More SPE changes

AmigaLabs · Jul 14, 2023 · 11d2e48 · 11d2e48
1 parent cf8c530
commit 11d2e48
Show file tree

Hide file tree

Showing 11 changed files with 191 additions and 27 deletions.
diff --git a/GNUmakefile.os4 b/GNUmakefile.os4
@@ -255,14 +255,14 @@ endef
 
 define COMPILE_INLINE
 $(VERBOSE)@$(MAKEDIR) $(@D)
-$(VERBOSE)echo -e "\rCompiling Static \033[0;33m[$(@D)]\033[0m \r\t\t\t\t\t\t\t\t\t\t\t \033[0;31m$(@F)\033[0m"
+$(VERBOSE)echo -e "\rCompiling Inline Static \033[0;33m[$(@D)]\033[0m \r\t\t\t\t\t\t\t\t\t\t\t \033[0;31m$(@F)\033[0m"
 $(VERBOSE)$(CC) $(CFLAGS_N) -Wno-missing-prototypes -Wno-missing-declarations -D__timespec_defined -o $@ -c $< $(LOG_COMMAND)
 endef
 
 define COMPILE_ASM
 $(VERBOSE)@$(MAKEDIR) $(@D)
 $(VERBOSE)echo -e "\rCompiling Static ASM \033[0;33m[$(@D)]\033[0m \r\t\t\t\t\t\t\t\t\t\t\t \033[0;31m$(@F)\033[0m"
-$(VERBOSE)$(CC)  -o $@ -c $< $(LOG_COMMAND)
+$(VERBOSE)$(CC) -o $@ -c $< $(LOG_COMMAND)
 endef
 
 define COMPILE_SHARED
@@ -297,8 +297,8 @@ endef
 
 define COMPILE_SPE_REG
 $(VERBOSE)@$(MAKEDIR) $(@D)
-$(VERBOSE)echo -e "\rAssembling with regnames \033[0;33m[$(@D)]\033[0m \r\t\t\t\t\t\t\t\t\t\t\t \033[0;31m$(@F)\033[0m"
-$(VERBOSE)$(CC) -mvrsave -mregnames -DVRSAVE -mspe -mcpu=8540 -mfloat-gprs=double -o $@ -c $< $(LOG_COMMAND)
+$(VERBOSE)echo -e "\rAssembling SPE with regnames \033[0;33m[$(@D)]\033[0m \r\t\t\t\t\t\t\t\t\t\t\t \033[0;31m$(@F)\033[0m"
+$(VERBOSE)$(CC) $(AFLAGS) -mvrsave -mregnames -DVRSAVE -mspe -mcpu=8540 -mfloat-gprs=double -o $@ -c $< $(LOG_COMMAND)
 endef
 
 define COMPILE_SPE

diff --git a/libc.gmk b/libc.gmk
@@ -215,8 +215,17 @@ C_SETJMP := \
     setjmp/longjmp.o \
     setjmp/setjmp.o \
     setjmp/sigjmp.o \
-    setjmp/siglongjmp.o \
-    setjmp/sigsetjmp.o
+    setjmp/siglongjmp.o
+
+ifdef SPE
+    $(info Using SPE sigssetjmp)
+    C_SETJMP := $(C_SETJMP)  \
+        setjmp/sigsetjmp_spe.o
+else
+    $(info Using NON-SPE sigsetjmp)
+    C_SETJMP := $(C_SETJMP)  \
+        setjmp/sigsetjmp.o
+endif
 
 C_SOCKET := \
     resolv/dns_parse.o \

diff --git a/libm.gmk b/libm.gmk
@@ -136,7 +136,6 @@ C_MATH := \
 	math/expm1.o \
 	math/expm1f.o \
 	math/expm1l.o \
-	math/fabs.o \
 	math/fabsf.o \
 	math/fabsl.o \
 	math/fdim.o \
@@ -293,6 +292,14 @@ C_MATH := \
 	math/truncf.o \
 	math/truncl.o
 
+ifdef SPE
+	C_MATH := $(C_MATH)\
+		math/fabs_spe.o
+else
+	C_MATH := $(C_MATH)\
+		math/fabs.o
+endif
+
 MATH_LIB := \
 	math/m.lib_rev.o
 

diff --git a/library/include/fenv.h b/library/include/fenv.h
@@ -66,11 +66,16 @@ __fe_dfl_env;
              FE_OVERFLOW | FE_UNDERFLOW) >> _FPUSW_SHIFT)
 
