R600/SI Allow same SGPR to be used for multiple operands

Instead of moving the first SGPR that is different than the first,
legalize the operand that requires the fewest moves if one
SGPR is used for multiple operands.

This saves extra moves and is also required for some instructions
which require that the same operand be used for multiple operands.

llvm-svn: 218532

2 files changed, 128 insertions, 5 deletions

diff --git a/llvm/lib/Target/R600/SIInstrInfo.cpp b/llvm/lib/Target/R600/SIInstrInfo.cpp
index ed8d9793af7..c748228f342 100644
--- a/llvm/lib/Target/R600/SIInstrInfo.cpp
+++ b/llvm/lib/Target/R600/SIInstrInfo.cpp
@@ -1390,10 +1390,12 @@ void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
   // Legalize VOP3
   if (isVOP3(MI->getOpcode())) {
     const MCInstrDesc &Desc = get(MI->getOpcode());
-    unsigned SGPRReg = AMDGPU::NoRegister;
 
     int VOP3Idx[3] = { Src0Idx, Src1Idx, Src2Idx };
 
+    // Find the one SGPR operand we are allowed to use.
+    unsigned SGPRReg = AMDGPU::NoRegister;
+
     for (const MachineOperand &MO : MI->implicit_operands()) {
       // We only care about reads.
       if (MO.isDef())
@@ -1410,8 +1412,9 @@ void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
       }
     }
 
-
     if (SGPRReg == AMDGPU::NoRegister) {
+      unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
+
       // First we need to consider the instruction's operand requirements before
       // legalizing. Some operands are required to be SGPRs, but we are still
       // bound by the constant bus requirement to only use one.
@@ -1422,9 +1425,33 @@ void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
         if (Idx == -1)
           break;
 
