literal
+    ; GCN: [[V_S_RCP_F16_e64_:%[0-9]+]]:sgpr_32 = V_S_RCP_F16_e64 1, 16960, 0, 0, implicit $mode, implicit $exec
+    %0:sgpr_32 = S_MOV_B32 16960
+    %1:sgpr_32 = V_S_RCP_F16_e64 1, %0:sgpr_32, 0, 0, implicit $mode, implicit $exec
+...
+
+---
+name: rsq_f16_imm
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: rsq_f16_imm
+    ; GCN: [[S_MOV_B32_:%[0-9]+]]:sgpr_32 = S_MOV_B32 15360
+    ; GCN-NEXT: [[V_S_RSQ_F16_e64_:%[0-9]+]]:sgpr_32 = V_S_RSQ_F16_e64 1, [[S_MOV_B32_]], 0, 0, implicit $mode, implicit $exec
+    %0:sgpr_32 = S_MOV_B32 15360
+    %1:sgpr_32 = V_S_RSQ_F16_e64 1, %0:sgpr_32, 0, 0, implicit $mode, implicit $exec
+...
+
+---
+name: rsq_f16_literal
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: rsq_f16_literal
+    ; GCN: [[V_S_RSQ_F16_e64_:%[0-9]+]]:sgpr_32 = V_S_RSQ_F16_e64 1, 16960, 0, 0, implicit $mode, implicit $exec
+    %0:sgpr_32 = S_MOV_B32 16960
+    %1:sgpr_32 = V_S_RSQ_F16_e64 1, %0:sgpr_32, 0, 0, implicit $mode, implicit $exec
+...
+
+---
+name: sqrt_f16_imm
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: sqrt_f16_imm
+    ; GCN: [[S_MOV_B32_:%[0-9]+]]:sgpr_32 = S_MOV_B32 15360
+    ; GCN-NEXT: [[V_S_SQRT_F16_e64_:%[0-9]+]]:sgpr_32 = V_S_SQRT_F16_e64 1, [[S_MOV_B32_]], 0, 0, implicit $mode, implicit $exec
+    %0:sgpr_32 = S_MOV_B32 15360
+    %1:sgpr_32 = V_S_SQRT_F16_e64 1, %0:sgpr_32, 0, 0, implicit $mode, implicit $exec
+...
+
+---
+name: sqrt_f16_literal
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: sqrt_f16_literal
+    ; GCN: [[V_S_SQRT_F16_e64_:%[0-9]+]]:sgpr_32 = V_S_SQRT_F16_e64 1, 16960, 0, 0, implicit $mode, implicit $exec
+    %0:sgpr_32 = S_MOV_B32 16960
+    %1:sgpr_32 = V_S_SQRT_F16_e64 1, %0:sgpr_32, 0, 0, implicit $mode, implicit $exec
+...