diff mbox series

[v2] Break false dependence for vpternlog by inserting vpxor or setting constraint of input operand to '0'

Message ID 20230711040130.3090884-1-hongtao.liu@intel.com
State New
Headers show
Series [v2] Break false dependence for vpternlog by inserting vpxor or setting constraint of input operand to '0' | expand

Commit Message

Liu, Hongtao July 11, 2023, 4:01 a.m. UTC
Here's updated patch.
1. use optimize_insn_for_speed_p instead of using optimize_function_for_speed_p.
2. explicitly move memory to dest register to avoid false dependence in one_cmpl pattern.


False dependency happens when destination is only updated by
pternlog. There is no false dependency when destination is also used
in source. So either a pxor should be inserted, or input operand
should be set with constraint '0'.

gcc/ChangeLog:

	PR target/110438
	PR target/110202
	* config/i386/predicates.md
	(int_float_vector_all_ones_operand): New predicate.
	* config/i386/sse.md (*vmov<mode>_constm1_pternlog_false_dep): New
	define_insn.
	(*<avx512>_cvtmask2<ssemodesuffix><mode>_pternlog_false_dep):
	Ditto.
	(*<avx512>_cvtmask2<ssemodesuffix><mode>_pternlog_false_dep):
	Ditto.
	(*<avx512>_cvtmask2<ssemodesuffix><mode>): Adjust to
	define_insn_and_split to avoid false dependence.
	(*<avx512>_cvtmask2<ssemodesuffix><mode>): Ditto.
	(<mask_codefor>one_cmpl<mode>2<mask_name>): Adjust constraint
	of operands 1 to '0' to avoid false dependence.
	(*andnot<mode>3): Ditto.
	(iornot<mode>3): Ditto.
	(*<nlogic><mode>3): Ditto.

gcc/testsuite/ChangeLog:

	* gcc.target/i386/pr110438.c: New test.
	* gcc.target/i386/pr100711.c: Adjust testcase.
---
 gcc/config/i386/predicates.md              |   8 +-
 gcc/config/i386/sse.md                     | 145 ++++++++++++++++++---
 gcc/testsuite/gcc.target/i386/pr100711-6.c |   2 +-
 gcc/testsuite/gcc.target/i386/pr110438.c   |  30 +++++
 4 files changed, 168 insertions(+), 17 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/i386/pr110438.c
diff mbox series

Patch

diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md
index 7ddbe01a6f9..37d20c6303a 100644
--- a/gcc/config/i386/predicates.md
+++ b/gcc/config/i386/predicates.md
@@ -1192,12 +1192,18 @@  (define_predicate "float_vector_all_ones_operand"
     return false;
 })
 
-/* Return true if operand is a vector constant that is all ones. */
+/* Return true if operand is an integral vector constant that is all ones. */
 (define_predicate "vector_all_ones_operand"
   (and (match_code "const_vector")
        (match_test "INTEGRAL_MODE_P (GET_MODE (op))")
        (match_test "op == CONSTM1_RTX (GET_MODE (op))")))
 
+/* Return true if operand is a vector constant that is all ones. */
+(define_predicate "int_float_vector_all_ones_operand"
+  (ior (match_operand 0 "vector_all_ones_operand")
+       (match_operand 0 "float_vector_all_ones_operand")
+       (match_test "op == constm1_rtx")))
+
 /* Return true if operand is an 128/256bit all ones vector
    that zero-extends to 256/512bit.  */
 (define_predicate "vector_all_ones_zero_extend_half_operand"
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 418c337a775..05485b1792d 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -1382,6 +1382,29 @@  (define_insn "mov<mode>_internal"
 	      ]
 	      (symbol_ref "true")))])
 
+; False dependency happens on destination register which is not really
+; used when moving all ones to vector register
+(define_split
+  [(set (match_operand:VMOVE 0 "register_operand")
+	(match_operand:VMOVE 1 "int_float_vector_all_ones_operand"))]
+  "TARGET_AVX512F && reload_completed
+  && (<MODE_SIZE> == 64 || EXT_REX_SSE_REG_P (operands[0]))
+  && optimize_insn_for_speed_p ()"
+  [(set (match_dup 0) (match_dup 2))
+   (parallel
+     [(set (match_dup 0) (match_dup 1))
+      (unspec [(match_dup 0)] UNSPEC_INSN_FALSE_DEP)])]
+  "operands[2] = CONST0_RTX (<MODE>mode);")
+
+(define_insn "*vmov<mode>_constm1_pternlog_false_dep"
+  [(set (match_operand:VMOVE 0 "register_operand" "=v")
+	(match_operand:VMOVE 1 "int_float_vector_all_ones_operand" "<sseconstm1>"))
+   (unspec [(match_operand:VMOVE 2 "register_operand" "0")] UNSPEC_INSN_FALSE_DEP)]
+   "TARGET_AVX512VL || <MODE_SIZE> == 64"
+   "vpternlogd\t{$0xFF, %0, %0, %0|%0, %0, %0, 0xFF}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix" "evex")])
+
 ;; If mem_addr points to a memory region with less than whole vector size bytes
 ;; of accessible memory and k is a mask that would prevent reading the inaccessible
 ;; bytes from mem_addr, add UNSPEC_MASKLOAD to prevent it to be transformed to vpblendd
@@ -9336,7 +9359,7 @@  (define_expand "<avx512>_cvtmask2<ssemodesuffix><mode>"
     operands[3] = CONST0_RTX (<MODE>mode);
   }")
 
