https://gcc.gnu.org/g:6783b7375fff65e917637e86688918e808682f14

commit r17-5901-g6783b7375fff65e917637e86688918e808682f14
Author: Sarvesh Chandra <[email protected]>
Date:   Thu Oct 29 18:21:53 2026 +0530

    i386: Add AVX512BMM bitreverse lowering [PR128372]
    
    Use vbitrevb for scalar bitreverse when AVX512BMM and AVX512VL are
    available, while retaining GFNI as the fallback without AVX512VL.
    
    The vector bitreverse optab has also been added; vectorizer support will be
    addressed separately.
    
    gcc/ChangeLog:
            PR target/128372
            * config/i386/i386-expand.cc (ix86_expand_bitreverse): Use vbitrevb 
only
            when AVX512BMM and AVX512VL are available; retain GFNI otherwise.
            * config/i386/i386-protos.h (ix86_expand_gfni_bitreverse): Rename 
to ...
            (ix86_expand_bitreverse): ... this.
            * config/i386/i386.md (bitreverse<mode>2): Require AVX512VL for BMM.
            * config/i386/sse.md (bitreverse<mode>2): New expander for 
VI1_AVX512VL.
    
    gcc/testsuite/ChangeLog:
            PR target/128372
            * gcc.target/i386/avx512bmm-bitreverse-1.c: Test BMM scalar 
lowering.
            * gcc.target/i386/gfni-bitreverse-scalar.c: Test the GFNI-only path.
    
    Co-authored-by: Venkataramanan Kumar <[email protected]>
    Signed-off-by: Sarvesh Chandra <[email protected]>

Diff:
---
 gcc/config/i386/i386-expand.cc                     | 32 +++++++++++++---------
 gcc/config/i386/i386-protos.h                      |  2 +-
 gcc/config/i386/i386.md                            |  4 +--
 gcc/config/i386/sse.md                             | 10 +++++++
 .../gcc.target/i386/avx512bmm-bitreverse-1.c       | 28 +++++++++++++++++++
 .../gcc.target/i386/gfni-bitreverse-scalar.c       | 30 ++++++++++++++++++++
 6 files changed, 90 insertions(+), 16 deletions(-)

diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
index 12ccc9a03dff..c02e7128b9db 100644
--- a/gcc/config/i386/i386-expand.cc
+++ b/gcc/config/i386/i386-expand.cc
@@ -30561,10 +30561,10 @@ ix86_expand_vector_bf2sf_with_vec_perm (rtx dest, rtx 
src)
   emit_move_insn (dest, lowpart_subreg (GET_MODE (dest), target, vperm_mode));
 }
 
-/* Implement bitreverse<mode>2 using gf2p8affineqb.  */
+/* Implement bitreverse<mode>2 using vbitrevb or gf2p8affineqb.  */
 
 void
-ix86_expand_gfni_bitreverse (rtx dest, rtx src)
+ix86_expand_bitreverse (rtx dest, rtx src)
 {
   machine_mode mode = GET_MODE (dest);
   rtx temp;
@@ -30611,18 +30611,24 @@ ix86_expand_gfni_bitreverse (rtx dest, rtx src)
     }
   src = temp;
   temp = gen_reg_rtx (V16QImode);
-  rtx src2 = gen_rtx_CONST_VECTOR (V16QImode,
-                                  gen_rtvec (16, GEN_INT (1), GEN_INT (2),
-                                             GEN_INT (4), GEN_INT (8),
-                                             GEN_INT (16), GEN_INT (32),
-                                             GEN_INT (64), GEN_INT (-128),
-                                             GEN_INT (1), GEN_INT (2),
-                                             GEN_INT (4), GEN_INT (8),
-                                             GEN_INT (16), GEN_INT (32),
-                                             GEN_INT (64), GEN_INT (-128)));
-  src2 = validize_mem (force_const_mem (V16QImode, src2));
   src = lowpart_subreg (V16QImode, src, GET_MODE (src));
-  emit_insn (gen_vgf2p8affineqb_v16qi (temp, src, src2, const0_rtx));
+  if (TARGET_AVX512BMM && TARGET_AVX512VL)
+    emit_insn (gen_avx512bmm_vbitrevb_v16qi (temp, src));
+  else if (TARGET_GFNI)
+    {
+      rtx src2 = gen_rtx_CONST_VECTOR (V16QImode,
+                                      gen_rtvec (16, GEN_INT (1), GEN_INT (2),
+                                                 GEN_INT (4), GEN_INT (8),
+                                                 GEN_INT (16), GEN_INT (32),
+                                                 GEN_INT (64), GEN_INT (-128),
+                                                 GEN_INT (1), GEN_INT (2),
+                                                 GEN_INT (4), GEN_INT (8),
+                                                 GEN_INT (16), GEN_INT (32),
+                                                 GEN_INT (64),
+                                                 GEN_INT (-128)));
+      src2 = validize_mem (force_const_mem (V16QImode, src2));
+      emit_insn (gen_vgf2p8affineqb_v16qi (temp, src, src2, const0_rtx));
+    }
   if (mode == QImode)
     {
       rtx temp1 = gen_reg_rtx (SImode);
diff --git a/gcc/config/i386/i386-protos.h b/gcc/config/i386/i386-protos.h
index 788ad443b44b..8ac329a716a8 100644
--- a/gcc/config/i386/i386-protos.h
+++ b/gcc/config/i386/i386-protos.h
@@ -267,7 +267,7 @@ extern rtx ix86_expand_ternlog (machine_mode mode, rtx op0, 
rtx op1, rtx op2,
                                int idx, rtx target);
 extern void ix86_expand_vector_sf2bf_with_vec_perm (rtx, rtx);
 extern void ix86_expand_vector_bf2sf_with_vec_perm (rtx, rtx);
-extern void ix86_expand_gfni_bitreverse (rtx, rtx);
+extern void ix86_expand_bitreverse (rtx, rtx);
 
