[to-be-committed,RISC-V] Eliminate unnecessary sign extension after inlined str[n]cmp

Message ID 4b819e72-26bb-47dc-a405-c239ab07db71@ventanamicro.com
State Committed
Delegated to: Jeff Law
Headers
Series [to-be-committed,RISC-V] Eliminate unnecessary sign extension after inlined str[n]cmp |

Checks

Context Check Description
rivoscibot/toolchain-ci-rivos-lint success Lint passed
rivoscibot/toolchain-ci-rivos-apply-patch success Patch applied
rivoscibot/toolchain-ci-rivos-build--linux-rv32gc_zba_zbb_zbc_zbs-ilp32d-non-multilib pending Build started
rivoscibot/toolchain-ci-rivos-build--newlib-rv64gcv-lp64d-multilib success Build passed
rivoscibot/toolchain-ci-rivos-build--linux-rv64gcv-lp64d-multilib success Build passed
rivoscibot/toolchain-ci-rivos-build--newlib-rv32imc_zba_zbb_zbc_zbs-ilp32d-non-multilib success Build passed
rivoscibot/toolchain-ci-rivos-build--newlib-rv64gc-lp64d-multilib success Build passed
rivoscibot/toolchain-ci-rivos-build--linux-rv64gc_zba_zbb_zbc_zbs-lp64d-non-multilib success Build passed
linaro-tcwg-bot/tcwg_gcc_build--master-arm success Build passed
linaro-tcwg-bot/tcwg_gcc_build--master-aarch64 success Build passed
rivoscibot/toolchain-ci-rivos-test success Testing passed
linaro-tcwg-bot/tcwg_gcc_check--master-aarch64 success Test passed
linaro-tcwg-bot/tcwg_gcc_check--master-arm success Test passed

Commit Message

Jeff Law July 10, 2024, 6:13 p.m. UTC
  This patch eliminates an unnecessary sign extension for scalar inlined 
string comparisons on rv64.

Conceptually this is pretty simple.  Prove all the paths which "return" 
a value from the inlined string comparison already have sign extended 
values.

FINAL_LABEL is the point after the calculation of the return value.  So 
if we have a jump to FINAL_LABEL, we must have a properly extended 
result value at that point.

Second we're going to arrange in the .md part of the expander to use an 
X mode temporary for the result.  After computing the result we will (if 
necessary) extract the low part of the result using a SUBREG tagged with 
the appropriate SUBREG_PROMOTED_* bits.


So with that background.

We find a jump to FINAL_LABEL in emit_strcmp_scalar_compare_byte.  Since 
we know the result is X mode, we can just emit the subtraction of the 
two chars in X mode and we'll have a properly sign extended result.

There's 4 jumps to final_label in emit_strcmp_scalar.

The first is just returning zero and needs trivial simplification to not 
force the result into SImode.

The second is after calling strcmp in the library.  The ABI mandates 
that value is sign extended, so there's nothing to do for that case.

The 3rd occurs after a call to 
emit_strcmp_scalar_result_calculation_nonul.  If we dive into that 
routine it needs simplificationq similar to what we did in 
emit_strcmp_scalar_compare_byte

The 4th occurs after a call to emit_strcmp_scalar_result_calculation 
which again needs trivial adjustment like we've done in the other routines.

Finally, at the end of expand_strcmp, just store the X mode result 
sitting in SUB to RESULT.


The net of all that is we know every path has its result properly 
extended to X mode.  Standard redundant extension removal will take care 
of the rest.

We've been running this within Ventana for about 6 months, so naturally 
it's been through various QA cycles, dhrystone, spec2017, etc.  It's 
also been through a build/test cycle in my tester.  Waiting on results 
from the pre-commit testing before moving forward.

jeff

ps.  I suspect memcmp could probably benefit from the same treatment.
gcc/
	* config/riscv/riscv-string.cc
	(emit_strcmp_scalar_compare_byte): Set RESULT directly rather
	than using a new temporary.
	(emit_strcmp_scalar_result_calculation_nonul): Likewise.
	(emit_strcmp_scalar_result_calculation): Likewise.
	(riscv_expand_strcmp_scalar): Use CONST0_RTX rather than
	generating a new node.
	(expand_strcmp): Copy directly from SUB to RESULT.
	* config/riscv/riscv.md (cmpstrnsi, cmpstrsi): Pass an X
	mode temporary to the expansion routines.  If necessary
	extract low part of the word to store in final result location.
  

