[AMDGPU] Recognize x & ((1 << y) - 1) pattern.

Summary:
As a followup for D48007.

Since we already handle `x << (bitwidth - y) >> (bitwidth - y)` pattern,
which does not have ub for both the edge cases (`y == 0`, `y == bitwidth`),
i think also handling a pattern that is ub for `y == bitwidth` should be fine.

Reviewers: nhaehnle, bogner, tstellar, arsenm

Reviewed By: arsenm

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

Tags: #amdgpu

Differential Revision: https://reviews.llvm.org/D48010

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@334816 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPUInstructions.td b/lib/Target/AMDGPU/AMDGPUInstructions.td
index 22a7a9c..65d72fc 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ b/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -126,6 +126,7 @@ def or_oneuse : HasOneUseBinOp<or>;
 def xor_oneuse : HasOneUseBinOp<xor>;
 } // Properties = [SDNPCommutative, SDNPAssociative]
 
+def add_oneuse : HasOneUseBinOp<add>;
 def sub_oneuse : HasOneUseBinOp<sub>;
 
 def srl_oneuse : HasOneUseBinOp<srl>;
@@ -682,6 +683,12 @@ multiclass BFEPattern <Instruction UBFE, Instruction SBFE, Instruction MOV> {
     (UBFE $src, $rshift, (MOV (i32 (IMMPopCount $mask))))
   >;
 
+  // x & ((1 << y) - 1)
+  def : AMDGPUPat <
+    (and i32:$src, (add_oneuse (shl_oneuse 1, i32:$width), -1)),
+    (UBFE $src, (i32 0), $width)
+  >;
+
   // x & (-1 >> (bitwidth - y))
   def : AMDGPUPat <
     (and i32:$src, (srl_oneuse -1, (sub 32, i32:$width))),

diff --git a/test/CodeGen/AMDGPU/extract-lowbits.ll b/test/CodeGen/AMDGPU/extract-lowbits.ll
index 150b064..70a3dc3 100644 (file)
--- a/test/CodeGen/AMDGPU/extract-lowbits.ll
+++ b/test/CodeGen/AMDGPU/extract-lowbits.ll
@@ -17,19 +17,11 @@
 ; ---------------------------------------------------------------------------- ;
 
 define i32 @bzhi32_a0(i32 %val, i32 %numlowbits) nounwind {
-; SI-LABEL: bzhi32_a0:
-; SI:       ; %bb.0:
-; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT:    v_bfm_b32_e64 v1, v1, 0
-; SI-NEXT:    v_and_b32_e32 v0, v1, v0
-; SI-NEXT:    s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bzhi32_a0:
-; VI:       ; %bb.0:
-; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT:    v_bfm_b32 v1, v1, 0
-; VI-NEXT:    v_and_b32_e32 v0, v1, v0
-; VI-NEXT:    s_setpc_b64 s[30:31]
+; GCN-LABEL: bzhi32_a0:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_bfe_u32 v0, v0, 0, v1
+; GCN-NEXT:    s_setpc_b64 s[30:31]
   %onebit = shl i32 1, %numlowbits
   %mask = add nsw i32 %onebit, -1
   %masked = and i32 %mask, %val
@@ -37,19 +29,11 @@ define i32 @bzhi32_a0(i32 %val, i32 %numlowbits) nounwind {
 }
 
 define i32 @bzhi32_a1_indexzext(i32 %val, i8 zeroext %numlowbits) nounwind {
-; SI-LABEL: bzhi32_a1_indexzext:
-; SI:       ; %bb.0:
-; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT:    v_bfm_b32_e64 v1, v1, 0
-; SI-NEXT:    v_and_b32_e32 v0, v1, v0
-; SI-NEXT:    s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bzhi32_a1_indexzext:
-; VI:       ; %bb.0:
-; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT:    v_bfm_b32 v1, v1, 0
-; VI-NEXT:    v_and_b32_e32 v0, v1, v0
-; VI-NEXT:    s_setpc_b64 s[30:31]
+; GCN-LABEL: bzhi32_a1_indexzext:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_bfe_u32 v0, v0, 0, v1
+; GCN-NEXT:    s_setpc_b64 s[30:31]
   %conv = zext i8 %numlowbits to i32
   %onebit = shl i32 1, %conv
   %mask = add nsw i32 %onebit, -1
@@ -58,19 +42,11 @@ define i32 @bzhi32_a1_indexzext(i32 %val, i8 zeroext %numlowbits) nounwind {
 }
 
 define i32 @bzhi32_a4_commutative(i32 %val, i32 %numlowbits) nounwind {
-; SI-LABEL: bzhi32_a4_commutative:
-; SI:       ; %bb.0:
-; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT:    v_bfm_b32_e64 v1, v1, 0
-; SI-NEXT:    v_and_b32_e32 v0, v0, v1
-; SI-NEXT:    s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bzhi32_a4_commutative:
-; VI:       ; %bb.0:
-; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT:    v_bfm_b32 v1, v1, 0
-; VI-NEXT:    v_and_b32_e32 v0, v0, v1
-; VI-NEXT:    s_setpc_b64 s[30:31]
+; GCN-LABEL: bzhi32_a4_commutative:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_bfe_u32 v0, v0, 0, v1
+; GCN-NEXT:    s_setpc_b64 s[30:31]
   %onebit = shl i32 1, %numlowbits
   %mask = add nsw i32 %onebit, -1
   %masked = and i32 %val, %mask ; swapped order