 
 #ifdef TREE_CODE
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index f24f9185c953..b1469babab6e 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -23972,9 +23972,9 @@
 (define_expand "bitreverse<mode>2"
   [(set (match_operand:SWIDWI 0 "register_operand")
        (bitreverse:SWIDWI (match_operand:SWIDWI 1 "nonimmediate_operand")))]
-  "TARGET_GFNI && TARGET_SSE2"
+  "(TARGET_GFNI || (TARGET_AVX512BMM && TARGET_AVX512VL)) && TARGET_SSE2"
 {
-  ix86_expand_gfni_bitreverse (operands[0], operands[1]);
+  ix86_expand_bitreverse (operands[0], operands[1]);
   DONE;
 })
 
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index ea710dddc6be..fda267fc477f 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -34577,6 +34577,16 @@
   [(set_attr "prefix" "evex")
    (set_attr "mode" "<sseinsnmode>")])
 
+(define_expand "bitreverse<mode>2"
+  [(set (match_operand:VI1_AVX512VL 0 "register_operand")
+       (bitreverse:VI1_AVX512VL
+         (match_operand:VI1_AVX512VL 1 "nonimmediate_operand")))]
+  "TARGET_AVX512BMM"
+{
+  emit_insn (gen_avx512bmm_vbitrevb_<mode> (operands[0], operands[1]));
+  DONE;
+})
+
 ;; AVX10V2AUX instructions
 ;; FP32 to FP8 single-source converts (VCVTPS2BF8, VCVTPS2BF8S,
 ;; VCVTPS2HF8, VCVTPS2HF8S, VCVTROPS2HF8, VCVTROPS2HF8S)
diff --git a/gcc/testsuite/gcc.target/i386/avx512bmm-bitreverse-1.c 
b/gcc/testsuite/gcc.target/i386/avx512bmm-bitreverse-1.c
new file mode 100644
index 000000000000..a2dd451e533e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx512bmm-bitreverse-1.c
@@ -0,0 +1,28 @@
+/* { dg-do compile } */
+/* { dg-options "-mavx512bmm -mavx512vl -mno-gfni -O2" } */
+/* { dg-final { scan-assembler-times "vbitrevb" 4 } } */
+/* { dg-final { scan-assembler-not "gf2p8affineqb" } } */
+
+unsigned char
+br8 (unsigned char x)
+{
+  return __builtin_bitreverse8 (x);
+}
+
+unsigned short
+br16 (unsigned short x)
+{
+  return __builtin_bitreverse16 (x);
+}
+
+unsigned int
+br32 (unsigned int x)
+{
+  return __builtin_bitreverse32 (x);
+}
+
+unsigned long long
+br64 (unsigned long long x)
+{
+  return __builtin_bitreverse64 (x);
+}
diff --git a/gcc/testsuite/gcc.target/i386/gfni-bitreverse-scalar.c 
b/gcc/testsuite/gcc.target/i386/gfni-bitreverse-scalar.c
new file mode 100644
index 000000000000..94c032f27b71
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/gfni-bitreverse-scalar.c
@@ -0,0 +1,30 @@
+/* { dg-do compile } */
+/* { dg-options "-mgfni -mno-avx512bmm -O2" } */
+/* { dg-final { scan-assembler-times "gf2p8affineqb" 4 } } */
+/* { dg-final { scan-assembler-not "vbitrevb" } } */
+/* { dg-final { scan-assembler-not "vgf2p8affineqb" } } */
+
+/* GFNI-only scalar bitreverse must retain the existing legacy GFNI path.  */
+unsigned char
+br8 (unsigned char x)
+{
+  return __builtin_bitreverse8 (x);
+}
+
+unsigned short
+br16 (unsigned short x)
+{
+  return __builtin_bitreverse16 (x);
+}
+
+unsigned int
+br32 (unsigned int x)
+{
+  return __builtin_bitreverse32 (x);
+}
+
+unsigned long long
+br64 (unsigned long long x)
+{
+  return __builtin_bitreverse64 (x);
+}

Reply via email to