[PULL v2 04/12] Hexagon (target/hexagon) properly handle denorm in arch_sf_recip_common

[Taylor Simpson <tsimpson@quicinc.com>]

The arch_sf_recip_common function was calling float32_getexp which
adjusts for denorm, but the we actually need the raw exponent bits.

This function is called from 3 instructions
    sfrecipa
    sffixupn
    sffixupd

Test cases added to tests/tcg/hexagon/fpstuff.c

Signed-off-by: Taylor Simpson <tsimpson@quicinc.com>
Message-Id: <20220210021556.9217-6-tsimpson@quicinc.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/hexagon/fma_emu.h    |  6 ++++-
 target/hexagon/arch.c       |  6 ++---
 tests/tcg/hexagon/fpstuff.c | 44 ++++++++++++++++++++++++++++++++++---
 3 files changed, 49 insertions(+), 7 deletions(-)

diff --git a/target/hexagon/fma_emu.h b/target/hexagon/fma_emu.h
index e3b99a8cf4..91591d6050 100644
--- a/target/hexagon/fma_emu.h
+++ b/target/hexagon/fma_emu.h
@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2021 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -24,6 +24,10 @@ static inline bool is_finite(float64 x)
 }
 
 int32_t float64_getexp(float64 f64);
+static inline uint32_t float32_getexp_raw(float32 f32)
+{
+    return extract32(f32, 23, 8);
+}
 int32_t float32_getexp(float32 f32);
 float32 infinite_float32(uint8_t sign);
 float32 internal_fmafx(float32 a, float32 b, float32 c,
diff --git a/target/hexagon/arch.c b/target/hexagon/arch.c
index 68a55b3bd4..da79b41c4d 100644
--- a/target/hexagon/arch.c
+++ b/target/hexagon/arch.c
@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2021 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -298,8 +298,8 @@ int arch_sf_recip_common(float32 *Rs, float32 *Rt, float32 *Rd, int *adjust,
     } else {
         PeV = 0x00;
         /* Basic checks passed */
-        n_exp = float32_getexp(RsV);
-        d_exp = float32_getexp(RtV);
+        n_exp = float32_getexp_raw(RsV);
+        d_exp = float32_getexp_raw(RtV);
         if ((n_exp - d_exp + SF_BIAS) <= SF_MANTBITS) {
             /* Near quotient underflow / inexact Q */
             PeV = 0x80;
diff --git a/tests/tcg/hexagon/fpstuff.c b/tests/tcg/hexagon/fpstuff.c
index 0dff429f4c..043f18fab3 100644
--- a/tests/tcg/hexagon/fpstuff.c
+++ b/tests/tcg/hexagon/fpstuff.c
@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2020-2021 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2020-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -38,6 +38,8 @@ const int SF_NaN_special =                0x7f800001;
 const int SF_ANY =                        0x3f800000;
 const int SF_HEX_NAN =                    0xffffffff;
 const int SF_small_neg =                  0xab98fba8;
+const int SF_denorm =                     0x00000001;
+const int SF_random =                     0x346001d6;
 
 const long long DF_NaN =                  0x7ff8000000000000ULL;
 const long long DF_ANY =                  0x3f80000000000000ULL;
@@ -250,10 +252,11 @@ static void check_dfminmax(void)
     check_fpstatus(usr, FPINVF);
 }
 
-static void check_recip_exception(void)
+static void check_sfrecipa(void)
 {
     int result;
     int usr;
+    int pred;
 
     /*
      * Check that sfrecipa doesn't set status bits when
@@ -329,6 +332,17 @@ static void check_recip_exception(void)
          : "r2", "p0", "usr");
     check32(result, 0x3f800000);
     check_fpstatus(usr, 0);
+
+    /*
+     * Check that sfrecipa properly handles denorm
+     */
+    asm (CLEAR_FPSTATUS
+         "%0,p0 = sfrecipa(%2, %3)\n\t"
+         "%1 = p0\n\t"
+         : "=r"(result), "=r"(pred) : "r"(SF_denorm), "r"(SF_random)
+         : "p0", "usr");
+    check32(result, 0x6a920001);
+    check32(pred, 0x80);
 }
 
 static void check_canonical_NaN(void)
@@ -455,6 +469,28 @@ static void check_invsqrta(void)
     check32(predval, 0x0);
 }
 
+static void check_sffixupn(void)
+{
+    int result;
+
+    /* Check that sffixupn properly deals with denorm */
+    asm volatile("%0 = sffixupn(%1, %2)\n\t"
+                 : "=r"(result)
+                 : "r"(SF_random), "r"(SF_denorm));
+    check32(result, 0x246001d6);
+}
+
+static void check_sffixupd(void)
+{
+    int result;
+
+    /* Check that sffixupd properly deals with denorm */
+    asm volatile("%0 = sffixupd(%1, %2)\n\t"
+                 : "=r"(result)
+                 : "r"(SF_denorm), "r"(SF_random));
+    check32(result, 0x146001d6);
+}
+
 static void check_float2int_convs()
 {
     int res32;
@@ -602,9 +638,11 @@ int main()
     check_compare_exception();
     check_sfminmax();
     check_dfminmax();
-    check_recip_exception();
+    check_sfrecipa();
     check_canonical_NaN();
     check_invsqrta();
+    check_sffixupn();
+    check_sffixupd();
     check_float2int_convs();
 
     puts(err ? "FAIL" : "PASS");
-- 
2.17.1