 #ifndef _SOFT_FLOAT
-#define    __mffs(__env)  __asm __volatile("mffs %0" : "=f" (*(__env)))
-#define    __mtfsf(__env) __asm __volatile("mtfsf 255,%0" : : "f" (__env))
+    #ifdef __SPE__
+        #define	__mffs(__env) __asm __volatile("mfspr %0, 512" : "=r" ((__env)->__bits.__reg))
+        #define	__mtfsf(__env) __asm __volatile("mtspr 512,%0;isync" :: "r" ((__env).__bits.__reg))
+    #else
+        #define	__mffs(__env) __asm __volatile("mffs %0" : "=f" ((__env)->__d))
+        #define	__mtfsf(__env) __asm __volatile("mtfsf 255,%0" :: "f" ((__env).__d))
+    #endif
 #else
-#define	__mffs(__env)
-#define	__mtfsf(__env)
+    #define	__mffs(__env)
+    #define	__mtfsf(__env)
 #endif
 
 union __fpscr {
@@ -142,7 +147,7 @@ fetestexcept(int __excepts) {
 
 __fenv_static inline int
 fegetround(void) {
-    union __fpscr __r = {0};
+    union __fpscr __r;
 
     __mffs(&__r.__d);
     return (__r.__bits.__reg & _ROUND_MASK);

diff --git a/library/math/fabs.c b/library/math/fabs.c
@@ -6,23 +6,14 @@
 #include "math_headers.h"
 #endif /* _MATH_HEADERS_H */
 
+#ifndef __SPE__
 inline static double
 __fabs(double x) {
-#ifndef _SOFT_FLOAT
     double res;
 
     __asm volatile("fabs %0, %1"
-                   : "=f"(res)
-                   : "f"(x));
-
-    return res;
-#else
-    uint32_t high;
-    GET_HIGH_WORD(high, x);
-    SET_HIGH_WORD(x, high & 0x7fffffff);
-
-    return x;
-#endif
+            : "=f"(res)
+            : "f"(x));
 }
 
 double
@@ -33,3 +24,4 @@ fabs(double x) {
 
     return (result);
 }
+#endif
diff --git a/library/math/fabs_spe.S b/library/math/fabs_spe.S
@@ -0,0 +1,13 @@
+/*
+ * $Id: math_fabs_spe.S,v 1.00 2023-07-13 12:04:23 clib2devs Exp $
+*/
+
+
+#include "ppc-asm.h"
+
+#ifdef __SPE__
+FUNC_START(fabs)
+	efdabs	f1, f1
+	blr
+FUNC_END(fabs)
+#endif
diff --git a/library/setjmp/longjmp.S b/library/setjmp/longjmp.S
@@ -96,7 +96,7 @@ FUNC_START(longjmp_altivec)
     bclr+	4,2
     li		3,1
     blr
-        FUNC_END(longjmp_altivec)
+FUNC_END(longjmp_altivec)
 #endif /* __SPE__ */
 
 FUNC_START(longjmp)

diff --git a/library/setjmp/siglongjmp.c b/library/setjmp/siglongjmp.c
@@ -6,6 +6,7 @@
 #include "signal_headers.h"
 #endif /* _SIGNAL_HEADERS_H */
 
+#ifndef __SPE__
 void siglongjmp(sigjmp_buf env, int val) {
     SHOWVALUE(val);
 
@@ -15,4 +16,5 @@ void siglongjmp(sigjmp_buf env, int val) {
 
     /* Call the machine-dependent function to restore machine state.  */
     _longjmp(env, val ?: 1);
-}
+}
+#endif
diff --git a/library/setjmp/sigsetjmp.S b/library/setjmp/sigsetjmp.S
@@ -16,10 +16,12 @@
    License along with the GNU C Library; if not, see
    <https://www.gnu.org/licenses/>.  */
 