-        if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass)) {
-          SGPRReg = MI->getOperand(Idx).getReg();
-          break;
+        const MachineOperand &MO = MI->getOperand(Idx);
+        if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass))
+          SGPRReg = MO.getReg();
+
+        if (MO.isReg() && RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
+          UsedSGPRs[i] = MO.getReg();
+      }
+
+      if (SGPRReg == AMDGPU::NoRegister) {
+        // We don't have a required SGPR operand, so we have a bit more freedom in
+        // selecting operands to move.
+
+        // Try to select the most used SGPR. If an SGPR is equal to one of the
+        // others, we choose that.
+        //
+        // e.g.
+        // V_FMA_F32 v0, s0, s0, s0 -> No moves
+        // V_FMA_F32 v0, s0, s1, s0 -> Move s1
+
+        if (UsedSGPRs[0] != AMDGPU::NoRegister) {
+          if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
+            SGPRReg = UsedSGPRs[0];
+        }
+
+        if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
+          if (UsedSGPRs[1] == UsedSGPRs[2])
+            SGPRReg = UsedSGPRs[1];
         }
       }
     }
diff --git a/llvm/test/CodeGen/R600/use-sgpr-multiple-times.ll b/llvm/test/CodeGen/R600/use-sgpr-multiple-times.ll
new file mode 100644
index 00000000000..5c007183354
--- /dev/null
+++ b/llvm/test/CodeGen/R600/use-sgpr-multiple-times.ll
@@ -0,0 +1,96 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+
+declare float @llvm.fma.f32(float, float, float) #1
+declare float @llvm.fmuladd.f32(float, float, float) #1
+declare i32 @llvm.AMDGPU.imad24(i32, i32, i32) #1
+
+
+; SI-LABEL: @test_sgpr_use_twice_binop:
+; SI: S_LOAD_DWORD [[SGPR:s[0-9]+]],
+; SI: V_ADD_F32_e64 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]]
+; SI: BUFFER_STORE_DWORD [[RESULT]]
+define void @test_sgpr_use_twice_binop(float addrspace(1)* %out, float %a) #0 {
+  %dbl = fadd float %a, %a
+  store float %dbl, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @test_sgpr_use_three_ternary_op:
+; SI: S_LOAD_DWORD [[SGPR:s[0-9]+]],
+; SI: V_FMA_F32 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]], [[SGPR]]
+; SI: BUFFER_STORE_DWORD [[RESULT]]
+define void @test_sgpr_use_three_ternary_op(float addrspace(1)* %out, float %a) #0 {
+  %fma = call float @llvm.fma.f32(float %a, float %a, float %a) #1
+  store float %fma, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @test_sgpr_use_twice_ternary_op_a_a_b:
+; SI: S_LOAD_DWORD [[SGPR0:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI: S_LOAD_DWORD [[SGPR1:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xc
+; SI: V_MOV_B32_e32 [[VGPR1:v[0-9]+]], [[SGPR1]]
+; SI: V_FMA_F32 [[RESULT:v[0-9]+]], [[SGPR0]], [[SGPR0]], [[VGPR1]]
+; SI: BUFFER_STORE_DWORD [[RESULT]]
+define void @test_sgpr_use_twice_ternary_op_a_a_b(float addrspace(1)* %out, float %a, float %b) #0 {
+  %fma = call float @llvm.fma.f32(float %a, float %a, float %b) #1
+  store float %fma, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @test_sgpr_use_twice_ternary_op_a_b_a:
+; SI: S_LOAD_DWORD [[SGPR0:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI: S_LOAD_DWORD [[SGPR1:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xc
+; SI: V_MOV_B32_e32 [[VGPR1:v[0-9]+]], [[SGPR1]]
+; SI: V_FMA_F32 [[RESULT:v[0-9]+]], [[SGPR0]], [[VGPR1]], [[SGPR0]]
+; SI: BUFFER_STORE_DWORD [[RESULT]]
+define void @test_sgpr_use_twice_ternary_op_a_b_a(float addrspace(1)* %out, float %a, float %b) #0 {
+  %fma = call float @llvm.fma.f32(float %a, float %b, float %a) #1
+  store float %fma, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @test_sgpr_use_twice_ternary_op_b_a_a:
+; SI: S_LOAD_DWORD [[SGPR0:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI: S_LOAD_DWORD [[SGPR1:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xc
+; SI: V_MOV_B32_e32 [[VGPR1:v[0-9]+]], [[SGPR1]]
+; SI: V_FMA_F32 [[RESULT:v[0-9]+]], [[VGPR1]], [[SGPR0]], [[SGPR0]]
+; SI: BUFFER_STORE_DWORD [[RESULT]]
+define void @test_sgpr_use_twice_ternary_op_b_a_a(float addrspace(1)* %out, float %a, float %b) #0 {
+  %fma = call float @llvm.fma.f32(float %b, float %a, float %a) #1
+  store float %fma, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @test_sgpr_use_twice_ternary_op_a_a_imm:
+; SI: S_LOAD_DWORD [[SGPR:s[0-9]+]]
+; SI: V_FMA_F32 [[RESULT:v[0-9]+]], [[SGPR]], [[SGPR]], 2.0
+; SI: BUFFER_STORE_DWORD [[RESULT]]
+define void @test_sgpr_use_twice_ternary_op_a_a_imm(float addrspace(1)* %out, float %a) #0 {
+  %fma = call float @llvm.fma.f32(float %a, float %a, float 2.0) #1
+  store float %fma, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @test_sgpr_use_twice_ternary_op_a_imm_a:
+; SI: S_LOAD_DWORD [[SGPR:s[0-9]+]]
+; SI: V_FMA_F32 [[RESULT:v[0-9]+]], [[SGPR]], 2.0, [[SGPR]]
+; SI: BUFFER_STORE_DWORD [[RESULT]]
+define void @test_sgpr_use_twice_ternary_op_a_imm_a(float addrspace(1)* %out, float %a) #0 {
+  %fma = call float @llvm.fma.f32(float %a, float 2.0, float %a) #1
+  store float %fma, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; Don't use fma since fma c, x, y is canonicalized to fma x, c, y
+; SI-LABEL: @test_sgpr_use_twice_ternary_op_imm_a_a:
+; SI: S_LOAD_DWORD [[SGPR:s[0-9]+]]
+; SI: V_MAD_I32_I24 [[RESULT:v[0-9]+]], 2, [[SGPR]], [[SGPR]]
+; SI: BUFFER_STORE_DWORD [[RESULT]]
+define void @test_sgpr_use_twice_ternary_op_imm_a_a(i32 addrspace(1)* %out, i32 %a) #0 {
+  %fma = call i32 @llvm.AMDGPU.imad24(i32 2, i32 %a, i32 %a) #1
+  store i32 %fma, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }