[PATCH][X86] Improved tablegen patters for matching TZCNT/LZCNT.
Andrea Di Biagio
andrea.dibiagio at gmail.com
Mon Dec 8 09:36:15 PST 2014
Hi Nadav, Quentin (and all).
This patch teaches ISel how to match a TZCNT/LZCNT from a conditional
move if the condition code is equal to X86_COND_NE.
Existing tablegen patterns only allowed to match a TZCNT/LZCNT from a
X86cond with condition code equal to X86_COND_E. To avoid introducing
extra rules, I added an 'ImmLeaf' definition to check if the condition
code is COND_E or COND_NE.
Example:
///
unsigned foo(unsigned A) {
unsigned bar = __builtin_clz(A);
return !A ? 32 : bar;
}
///
Before, (on targets that feature LZCNT), we generated the following
sub-optimal assembly:
lzcntl %edi, %ecx
movl $32, %eax
cmovael %ecx, %eax
retq
Now we simply generate:
lzcntl %edi, %eax
retq
Please let me know if ok to submit.
Thanks,
Andrea
-------------- next part --------------
Index: lib/Target/X86/X86InstrInfo.td
===================================================================
--- lib/Target/X86/X86InstrInfo.td (revision 223666)
+++ lib/Target/X86/X86InstrInfo.td (working copy)
@@ -803,6 +803,11 @@
def X86_COND_P : PatLeaf<(i8 14)>; // alt. COND_PE
def X86_COND_S : PatLeaf<(i8 15)>;
+// Predicate used to help when pattern matching LZCNT/TZCNT.
+def X86_COND_E_OR_NE : ImmLeaf<i8, [{
+ return (Imm == X86::COND_E) || (Imm == X86::COND_NE);
+}]>;
+
let FastIselShouldIgnore = 1 in { // FastIsel should ignore all simm8 instrs.
def i16immSExt8 : ImmLeaf<i16, [{ return Imm == (int8_t)Imm; }]>;
def i32immSExt8 : ImmLeaf<i32, [{ return Imm == (int8_t)Imm; }]>;
@@ -1981,41 +1986,41 @@
}
let Predicates = [HasLZCNT] in {
- def : Pat<(X86cmov (ctlz GR16:$src), (i16 16), (X86_COND_E),
+ def : Pat<(X86cmov (ctlz GR16:$src), (i16 16), (X86_COND_E_OR_NE),
(X86cmp GR16:$src, (i16 0))),
(LZCNT16rr GR16:$src)>;
- def : Pat<(X86cmov (ctlz GR32:$src), (i32 32), (X86_COND_E),
+ def : Pat<(X86cmov (ctlz GR32:$src), (i32 32), (X86_COND_E_OR_NE),
(X86cmp GR32:$src, (i32 0))),
(LZCNT32rr GR32:$src)>;
- def : Pat<(X86cmov (ctlz GR64:$src), (i64 64), (X86_COND_E),
+ def : Pat<(X86cmov (ctlz GR64:$src), (i64 64), (X86_COND_E_OR_NE),
(X86cmp GR64:$src, (i64 0))),
(LZCNT64rr GR64:$src)>;
- def : Pat<(X86cmov (i16 16), (ctlz GR16:$src), (X86_COND_E),
+ def : Pat<(X86cmov (i16 16), (ctlz GR16:$src), (X86_COND_E_OR_NE),
(X86cmp GR16:$src, (i16 0))),
(LZCNT16rr GR16:$src)>;
- def : Pat<(X86cmov (i32 32), (ctlz GR32:$src), (X86_COND_E),
+ def : Pat<(X86cmov (i32 32), (ctlz GR32:$src), (X86_COND_E_OR_NE),
(X86cmp GR32:$src, (i32 0))),
(LZCNT32rr GR32:$src)>;
- def : Pat<(X86cmov (i64 64), (ctlz GR64:$src), (X86_COND_E),
+ def : Pat<(X86cmov (i64 64), (ctlz GR64:$src), (X86_COND_E_OR_NE),
(X86cmp GR64:$src, (i64 0))),
(LZCNT64rr GR64:$src)>;
- def : Pat<(X86cmov (ctlz (loadi16 addr:$src)), (i16 16), (X86_COND_E),
+ def : Pat<(X86cmov (ctlz (loadi16 addr:$src)), (i16 16), (X86_COND_E_OR_NE),
(X86cmp (loadi16 addr:$src), (i16 0))),
(LZCNT16rm addr:$src)>;
- def : Pat<(X86cmov (ctlz (loadi32 addr:$src)), (i32 32), (X86_COND_E),
+ def : Pat<(X86cmov (ctlz (loadi32 addr:$src)), (i32 32), (X86_COND_E_OR_NE),
(X86cmp (loadi32 addr:$src), (i32 0))),
(LZCNT32rm addr:$src)>;
- def : Pat<(X86cmov (ctlz (loadi64 addr:$src)), (i64 64), (X86_COND_E),
+ def : Pat<(X86cmov (ctlz (loadi64 addr:$src)), (i64 64), (X86_COND_E_OR_NE),
(X86cmp (loadi64 addr:$src), (i64 0))),
(LZCNT64rm addr:$src)>;
- def : Pat<(X86cmov (i16 16), (ctlz (loadi16 addr:$src)), (X86_COND_E),
+ def : Pat<(X86cmov (i16 16), (ctlz (loadi16 addr:$src)), (X86_COND_E_OR_NE),
(X86cmp (loadi16 addr:$src), (i16 0))),
(LZCNT16rm addr:$src)>;
- def : Pat<(X86cmov (i32 32), (ctlz (loadi32 addr:$src)), (X86_COND_E),
+ def : Pat<(X86cmov (i32 32), (ctlz (loadi32 addr:$src)), (X86_COND_E_OR_NE),
(X86cmp (loadi32 addr:$src), (i32 0))),
(LZCNT32rm addr:$src)>;
- def : Pat<(X86cmov (i64 64), (ctlz (loadi64 addr:$src)), (X86_COND_E),
+ def : Pat<(X86cmov (i64 64), (ctlz (loadi64 addr:$src)), (X86_COND_E_OR_NE),
(X86cmp (loadi64 addr:$src), (i64 0))),
(LZCNT64rm addr:$src)>;
}
@@ -2097,41 +2102,41 @@
}
let Predicates = [HasBMI] in {
- def : Pat<(X86cmov (cttz GR16:$src), (i16 16), (X86_COND_E),
+ def : Pat<(X86cmov (cttz GR16:$src), (i16 16), (X86_COND_E_OR_NE),
(X86cmp GR16:$src, (i16 0))),
(TZCNT16rr GR16:$src)>;
- def : Pat<(X86cmov (cttz GR32:$src), (i32 32), (X86_COND_E),
+ def : Pat<(X86cmov (cttz GR32:$src), (i32 32), (X86_COND_E_OR_NE),
(X86cmp GR32:$src, (i32 0))),
(TZCNT32rr GR32:$src)>;
- def : Pat<(X86cmov (cttz GR64:$src), (i64 64), (X86_COND_E),
+ def : Pat<(X86cmov (cttz GR64:$src), (i64 64), (X86_COND_E_OR_NE),
(X86cmp GR64:$src, (i64 0))),
(TZCNT64rr GR64:$src)>;
- def : Pat<(X86cmov (i16 16), (cttz GR16:$src), (X86_COND_E),
+ def : Pat<(X86cmov (i16 16), (cttz GR16:$src), (X86_COND_E_OR_NE),
(X86cmp GR16:$src, (i16 0))),
(TZCNT16rr GR16:$src)>;
- def : Pat<(X86cmov (i32 32), (cttz GR32:$src), (X86_COND_E),
+ def : Pat<(X86cmov (i32 32), (cttz GR32:$src), (X86_COND_E_OR_NE),
