R600/SI: Make more unordered comparisons legal

This saves a second compare and an and / or by using
the unordered comparison instructions.

llvm-svn: 224066

6 files changed, 33 insertions, 54 deletions

diff --git a/llvm/lib/Target/R600/AMDGPUInstructions.td b/llvm/lib/Target/R600/AMDGPUInstructions.td
index c746d2e1f7e..4e536c37b0b 100644
--- a/llvm/lib/Target/R600/AMDGPUInstructions.td
+++ b/llvm/lib/Target/R600/AMDGPUInstructions.td
@@ -103,7 +103,7 @@ def COND_O : PatLeaf <(cond), [{return N->get() == ISD::SETO;}]>;
 def COND_UO : PatLeaf <(cond), [{return N->get() == ISD::SETUO;}]>;
 
 //===----------------------------------------------------------------------===//
-// PatLeafs for unsigned comparisons
+// PatLeafs for unsigned / unordered comparisons
 //===----------------------------------------------------------------------===//
 
 def COND_UEQ : PatLeaf <(cond), [{return N->get() == ISD::SETUEQ;}]>;
diff --git a/llvm/lib/Target/R600/SIISelLowering.cpp b/llvm/lib/Target/R600/SIISelLowering.cpp
index 03068462290..197e368ab38 100644
--- a/llvm/lib/Target/R600/SIISelLowering.cpp
+++ b/llvm/lib/Target/R600/SIISelLowering.cpp
@@ -63,16 +63,7 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
 
   // Condition Codes
   setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
-  setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
-  setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
-  setCondCodeAction(ISD::SETULE, MVT::f32, Expand);
-  setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
-
   setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
-  setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
-  setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
-  setCondCodeAction(ISD::SETULE, MVT::f64, Expand);
-  setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
 
   setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
   setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
diff --git a/llvm/lib/Target/R600/SIInstructions.td b/llvm/lib/Target/R600/SIInstructions.td
index 83861799cc1..c12ffe78bbc 100644
--- a/llvm/lib/Target/R600/SIInstructions.td
+++ b/llvm/lib/Target/R600/SIInstructions.td
@@ -510,12 +510,12 @@ defm V_CMP_LG_F32 : VOPC_F32 <vopc<0x5, 0x45>, "v_cmp_lg_f32", COND_ONE>;
 defm V_CMP_GE_F32 : VOPC_F32 <vopc<0x6, 0x46>, "v_cmp_ge_f32", COND_OGE>;
 defm V_CMP_O_F32 : VOPC_F32 <vopc<0x7, 0x47>, "v_cmp_o_f32", COND_O>;
 defm V_CMP_U_F32 : VOPC_F32 <vopc<0x8, 0x48>, "v_cmp_u_f32", COND_UO>;
-defm V_CMP_NGE_F32 : VOPC_F32 <vopc<0x9, 0x49>, "v_cmp_nge_f32">;
+defm V_CMP_NGE_F32 : VOPC_F32 <vopc<0x9, 0x49>, "v_cmp_nge_f32",  COND_ULT>;
 defm V_CMP_NLG_F32 : VOPC_F32 <vopc<0xa, 0x4a>, "v_cmp_nlg_f32">;
-defm V_CMP_NGT_F32 : VOPC_F32 <vopc<0xb, 0x4b>, "v_cmp_ngt_f32">;
-defm V_CMP_NLE_F32 : VOPC_F32 <vopc<0xc, 0x4c>, "v_cmp_nle_f32">;
+defm V_CMP_NGT_F32 : VOPC_F32 <vopc<0xb, 0x4b>, "v_cmp_ngt_f32", COND_ULE>;
+defm V_CMP_NLE_F32 : VOPC_F32 <vopc<0xc, 0x4c>, "v_cmp_nle_f32", COND_UGT>;
 defm V_CMP_NEQ_F32 : VOPC_F32 <vopc<0xd, 0x4d>, "v_cmp_neq_f32", COND_UNE>;
-defm V_CMP_NLT_F32 : VOPC_F32 <vopc<0xe, 0x4e>, "v_cmp_nlt_f32">;
+defm V_CMP_NLT_F32 : VOPC_F32 <vopc<0xe, 0x4e>, "v_cmp_nlt_f32", COND_UGE>;
 defm V_CMP_TRU_F32 : VOPC_F32 <vopc<0xf, 0x4f>, "v_cmp_tru_f32">;
 
