[llvm] [X86] Select BLSI for i8 operands (#202344) (PR #204746)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Jun 19 00:45:56 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-x86
Author: Harishankar.1 (Harishankar14)
<details>
<summary>Changes</summary>
Adds a tablegen pattern to select BLSI for `and (neg x), x` at i8.
Fixes #<!-- -->202344
---
Full diff: https://github.com/llvm/llvm-project/pull/204746.diff
2 Files Affected:
- (modified) llvm/lib/Target/X86/X86InstrCompiler.td (+8)
- (modified) llvm/test/CodeGen/X86/bmi.ll (+28-21)
``````````diff
diff --git a/llvm/lib/Target/X86/X86InstrCompiler.td b/llvm/lib/Target/X86/X86InstrCompiler.td
index c7ea2985d1353..f87326ecdfc4d 100644
--- a/llvm/lib/Target/X86/X86InstrCompiler.td
+++ b/llvm/lib/Target/X86/X86InstrCompiler.td
@@ -2328,3 +2328,11 @@ def : Pat<(cttz_zero_poison (loadi64 addr:$src)), (BSF64rm (i64 (IMPLICIT_DEF)),
let Predicates = [HasMOVBE] in {
def : Pat<(bswap GR16:$src), (ROL16ri GR16:$src, (i8 8))>;
}
+// BLSI {32/64-bit operands}. For an i8/i16 value, the lowest
+// set bit is unchanged by any garbage in the upper bits.
+// So Extending it using IMPLICIT_DEF,INSERT_SUBREG and extract.
+let Predicates = [HasBMI] in {
+ def : Pat<(and GR8:$src, (ineg_su GR8:$src)),
+ (EXTRACT_SUBREG (BLSI32rr (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR8:$src, sub_8bit)), sub_8bit)>;
+
+}
\ No newline at end of file
diff --git a/llvm/test/CodeGen/X86/bmi.ll b/llvm/test/CodeGen/X86/bmi.ll
index f37176d7c0a37..2c1b3dfbbbb4e 100644
--- a/llvm/test/CodeGen/X86/bmi.ll
+++ b/llvm/test/CodeGen/X86/bmi.ll
@@ -632,46 +632,53 @@ entry:
ret i32 %and
}
-define i32 @blsi32(i32 %x) {
-; X86-LABEL: blsi32:
+define i8 @blsi8_trunc(i32 %x) {
+; X86-LABEL: blsi8_trunc:
; X86: # %bb.0:
-; X86-NEXT: blsil {{[0-9]+}}(%esp), %eax
+; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax
+; X86-NEXT: blsil %eax, %eax
+; X86-NEXT: # kill: def $al killed $al killed $eax
; X86-NEXT: retl
;
-; X64-LABEL: blsi32:
+; X64-LABEL: blsi8_trunc:
; X64: # %bb.0:
; X64-NEXT: blsil %edi, %eax
+; X64-NEXT: # kill: def $al killed $al killed $eax
; X64-NEXT: retq
;
-; EGPR-LABEL: blsi32:
+; EGPR-LABEL: blsi8_trunc:
; EGPR: # %bb.0:
-; EGPR-NEXT: blsil %edi, %eax # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0xf3,0xdf]
+; EGPR-NEXT: blsil %edi, %eax # encoding: [0xc4,0xe2,0x78,0xf3,0xdf]
+; EGPR-NEXT: # kill: def $al killed $al killed $eax
; EGPR-NEXT: retq # encoding: [0xc3]
- %tmp = sub i32 0, %x
- %tmp2 = and i32 %x, %tmp
- ret i32 %tmp2
+ %t = trunc i32 %x to i8
+ %neg = sub i8 0, %t
+ %and = and i8 %t, %neg
+ ret i8 %and
}
-define i32 @blsi32_load(ptr %x) {
-; X86-LABEL: blsi32_load:
+define i16 @blsi16_trunc(i32 %x) {
+; X86-LABEL: blsi16_trunc:
; X86: # %bb.0:
-; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: blsil (%eax), %eax
+; X86-NEXT: blsil {{[0-9]+}}(%esp), %eax
+; X86-NEXT: # kill: def $ax killed $ax killed $eax
; X86-NEXT: retl
;
-; X64-LABEL: blsi32_load:
+; X64-LABEL: blsi16_trunc:
; X64: # %bb.0:
-; X64-NEXT: blsil (%rdi), %eax
+; X64-NEXT: blsil %edi, %eax
+; X64-NEXT: # kill: def $ax killed $ax killed $eax
; X64-NEXT: retq
;
-; EGPR-LABEL: blsi32_load:
+; EGPR-LABEL: blsi16_trunc:
; EGPR: # %bb.0:
-; EGPR-NEXT: blsil (%rdi), %eax # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0xf3,0x1f]
+; EGPR-NEXT: blsil %edi, %eax # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0xf3,0xdf]
+; EGPR-NEXT: # kill: def $ax killed $ax killed $eax
; EGPR-NEXT: retq # encoding: [0xc3]
- %x1 = load i32, ptr %x
- %tmp = sub i32 0, %x1
- %tmp2 = and i32 %x1, %tmp
- ret i32 %tmp2
+ %t = trunc i32 %x to i16
+ %neg = sub i16 0, %t
+ %and = and i16 %t, %neg
+ ret i16 %and
}
define i32 @blsi32_z(i32 %a, i32 %b) nounwind {
``````````
</details>
https://github.com/llvm/llvm-project/pull/204746
More information about the llvm-commits
mailing list