AMDGPU: Add a fast path for icmp.i1(src, false, NE)

Summary:
This allows moving the condition from the intrinsic to the standard ICmp
opcode, so that LLVM can do simplifications on it. The icmp.i1 intrinsic
is an identity for retrieving the SGPR mask.

And we can also get the mask from and i1, or i1, xor i1.

Reviewers: arsenm, nhaehnle

Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits

Differential Revision: https://reviews.llvm.org/D52060

5 files changed, 226 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 9e167378d09..0ba92164709 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -5355,6 +5355,11 @@ SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
                        Denominator, Numerator);
   }
   case Intrinsic::amdgcn_icmp: {
+    // There is a Pat that handles this variant, so return it as-is.
+    if (Op.getOperand(1).getValueType() == MVT::i1 &&
+        Op.getConstantOperandVal(2) == 0 &&
+        Op.getConstantOperandVal(3) == ICmpInst::Predicate::ICMP_NE)
+      return Op;
     return lowerICMPIntrinsic(*this, Op.getNode(), DAG);
   }
   case Intrinsic::amdgcn_fcmp: {
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 492f0f4d6bc..b6b00c2e425 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -583,6 +583,11 @@ def : Pat <
 
   // TODO: we could add more variants for other types of conditionals
 
+def : Pat <
+  (int_amdgcn_icmp i1:$src, (i1 0), (i32 33)),
+  (COPY $src) // Return the SGPRs representing i1 src
+>;
+
 //===----------------------------------------------------------------------===//
 // VOP1 Patterns
 //===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index dbbe32b5636..159865a01c0 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -3760,6 +3760,11 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
         // Promote to next legal integer type.
         unsigned Width = CmpType->getBitWidth();
         unsigned NewWidth = Width;
+
+        // Don't do anything for i1 comparisons.
+        if (Width == 1)
+          break;
+
         if (Width <= 16)
           NewWidth = 16;
         else if (Width <= 32)
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.icmp.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.icmp.ll
index 584fb43fd8e..06ac7da1288 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.icmp.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.icmp.ll
@@ -4,6 +4,7 @@
 declare i64 @llvm.amdgcn.icmp.i32(i32, i32, i32) #0
 declare i64 @llvm.amdgcn.icmp.i64(i64, i64, i32) #0
 declare i64 @llvm.amdgcn.icmp.i16(i16, i16, i32) #0
+declare i64 @llvm.amdgcn.icmp.i1(i1, i1, i32) #0
 
 ; No crash on invalid input
 ; GCN-LABEL: {{^}}v_icmp_i32_dynamic_cc:
@@ -314,4 +315,21 @@ define amdgpu_kernel void @v_icmp_i16_sle(i64 addrspace(1)* %out, i16 %src) {
   ret void
 }
 
+; GCN-LABEL: {{^}}v_icmp_i1_ne0:
+; GCN: v_cmp_gt_u32_e64 s[[C0:\[[0-9]+:[0-9]+\]]],
+; GCN: v_cmp_gt_u32_e64 s[[C1:\[[0-9]+:[0-9]+\]]],
+; GCN: s_and_b64 s[[SRC:\[[0-9]+:[0-9]+\]]], s[[C0]], s[[C1]]
+; SI-NEXT: s_mov_b32 s{{[0-9]+}}, -1
+; GCN-NEXT: v_mov_b32_e32
+; GCN-NEXT: v_mov_b32_e32
+; GCN-NEXT: {{global|flat|buffer}}_store_dwordx2
+define amdgpu_kernel void @v_icmp_i1_ne0(i64 addrspace(1)* %out, i32 %a, i32 %b) {
+  %c0 = icmp ugt i32 %a, 1
+  %c1 = icmp ugt i32 %b, 2
+  %src = and i1 %c0, %c1
+  %result = call i64 @llvm.amdgcn.icmp.i1(i1 %src, i1 false, i32 33)
+  store i64 %result, i64 addrspace(1)* %out
+  ret void
+}
+
 attributes #0 = { nounwind readnone convergent }
diff --git a/llvm/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll b/llvm/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
index 0a179d1f96d..b8e19e2bc94 100644
--- a/llvm/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
+++ b/llvm/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
@@ -1406,6 +1406,7 @@ define float @fmed3_0_1_undef_f32() {
 
 declare i64 @llvm.amdgcn.icmp.i32(i32, i32, i32) nounwind readnone convergent
 declare i64 @llvm.amdgcn.icmp.i64(i64, i64, i32) nounwind readnone convergent
+declare i64 @llvm.amdgcn.icmp.i1(i1, i1, i32) nounwind readnone convergent
 
