py/parse: Work around xtensa esp-2020r3 compiler bug.

dpgeorge · dpgeorge · commit f63b4f85aae1 · 2022-06-09T13:56:30.000+10:00
This commit works around a bug in xtensa-esp32-elf-gcc version esp-2020r3. The bug is in generation of loop constructs. The below code is generated by the xtensa-esp32 compiler. The first extract is the buggy machine code and the second extract is the corrected machine code. The test `basics/logic_constfolding.py` fails with the first code and succeeds with the second. Disassembly of section .text.push_result_rule: 00000000 <push_result_rule>: ... d6: 209770 or a9, a7, a7 d9: 178976 loop a9, f4 <push_result_rule+0xf4> d9: R_XTENSA_SLOT0_OP .text.push_result_rule+0xf4 dc: 030190 rsr.lend a9 df: 130090 wsr.lbeg a9 e2: a8c992 addi a9, a9, -88 e5: 06d992 addmi a9, a9, 0x600 e8: 130190 wsr.lend a9 eb: 002000 isync ee: 030290 rsr.lcount a9 f1: 01c992 addi a9, a9, 1 f4: 1494e7 bne a4, a14, 10c <push_result_rule+0x10c> f4: R_XTENSA_SLOT0_OP .text.push_result_rule+0x10c Disassembly of section .text.push_result_rule: 00000000 <push_result_rule>: ... d6: 209770 or a9, a7, a7 d9: 178976 loop a9, f4 <push_result_rule+0xf4> d9: R_XTENSA_SLOT0_OP .text.push_result_rule+0xf4 dc: 030190 rsr.lend a9 df: 130090 wsr.lbeg a9 e2: 000091 l32r a9, fffc00e4 <push_result_rule+0xfffc00e4> e2: R_XTENSA_SLOT0_OP .literal.push_result_rule+0x18 e5: 0020f0 nop e8: 130190 wsr.lend a9 eb: 002000 isync ee: 030290 rsr.lcount a9 f1: 01c992 addi a9, a9, 1 f4: 1494e7 bne a4, a14, 10c <push_result_rule+0x10c> f4: R_XTENSA_SLOT0_OP .text.push_result_rule+0x10c Work done in collaboration with @jimmo. Signed-off-by: Damien George <damien@micropython.org>
diff --git a/ports/esp32/mpconfigport.h b/ports/esp32/mpconfigport.h
@@ -30,6 +30,13 @@
 #define MICROPY_EMIT_XTENSAWIN              (1)
 #endif
 
+// workaround for xtensa-esp32-elf-gcc esp-2020r3, which can generate wrong code for loops
+// see https://github.com/espressif/esp-idf/issues/9130
+// this was fixed in newer versions of the compiler by:
+//   "gas: use literals/const16 for xtensa loop relaxation"
+//   https://github.com/jcmvbkbc/binutils-gdb-xtensa/commit/403b0b61f6d4358aee8493cb1d11814e368942c9
+#define MICROPY_COMP_CONST_FOLDING_COMPILER_WORKAROUND (1)
+
 // optimisations
 #define MICROPY_OPT_COMPUTED_GOTO           (1)
 
diff --git a/py/parse.c b/py/parse.c
@@ -639,6 +639,11 @@ STATIC MP_DEFINE_CONST_MAP(mp_constants_map, mp_constants_table);
 STATIC void push_result_rule(parser_t *parser, size_t src_line, uint8_t rule_id, size_t num_args);
 
 #if MICROPY_COMP_CONST_FOLDING
+#if MICROPY_COMP_CONST_FOLDING_COMPILER_WORKAROUND
+// Some versions of the xtensa-esp32-elf-gcc compiler generate wrong code if this
+// function is static, so provide a hook for them to work around this problem.
+MP_NOINLINE
+#endif
 STATIC bool fold_logical_constants(parser_t *parser, uint8_t rule_id, size_t *num_args) {
     if (rule_id == RULE_or_test
         || rule_id == RULE_and_test) {