(X86cmp GR32:$src, (i32 0))),
(TZCNT32rr GR32:$src)>;
- def : Pat<(X86cmov (i64 64), (cttz GR64:$src), (X86_COND_E),
+ def : Pat<(X86cmov (i64 64), (cttz GR64:$src), (X86_COND_E_OR_NE),
(X86cmp GR64:$src, (i64 0))),
(TZCNT64rr GR64:$src)>;
- def : Pat<(X86cmov (cttz (loadi16 addr:$src)), (i16 16), (X86_COND_E),
+ def : Pat<(X86cmov (cttz (loadi16 addr:$src)), (i16 16), (X86_COND_E_OR_NE),
(X86cmp (loadi16 addr:$src), (i16 0))),
(TZCNT16rm addr:$src)>;
- def : Pat<(X86cmov (cttz (loadi32 addr:$src)), (i32 32), (X86_COND_E),
+ def : Pat<(X86cmov (cttz (loadi32 addr:$src)), (i32 32), (X86_COND_E_OR_NE),
(X86cmp (loadi32 addr:$src), (i32 0))),
(TZCNT32rm addr:$src)>;
- def : Pat<(X86cmov (cttz (loadi64 addr:$src)), (i64 64), (X86_COND_E),
+ def : Pat<(X86cmov (cttz (loadi64 addr:$src)), (i64 64), (X86_COND_E_OR_NE),
(X86cmp (loadi64 addr:$src), (i64 0))),
(TZCNT64rm addr:$src)>;
- def : Pat<(X86cmov (i16 16), (cttz (loadi16 addr:$src)), (X86_COND_E),
+ def : Pat<(X86cmov (i16 16), (cttz (loadi16 addr:$src)), (X86_COND_E_OR_NE),
(X86cmp (loadi16 addr:$src), (i16 0))),
(TZCNT16rm addr:$src)>;
- def : Pat<(X86cmov (i32 32), (cttz (loadi32 addr:$src)), (X86_COND_E),
+ def : Pat<(X86cmov (i32 32), (cttz (loadi32 addr:$src)), (X86_COND_E_OR_NE),
(X86cmp (loadi32 addr:$src), (i32 0))),
(TZCNT32rm addr:$src)>;
- def : Pat<(X86cmov (i64 64), (cttz (loadi64 addr:$src)), (X86_COND_E),
+ def : Pat<(X86cmov (i64 64), (cttz (loadi64 addr:$src)), (X86_COND_E_OR_NE),
(X86cmp (loadi64 addr:$src), (i64 0))),
(TZCNT64rm addr:$src)>;
}
Index: test/CodeGen/X86/lzcnt-tzcnt.ll
===================================================================
--- test/CodeGen/X86/lzcnt-tzcnt.ll (revision 223666)
+++ test/CodeGen/X86/lzcnt-tzcnt.ll (working copy)
@@ -437,7 +437,138 @@
; CHECK: tzcnt
; CHECK-NEXT: ret
+define i16 @test1b_ctlz(i16 %v) {
+ %cnt = tail call i16 @llvm.ctlz.i16(i16 %v, i1 true)
+ %tobool = icmp ne i16 %v, 0
+ %cond = select i1 %tobool, i16 16, i16 %cnt
+ ret i16 %cond
+}
+; CHECK-LABEL: test1b_ctlz
+; CHECK: lzcnt
+; CHECK-NEXT: ret
+
+define i32 @test2b_ctlz(i32 %v) {
+ %cnt = tail call i32 @llvm.ctlz.i32(i32 %v, i1 true)
+ %tobool = icmp ne i32 %v, 0
+ %cond = select i1 %tobool, i32 32, i32 %cnt
+ ret i32 %cond
+}
+; CHECK-LABEL: test2b_ctlz
+; CHECK: lzcnt
+; CHECK-NEXT: ret
+
+
+define i64 @test3b_ctlz(i64 %v) {
+ %cnt = tail call i64 @llvm.ctlz.i64(i64 %v, i1 true)
+ %tobool = icmp ne i64 %v, 0
+ %cond = select i1 %tobool, i64 64, i64 %cnt
+ ret i64 %cond
