[llvm] r203281 - R600/SI: Using SGPRs is illegal for instructions that read carry-out from VCC

[Tom Stellard]

Author: tstellar
Date: Fri Mar  7 14:12:39 2014
New Revision: 203281

URL: http://llvm.org/viewvc/llvm-project?rev=203281&view=rev
Log:
R600/SI: Using SGPRs is illegal for instructions that read carry-out from VCC

Reviewed-by: Michel Dänzer <michel.daenzer at amd.com>

Modified:
    llvm/trunk/lib/Target/R600/SIInstrInfo.td
    llvm/trunk/lib/Target/R600/SIInstructions.td
    llvm/trunk/test/CodeGen/R600/add.ll

Modified: llvm/trunk/lib/Target/R600/SIInstrInfo.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/SIInstrInfo.td?rev=203281&r1=203280&r2=203281&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/SIInstrInfo.td (original)
+++ llvm/trunk/lib/Target/R600/SIInstrInfo.td Fri Mar  7 14:12:39 2014
@@ -298,10 +298,10 @@ multiclass VOP2_64 <bits<6> op, string o
   : VOP2_Helper <op, VReg_64, VSrc_64, opName, pattern, revOp>;
 
 multiclass VOP2b_32 <bits<6> op, string opName, list<dag> pattern,
-                     string revOp = opName> {
+                     RegisterClass src0_rc, string revOp = opName> {
 
   def _e32 : VOP2 <
-    op, (outs VReg_32:$dst), (ins VSrc_32:$src0, VReg_32:$src1),
+    op, (outs VReg_32:$dst), (ins src0_rc:$src0, VReg_32:$src1),
     opName#"_e32 $dst, $src0, $src1", pattern
   >, VOP <opName>, VOP2_REV<revOp#"_e32", !eq(revOp, opName)>;
 

Modified: llvm/trunk/lib/Target/R600/SIInstructions.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/SIInstructions.td?rev=203281&r1=203280&r2=203281&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/SIInstructions.td (original)
+++ llvm/trunk/lib/Target/R600/SIInstructions.td Fri Mar  7 14:12:39 2014
@@ -1007,14 +1007,16 @@ defm V_MBCNT_HI_U32_B32 : VOP2_32 <0x000
 let isCommutable = 1, Defs = [VCC] in { // Carry-out goes to VCC
 // No patterns so that the scalar instructions are always selected.
 // The scalar versions will be replaced with vector when needed later.
-defm V_ADD_I32 : VOP2b_32 <0x00000025, "V_ADD_I32", []>;
-defm V_SUB_I32 : VOP2b_32 <0x00000026, "V_SUB_I32", []>;
-defm V_SUBREV_I32 : VOP2b_32 <0x00000027, "V_SUBREV_I32", [], "V_SUB_I32">;
+defm V_ADD_I32 : VOP2b_32 <0x00000025, "V_ADD_I32", [], VSrc_32>;
+defm V_SUB_I32 : VOP2b_32 <0x00000026, "V_SUB_I32", [], VSrc_32>;
+defm V_SUBREV_I32 : VOP2b_32 <0x00000027, "V_SUBREV_I32", [], VSrc_32,
+                              "V_SUB_I32">;
 
 let Uses = [VCC] in { // Carry-in comes from VCC
-defm V_ADDC_U32 : VOP2b_32 <0x00000028, "V_ADDC_U32", []>;
-defm V_SUBB_U32 : VOP2b_32 <0x00000029, "V_SUBB_U32", []>;
-defm V_SUBBREV_U32 : VOP2b_32 <0x0000002a, "V_SUBBREV_U32", [], "V_SUBB_U32">;
+defm V_ADDC_U32 : VOP2b_32 <0x00000028, "V_ADDC_U32", [], VReg_32>;
+defm V_SUBB_U32 : VOP2b_32 <0x00000029, "V_SUBB_U32", [], VReg_32>;
+defm V_SUBBREV_U32 : VOP2b_32 <0x0000002a, "V_SUBBREV_U32", [], VReg_32,
+                               "V_SUBB_U32">;
 } // End Uses = [VCC]
 } // End isCommutable = 1, Defs = [VCC]
 

Modified: llvm/trunk/test/CodeGen/R600/add.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/add.ll?rev=203281&r1=203280&r2=203281&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/R600/add.ll (original)
+++ llvm/trunk/test/CodeGen/R600/add.ll Fri Mar  7 14:12:39 2014
@@ -125,3 +125,18 @@ entry:
   store i64 %0, i64 addrspace(1)* %out
   ret void
 }
+
+; The V_ADDC_U32 and V_ADD_I32 instruction can't read SGPRs, because they
+; use VCC.  The test is designed so that %a will be stored in an SGPR and
+; %0 will be stored in a VGPR, so the comiler will be forced to copy %a
+; to a VGPR before doing the add.
+
+; FUNC-LABEL: @add64_sgpr_vgpr
+; SI-CHECK-NOT: V_ADDC_U32_e32 s
+define void @add64_sgpr_vgpr(i64 addrspace(1)* %out, i64 %a, i64 addrspace(1)* %in) {
+entry:
+  %0 = load i64 addrspace(1)* %in
+  %1 = add i64 %a, %0
+  store i64 %1, i64 addrspace(1)* %out
+  ret void
+}