 let hasSideEffects = 1 in {
@@ -548,12 +548,12 @@ defm V_CMP_LG_F64 : VOPC_F64 <vopc<0x25, 0x65>, "v_cmp_lg_f64", COND_ONE>;
 defm V_CMP_GE_F64 : VOPC_F64 <vopc<0x26, 0x66>, "v_cmp_ge_f64", COND_OGE>;
 defm V_CMP_O_F64 : VOPC_F64 <vopc<0x27, 0x67>, "v_cmp_o_f64", COND_O>;
 defm V_CMP_U_F64 : VOPC_F64 <vopc<0x28, 0x68>, "v_cmp_u_f64", COND_UO>;
-defm V_CMP_NGE_F64 : VOPC_F64 <vopc<0x29, 0x69>, "v_cmp_nge_f64">;
+defm V_CMP_NGE_F64 : VOPC_F64 <vopc<0x29, 0x69>, "v_cmp_nge_f64", COND_ULT>;
 defm V_CMP_NLG_F64 : VOPC_F64 <vopc<0x2a, 0x6a>, "v_cmp_nlg_f64">;
-defm V_CMP_NGT_F64 : VOPC_F64 <vopc<0x2b, 0x6b>, "v_cmp_ngt_f64">;
-defm V_CMP_NLE_F64 : VOPC_F64 <vopc<0x2c, 0x6c>, "v_cmp_nle_f64">;
+defm V_CMP_NGT_F64 : VOPC_F64 <vopc<0x2b, 0x6b>, "v_cmp_ngt_f64", COND_ULE>;
+defm V_CMP_NLE_F64 : VOPC_F64 <vopc<0x2c, 0x6c>, "v_cmp_nle_f64", COND_UGT>;
 defm V_CMP_NEQ_F64 : VOPC_F64 <vopc<0x2d, 0x6d>, "v_cmp_neq_f64", COND_UNE>;
-defm V_CMP_NLT_F64 : VOPC_F64 <vopc<0x2e, 0x6e>, "v_cmp_nlt_f64">;
+defm V_CMP_NLT_F64 : VOPC_F64 <vopc<0x2e, 0x6e>, "v_cmp_nlt_f64", COND_UGE>;
 defm V_CMP_TRU_F64 : VOPC_F64 <vopc<0x2f, 0x6f>, "v_cmp_tru_f64">;
 
 let hasSideEffects = 1 in {
diff --git a/llvm/test/CodeGen/R600/fcmp64.ll b/llvm/test/CodeGen/R600/fcmp64.ll
index dc24443504d..73cee669c43 100644
--- a/llvm/test/CodeGen/R600/fcmp64.ll
+++ b/llvm/test/CodeGen/R600/fcmp64.ll
@@ -1,7 +1,7 @@
 ; RUN: llc < %s -march=r600 -mcpu=tahiti -verify-machineinstrs | FileCheck %s
 
 ; CHECK-LABEL: {{^}}flt_f64:
-; CHECK: v_cmp_lt_f64_e64 {{s[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+]}}
+; CHECK: v_cmp_nge_f64_e32 vcc, {{v[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+]}}
 define void @flt_f64(i32 addrspace(1)* %out, double addrspace(1)* %in1,
                      double addrspace(1)* %in2) {
    %r0 = load double addrspace(1)* %in1
@@ -13,7 +13,7 @@ define void @flt_f64(i32 addrspace(1)* %out, double addrspace(1)* %in1,
 }
 
 ; CHECK-LABEL: {{^}}fle_f64:
-; CHECK: v_cmp_le_f64_e64 {{s[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+]}}
+; CHECK: v_cmp_ngt_f64_e32 vcc, {{v[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+]}}
 define void @fle_f64(i32 addrspace(1)* %out, double addrspace(1)* %in1,
                      double addrspace(1)* %in2) {
    %r0 = load double addrspace(1)* %in1
@@ -25,7 +25,7 @@ define void @fle_f64(i32 addrspace(1)* %out, double addrspace(1)* %in1,
 }
 
 ; CHECK-LABEL: {{^}}fgt_f64:
-; CHECK: v_cmp_gt_f64_e64 {{s[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+]}}
+; CHECK: v_cmp_nle_f64_e32 vcc, {{v[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+]}}
 define void @fgt_f64(i32 addrspace(1)* %out, double addrspace(1)* %in1,
                      double addrspace(1)* %in2) {
    %r0 = load double addrspace(1)* %in1
@@ -37,7 +37,7 @@ define void @fgt_f64(i32 addrspace(1)* %out, double addrspace(1)* %in1,
 }
 