 ; Make sure there's no crash for invalid input
 ; CHECK-LABEL: @invalid_nonconstant_icmp_code(
@@ -1815,6 +1816,198 @@ define i64 @fold_icmp_ne_0_zext_icmp_ult_i16(i16 %a, i16 %b) {
   ret i64 %mask
 }
 
+; 1-bit NE comparisons
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_eq_i1(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_eq_i1(i32 %a, i32 %b) {
+  %cmp = icmp eq i32 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_ne_i1(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_ne_i1(i32 %a, i32 %b) {
+  %cmp = icmp ne i32 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_sle_i1(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_sle_i1(i32 %a, i32 %b) {
+  %cmp = icmp sle i32 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_ugt_i64(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_ugt_i64(i64 %a, i64 %b) {
+  %cmp = icmp ugt i64 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_ult_swap_i64(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_ult_swap_i64(i64 %a, i64 %b) {
+  %cmp = icmp ugt i64 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 false, i1 %cmp, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_fcmp_oeq_f32(
+; CHECK-NEXT: fcmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_fcmp_oeq_f32(float %a, float %b) {
+  %cmp = fcmp oeq float %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_fcmp_une_f32(
+; CHECK-NEXT: fcmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_fcmp_une_f32(float %a, float %b) {
+  %cmp = fcmp une float %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_fcmp_olt_f64(
+; CHECK-NEXT: fcmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_fcmp_olt_f64(double %a, double %b) {
+  %cmp = fcmp olt double %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_eq_i4(
+; CHECK-NEXT: icmp
+; CHECK: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_eq_i4(i4 %a, i4 %b) {
+  %cmp = icmp eq i4 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_eq_i8(
+; CHECK-NEXT: icmp
+; CHECK: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_eq_i8(i8 %a, i8 %b) {
+  %cmp = icmp eq i8 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_eq_i16(
+; CHECK-NEXT: icmp
+; CHECK: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_eq_i16(i16 %a, i16 %b) {
+  %cmp = icmp eq i16 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_eq_i36(
+; CHECK-NEXT: icmp
+; CHECK: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_eq_i36(i36 %a, i36 %b) {
+  %cmp = icmp eq i36 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_eq_i128(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_eq_i128(i128 %a, i128 %b) {
+  %cmp = icmp eq i128 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_fcmp_oeq_f16(
+; CHECK-NEXT: fcmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_fcmp_oeq_f16(half %a, half %b) {
+  %cmp = fcmp oeq half %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_fcmp_oeq_f128(
+; CHECK-NEXT: fcmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_fcmp_oeq_f128(fp128 %a, fp128 %b) {
+;
+  %cmp = fcmp oeq fp128 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_slt_i4(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_slt_i4(i4 %a, i4 %b) {
+  %cmp = icmp slt i4 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_slt_i8(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_slt_i8(i8 %a, i8 %b) {
+  %cmp = icmp slt i8 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_slt_i16(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_slt_i16(i16 %a, i16 %b) {
+  %cmp = icmp slt i16 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_ult_i4(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_ult_i4(i4 %a, i4 %b) {
+  %cmp = icmp ult i4 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_ult_i8(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_ult_i8(i8 %a, i8 %b) {
+  %cmp = icmp ult i8 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
+; CHECK-LABEL: @fold_icmp_i1_ne_0_icmp_ult_i16(
+; CHECK-NEXT: icmp
+; CHECK-NEXT: call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+define i64 @fold_icmp_i1_ne_0_icmp_ult_i16(i16 %a, i16 %b) {
+  %cmp = icmp ult i16 %a, %b
+  %mask = call i64 @llvm.amdgcn.icmp.i1(i1 %cmp, i1 false, i32 33)
+  ret i64 %mask
+}
+
 ; --------------------------------------------------------------------
 ; llvm.amdgcn.fcmp
 ; --------------------------------------------------------------------