[llvm] 833996c - AMDGPU: Fix backwards s_cselect_* operands

[Matt Arsenault via llvm-commits]

Author: Matt Arsenault
Date: 2020-05-27T09:26:09-04:00
New Revision: 833996cef1381115b0077ab5694e189463f5d02e

URL: https://github.com/llvm/llvm-project/commit/833996cef1381115b0077ab5694e189463f5d02e
DIFF: https://github.com/llvm/llvm-project/commit/833996cef1381115b0077ab5694e189463f5d02e.diff

LOG: AMDGPU: Fix backwards s_cselect_* operands

The vector equivalent has backwards operands, but the scalar version
does not. The passes that use these hooks aren't enabled by default,
so this doesn't really change anything.

Added: 
    

Modified: 
    llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
    llvm/test/CodeGen/AMDGPU/early-if-convert.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 1d0397afd771..fd1da238a876 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -2348,14 +2348,17 @@ void SIInstrInfo::insertSelect(MachineBasicBlock &MBB,
   unsigned DstSize = RI.getRegSizeInBits(*DstRC);
 
   if (DstSize == 32) {
-    unsigned SelOp = Pred == SCC_TRUE ?
-      AMDGPU::S_CSELECT_B32 : AMDGPU::V_CNDMASK_B32_e32;
-
-    // Instruction's operands are backwards from what is expected.
-    MachineInstr *Select =
-      BuildMI(MBB, I, DL, get(SelOp), DstReg)
-      .addReg(FalseReg)
-      .addReg(TrueReg);
+    MachineInstr *Select;
+    if (Pred == SCC_TRUE) {
+      Select = BuildMI(MBB, I, DL, get(AMDGPU::S_CSELECT_B32), DstReg)
+        .addReg(TrueReg)
+        .addReg(FalseReg);
+    } else {
+      // Instruction's operands are backwards from what is expected.
+      Select = BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e32), DstReg)
+        .addReg(FalseReg)
+        .addReg(TrueReg);
+    }
 
     preserveCondRegFlags(Select->getOperand(3), Cond[1]);
     return;
@@ -2364,8 +2367,8 @@ void SIInstrInfo::insertSelect(MachineBasicBlock &MBB,
   if (DstSize == 64 && Pred == SCC_TRUE) {
     MachineInstr *Select =
       BuildMI(MBB, I, DL, get(AMDGPU::S_CSELECT_B64), DstReg)
-      .addReg(FalseReg)
-      .addReg(TrueReg);
+      .addReg(TrueReg)
+      .addReg(FalseReg);
 
     preserveCondRegFlags(Select->getOperand(3), Cond[1]);
     return;
@@ -2416,10 +2419,19 @@ void SIInstrInfo::insertSelect(MachineBasicBlock &MBB,
 
     unsigned SubIdx = SubIndices[Idx];
 
-    MachineInstr *Select =
-      BuildMI(MBB, I, DL, get(SelOp), DstElt)
-      .addReg(FalseReg, 0, SubIdx)
-      .addReg(TrueReg, 0, SubIdx);
+    MachineInstr *Select;
+    if (SelOp == AMDGPU::V_CNDMASK_B32_e32) {
+      Select =
+        BuildMI(MBB, I, DL, get(SelOp), DstElt)
+        .addReg(FalseReg, 0, SubIdx)
+        .addReg(TrueReg, 0, SubIdx);
+    } else {
+      Select =
+        BuildMI(MBB, I, DL, get(SelOp), DstElt)
+        .addReg(TrueReg, 0, SubIdx)
+        .addReg(FalseReg, 0, SubIdx);
+    }
+
     preserveCondRegFlags(Select->getOperand(3), Cond[1]);
     fixImplicitOperands(*Select);
 

diff  --git a/llvm/test/CodeGen/AMDGPU/early-if-convert.ll b/llvm/test/CodeGen/AMDGPU/early-if-convert.ll
index 66d5411cd978..41ed70bff160 100644
--- a/llvm/test/CodeGen/AMDGPU/early-if-convert.ll
+++ b/llvm/test/CodeGen/AMDGPU/early-if-convert.ll
@@ -246,7 +246,7 @@ endif:
 ; GCN: s_load_dword [[VAL:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x0
 ; GCN: s_add_i32 [[ADD:s[0-9]+]], [[VAL]], [[VAL]]
 ; GCN: s_cmp_lg_u32 s{{[0-9]+}}, 1
-; GCN-NEXT: s_cselect_b32 [[SELECT:s[0-9]+]], [[ADD]], [[VAL]]
+; GCN-NEXT: s_cselect_b32 [[SELECT:s[0-9]+]], [[VAL]], [[ADD]]
 define amdgpu_kernel void @test_scc1_sgpr_ifcvt_triangle(i32 addrspace(4)* %in, i32 %cond) #0 {
 entry:
   %v = load i32, i32 addrspace(4)* %in
@@ -362,7 +362,7 @@ endif:
 
 ; GCN-LABEL: {{^}}uniform_if_swap_br_targets_scc_constant_select:
 ; GCN: s_cmp_lg_u32 s{{[0-9]+}}, 0
-; GCN: s_cselect_b32 s{{[0-9]+}}, 1, 0{{$}}
+; GCN: s_cselect_b32 s{{[0-9]+}}, 0, 1{{$}}
 define amdgpu_kernel void @uniform_if_swap_br_targets_scc_constant_select(i32 %cond, i32 addrspace(1)* %out) {
 entry:
   %cmp0 = icmp eq i32 %cond, 0
@@ -383,7 +383,7 @@ done:
 ; GCN-LABEL: {{^}}ifcvt_undef_scc:
 ; GCN: {{^}}; %bb.0:
 ; GCN-NEXT: s_load_dwordx2
-; GCN-NEXT: s_cselect_b32 s{{[0-9]+}}, 1, 0
+; GCN-NEXT: s_cselect_b32 s{{[0-9]+}}, 0, 1{{$}}
 define amdgpu_kernel void @ifcvt_undef_scc(i32 %cond, i32 addrspace(1)* %out) {
 entry:
   br i1 undef, label %else, label %if