-(define_insn "*<avx512>_cvtmask2<ssemodesuffix><mode>"
+(define_insn_and_split "*<avx512>_cvtmask2<ssemodesuffix><mode>"
   [(set (match_operand:VI48_AVX512VL 0 "register_operand" "=v,v")
 	(vec_merge:VI48_AVX512VL
 	  (match_operand:VI48_AVX512VL 2 "vector_all_ones_operand")
@@ -9346,11 +9369,35 @@  (define_insn "*<avx512>_cvtmask2<ssemodesuffix><mode>"
   "@
    vpmovm2<ssemodesuffix>\t{%1, %0|%0, %1}
    vpternlog<ssemodesuffix>\t{$0x81, %0, %0, %0%{%1%}%{z%}|%0%{%1%}%{z%}, %0, %0, 0x81}"
+  "&& !TARGET_AVX512DQ && reload_completed
+   && optimize_function_for_speed_p (cfun)"
+  [(set (match_dup 0) (match_dup 4))
+   (parallel
+    [(set (match_dup 0)
+	  (vec_merge:VI48_AVX512VL
+	    (match_dup 2)
+	    (match_dup 3)
+	    (match_dup 1)))
+     (unspec [(match_dup 0)] UNSPEC_INSN_FALSE_DEP)])]
+  "operands[4] = CONST0_RTX (<MODE>mode);"
   [(set_attr "isa" "avx512dq,*")
    (set_attr "length_immediate" "0,1")
    (set_attr "prefix" "evex")
    (set_attr "mode" "<sseinsnmode>")])
 
+(define_insn "*<avx512>_cvtmask2<ssemodesuffix><mode>_pternlog_false_dep"
+  [(set (match_operand:VI48_AVX512VL 0 "register_operand" "=v")
+	(vec_merge:VI48_AVX512VL
+	  (match_operand:VI48_AVX512VL 2 "vector_all_ones_operand")
+	  (match_operand:VI48_AVX512VL 3 "const0_operand")
+	  (match_operand:<avx512fmaskmode> 1 "register_operand" "Yk")))
+   (unspec [(match_operand:VI48_AVX512VL 4 "register_operand" "0")] UNSPEC_INSN_FALSE_DEP)]
+  "TARGET_AVX512F && !TARGET_AVX512DQ"
+  "vpternlog<ssemodesuffix>\t{$0x81, %0, %0, %0%{%1%}%{z%}|%0%{%1%}%{z%}, %0, %0, 0x81}"
+  [(set_attr "length_immediate" "1")
+   (set_attr "prefix" "evex")
+   (set_attr "mode" "<sseinsnmode>")])
+
 (define_expand "extendv2sfv2df2"
   [(set (match_operand:V2DF 0 "register_operand")
 	(float_extend:V2DF
@@ -17166,14 +17213,82 @@  (define_expand "one_cmpl<mode>2"
     operands[2] = force_reg (<MODE>mode, operands[2]);
 })
 