+#ifndef __SPE__
 #define __NO_VMX__
 #undef __sigsetjmp_symbol
 #undef __sigjmp_save_symbol
 #undef JB_SIZE
 #define __sigsetjmp_symbol __sigsetjmp
 #define __sigjmp_save_symbol __sigjmp_save
-#include "sigsetjmp-common.S"
+#include "sigsetjmp-common.S"
+#endif
diff --git a/library/setjmp/sigsetjmp_spe.S b/library/setjmp/sigsetjmp_spe.S
@@ -0,0 +1,130 @@
+/*
+ * $Id: stdlib_sigsetjmp_spe.S,v 1.0 2023-07-13 13:50:17 clib2devs Exp $
+ */
+
+/*
+ * C library -- sigsetjmp, siglongjmp
+ *
+ *      siglongjmp(a,v)
+ * will generate a "return(v?v:1)" from the last call to
+ *      sigsetjmp(a, savemask)
+ * by restoring registers from the stack.
+ * The previous signal state is restored if savemask is non-zero
+ *
+ * jmpbuf layout:
+ *     +------------+
+ *     |  savemask  |
+ *     +------------+
+ *     | sig state  |
+ *     |            |
+ *     | (4 words)  |
+ *     |            |
+ *     +------------+
+ *     | saved regs |
+ *     |    ...     |
+ */
+
+#include "ppc-asm.h"
+
+#ifdef __SPE__
+FUNC_START(_setjmp)
+    li r4,0			/* Set second argument to 0.  */
+    b __sigsetjmp@local
+FUNC_END(_setjmp)
+
+FUNC_START(__sigsetjmp)
+	mr	r6,r3
+	stw	r4,0(r3)
+	or.	r7,r4,r4
+	beq	1f
+	li	r3,1			/* SIG_BLOCK, but doesn't matter */
+					    /*            since set == NULL  */
+	li	r4,0			/* set = NULL */
+	mr	r5,r6			/* &oset */
+	addi	r5,r5,4
+	bl sigprocmask    /* sigprocmask(SIG_BLOCK, NULL, &oset)*/
+	sc				        /* assume no error       XXX */
+1:
+	mflr	r11
+	mfcr	r12
+	mr	r10,r1
+	mr	r9,r2
+
+	/* FPRs */
+	evstdd	r9,24+0*8(r6)
+	evstdd	r10,24+1*8(r6)
+	evstdd	r11,24+2*8(r6)
+	evstdd	r12,24+3*8(r6)
+	evstdd	r13,24+4*8(r6)
+	evstdd	r14,24+5*8(r6)
+	evstdd	r15,24+6*8(r6)
+	evstdd	r16,24+7*8(r6)
+	evstdd	r17,24+8*8(r6)
+	evstdd	r18,24+9*8(r6)
+	evstdd	r19,24+10*8(r6)
+	evstdd	r20,24+11*8(r6)
+	evstdd	r21,24+12*8(r6)
+	evstdd	r22,24+13*8(r6)
+	evstdd	r23,24+14*8(r6)
+	evstdd	r24,24+15*8(r6)
+	evstdd	r25,24+16*8(r6)
+	evstdd	r26,24+17*8(r6)
+	evstdd	r27,24+18*8(r6)
+	evstdd	r28,24+19*8(r6)
+	evstdd	r29,24+20*8(r6)
+	evstdd	r30,24+21*8(r6)
+	evstdd	r31,24+22*8(r6)
+
+	li	r3,0
+	blr
+FUNC_END(__sigsetjmp)
+
+FUNC_START(siglongjmp)
+
+	/* FPRs */
+	evldd	r9,24+0*8(r6)
+	evldd	r10,24+1*8(r6)
+	evldd	r11,24+2*8(r6)
+	evldd	r12,24+3*8(r6)
+	evldd	r13,24+4*8(r6)
+	evldd	r14,24+5*8(r6)
+	evldd	r15,24+6*8(r6)
+	evldd	r16,24+7*8(r6)
+	evldd	r17,24+8*8(r6)
+	evldd	r18,24+9*8(r6)
+	evldd	r19,24+10*8(r6)
+	evldd	r20,24+11*8(r6)
+	evldd	r21,24+12*8(r6)
+	evldd	r22,24+13*8(r6)
+	evldd	r23,24+14*8(r6)
+	evldd	r24,24+15*8(r6)
+	evldd	r25,24+16*8(r6)
+	evldd	r26,24+17*8(r6)
+	evldd	r27,24+18*8(r6)
+	evldd	r28,24+19*8(r6)
+	evldd	r29,24+20*8(r6)
+	evldd	r30,24+21*8(r6)
+	evldd	r31,24+22*8(r6)
+
+	lwz	r7,0(r3)
+	mr	r6,r4
+	mtlr	r11
+	mtcr	r12
+	mr	r1,r10
+	or.	r7,r7,r7
+	beq	1f
+	mr	r4,r3
+	li	r3,3			/* SIG_SETMASK */
+	addi	r4,r4,4	/* &set */
+	li	r5,0			/* oset = NULL */
+	bl sigprocmask	/* sigprocmask(SIG_SET, &set, NULL) */
+	sc				    /* assume no error       XXX */
+1:
+	or.	r3,r6,r6
+	bnelr
+	li	r3,1
+	blr
+FUNC_END(siglongjmp)
+
+	.section .note.GNU-stack,"",%progbits
+#endif
diff --git a/shared.gmk b/shared.gmk
@@ -35,6 +35,10 @@ $(OUT_SHARED_LIB)/cpu/generic/%.o : $(LIB_DIR)/cpu/generic/%.S
 
 $(OUT_SHARED_LIB)/cpu/spe/%.o : $(LIB_DIR)/cpu/spe/%.S
 	$(VERBOSE)$(COMPILE_SPE_REG)
+$(OUT_SHARED_LIB)/setjmp/sigsetjmp_spe.o : $(LIB_DIR)/setjmp/sigsetjmp_spe.S
+	$(VERBOSE)$(COMPILE_SPE_REG)
+$(OUT_SHARED_LIB)/math/fabs_spe.o : $(LIB_DIR)/math/fabs_spe.S
+	$(VERBOSE)$(COMPILE_SPE_REG)
 
 $(OUT_SHARED_LIB)/%.o : $(LIB_DIR)/%.S
 	$(VERBOSE)$(COMPILE_ASM)