[gcc(refs/vendors/ibm/heads/perf)] amdgcn: fix ICE on subreg of BI reg.

[gcc(refs/vendors/ibm/heads/perf)] amdgcn: fix ICE on subreg of BI reg.

[Jiu Fu Guo]

https://gcc.gnu.org/g:82863a5cc99143f93203b2caed3750c7313a87ea

commit 82863a5cc99143f93203b2caed3750c7313a87ea
Author: Andrew Stubbs <ams@codesourcery.com>
Date:   Wed Feb 26 16:45:29 2020 +0000

    amdgcn: fix ICE on subreg of BI reg.
    
    BImode usually only requires one bit, but instructions that write to VCC also
    clobber the reset of the DImode register pair, so gcn_class_max_nregs reports
    that two registers are needed for BImode.  Paradoxically, accessing VCC via
    SImode is therefore uses fewer registers than accessing via BImode.
    
    The LRA checking code takes this into account, but the DF liveness data also
    looks at the subreg, so it says (subreg:SI (reg:BI VCC) 0) only makes the low
    part live.  Both are "correct", but they disagree, which causes an ICE.
    
    This doesn't happen when writing conditions to VCC; it happens when accessing
    VCC_LO via a regular move to a regular SImode register.
    
    If we transform the subregs so that BImode is always the outer mode then it
    basically means the same thing, except that now both LRA and DF calculate nregs
    the same, and ICE goes away.
    
    As soon as LRA is done the subregs all evaporate anyway.
    
    2020-02-27  Andrew Stubbs  <ams@codesourcery.com>
    
            gcc/
            * config/gcn/gcn.md (mov<mode>): Add transformations for BI subregs.

Diff:
---
 gcc/ChangeLog         |  4 ++++
 gcc/config/gcn/gcn.md | 25 +++++++++++++++++++++++++
 2 files changed, 29 insertions(+)

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index beb56bc0130..b17d255ac8a 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,7 @@
+2020-02-27  Andrew Stubbs  <ams@codesourcery.com>
+
+	* config/gcn/gcn.md (mov<mode>): Add transformations for BI subregs.
+
 2020-02-27  Mark Williams  <mwilliams@fb.com>
 
 	* dwarf2out.c (file_name_acquire): Call remap_debug_filename.
diff --git a/gcc/config/gcn/gcn.md b/gcc/config/gcn/gcn.md
index b527d9a7a8b..d8b49dfd640 100644
--- a/gcc/config/gcn/gcn.md
+++ b/gcc/config/gcn/gcn.md
@@ -395,6 +395,31 @@
 	(match_operand:MOV_MODE 1 "general_operand"))]
   ""
   {
+    if (SUBREG_P (operands[1])
+	&& GET_MODE (operands[1]) == SImode
+	&& GET_MODE (SUBREG_REG (operands[1])) == BImode)
+    {
+      /* (reg:BI VCC) has nregs==2 to ensure it gets clobbered as a whole,
+	 but (subreg:SI (reg:BI VCC)) doesn't, which causes the LRA liveness
+	 checks to assert.  Transform this:
+	   (set (reg:SI) (subreg:SI (reg:BI)))
+	 to this:
+	   (set (subreg:BI (reg:SI)) (reg:BI))  */
+      operands[0] = gen_rtx_SUBREG (BImode, operands[0], 0);
+      operands[1] = SUBREG_REG (operands[1]);
+    }
+    if (SUBREG_P (operands[0])
+	&& GET_MODE (operands[0]) == SImode
+	&& GET_MODE (SUBREG_REG (operands[0])) == BImode)
+      {
+	/* Likewise, transform this:
+	     (set (subreg:SI (reg:BI)) (reg:SI))
+	   to this:
+	     (set (reg:BI) (subreg:BI (reg:SI))) */
+	operands[0] = SUBREG_REG (operands[0]);
+	operands[1] = gen_rtx_SUBREG (BImode, operands[1], 0);
+      }
+
     if (MEM_P (operands[0]))
       operands[1] = force_reg (<MODE>mode, operands[1]);