-(define_insn "<mask_codefor>one_cmpl<mode>2<mask_name>"
-  [(set (match_operand:VI 0 "register_operand" "=v,v")
-	(xor:VI (match_operand:VI 1 "bcst_vector_operand" "vBr,m")
-		(match_operand:VI 2 "vector_all_ones_operand" "BC,BC")))]
+(define_insn_and_split "<mask_codefor>one_cmpl<mode>2<mask_name>"
+  [(set (match_operand:VI 0 "register_operand" "=v,v,v")
+	(xor:VI (match_operand:VI 1 "bcst_vector_operand"     " 0, m,Br")
+		(match_operand:VI 2 "vector_all_ones_operand" "BC,BC,BC")))]
   "TARGET_AVX512F
    && (!<mask_applied>
        || <ssescalarmode>mode == SImode
        || <ssescalarmode>mode == DImode)"
+{
+  if (!<mask_applied> && which_alternative
+      && optimize_insn_for_speed_p ())
+    return "#";
+
+  if (TARGET_AVX512VL)
+    return "vpternlog<ternlogsuffix>\t{$0x55, %1, %0, %0<mask_operand3>|%0<mask_operand3>, %0, %1, 0x55}";
+  else
+    return "vpternlog<ternlogsuffix>\t{$0x55, %g1, %g0, %g0<mask_operand3>|%g0<mask_operand3>, %g0, %g1, 0x55}";
+}
+  "&& reload_completed && !REG_P (operands[1]) && !<mask_applied>
+   && optimize_insn_for_speed_p ()"
+  [(set (match_dup 0) (match_dup 3))
+   (parallel
+     [(set (match_dup 0)
+	   (xor:VI (match_dup 1) (match_dup 2)))
+      (unspec [(match_dup 0)] UNSPEC_INSN_FALSE_DEP)])]
+{
+  if (MEM_P (operands[1]))
+   {
+     operands[3] = operands[1];
+     operands[1] = operands[0];
+   }
+  else
+    {
+      if (GET_MODE_SIZE (<ssescalarmode>mode) < 4)
+	{
+	  if (<MODE_SIZE> == 64 ? TARGET_AVX512BW
+	     : (TARGET_AVX512BW && TARGET_AVX512VL)
+	       || !EXT_REX_SSE_REG_P (operands[0]))
+	    {
+	      operands[3] = operands[1];
+	      operands[1] = operands[0];
+	    }
+	  else
+	    operands[3] = CONST0_RTX (<MODE>mode);
+	}
+       else
+	 {
+	    if (<MODE_SIZE> == 64 || TARGET_AVX512VL
+		|| !EXT_REX_SSE_REG_P (operands[0]))
+	      {
+		operands[3] = operands[1];
+		operands[1] = operands[0];
+	      }
+	    else
+	      operands[3] = CONST0_RTX (<MODE>mode);
+	 }
+    }
+}
+  [(set_attr "type" "sselog")
+   (set_attr "prefix" "evex")
+   (set (attr "mode")
+        (if_then_else (match_test "TARGET_AVX512VL")
+		      (const_string "<sseinsnmode>")
+		      (const_string "XI")))
+   (set (attr "enabled")
+	(if_then_else (eq_attr "alternative" "1")
+		      (symbol_ref "<MODE_SIZE> == 64 || TARGET_AVX512VL")
+		      (const_int 1)))])
+
+(define_insn "*one_cmpl<mode>2_pternlog_false_dep"
+  [(set (match_operand:VI 0 "register_operand" "=v,v,v")
+	(xor:VI (match_operand:VI 1 "bcst_vector_operand"     " 0, m,Br")
+		(match_operand:VI 2 "vector_all_ones_operand" "BC,BC,BC")))
+   (unspec [(match_operand:VI 3 "register_operand" "0,0,0")]
+     UNSPEC_INSN_FALSE_DEP)]
+  "TARGET_AVX512F"
 {
   if (TARGET_AVX512VL)
     return "vpternlog<ternlogsuffix>\t{$0x55, %1, %0, %0<mask_operand3>|%0<mask_operand3>, %0, %1, 0x55}";
@@ -17226,7 +17341,7 @@  (define_insn "*andnot<mode>3"
   [(set (match_operand:VI 0 "register_operand" "=x,x,v,v,v")
 	(and:VI
 	  (not:VI (match_operand:VI 1 "bcst_vector_operand" "0,x,v,m,Br"))
-	  (match_operand:VI 2 "bcst_vector_operand" "xBm,xm,vmBr,v,v")))]
+	  (match_operand:VI 2 "bcst_vector_operand" "xBm,xm,vmBr,0,0")))]
   "TARGET_SSE
    && (register_operand (operands[1], <MODE>mode)
        || register_operand (operands[2], <MODE>mode))"
@@ -17685,8 +17800,8 @@  (define_insn "*iornot<mode>3"
   [(set (match_operand:VI 0 "register_operand" "=v,v,v,v")
 	(ior:VI
 	  (not:VI
-	    (match_operand:VI 1 "bcst_vector_operand" "v,Br,v,m"))
-	  (match_operand:VI 2 "bcst_vector_operand" "vBr,v,m,v")))]
+	    (match_operand:VI 1 "bcst_vector_operand" "0,m,  0,vBr"))
+	  (match_operand:VI 2 "bcst_vector_operand"   "m,0,vBr,  0")))]
   "(<MODE_SIZE> == 64 || TARGET_AVX512VL
     || (TARGET_AVX512F && !TARGET_PREFER_AVX256))
    && (register_operand (operands[1], <MODE>mode)
@@ -17710,7 +17825,7 @@  (define_insn "*iornot<mode>3"
 		      (const_string "<sseinsnmode>")
 		      (const_string "XI")))
    (set (attr "enabled")
-	(if_then_else (eq_attr "alternative" "2,3")
+	(if_then_else (eq_attr "alternative" "0,1")
 		      (symbol_ref "<MODE_SIZE> == 64 || TARGET_AVX512VL")
 		      (const_string "*")))])
 