 ; CHECK-LABEL: {{^}}fge_f64:
-; CHECK: v_cmp_ge_f64_e64 {{s[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+]}}
+; CHECK: v_cmp_nlt_f64_e32 vcc, {{v[[0-9]+:[0-9]+], v[[0-9]+:[0-9]+]}}
 define void @fge_f64(i32 addrspace(1)* %out, double addrspace(1)* %in1,
                      double addrspace(1)* %in2) {
    %r0 = load double addrspace(1)* %in1
diff --git a/llvm/test/CodeGen/R600/setcc.ll b/llvm/test/CodeGen/R600/setcc.ll
index d2842d440cb..2b19fcf7868 100644
--- a/llvm/test/CodeGen/R600/setcc.ll
+++ b/llvm/test/CodeGen/R600/setcc.ll
@@ -145,10 +145,8 @@ entry:
 ; FUNC-LABEL: {{^}}f32_ugt:
 ; R600: SETGE
 ; R600: SETE_DX10
-; SI: v_cmp_u_f32
-; SI: v_cmp_gt_f32
-; SI: s_or_b64
-; SI: v_cndmask_b32
+; SI: v_cmp_nle_f32_e32 vcc
+; SI-NEXT: v_cndmask_b32_e64 {{v[0-9]+}}, 0, -1, vcc
 define void @f32_ugt(i32 addrspace(1)* %out, float %a, float %b) {
 entry:
   %0 = fcmp ugt float %a, %b
@@ -160,10 +158,9 @@ entry:
 ; FUNC-LABEL: {{^}}f32_uge:
 ; R600: SETGT
 ; R600: SETE_DX10
-; SI: v_cmp_u_f32
-; SI: v_cmp_ge_f32
-; SI: s_or_b64
-; SI: v_cndmask_b32
+
+; SI: v_cmp_nlt_f32_e32 vcc
+; SI-NEXT: v_cndmask_b32_e64 {{v[0-9]+}}, 0, -1, vcc
 define void @f32_uge(i32 addrspace(1)* %out, float %a, float %b) {
 entry:
   %0 = fcmp uge float %a, %b
@@ -175,10 +172,9 @@ entry:
 ; FUNC-LABEL: {{^}}f32_ult:
 ; R600: SETGE
 ; R600: SETE_DX10
-; SI: v_cmp_u_f32
-; SI: v_cmp_lt_f32
-; SI: s_or_b64
-; SI: v_cndmask_b32
+
+; SI: v_cmp_nge_f32_e32 vcc
+; SI-NEXT: v_cndmask_b32_e64 {{v[0-9]+}}, 0, -1, vcc
 define void @f32_ult(i32 addrspace(1)* %out, float %a, float %b) {
 entry:
   %0 = fcmp ult float %a, %b
@@ -190,10 +186,9 @@ entry:
 ; FUNC-LABEL: {{^}}f32_ule:
 ; R600: SETGT
 ; R600: SETE_DX10
-; SI: v_cmp_u_f32
-; SI: v_cmp_le_f32
-; SI: s_or_b64
-; SI: v_cndmask_b32
+
+; SI: v_cmp_ngt_f32_e32 vcc
+; SI-NEXT: v_cndmask_b32_e64 {{v[0-9]+}}, 0, -1, vcc
 define void @f32_ule(i32 addrspace(1)* %out, float %a, float %b) {
 entry:
   %0 = fcmp ule float %a, %b
diff --git a/llvm/test/CodeGen/R600/setcc64.ll b/llvm/test/CodeGen/R600/setcc64.ll
index 6a1c5ef1af0..d9e982e09b2 100644
--- a/llvm/test/CodeGen/R600/setcc64.ll
+++ b/llvm/test/CodeGen/R600/setcc64.ll
@@ -91,10 +91,9 @@ entry:
 }
 
 ; FUNC-LABEL: {{^}}f64_ugt:
-; SI: v_cmp_u_f64
-; SI: v_cmp_gt_f64
-; SI: s_or_b64
-; SI: v_cndmask_b32
+
+; SI: v_cmp_nle_f64_e32 vcc
+; SI-NEXT: v_cndmask_b32_e64 {{v[0-9]+}}, 0, -1, vcc
 define void @f64_ugt(i32 addrspace(1)* %out, double %a, double %b) {
 entry:
   %0 = fcmp ugt double %a, %b
@@ -104,10 +103,8 @@ entry:
 }
 
 ; FUNC-LABEL: {{^}}f64_uge:
-; SI: v_cmp_u_f64
-; SI: v_cmp_ge_f64
-; SI: s_or_b64
-; SI: v_cndmask_b32
+; SI: v_cmp_nlt_f64_e32 vcc
+; SI-NEXT: v_cndmask_b32_e64 {{v[0-9]+}}, 0, -1, vcc
 define void @f64_uge(i32 addrspace(1)* %out, double %a, double %b) {
 entry:
   %0 = fcmp uge double %a, %b
@@ -117,10 +114,8 @@ entry:
 }
 
 ; FUNC-LABEL: {{^}}f64_ult:
-; SI: v_cmp_u_f64
-; SI: v_cmp_lt_f64
-; SI: s_or_b64
-; SI: v_cndmask_b32
+; SI: v_cmp_nge_f64_e32 vcc
+; SI-NEXT: v_cndmask_b32_e64 {{v[0-9]+}}, 0, -1, vcc
 define void @f64_ult(i32 addrspace(1)* %out, double %a, double %b) {
 entry:
   %0 = fcmp ult double %a, %b
@@ -130,10 +125,8 @@ entry:
 }
 
 ; FUNC-LABEL: {{^}}f64_ule:
-; SI: v_cmp_u_f64
-; SI: v_cmp_le_f64
-; SI: s_or_b64
-; SI: v_cndmask_b32
+; SI: v_cmp_ngt_f64_e32 vcc
+; SI-NEXT: v_cndmask_b32_e64 {{v[0-9]+}}, 0, -1, vcc
 define void @f64_ule(i32 addrspace(1)* %out, double %a, double %b) {
 entry:
   %0 = fcmp ule double %a, %b