From 70e65633d892765bcbaad3493e5b690abd5402f2 Mon Sep 17 00:00:00 2001
From: Siddhesh Poyarekar <siddhesh@sourceware.org>
Date: Thu, 28 Mar 2019 09:19:34 +0530
Subject: [PATCH 71/72] aarch64: better float to unsigned int conversion

A straight float to unsigned conversion has a limited range of (-1.0,
UTYPE_MAX) which should be fine in general but for the sake of
consistency across the interpreter and the JIT compiler, it is
necessary to work a wee bit harder to expand this range to (TYPE_MIN,
UTYPE_MAX), which can be done with a simple range check.  This adds a
couple of branches but only one of the branches should have a
noticeable performance impact on most processors with branch
predictors, and that too only if the input number varies wildly in
range.

This currently works only for 64-bit conversions, 32-bit is still WIP.
---
 src/lj_asm_arm64.h    | 30 ++++++++++++++++++++++--------
 src/lj_target_arm64.h |  1 +
 2 files changed, 23 insertions(+), 8 deletions(-)

diff --git a/src/lj_asm_arm64.h b/src/lj_asm_arm64.h
index 42a4fae..c72144a 100644
--- a/src/lj_asm_arm64.h
+++ b/src/lj_asm_arm64.h
@@ -594,14 +594,28 @@ static void asm_conv(ASMState *as, IRIns *ir)
     } else {
       Reg left = ra_alloc1(as, lref, RSET_FPR);
       Reg dest = ra_dest(as, ir, RSET_GPR);
-      A64Ins ai = irt_is64(ir->t) ?
-	(st == IRT_NUM ?
-	 (irt_isi64(ir->t) ? A64I_FCVT_S64_F64 : A64I_FCVT_U64_F64) :
-	 (irt_isi64(ir->t) ? A64I_FCVT_S64_F32 : A64I_FCVT_U64_F32)) :
-	(st == IRT_NUM ?
-	 (irt_isint(ir->t) ? A64I_FCVT_S32_F64 : A64I_FCVT_U32_F64) :
-	 (irt_isint(ir->t) ? A64I_FCVT_S32_F32 : A64I_FCVT_U32_F32));
-      emit_dn(as, ai, dest, (left & 31));
+
+      A64Ins ai_signed = st == IRT_NUM ?
+        (irt_is64(ir->t) ? A64I_FCVT_S64_F64 : A64I_FCVT_S32_F64) :
+        (irt_is64(ir->t) ? A64I_FCVT_S64_F32 : A64I_FCVT_S32_F32);
+
+      if (irt_isi64(ir->t) || irt_isint(ir->t))
+        emit_dn(as, ai_signed, dest, (left & 31));
+      else {
+        A64Ins ai_unsigned = st == IRT_NUM ?
+          (irt_is64(ir->t) ? A64I_FCVT_U64_F64 : A64I_FCVT_U32_F64) :
+          (irt_is64(ir->t) ? A64I_FCVT_U64_F32 : A64I_FCVT_U32_F32);
+
+        MCLabel l_done = emit_label(as);
+        emit_dn(as, ai_unsigned, dest, (left & 31));
+        MCLabel l_signed = emit_label(as);
+	emit_jmp(as, l_done);
+        emit_dn(as, ai_signed, dest, (left & 31));
+        /* The valid range for float to unsigned int conversion is (-1.0,
+           UINT{,64}_MAX-1), but we just compare with 0 to save a load. */
+        emit_cond_branch(as, CC_PL, l_signed);
+        emit_nm(as, st == IRT_NUM ? A64I_FCMPZd : A64I_FCMPZs, left & 31, 0);
+      }
     }
   } else if (st >= IRT_I8 && st <= IRT_U16) { /* Extend to 32 bit integer. */
     Reg dest = ra_dest(as, ir, RSET_GPR);
diff --git a/src/lj_target_arm64.h b/src/lj_target_arm64.h
index a207a2b..2f8357f 100644
--- a/src/lj_target_arm64.h
+++ b/src/lj_target_arm64.h
@@ -279,6 +279,7 @@ typedef enum A64Ins {
   A64I_STPs = 0x2d000000,
   A64I_STPd = 0x6d000000,
   A64I_FCMPd = 0x1e602000,
+  A64I_FCMPZs = 0x1e202008,
   A64I_FCMPZd = 0x1e602008,
   A64I_FCSELd = 0x1e600c00,
   A64I_FRINTMd = 0x1e654000,
-- 
2.20.1