@@ -17718,8 +17833,8 @@  (define_insn "*xnor<mode>3"
   [(set (match_operand:VI 0 "register_operand" "=v,v")
 	(not:VI
 	  (xor:VI
-	    (match_operand:VI 1 "bcst_vector_operand" "%v,v")
-	    (match_operand:VI 2 "bcst_vector_operand" "vBr,m"))))]
+	    (match_operand:VI 1 "bcst_vector_operand" "%0, 0")
+	    (match_operand:VI 2 "bcst_vector_operand" " m,vBr"))))]
   "(<MODE_SIZE> == 64 || TARGET_AVX512VL
     || (TARGET_AVX512F && !TARGET_PREFER_AVX256))
    && (register_operand (operands[1], <MODE>mode)
@@ -17738,7 +17853,7 @@  (define_insn "*xnor<mode>3"
 		      (const_string "<sseinsnmode>")
 		      (const_string "XI")))
    (set (attr "enabled")
-	(if_then_else (eq_attr "alternative" "1")
+	(if_then_else (eq_attr "alternative" "0")
 		      (symbol_ref "<MODE_SIZE> == 64 || TARGET_AVX512VL")
 		      (const_string "*")))])
 
@@ -17749,8 +17864,8 @@  (define_code_attr ternlog_nlogic [(and "0x11") (ior "0x77")])
 (define_insn "*<nlogic><mode>3"
   [(set (match_operand:VI 0 "register_operand" "=v,v")
 	(andor:VI
-	  (not:VI (match_operand:VI 1 "bcst_vector_operand" "%v,v"))
-	  (not:VI (match_operand:VI 2 "bcst_vector_operand" "vBr,m"))))]
+	  (not:VI (match_operand:VI 1 "bcst_vector_operand" "%0, 0"))
+	  (not:VI (match_operand:VI 2 "bcst_vector_operand" "m,vBr"))))]
   "(<MODE_SIZE> == 64 || TARGET_AVX512VL
     || (TARGET_AVX512F && !TARGET_PREFER_AVX256))
    && (register_operand (operands[1], <MODE>mode)
@@ -17769,7 +17884,7 @@  (define_insn "*<nlogic><mode>3"
 		      (const_string "<sseinsnmode>")
 		      (const_string "XI")))
    (set (attr "enabled")
-	(if_then_else (eq_attr "alternative" "1")
+	(if_then_else (eq_attr "alternative" "0")
 		      (symbol_ref "<MODE_SIZE> == 64 || TARGET_AVX512VL")
 		      (const_string "*")))])
 
diff --git a/gcc/testsuite/gcc.target/i386/pr100711-6.c b/gcc/testsuite/gcc.target/i386/pr100711-6.c
index 7142a98f537..808507471c9 100644
--- a/gcc/testsuite/gcc.target/i386/pr100711-6.c
+++ b/gcc/testsuite/gcc.target/i386/pr100711-6.c
@@ -1,5 +1,5 @@ 
 /* { dg-do compile } */
-/* { dg-options "-mavx512f -mno-avx512vl -mprefer-vector-width=512 -O2" } */
+/* { dg-options "-mavx512f -mno-avx512vl -mprefer-vector-width=512 -Os" } */
 
 typedef int v16si __attribute__ ((vector_size (64)));
 typedef long long v8di __attribute__((vector_size (64)));
diff --git a/gcc/testsuite/gcc.target/i386/pr110438.c b/gcc/testsuite/gcc.target/i386/pr110438.c
new file mode 100644
index 00000000000..11b8cc59fd2
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr110438.c
@@ -0,0 +1,30 @@ 
+/* { dg-do compile } */
+/* { dg-options "-mavx512f -O2 -ftree-vectorize -mno-avx512dq -dp -mprefer-vector-width=512" } */
+/* { dg-final { scan-assembler-times {cvtmask2.*_pternlog} "1" } } */
+/* { dg-final { scan-assembler-times {constm1_pternlog} "1" } } */
+/* { dg-final { scan-assembler-not {(?n)vpternlogd.*\(} } } */
+
+
+#include <immintrin.h>
+
+__m512i g(void)
+{
+  return (__m512i){ 0 } - 1;
+}
+
+__m512i g1(__m512i* a)
+{
+  return ~(*a);
+}
+
+void
+foo (int* a, int* __restrict b)
+{
+  for (int i = 0; i != 16; i++)
+    {
+      if (b[i])
+	a[i] = -1;
+      else
+	a[i] = 0;
+    }
+}