+}
+; CHECK-LABEL: test3b_ctlz
+; CHECK: lzcnt
+; CHECK-NEXT: ret
+
+
+define i16 @test4b_ctlz(i16 %v) {
+ %cnt = tail call i16 @llvm.ctlz.i16(i16 %v, i1 true)
+ %tobool = icmp ne i16 %v, 0
+ %cond = select i1 %tobool, i16 %cnt, i16 16
+ ret i16 %cond
+}
+; CHECK-LABEL: test4b_ctlz
+; CHECK: lzcnt
+; CHECK-NEXT: ret
+
+
+define i32 @test5b_ctlz(i32 %v) {
+ %cnt = tail call i32 @llvm.ctlz.i32(i32 %v, i1 true)
+ %tobool = icmp ne i32 %v, 0
+ %cond = select i1 %tobool, i32 %cnt, i32 32
+ ret i32 %cond
+}
+; CHECK-LABEL: test5b_ctlz
+; CHECK: lzcnt
+; CHECK-NEXT: ret
+
+
+define i64 @test6b_ctlz(i64 %v) {
+ %cnt = tail call i64 @llvm.ctlz.i64(i64 %v, i1 true)
+ %tobool = icmp ne i64 %v, 0
+ %cond = select i1 %tobool, i64 %cnt, i64 64
+ ret i64 %cond
+}
+; CHECK-LABEL: test6b_ctlz
+; CHECK: lzcnt
+; CHECK-NEXT: ret
+
+
+define i16 @test1b_cttz(i16 %v) {
+ %cnt = tail call i16 @llvm.cttz.i16(i16 %v, i1 true)
+ %tobool = icmp ne i16 %v, 0
+ %cond = select i1 %tobool, i16 16, i16 %cnt
+ ret i16 %cond
+}
+; CHECK-LABEL: test1b_cttz
+; CHECK: tzcnt
+; CHECK-NEXT: ret
+
+
+define i32 @test2b_cttz(i32 %v) {
+ %cnt = tail call i32 @llvm.cttz.i32(i32 %v, i1 true)
+ %tobool = icmp ne i32 %v, 0
+ %cond = select i1 %tobool, i32 32, i32 %cnt
+ ret i32 %cond
+}
+; CHECK-LABEL: test2b_cttz
+; CHECK: tzcnt
+; CHECK-NEXT: ret
+
+
+define i64 @test3b_cttz(i64 %v) {
+ %cnt = tail call i64 @llvm.cttz.i64(i64 %v, i1 true)
+ %tobool = icmp ne i64 %v, 0
+ %cond = select i1 %tobool, i64 64, i64 %cnt
+ ret i64 %cond
+}
+; CHECK-LABEL: test3b_cttz
+; CHECK: tzcnt
+; CHECK-NEXT: ret
+
+
+define i16 @test4b_cttz(i16 %v) {
+ %cnt = tail call i16 @llvm.cttz.i16(i16 %v, i1 true)
+ %tobool = icmp ne i16 %v, 0
+ %cond = select i1 %tobool, i16 %cnt, i16 16
+ ret i16 %cond
+}
+; CHECK-LABEL: test4b_cttz
+; CHECK: tzcnt
+; CHECK-NEXT: ret
+
+
+define i32 @test5b_cttz(i32 %v) {
+ %cnt = tail call i32 @llvm.cttz.i32(i32 %v, i1 true)
+ %tobool = icmp ne i32 %v, 0
+ %cond = select i1 %tobool, i32 %cnt, i32 32
+ ret i32 %cond
+}
+; CHECK-LABEL: test5b_cttz
+; CHECK: tzcnt
+; CHECK-NEXT: ret
+
+
+define i64 @test6b_cttz(i64 %v) {
+ %cnt = tail call i64 @llvm.cttz.i64(i64 %v, i1 true)
+ %tobool = icmp ne i64 %v, 0
+ %cond = select i1 %tobool, i64 %cnt, i64 64
+ ret i64 %cond
+}
+; CHECK-LABEL: test6b_cttz
+; CHECK: tzcnt
+; CHECK-NEXT: ret
+
+
declare i64 @llvm.cttz.i64(i64, i1)
declare i32 @llvm.cttz.i32(i32, i1)
declare i16 @llvm.cttz.i16(i16, i1)
More information about the llvm-commits
mailing list