Patch

diff --git a/gcc/config/riscv/riscv-string.cc b/gcc/config/riscv/riscv-string.cc
index d9380033175..29e38a1c8b7 100644
--- a/gcc/config/riscv/riscv-string.cc
+++ b/gcc/config/riscv/riscv-string.cc
@@ -140,9 +140,7 @@  static void
 emit_strcmp_scalar_compare_byte (rtx result, rtx data1, rtx data2,
 				 rtx final_label)
 {
-  rtx tmp = gen_reg_rtx (Xmode);
-  do_sub3 (tmp, data1, data2);
-  emit_insn (gen_movsi (result, gen_lowpart (SImode, tmp)));
+  do_sub3 (result, data1, data2);
   emit_jump_insn (gen_jump (final_label));
   emit_barrier (); /* No fall-through.  */
 }
@@ -310,8 +308,7 @@  emit_strcmp_scalar_result_calculation_nonul (rtx result, rtx data1, rtx data2)
   rtx tmp = gen_reg_rtx (Xmode);
   emit_insn (gen_slt_3 (LTU, Xmode, Xmode, tmp, data1, data2));
   do_neg2 (tmp, tmp);
-  do_ior3 (tmp, tmp, const1_rtx);
-  emit_insn (gen_movsi (result, gen_lowpart (SImode, tmp)));
+  do_ior3 (result, tmp, const1_rtx);
 }
 
 /* strcmp-result calculation.
@@ -367,9 +364,7 @@  emit_strcmp_scalar_result_calculation (rtx result, rtx data1, rtx data2,
   unsigned int shiftr = (xlen - 1) * BITS_PER_UNIT;
   do_lshr3 (data1, data1, GEN_INT (shiftr));
   do_lshr3 (data2, data2, GEN_INT (shiftr));
-  rtx tmp = gen_reg_rtx (Xmode);
-  do_sub3 (tmp, data1, data2);
-  emit_insn (gen_movsi (result, gen_lowpart (SImode, tmp)));
+  do_sub3 (result, data1, data2);
 }
 
 /* Expand str(n)cmp using Zbb/TheadBb instructions.
@@ -444,7 +439,7 @@  riscv_expand_strcmp_scalar (rtx result, rtx src1, rtx src2,
   /* All compared and everything was equal.  */
   if (ncompare)
     {
-      emit_insn (gen_rtx_SET (result, gen_rtx_CONST_INT (SImode, 0)));
+      emit_insn (gen_rtx_SET (result, CONST0_RTX (GET_MODE (result))));
       emit_jump_insn (gen_jump (final_label));
       emit_barrier (); /* No fall-through.  */
     }
@@ -1544,7 +1537,7 @@  expand_strcmp (rtx result, rtx src1, rtx src2, rtx nbytes,
   if (with_length)
     emit_label (done);
 
-  emit_insn (gen_movsi (result, gen_lowpart (SImode, sub)));
+  emit_move_insn (result, sub);
   return true;
 }
 
diff --git a/gcc/config/riscv/riscv.md b/gcc/config/riscv/riscv.md
index 6032a65fc68..9c6382a2b4e 100644
--- a/gcc/config/riscv/riscv.md
+++ b/gcc/config/riscv/riscv.md
@@ -4345,9 +4345,19 @@  (define_expand "cmpstrnsi"
   "riscv_inline_strncmp && !optimize_size
     && (TARGET_ZBB || TARGET_XTHEADBB || TARGET_VECTOR)"
 {
-  if (riscv_expand_strcmp (operands[0], operands[1], operands[2],
+  rtx temp = gen_reg_rtx (word_mode);
+  if (riscv_expand_strcmp (temp, operands[1], operands[2],
                            operands[3], operands[4]))
-    DONE;
+    {
+      if (TARGET_64BIT)
+	{
+	  temp = gen_lowpart (SImode, temp);
+	  SUBREG_PROMOTED_VAR_P (temp) = 1;
+	  SUBREG_PROMOTED_SET (temp, SRP_SIGNED);
+	}
+      emit_move_insn (operands[0], temp);
+      DONE;
+    }
   else
     FAIL;
 })
@@ -4366,9 +4376,19 @@  (define_expand "cmpstrsi"
   "riscv_inline_strcmp && !optimize_size
     && (TARGET_ZBB || TARGET_XTHEADBB || TARGET_VECTOR)"
 {
-  if (riscv_expand_strcmp (operands[0], operands[1], operands[2],
+  rtx temp = gen_reg_rtx (word_mode);
+  if (riscv_expand_strcmp (temp, operands[1], operands[2],
                            NULL_RTX, operands[3]))
-    DONE;
+    {
+      if (TARGET_64BIT)
+	{
+	  temp = gen_lowpart (SImode, temp);
+	  SUBREG_PROMOTED_VAR_P (temp) = 1;
+	  SUBREG_PROMOTED_SET (temp, SRP_SIGNED);
+	}
+      emit_move_insn (operands[0], temp);
+      DONE;
+    }
   else
     FAIL;
 })